feat: 诊断（任务书 18 第一块）：诊断数学、四种提交、一次运行跑多个实验、账本与材料、flow 诊断各步
消融默认关，关掉时解释阶段一并跳过；选诊断实验用信息价值减成本。 eval/diagnosis/research.json 为第三块配置。vocab 字段说明的出口文字随提示词一起改，见下一次提交的 prompts/ 与同文件改动（本次含在内）。 Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
This commit is contained in:
+64
-19
@@ -158,7 +158,7 @@
|
||||
|
||||
Jev 是一个判别器，和程序检查地位相同，不单独成一种记录：它的原始概率和结论，记在它判断的那件事里（第 3 种的提交和退回、第 8 种的判重合并）。
|
||||
|
||||
**写入时由 ledger 拒绝的几种情况**：阶段已关闭还要写提交；要替换的那一份不是本人在这个阶段刚交的；没有"确认启动"就写运行结果（"确认没启动"这种结果除外）；这次消融已经有格启动了才冻结解释。
|
||||
**写入时由 ledger 拒绝的几种情况**：阶段已关闭还要写提交；要替换的那一份不是本人在这个阶段刚交的；没有"确认启动"就写运行结果（"确认没启动"这种结果除外）；这次消融已经有格启动了才冻结解释；这次诊断运行已经确认启动了才冻结诊断预测。
|
||||
|
||||
---
|
||||
|
||||
@@ -245,8 +245,8 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
**名单**
|
||||
|
||||
- `ROLES`、`Role`：三个角色，研究者、实验员、核验员。
|
||||
- `STAGES`、`Stage`：有人交提交的各个阶段，按流程先后排；程序自己做的步骤不在里面。
|
||||
- `SUBMISSION_KINDS`、`SubmissionKind`：每种提交属于哪个阶段、由哪个角色交，各模块都从这张表取。
|
||||
- `STAGES`、`Stage`：有人交提交的各个阶段，按流程先后排；程序自己做的步骤不在里面。诊断有四个：诊断解释与实验、诊断预测、诊断实现（含修复）、诊断核验（任务书 18）。
|
||||
- `SUBMISSION_KINDS`、`SubmissionKind`：每种提交属于哪个阶段、由哪个角色交，各模块都从这张表取。诊断加五种：`diagnosisProposal`、`diagnosisPrediction`、`diagnosisImplementation`、`diagnosisRepair`、`diagnosisVerification`；实现和修复同一个字段定义，写法同正向的 `implementation`、`repair`。
|
||||
|
||||
**指标**
|
||||
|
||||
@@ -276,7 +276,7 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
**Jev**
|
||||
|
||||
- `JevJudgment`（字段定义）：Jev 对一个问题的一次判断：问题名、问题全文的哈希（措辞、示例、阈值一起算）、阈值、原始概率、结论（概率是否超过阈值）、Jev 返回的请求编号。jev 交出它，ledger 在提交和退回记录里原样存它，两个模块互不引用，所以放在这里。
|
||||
- `Checklist`、`ChecklistOf<K>`（类型）：对号清单，materials 为一次任务算出、submissions 拿来判收不收；按提交种类区分，每种只放它要用的（入口和讨论的入口个数范围；认领的入口编号和至少排几个；提案的 M、因子数上限、能延续的因子、能引用的假设专用指标（全名）、已占用的指标编号（任务的指标和两个保留键）；两次预测的候选；挑刺的原提案和它的因子；回应、统一的挑刺编号（统一另有被合并的原提案）；实现、修复、核验的条款、因子、保障（核验另有没有归属的改动）；解释的格和 M；消融实现的格、要关的因子、全部因子；消融核验的格和全部因子；收尾无）。用户 2026-10-03 决定：vocab 只放这一个，任务材料的其余类型放 materials。
|
||||
- `Checklist`、`ChecklistOf<K>`（类型）：对号清单，materials 为一次任务算出、submissions 拿来判收不收；按提交种类区分，每种只放它要用的（入口和讨论的入口个数范围；认领的入口编号和至少排几个；提案的 M、因子数上限、能延续的因子、能引用的假设专用指标（全名）、已占用的指标编号（任务的指标和两个保留键）；两次预测的候选；挑刺的原提案和它的因子；回应、统一的挑刺编号（统一另有被合并的原提案）；实现、修复、核验的条款、因子、保障（核验另有没有归属的改动）；解释的格和 M；消融实现的格、要关的因子、全部因子；消融核验的格和全部因子；收尾无；诊断解释与实验的解释上限 M、实验上限 A 和已占用的指标编号；诊断预测的自己的解释编号和合并后的实验编号；诊断实现、修复、核验的选中实验编号）。用户 2026-10-03 决定：vocab 只放这一个，任务材料的其余类型放 materials。
|
||||
|
||||
### rules
|
||||
|
||||
@@ -376,6 +376,21 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
- `BetResult`（类型）：押的因子的结果：更新后的成分权重、权重是否无从更新、各因子的方向收益（实测换算的或估计的）、挑出的因子及其预测收益、押或不押的两种情况、关掉反而更好的格。
|
||||
- `chooseBet(input)`：押的因子，用有效格的实测更新各成分的权重，再挑方向收益最大的因子；有格关掉反而更好或估计收益不为正时不押，实测落在所有成分之外时保持起始权重（6.8）。
|
||||
|
||||
**诊断（6.12，4.3 诊 3、诊 4）**
|
||||
|
||||
成分、权重更新、精确积分、部分格网格、价值减成本的公式只在 `value.ts` 写一份，消融和诊断共用，差别只在"实测 → 收益"的换算：消融是撑几成乘剩余空间 h，没有"都不走"；诊断是改进量 x − b，h 取 1，另有收益为 0 的"都不走"（`floor = 0`）。消融的数字逐位不变（任务书 18 第 1.2 节）。
|
||||
|
||||
- `DiagnosisExplanation`（类型）：一条诊断解释，带作者和它对每个合并后的实验给的三个数。
|
||||
- `DiagnosisModelInput`、`DiagnosisModel`（类型）：成分模型的输入和结果（因子换成实验、撑几成换成改进量）。
|
||||
- `diagnosisModel(input)`：建成分模型，每条解释一个成分、开放预留一个成分，每个实验自己的分布和开放预留（6.3"诊断"、6.12"成分与权重"）。
|
||||
- `DiagnosisParams`、`DiagnosisCombo`（类型）：余额、每批上限、u、衰减和 D；一个实验组合的 VOI、R̂、m′、价值、成本、净值。
|
||||
- `evaluateDiagnosisCombos(model, params)`：列出可选组合并算价值、成本、净值：余额不到 2 次没有组合；实验数不超过 D 时只有"全做"，超过 D 时是所有恰好 D 个实验的组合。R̂ = ⌈(余额 − 1) ÷ 每批上限⌉，成本按一次运行算，ĝ₀ = max(0, max_d E(g_d))。
|
||||
- `DiagnosisRuleInput`、`DiagnosisDecision`、`SkippedReason`（类型）：选诊断实验的输入和结果，结果含 6.12 要求附上的全部数字：参数、成分与起始权重、各实验的事前期望收益、ĝ₀、全部组合的价值成本净值、做不做、选中的实验及其净值、每个实验做不做和没做的原因（放不下、信息价值不足以抵过成本）以及含它的净值最大组合。
|
||||
- `selectDiagnosis(input)`：选诊断实验规则"信息价值减成本"：取净值最大且 > 1e-12 的组合，差不到 1e-12 的按实验编号比（和选格同一条挑法），都不大于 1e-12 时不诊断。
|
||||
- `ExperimentOutcome`（类型）：一个实验的结局：有效分数、证据不足、没做（和消融一格同形）。
|
||||
- `SettleDiagnosisInput`、`DiagnosisDirection`（类型）：一次诊断结算的输入；诊断推出的方向（更新后的权重、各实验的收益、收益最大的实验、有没有指出方向）。
|
||||
- `settleDiagnosis(input)`：一次诊断的结算：每个实验的收益和合并范围，有效实验的世界比特、惊讶、分歧量；每条解释在每个有效实验上的确认或否定（用 `judgeExplanations`）；诊断推出的方向，最大收益不为正时 `found` 为 false。证据不足和没做的实验不判、不记账。
|
||||
|
||||
**信息账（6.9）**
|
||||
|
||||
这一组是方法里"唯一的尺子"，不随选批、选格规则变，一律在分数原值上从本批基线 b 起按格宽切小格。
|
||||
@@ -397,9 +412,9 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
- `runBalance(total, runs)`：从运行记录数出已扣次数、占用名额和余额；确认启动的扣一次，已预留的先占着名额，确认没启动的退还。只数运行次数，不算计算时间。
|
||||
- `forwardBatchSize(batchCap, balance)`：本轮正向批量 k = min(每批上限, 余额)。
|
||||
- `ablationCellLimit(balance)`：消融最多测"余额 − 1"格，余额为 0 或 1 时是 0。
|
||||
- `Phase`（类型）：正向或消融。
|
||||
- `Phase`（类型）：正向、诊断或消融。
|
||||
- `Tally`（类型）：一个人的战绩：总数、中位数偏高几次、偏低几次、被穿过上沿几次、下沿几次、包住几次、落点四段计数。
|
||||
- `trackRecord(samples)`：按研究者、按正向与消融分开数战绩，只用来显示，不改任何权重（6.9、6.10）。
|
||||
- `trackRecord(samples)`：按研究者、按正向、诊断与消融分开数战绩，结果按研究者编号、正向、诊断、消融排，只用来显示，不改任何权重（6.9、6.10）。
|
||||
- `authorVsMerged(items, direction)`："作者自评 vs 大家合并"，给出作者比大家乐观多少，以及实测落在两个范围的哪一段；只收已经跑过、有有效成绩的候选（6.10 第 4 条）。
|
||||
|
||||
**入口分配、种子与规则登记表（4.2、第九部分、第十一部分）**
|
||||
@@ -411,8 +426,10 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
- `seededSample(items, size, seed)`：用种子从一组编号里不放回地均匀抽 size 个，结果只取决于种子和编号集合，与传入次序无关；候选和因子都这样抽。
|
||||
- `BATCH_RULES`、`BatchRuleName`：选批规则登记表，名字 → 函数和版本号，两条：批内最好期望、均匀随机。
|
||||
- `CELL_RULES`、`CellRuleName`：选格规则登记表，四条：信息价值减成本、全测、随机抽同样多的格、都不测。
|
||||
- `DIAGNOSIS_RULES`、`DiagnosisRuleName`：选诊断实验规则登记表，只有一条：信息价值减成本（6.12），配置里不选，步骤清单的 `selectDiagnosis` 一步带它的名字和版本。
|
||||
- `runBatchRule(name, input)`：按名字跑一条选批规则，结果带上规则名和版本，flow 写决定记录时从这里取。
|
||||
- `runCellRule(name, input)`：按名字跑一条选格规则，结果带上规则名和版本。
|
||||
- `runDiagnosisRule(name, input)`：按名字跑选诊断实验规则，结果带上规则名和版本。
|
||||
|
||||
### ledger
|
||||
|
||||
@@ -420,9 +437,9 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
|
||||
**打开与写入**
|
||||
|
||||
- `Ledger.open(path)`：打开一份研究的账本，新文件就建表，旧文件核对表结构版本号（现为 4），对不上就报错。
|
||||
- `Ledger.open(path)`：打开一份研究的账本，新文件就建表，旧文件核对表结构版本号（现为 5），对不上就报错。
|
||||
- `ledger.close()`：关闭数据库。
|
||||
- `ledger.append(kind, content)`：追加一条记录并返回它。先按种类检查内容，再做第四部分说的四种拒收检查（第 1 条是只收开着的阶段的提交，即有开阶段记录、没有关阶段记录），都在同一个事务里。有自己编号、按定义只发生一次的记录按编号加了唯一约束，重复写就报错：阶段开、阶段关、运行登记、确认启动（"结果是没启动"也算在这一条里，一次运行两样只能有一样）、运行结果、模型调用开始、模型调用结束、消融结算（按轮次，一轮一条）；其余记录照写。记录之间的先后（例如结果前要先有登记）不查。写提案时，作者自评从提案里拆出来，在同一个事务里另存成一条封存。
|
||||
- `ledger.append(kind, content)`：追加一条记录并返回它。先按种类检查内容，再做第四部分说的五种拒收检查（第 1 条是只收开着的阶段的提交，即有开阶段记录、没有关阶段记录；第 5 条是这一轮的诊断运行已经确认启动，就不能再关这一轮的诊断预测阶段），都在同一个事务里。有自己编号、按定义只发生一次的记录按编号加了唯一约束，重复写就报错：阶段开、阶段关、运行登记、确认启动（"结果是没启动"也算在这一条里，一次运行两样只能有一样）、运行结果、模型调用开始、模型调用结束、消融结算、诊断结算（都按轮次，一轮一条）；其余记录照写。记录之间的先后（例如结果前要先有登记）不查。写提案时，作者自评从提案里拆出来，在同一个事务里另存成一条封存。
|
||||
- `ledger.transaction(fn)`：里面的几次追加要么全部写进去，要么一条都不写；fn 必须是同步函数，返回值类型排除了 Promise，异步函数在类型检查时就被挡住；不能嵌套。
|
||||
|
||||
**读取**
|
||||
@@ -435,7 +452,7 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
|
||||
**类型**
|
||||
|
||||
- `Kind`：记录的种类，共 19 种：`config`（研究配置）、`stageOpen` / `stageClose`（阶段开、关）、`submission`（提交的每一版，外层带收下时用到的 `JevJudgment` 数组，没用到 Jev 的为 null）、`rejection`（退回，带对象 `target`（与提交的 target 同义，没有对象的写 null；同一人在同一阶段可能对几个对象各做一次同种任务，materials 按它找本人这个对象的上一次退回）、原样内容和理由；Jev 退回的，原样存这次用到的一条或几条 `JevJudgment`，程序检查退回的这一项写 null）、`note`（私人笔记，每次修改追加一条，以最后写入的一条为准，不带版本号）、`protocol`（冻结协议）、`seal` / `unseal`（作者自评的封存、解封，都用记录序号指明是哪一条）、`code`（冻结代码）、`runPlanned` / `runStarted` / `runResult`（运行的登记要跑、确认启动、结果）、`decision`（决定）、`settlement`（一轮的消融结算，按轮次唯一）、`callStarted` / `callUsage` / `callEnded`（模型调用的开始，附这次任务的全部材料；每次回复的用量；结束，带墙钟时间（秒，`Measured`：中断后恢复时补写的结束记录，结局记出错、原因写"进程中断"，墙钟时间记测不到），结局与 agents 的会话结局同名：交了、没交（带原因）、出错（带原因）、被停下）、`event`（事件）。
|
||||
- `Kind`：记录的种类，共 20 种：`config`（研究配置）、`stageOpen` / `stageClose`（阶段开、关）、`submission`（提交的每一版，外层带收下时用到的 `JevJudgment` 数组，没用到 Jev 的为 null）、`rejection`（退回，带对象 `target`（与提交的 target 同义，没有对象的写 null；同一人在同一阶段可能对几个对象各做一次同种任务，materials 按它找本人这个对象的上一次退回）、原样内容和理由；Jev 退回的，原样存这次用到的一条或几条 `JevJudgment`，程序检查退回的这一项写 null）、`note`（私人笔记，每次修改追加一条，以最后写入的一条为准，不带版本号）、`protocol`（冻结协议）、`seal` / `unseal`（作者自评的封存、解封，都用记录序号指明是哪一条）、`code`（冻结代码）、`runPlanned` / `runStarted` / `runResult`（运行的登记要跑、确认启动、结果；登记的用途有基线、正向、消融、诊断四种）、`decision`（决定）、`settlement`（一轮的消融结算，按轮次唯一）、`diagnosisSettlement`（一轮的诊断结算，按轮次唯一，和消融结算各管各的）、`callStarted` / `callUsage` / `callEnded`（模型调用的开始，附这次任务的全部材料；每次回复的用量；结束，带墙钟时间（秒，`Measured`：中断后恢复时补写的结束记录，结局记出错、原因写"进程中断"，墙钟时间记测不到），结局与 agents 的会话结局同名：交了、没交（带原因）、出错（带原因）、被停下）、`event`（事件）。
|
||||
- `AppendKind`：调用方能追加的种类，除了 `seal`（封存只由 ledger 在写提案时一并写）。
|
||||
- `Content<K>`：某种记录存下的内容。
|
||||
- `Input<K>`：追加时交的内容；只有提交和存下的不同：提案带着作者自评。
|
||||
@@ -454,7 +471,7 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
|
||||
**字段定义**
|
||||
|
||||
- `ResearchConfig`（字段定义）：research.json 的全部字段，全部必填，不认识的字段报错。每个字段带中文说明（`description`）和类别（`category`：用户配置或方法参数），配置对话照着它问，不另写一份。`models` 里除了各角色的模型，还有 Jev 用的模型名（`models.jev`，用户配置）；Jev 的密钥不进配置，从环境变量读。流程是两选一：完整流程写四个开关、选批规则、它才用得到的 λ、u、衰减；朴素循环只写 `preset`，别的都写不进去。
|
||||
- `ResearchConfig`（字段定义）：research.json 的全部字段，全部必填，不认识的字段报错。每个字段带中文说明（`description`）和类别（`category`：用户配置或方法参数），配置对话照着它问，不另写一份。`models` 里除了各角色的模型，还有 Jev 用的模型名（`models.jev`，用户配置）；Jev 的密钥不进配置，从环境变量读。流程是两选一：完整流程写开关（含诊断：关，或开并写每人每轮诊断实验上限 A 和一次诊断最多做几个实验 D，两个都是方法参数）、选批规则、它才用得到的 λ、u、衰减；朴素循环只写 `preset`，别的都写不进去。
|
||||
- 研究编号 `studyId` 只许小写字母、数字和 `._-`，首字符是小写字母或数字：它会进 Docker 镜像名，镜像名只认小写。（来源：用户 2026-10-03 同意，任务书 13）
|
||||
- `task`、`promptDir` 必须是绝对路径：从别的目录接着跑一份研究时，相对路径会读到别处，而任务目录没有哈希核对挡着。（来源：用户 2026-10-03 同意，任务书 13）
|
||||
- `FIELD_CATEGORIES`、`FieldCategory`：字段的两种类别，用户配置、方法参数。
|
||||
@@ -467,7 +484,7 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
- `Step<S>`（类型）：一步：编号、规则（只有选批、选格有，其余为空）、执行或跳过及原因。
|
||||
- `RuleRef`（类型）：决策点用的规则，名字和版本。
|
||||
- `STUDY_STEPS`、`StudyStepId`：整份研究的步骤，按先后排。
|
||||
- `ROUND_STEPS`、`RoundStepId`：每一轮的步骤，按先后排；比第三部分拆得更细（例如"回应与统一"和"协议冻结"分成两步，挑刺与回应关掉时协议照样冻结）。
|
||||
- `ROUND_STEPS`、`RoundStepId`：每一轮的步骤，按先后排；诊断七步（`diagnosisProposal`、`diagnosisDedupe`、`diagnosisPrediction`、`selectDiagnosis`、`diagnosisImplementation`、`diagnosisRun`、`diagnosisSettle`）排在 `carryOverPool` 与 `proposal` 之间，`selectDiagnosis` 带规则名和版本；诊断关掉时七步都按配置跳过，消融关掉时 `explanation` 也跳过；比第三部分拆得更细（例如"回应与统一"和"协议冻结"分成两步，挑刺与回应关掉时协议照样冻结）。
|
||||
|
||||
### runner
|
||||
|
||||
@@ -528,13 +545,14 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
|
||||
下面是 `src/materials/index.ts` 导出的全部内容。谁能看什么每种任务一个函数，对着方法.md 第三部分、4.2、4.3 写，不写通用判定（用户 2026-10-02）。materials 只读账本，不写；任务材料由 flow 整份存进"模型调用开始"记录。依赖只有 ledger、rules、vocab；代码比较结果、运行产物、提示词由 flow 传进来。
|
||||
|
||||
- `buildMaterial(req)`：生成一份任务材料：给模型看的各节、可查询的记录、拼好的第一条消息、对号清单、能读的代码版本。研究者的材料按"任务说明 → 你自己的 → 公共材料 → 现在交什么"排；实验员、核验员的按"任务说明 → 分给你的材料 → 现在交什么"排，不给信念本和成绩。同一人重新发起同一任务时，最前面放上一次被退回的原因：按阶段、作者、提交种类、对象找本人最后一条退回记录。"现在交什么"由 vocab `SUBMISSION_SCHEMAS` 的字段说明生成。几种任务的取法：
|
||||
- `buildMaterial(req)`：生成一份任务材料：给模型看的各节、可查询的记录、拼好的第一条消息、对号清单、能读的代码版本。研究者的材料按"任务说明 → 你自己的 → 公共材料 → 现在交什么"排；实验员、核验员的按"任务说明 → 分给你的材料 → 现在交什么"排，不给信念本和成绩。同一人重新发起同一任务时，最前面放上一次被退回的原因：按阶段、作者、提交种类、对象找本人最后一条退回记录。"现在交什么"由 vocab `SUBMISSION_SCHEMAS` 的字段说明生成。`StudyFacts` 多一个 `diagnosisExperimentLimit`（每人每轮诊断实验上限 A，诊断关掉为 null），诊 1 的对号清单用它。几种任务的取法：
|
||||
- 讨论、认领：展示讨论原文时，原文引用的入口编号已经作废的（讨论时整份替换过清单），在它后面就地标"（已作废）"，读的人才分得清哪些编号还有效（用户 2026-10-03 同意，任务书 14 第 2.5 节）。前后紧挨字母、数字或"-"的不算，`e-x-1` 不会匹配到 `e-x-12` 里。
|
||||
- 第二次预测：冻结协议不带答复，"挑刺与答复"从这个候选已公开的最后一份回应或统一里读（回应另给答复正文）；留池候选读它冻结那一轮的（用户 2026-10-03 定）。
|
||||
- 核验：不放 Jev 对这份实现的判断。能走到核验说明 Jev 已判过，Jev 是判别条件，不是给核验员的提示。
|
||||
- 消融实现：这一格已有公开的消融核验（核验不通过、重做）时，和修复一样从上一次交的这一格代码接着改（用户 2026-10-03 定）：工作目录写"上一次交的这一格代码的副本"，能读的版本里加上它，完整代码照旧列出作对照；另放"退回意见"：最近一次核验里关错了的因子和核验员写的依据，写法同修复的退回意见。
|
||||
- 诊断（方法.md 第三部分、4.3 诊 1–诊 4）：诊 1 放当前基线的代码、产出它的那次运行的结果和原始答案（产物放进查询，flow 经 `artifacts` 传进来）、信念本、上一轮收尾，第 1 轮的第零轮记录在研究记录里；诊 2 放合并后的诊断实验（标明作者）和全部诊断解释，自己的解释另列，别人在这一步交的数一概不读；诊断实现、修复、核验只看选中的实验原文和当前基线代码（实现、修复另告诉原始产物已只读放进工作目录，位置在任务说明里；修复另放退回意见，只修没通过的实验；核验放当前基线到诊断程序的改动和位置表），不给信念本和成绩。从 ② 提案起的全部研究者任务，公共材料里在"基线"之后加"本轮诊断结果"一节（每个实验的做法、实测、用时、结局，每条解释确认或否定，推出的方向；本轮没诊断写原因：余额不到 2 次、配置关掉、信息价值不抵成本、选中的实验一个都没剩下）；诊断运行在诊断结算写入后公开，诊断程序进"能读的代码"，各人给的三个数不进任何记录或查询。实现、修复正向候选的实验员，本轮有诊断时多一个只读的诊断程序目录，注明"冻结协议引用诊断实验时作参考"。信念本加"诊断历史"，战绩加"诊断"一档；诊断实验的实测不算"手里最好"（诊断运行的运行记录不带判分，逐个实验的结果只在诊断结算里）。
|
||||
- 指标：冻结之后一律用全名"候选编号.短名"：已登记指标清单、对号清单的 `reusableMetrics`、实验员材料的"要实现的指标定义"（候选代码在结果文件里也用全名当键）。提案、回应、统一的对号清单里，已占用的指标编号只有任务的指标、`answer-path` 和 `safeguard-triggers` 两个保留键（用户 2026-10-03 定）；已登记的全名带 `.`，新登记的短名不带，不用列进来。
|
||||
- `Board`（类）：materials 读账本的那一份，构造参数是账本记录数组。flow 的"下一步做什么"是只拿得到记录数组的纯函数，它和 flow 的三个程序步骤都经这一份读账本，不另写一份（任务书 14 第 2.4 节）。flow 用到的：`stages`、`stage`（阶段的轮次、名单、关没关）、`published`（已公开的某种提交，最后一版为准）、`decision`（某轮某种决定）、`proposal`、`proposalsOf`、`authorsOf`（候选合并了哪些原提案、作者按编号排）、`answerOf`（冻结出候选的最后一份回应或统一）、`protocol`、`protocols`、`carried`（留池候选）、`baselineOf`（第 r 轮的本批基线）、`baselineScore`、`runStates`。任务书 14 写的是导出 `baselineOf`、`baselineScore` 两个函数，实际导出了整个类，原因是纯函数只拿得到记录数组，要用的远不止这两样（这一处偏离待用户确认）。
|
||||
- `Board`（类）：materials 读账本的那一份，构造参数是账本记录数组。flow 的"下一步做什么"是只拿得到记录数组的纯函数，它和 flow 的三个程序步骤都经这一份读账本，不另写一份（任务书 14 第 2.4 节）。flow 用到的：`stages`、`stage`（阶段的轮次、名单、关没关）、`published`（已公开的某种提交，最后一版为准）、`decision`（某轮某种决定）、`proposal`、`proposalsOf`、`authorsOf`（候选合并了哪些原提案、作者按编号排）、`answerOf`（冻结出候选的最后一份回应或统一）、`protocol`、`protocols`、`carried`（留池候选）、`baselineOf`（第 r 轮的本批基线）、`baselineScore`、`runStates`；诊断另有：`runStatesBefore(round)`（这一轮之前各次运行的状态，算诊断开始时的余额）、`diagnosisSettlement(round)`、`diagnosisSettlements()`、`diagnosisDone(round)`（提案阶段开了，就是诊断这一段走完了或本轮不诊断）、`diagnosisExplanations(round)`（已公开的诊断解释：编号、作者、正文）、`diagnosisExperiments(round)`（合并后的诊断实验：`MergedExperiment`，编号、全部作者、正文取作者里编号最小的那一份、登记的指标已改全名"实验编号.短名"）。任务书 14 写的是导出 `baselineOf`、`baselineScore` 两个函数，实际导出了整个类，原因是纯函数只拿得到记录数组，要用的远不止这两样（这一处偏离待用户确认）。
|
||||
- `protocolMetrics(board, protocol)`：把冻结协议自定义及引用的指标合成原登记定义，实验员材料与 flow 问 Jev、正式运行共用（来源：任务书 16 第 2.2、2.4 节）。
|
||||
- `finalEntries(ledger)`：按账本取的最终有效的入口（各人最后一版入口清单里的每个入口，带作者和正文）；flow 分配入口时用，和认领的对号清单是同一份（任务书 12 第三部分）。
|
||||
- `queryMaterial(material, id)`：只在这份材料的记录里按编号找，交回全文；找不到交回 null（agents 的查询回调据此答"看不到"）。
|
||||
@@ -542,12 +560,13 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
- `beliefBook(ledger, study, round)`：信念本全文（方法.md 6.10 五条），数字全部用 rules 现算，只写已公开的记录；`round` 是读者所在轮次，导出时写 null。选批历史的每次运行后面原样写出任务的附加记录（判分结果的 `extra`，合法化扣掉的分在里面，字段由任务定）。消融格记"证据不足"时，有运行的原因取判分，没有运行的（核验不通过、重做次数用完）原因取这一格的 `CELL_NOT_RUN` 事件。
|
||||
- `TaskMaterial`、`MaterialSection`、`MaterialRecord`、`ReadableVersion`（类型）：任务材料，以及其中的一节、一条可查询记录（编号、标题、全文）、一个能读的代码版本（版本号、说明、取出目录）。
|
||||
- `TaskRequest`、`StudyFacts`、`CodeChanges`、`Prompts`、`PromptFile`（类型）：flow 传进来的东西，全部必填；`CodeChanges` 与 workspace 的比较结果同形（照 jev `codeHunks` 的先例，不引用 workspace）。
|
||||
- 账本里只存成 Json 的内容的形状（用户 2026-10-03 决定 2：materials 定义，flow 照着写，ledger 不改）：`StudyRecord`（研究配置记录：配置、步骤清单、提示词每个文件的哈希、镜像编号、任务事实、种子；配置和任务事实存 Json，flow 读回时用 `checkConfig`、`checkTaskFacts` 再检查，任务书 12 第三部分）、`FrozenProtocol`（冻结协议：方法.md 4.3 ⑥ 列的因子、保障、指标、正文、实现条款，加轮次、出发的基线、合并了哪些原提案；不带答复，主体由 submissions 的 `protocolBody` 拼，指标编号是全名）、`LocationTable`（代码记录的位置表）、`DecisionKind`、`DECISION_KINDS`、`DecisionShapes`（六种决定：入口分配、判重合并、选批、采纳、选格、押的因子，输入输出用 rules 的类型；没采纳时也写一条采纳决定）、`DedupOutput`（判重合并的输出）、`Settlement`（结算）。
|
||||
- 账本里只存成 Json 的内容的形状（用户 2026-10-03 决定 2：materials 定义，flow 照着写，ledger 不改）：`StudyRecord`（研究配置记录：配置、步骤清单、提示词每个文件的哈希、镜像编号、任务事实、种子；配置和任务事实存 Json，flow 读回时用 `checkConfig`、`checkTaskFacts` 再检查，任务书 12 第三部分）、`FrozenProtocol`（冻结协议：方法.md 4.3 ⑥ 列的因子、保障、指标、正文、实现条款，加轮次、出发的基线、合并了哪些原提案；不带答复，主体由 submissions 的 `protocolBody` 拼，指标编号是全名）、`LocationTable`（代码记录的位置表）、`DecisionKind`、`DECISION_KINDS`、`DecisionShapes`（八种决定：入口分配、判重合并、选批、采纳、选格、押的因子、诊断实验判重合并（`diagnosisDedup`）、选诊断实验（`selectDiagnosis`），输入输出用 rules 的类型；没采纳时也写一条采纳决定）、`DedupOutput`（判重合并的输出）、`DiagnosisDedupOutput`（诊断实验判重合并的输出：合并后的每个实验由哪几个作者的哪几个实验合成，第一个的正文是合并后的正文）、`Settlement`（消融结算）、`DiagnosisSettlement`（诊断结算：运行编号、每个跑过的实验单独判分的结果（与 tasks 的 `DiagnosisJudged` 同形）、核验不通过而拿掉的实验和原因、`settleDiagnosis` 的输入输出）、`DiagnosisLocationTable`（诊断程序的位置表，每个选中的实验一项）。诊断运行的运行记录里 `judged` 写 null。
|
||||
- `freezeProtocol(body, candidate, round, baseline, proposals)`：flow 写冻结协议时调用。输入 submissions 交回的冻结协议主体、候选编号、轮次、基线、合并了哪些原提案，交回 `FrozenProtocol`；新登记的假设专用指标改用全名"候选编号.短名"（例如 `r2-c1.restart-gain`），引用的本来就是全名，照抄。加前缀只在这一处：同一轮的研究者互相看不到提案，会起同一个短名、定义却不同，候选编号判重合并之后才有，所以只能在冻结时加（用户 2026-10-03 定）。
|
||||
- `experimenterId(candidate)`、`verifierId(candidate)`：每个候选固定一位实验员、一位核验员，修复及再次核验沿用编号（来源：任务书 16 第 2.4 节）。
|
||||
- `FORWARD_NOT_RUN`、`forwardNotRunEvent(候选, 原因)`：正向候选修复用完仍未通过核验时的“不跑”事件，`objects` 只有该候选；不扣次（来源：任务书 16 第 2.2 节）。
|
||||
- 事件：`CELL_NOT_RUN`、`cellNotRunEvent(格, 原因)`：一格消融核验不通过、重做次数用完而不跑时，flow 用它写事件记录（`objects` 只有这一格），结算记证据不足（方法.md 4.3 ⑯、6.6）。
|
||||
- 编号规则：`ORIGINAL_BASELINE`、`researcherId`（第 i 位研究者，从 0 数，编号 `researcher-<i+1>`，任务书 12 第三部分）、`entryId`、`proposalId`、`critiqueId`、`candidateId`、`factorId`、`cellId`、`splitCell`、`explanationId`、`callId`，以及协议内的局部编号 `factorKey`（p1…）、`clauseKey`（c1…）。入口编号按作者一直往后编，讨论时整份替换也不复用旧编号。
|
||||
- 事件：`DIAGNOSIS_NOT_RUN`、`diagnosisNotRunEvent(轮次, 原因)`：选中的诊断实验核验都没通过、修复用完而一个都不剩时，flow 写这条事件（`objects` 只有本轮诊断的编号），不跑、不扣次；材料里"本轮没有诊断"的原因从它读。
|
||||
- 编号规则：`ORIGINAL_BASELINE`、`researcherId`（第 i 位研究者，从 0 数，编号 `researcher-<i+1>`，任务书 12 第三部分）、`entryId`、`proposalId`、`critiqueId`、`candidateId`、`factorId`、`cellId`、`splitCell`、`explanationId`，诊断的 `diagnosisId(轮次)`（本轮诊断：诊断程序的冻结代码、诊断运行、诊断实现与核验的对象）、`diagnosisKey(i)`（合并后第 i 个诊断实验，d1…）、`diagnosisExplanationId(轮次, 作者, i)`（诊断解释，`r1-researcher-1-dx1`），`callId`，以及协议内的局部编号 `factorKey`（p1…）、`clauseKey`（c1…）。入口编号按作者一直往后编，讨论时整份替换也不复用旧编号。
|
||||
|
||||
### agents
|
||||
|
||||
@@ -563,12 +582,15 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
|
||||
### jev
|
||||
|
||||
下面是 `src/jev/index.ts` 导出的全部内容。九问的问法和阈值照参考项目，只改了说法和字段名；阈值没有在新项目的材料上校准过，等真实研究的数据（任务书 08 第三部分）。模型名由调用方从 config 的 `models.jev` 传入，密钥从环境变量 `TYPESAFE_API_KEY` 读进请求头。调不通时在发请求那一处有限次重试，仍不通就报错；回答读不懂报错并带请求编号。
|
||||
下面是 `src/jev/index.ts` 导出的全部内容。九问的问法和阈值照参考项目，只改了说法和字段名；诊断实验另有四问（判重一问，冻结检查三问），照判重、冻结条款的写法，阈值暂用它们的值；阈值都没有在新项目的材料上校准过，等真实研究的数据（任务书 08 第三部分、任务书 18 第三块）。模型名由调用方从 config 的 `models.jev` 传入，密钥从环境变量 `TYPESAFE_API_KEY` 读进请求头。调不通时在发请求那一处有限次重试，仍不通就报错；回答读不懂报错并带请求编号。
|
||||
|
||||
- `judgeSameChange(model, a, b)`：③ 判重，两条提案的因子清单是不是同一个改法，一对一次请求，交回一条 `JevJudgment`。
|
||||
- `checkProtocol(model, protocol)`：⑥ 查冻结条款，三问（改动了评测、自相矛盾、太含糊）各一次请求，交回 `JevCheck`。
|
||||
- `checkHunk(model, hunk, protocol)`：⑨ 查一段代码改动，五问（不固定种子的随机数、没登记就读时钟、算登记指标的方式与定义不符、比登记的停止条件更早结束、测不到的量写成 0）同一份材料一次请求，交回 `JevCheck`。对照的指标取自冻结协议，编号是全名。
|
||||
- `judgeSameExperiment(model, a, b)`：诊 2 诊断实验判重，两个实验（名称、动哪一处、怎么动）是不是同一个实验，一对一次请求，交回一条 `JevJudgment`；阈值暂用判重的 0.3，待校准。
|
||||
- `checkExperiment(model, experiment)`：诊 1 查一个诊断实验，改动了评测、自相矛盾、太含糊三问各一次请求，用冻结三问的诊断变体（材料是"一个诊断实验"，不是冻结协议），交回 `JevCheck`；阈值暂用冻结三问的值，待校准。诊断程序的代码改动直接用 `checkHunk` 的五问。
|
||||
- `codeHunks(comparison)`：把 workspace 的比较结果（文件清单和差异原文）切成差异块；没有 `@@` 行的改动（二进制、只改可执行位、空文件增删）不出块，由程序检查。
|
||||
- `DedupExperiment`、`DiagnosisExperiment`（类型）：诊断实验判重和冻结检查收的输入。
|
||||
- `DedupFactor`、`FrozenProtocol`、`ProtocolMetric`、`ProtocolRules`、`CodeHunk`（类型）：三个函数收的带类型输入，取自 vocab 的字段定义；材料由 jev 自己拼。`DedupFactor` 是因子的名称、做法和关键设定：问题写着"设定不同就不算同一个改法"，所以关键设定要给 Jev 看，判重的材料说明里也用一句话讲了 `keySettings` 是什么。
|
||||
- `JevCheck`（类型）：每一问的判断，以及有没有任何一问判"是"。
|
||||
|
||||
@@ -576,12 +598,13 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
|
||||
下面是 `src/submissions/index.ts` 导出的全部内容。submissions 判一份提交收下还是退回：先做程序检查（vocab 字段定义管不到的跨字段规则，问题一次列全），有一项不过就退回、不问 Jev；全过了再问 Jev，几问问完、有一问判"是"就退回。它不读也不写账本，只依赖 jev 和 vocab；要 workspace、tasks、materials 算的东西由 flow 算好传进来。Jev 调不通、回答读不懂时 jev 报错，这里不接住，flow 按内部错误停下。
|
||||
|
||||
- `receive(incoming)`：判一份提交，交回 `Verdict`。回应、统一先拼好冻结协议主体，程序检查、问 Jev、交回都用这同一份；回应、统一问冻结条款三问，实现、修复对每个差异块问代码改动五问，其余种类不问 Jev。Jev 退回的原因只列判"是"的几问，用 jev 的英文问题名（用户 2026-10-03 定），差异块附原文。
|
||||
- `receive(incoming)`：判一份提交，交回 `Verdict`。回应、统一先拼好冻结协议主体，程序检查、问 Jev、交回都用这同一份；回应、统一问冻结条款三问，实现、修复（含诊断程序）对每个差异块问代码改动五问，诊断解释与实验对每个实验问诊断变体的冻结三问，其余种类不问 Jev。诊断另有四条程序检查：解释 1 至 M 条、实验 1 至 A 个；诊断预测覆盖自己每条解释 × 每个合并后的实验、不多不少、三个数严格递增；诊断实验登记的指标只能是"缺了只记缺失"；诊断核验的逐个实验结果覆盖全部选中的实验。Jev 退回的原因只列判"是"的几问，用 jev 的英文问题名（用户 2026-10-03 定），差异块附原文。
|
||||
- `protocolBody(from)`：拼冻结协议主体，只在这一处拼。回应：因子、保障、指标、正文四项，`changes` 写了哪项就整项换新，没写的用原提案的，实现条款取回应自己的；统一：直接用统一提交自己的字段。答复不进协议（用户 2026-10-03 定）。
|
||||
- `verificationOutcome(v)`：一份已收下的核验通过没有：任何条款、因子、保障没落实，或任何没有归属的改动判为越界，就不通过；四类分开列（方法.md 4.3 ⑨）。
|
||||
- `ablationOutcome(v, off, alsoOff)`：一份已收下的消融核验通过没有：要关的因子和实验员报的一起关的因子都写了"关着"、其余都写了"照留"才通过；分开列该关却照留的和该留却关了的（方法.md 4.3 ⑯）。
|
||||
- `diagnosisOutcome(v)`：一份已收下的诊断核验逐个实验通过没有：`followed` 或 `nothingElse` 任一为 false 的实验不通过，分开列没照写法做的和动了写法以外地方的（方法.md 4.3 诊 4）。
|
||||
- `Incoming`（类型）：交给 `receive` 的一份提交，按提交种类区分：内容、这次任务的对号清单、提交之后才知道的东西（`KnownOf`）、Jev 的模型名。
|
||||
- `KnownOf<K>`（类型）：提交之后才知道的东西，没有的种类为 null：提案是本人本阶段已收下几条，回应是作者那条原提案，实现和修复是 `ForwardCode`，消融实现是 `AblationCode`（交了"跑不起来"时为 null）。
|
||||
- `KnownOf<K>`（类型）：提交之后才知道的东西，没有的种类为 null：提案是本人本阶段已收下几条，回应是作者那条原提案，实现和修复是 `ForwardCode`，消融实现是 `AblationCode`（交了"跑不起来"时为 null），诊断实现和修复也是 `ForwardCode`。
|
||||
- `ForwardCode`（类型）：实现、修复的代码：存不成版本时带路径和原因；存成了带改了的不许改的文件、和本批基线比出的差异块、给 Jev 对照的协议条款。
|
||||
- `AblationCode`（类型）：消融实现的代码：存不成版本时带路径和原因；存成了只带改了的不许改的文件。
|
||||
- `OriginalProposal`（类型）：回应拼协议时叠在下面的原提案：因子、保障、指标、正文，作者自评不进协议。
|
||||
@@ -590,7 +613,7 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
|
||||
### flow
|
||||
|
||||
下面是 `src/flow/index.ts` 导出的全部内容，以后 plugin 也只用它。flow 只经各模块的 `index.ts` 用它们；正式代码里没有为检验写的参数、环境变量、回放模式或假实现（CLAUDE.md"测试"）。这一节对应 flow 第 1–4 部分（任务书 12、14、16、17）：骨架、研究开始、第零轮、第 r 轮从开轮到收尾、一轮接一轮直到余额用完；余额为 0 时走到 `export` 停下，原因"导出还没写"（导出还没写）。轮次按"第 r 轮"写，不写死轮数。
|
||||
下面是 `src/flow/index.ts` 导出的全部内容，以后 plugin 也只用它。flow 只经各模块的 `index.ts` 用它们；正式代码里没有为检验写的参数、环境变量、回放模式或假实现（CLAUDE.md"测试"）。这一节对应 flow 第 1–5 部分（任务书 12、14、16、17、18）：骨架、研究开始、第零轮、第 r 轮从开轮（含诊断）到收尾、一轮接一轮直到余额用完；余额为 0 时走到 `export` 停下，原因"导出还没写"（导出还没写）。轮次按"第 r 轮"写，不写死轮数。
|
||||
|
||||
- `runStudy(input)`：推进一份研究，直到停下。输入三样，都必填：研究目录；配置（新研究交 `checkConfig` 检查过的配置，接着跑一份已有的研究写 null，以账本里冻结的那份为准；两者对不上就报错）；"停下"信号。每次启动走同一条路：打开账本；新研究先冻结配置，本机已有 `inforsi-<研究编号>` 的镜像就说明研究编号用过了，报错停下，要换一个编号（编号重复的新研究会把旧研究遗留的同名容器当成自己的运行）；已有研究的，核对冻结的配置、镜像还在、提示词哈希没变（开发入口开着时提示词可以变）；给开始了没结束的模型调用补写结束（出错、"进程中断"、墙钟时间测不到）；按容器名问 Docker 核对登记了没结果的正式运行（没启动的写"没启动"退还名额，启动了的补写确认启动、等它、判分、写结果，问不到的写"状态未知"事件并停下）；然后进主循环：读账本 → 问"下一步做什么" → 做这一步 → 回到开头。主循环里 `docker run` 没启动的，写完"没启动"结果（退还名额、记下 Docker 报错原文）就停下，停下原因带上报错原文，本进程不重跑；下次启动时照旧起新编号重跑。恢复时补写的"没启动"不停下。
|
||||
- `StudyInput`（类型）：上面三样输入。
|
||||
@@ -607,6 +630,12 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
| `entryDraft`、`discussion1`、`discussion2`、`claim` | 这个阶段的关闭记录（`stageClose`）；阶段编号 `r0-<步骤>` |
|
||||
| `assignEntries` 分配 | 第 0 轮的入口分配决定（`decision`，种类 `entries`） |
|
||||
| `openRound`、`carryOverPool` | 不写记录：本轮名单和提案上限就是冻结的配置（方法.md 4.3 ①），留池候选由 materials 从记录算出（`Board.carried`） |
|
||||
| `diagnosisProposal`、`diagnosisPrediction` 诊 1、诊 2 | 这一轮这个阶段的关闭记录；阶段编号 `r<轮>-diagnosisProposal`、`r<轮>-diagnosisPrediction`。余额只算本轮开始之前的运行，不到 2 次时本轮不诊断：诊 1–诊 4 不开阶段、不写任何记录，材料里的"本轮没有诊断：余额只剩……"从余额现算 |
|
||||
| `diagnosisDedupe` 诊 2 判重合并 | 本轮的诊断实验判重合并决定（`decision`，种类 `diagnosisDedup`） |
|
||||
| `selectDiagnosis` 诊 3 | 本轮的选诊断实验决定（`decision`，种类 `selectDiagnosis`）；定了不诊断也写，之后诊 4 的几步都不做 |
|
||||
| `diagnosisImplementation` 诊 4 实现、核验、修复 | 最近一次核验里至少有一个实验通过（不通过的拿掉）；或修复次数用完、一个都不剩而且有"不跑"事件（`diagnosisNotRun`，`objects` 是本轮诊断）。阶段编号 `r<轮>-diagnosisImplementation-1`、`r<轮>-diagnosisVerification-1`、`r<轮>-diagnosisRepair-1`、`r<轮>-diagnosisVerification-2`……第几次从账本数，不另存 |
|
||||
| `diagnosisRun` 诊 4 运行 | 本轮诊断运行最近一次有结果，而且不是"没启动"；没有诊断、或实验一个都没剩下时不登记 |
|
||||
| `diagnosisSettle` 诊 4 结算 | 本轮的诊断结算记录（`diagnosisSettlement`，按轮次唯一）；没有诊断、或实验一个都没剩下时不写 |
|
||||
| `proposal`、`firstPrediction`、`critique`、`response`、`secondPrediction` | 这一轮这个阶段的关闭记录（`stageClose`）；阶段编号 `r<轮>-<阶段>` |
|
||||
| `dedupe` 判重合并 | 本轮的判重合并决定（`decision`，种类 `dedup`） |
|
||||
| `freezeProtocol` 冻结协议 | 本轮每个新候选都有冻结协议记录（`protocol`）；全部候选在同一个事务里写 |
|
||||
@@ -642,6 +671,10 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
| 实现 | 每个入选候选一个实验员，从本批基线取工作目录，不同候选同时实现 | 候选 |
|
||||
| 核验 | 每个待核验候选一个核验员；首次核验本批全部，之后只核验刚修复的候选 | 候选 |
|
||||
| 修复 | 只给核验不通过的候选，交原实验员，从上次提交的代码接着改 | 候选 |
|
||||
| 诊断解释与实验、诊断预测 | 每位研究者一个任务 | 无 |
|
||||
| 诊断实现 | 一位实验员（`experimenter-<本轮诊断>`），把选中的全部实验写进一份诊断程序，从当前基线取工作目录，再放进只读的原始产物 | 本轮诊断（`r<轮>-diagnosis`） |
|
||||
| 诊断核验 | 一位核验员（`verifier-<本轮诊断>`），逐个实验交结果；每次核验都覆盖全部选中的实验 | 本轮诊断 |
|
||||
| 诊断修复 | 只在最近一次核验有实验不通过时，交原实验员，从上次提交的诊断程序接着改，只修不通过的实验；修复后再核验，次数用同一个 `maxRepairs` | 本轮诊断 |
|
||||
| 解释 | 每位研究者一个任务 | 无 |
|
||||
| 消融实现 | 每个选中的格一个实验员；第一次从被采纳的完整代码取工作目录，重做时从上一次交的这一格代码接着改；不同格同时实现 | 消融格 |
|
||||
| 消融核验 | 每个交了代码的格一个核验员；首次核验全部，之后只核验刚重做的格 | 消融格 |
|
||||
@@ -680,6 +713,18 @@ Jev 是一个判别器，和程序检查地位相同，不单独成一种记录
|
||||
- 通过核验的候选用既有 `formalRun`，用途 `forward`；先登记占名额，确认启动才扣次，再独立判分记结果。**一批内一个接一个跑**，避免抢 CPU 改变按秒停下的程序成绩。没启动的处理同基线：本进程停下，下次启动起新编号。
|
||||
- 本批全部有结果或“不跑”事件后，用 rules 的 `adopt`（内部用 `adoptionLine`），只传有效实测分数、本批基线分数和任务的方向、ε、已知最好；写种类 `adopt` 的决定，即使 `adopted` 为 null 也写。materials 据此公开本轮正向运行。
|
||||
|
||||
**诊断**（来源：任务书 18 第 1.6 节，方法.md 4.3 诊 1–诊 4、6.12）：
|
||||
|
||||
- 做不做：余额（`runBalance`，只算本轮开始之前的运行，用 `Board.runStatesBefore`）不到 2 次时，诊 1–诊 4 都不做。余额不算本轮诊断运行自己扣的次数，所以诊断这一段的后几步不会因为扣次后余额变小被误算成"不诊断"。诊断关着时这几步按配置跳过，和此前一致。
|
||||
- 诊 2 判重合并：本轮已公开的诊断实验按作者编号、第几个排好，两两问一次 Jev（jev 的 `judgeSameExperiment`），全部问完才写决定（种类 `diagnosisDedup`，规则 `jevSameExperiment` 版本 1，输入是每一对的 Jev 判断）。判为相同的连通成组（同判重合并），每组一个实验，编号 `d1`、`d2`……，组内第一个（作者编号最小、再按第几个）的正文就是合并后的正文。
|
||||
- 诊 3 选实验：输入照 rules 的 `DiagnosisRuleInput` 拼（余额取本轮开始之前，种子不用，写 `null`），规则取步骤清单里 `selectDiagnosis` 那一步的，调 `runDiagnosisRule`；定了不诊断也写一条决定。
|
||||
- 诊 4 实现、核验、修复：对象都是本轮诊断（`diagnosisId`）。实验员从当前基线取工作目录，tasks 的 `placeBaselineArtifact` 把产出当前基线那次运行的原始产物只读地放进去；存版本时用 tasks 的 `diagnosisUnversionedNames`（多跳过这个目录）。提交时相对当前基线比较、切块，Jev 对照的是选中实验的原文和全名指标；收下的提交和代码版本、位置表在同一事务写入。核验员逐个实验交结果，用 submissions 的 `diagnosisOutcome` 判；核验、修复的材料差异都是当前基线到诊断程序。
|
||||
- 拿掉实验：修复次数用完后最近一次核验仍不通过的实验拿掉，结算里记"证据不足：核验不通过"和原因。诊断程序的代码不动（核验员核过的就是这一份），运行时只调通过的实验的编号，每个实验分到总时限 ÷ 跑的实验数；拿掉的实验的代码留在程序里不被调用。判"通过"以最近一次核验为准，修复前通过过、修复后又不通过的实验也拿掉。一个都不剩时用 `diagnosisNotRunEvent` 写事件，不登记运行、不扣次、不结算。
|
||||
- 诊断运行：用 `formalRun` 同一个登记、启动、扣次入口（formal-run.ts 的 `diagnosisFormalRun`，用途 `diagnosis`，对象本轮诊断），一次运行一次扣次；没启动的处理同正向。运行记录的 `judged` 写 null，逐个实验的判分在结算前由 `judgeDiagnosisRun` 从运行目录做。诊断运行编号 `r<轮>-diagnosis-<第几次>`。
|
||||
- 诊断结算：逐个实验判分，再照 `SettleDiagnosisInput` 调 `settleDiagnosis` 写一条诊断结算（运行编号、逐个实验的判分、拿掉的实验和原因、输入输出）。每个合并后的实验的结局：跑了且有效分数是有效；跑了没有效分数、或核验不通过拿掉的是证据不足；没选中的是没做。诊断结算写下、诊断运行公开之后才开 ② 提案。
|
||||
- 恢复：和其他运行走同一条恢复路径。登记了还没结果的诊断运行，按容器名问 Docker：没启动的写"没启动"退还名额，已启动的补写确认启动（扣次）、等它、写结果（`judged` 为 null）；只核对、不重跑。结算前被打断，恢复后从运行目录重新判分。
|
||||
- 诊 1 的材料要产出当前基线那次运行的产物（`metrics.json`、`solution.json`、日志），flow 读运行目录传给 materials。
|
||||
|
||||
**消融、收尾与研究结束**（来源：任务书 17 第 2.2、2.3 节，方法.md 4.3 ⑬–⑱、4.4、第七部分）：
|
||||
|
||||
- 采纳了而且被采纳改法有 2 个以上因子才走消融；否则 ⑬–⑰ 不写记录（从采纳决定和冻结协议算得出），直接到收尾。
|
||||
|
||||
@@ -0,0 +1,34 @@
|
||||
{
|
||||
"studyId": "diag-eval",
|
||||
"task": "/home/iomgaa/Projects/InfoRSI/tasks/circle_packing",
|
||||
"models": {
|
||||
"researchers": [
|
||||
{ "provider": "iomgaa", "model": "glm-5.3", "thinking": "medium" },
|
||||
{ "provider": "iomgaa", "model": "glm-5.3", "thinking": "medium" },
|
||||
{ "provider": "iomgaa", "model": "glm-5.3", "thinking": "medium" }
|
||||
],
|
||||
"experimenter": { "provider": "iomgaa", "model": "glm-5.3", "thinking": "medium" },
|
||||
"verifier": { "provider": "iomgaa", "model": "glm-5.3", "thinking": "medium" },
|
||||
"jev": "jev-latest"
|
||||
},
|
||||
"proposalLimit": 1,
|
||||
"batchLimit": 2,
|
||||
"runs": 13,
|
||||
"factorLimit": 3,
|
||||
"maxRepairs": 1,
|
||||
"lightChecks": { "maxCount": 2, "seconds": 5 },
|
||||
"promptDir": "/home/iomgaa/Projects/InfoRSI/prompts",
|
||||
"devEntry": false,
|
||||
"flow": {
|
||||
"preset": "full",
|
||||
"roundZero": { "enabled": true, "discussions": 2 },
|
||||
"firstPrediction": true,
|
||||
"critique": true,
|
||||
"diagnosis": { "enabled": true, "perResearcher": 2, "maxExperiments": 4 },
|
||||
"ablation": { "enabled": false },
|
||||
"batchRule": "bestExpected",
|
||||
"openReserve": 0.1,
|
||||
"knowledgeUse": 0.5,
|
||||
"decay": 0.5
|
||||
}
|
||||
}
|
||||
@@ -24,6 +24,7 @@
|
||||
"roundZero": { "enabled": true, "discussions": 2 },
|
||||
"firstPrediction": true,
|
||||
"critique": true,
|
||||
"diagnosis": { "enabled": false },
|
||||
"ablation": { "enabled": true, "cellRule": "valueMinusCost" },
|
||||
"batchRule": "bestExpected",
|
||||
"openReserve": 0.1,
|
||||
|
||||
+17
-4
@@ -12,6 +12,7 @@ export const FIELD_CATEGORIES = ["用户配置", "方法参数"] as const;
|
||||
export type FieldCategory = (typeof FIELD_CATEGORIES)[number];
|
||||
|
||||
const CLOSED = { additionalProperties: false } as const;
|
||||
const PositiveInteger = Type.Integer({ minimum: 1 });
|
||||
/** 文字字段：不能为空，也不能只有空白。 */
|
||||
const TEXT = { pattern: "\\S" };
|
||||
/**
|
||||
@@ -68,7 +69,7 @@ const RoundZero = Type.Union([
|
||||
/** 消融（方法.md 第十一部分"步骤开关"）：关，或开并写选格规则。 */
|
||||
const Ablation = Type.Union([
|
||||
Type.Object(
|
||||
{ enabled: method(Type.Literal(false), "消融关：不列消融格；解释照样写文字，但不给每格的三个数") },
|
||||
{ enabled: method(Type.Literal(false), "消融关：不列消融格，也不写解释（解释就是对消融格的预测，方法.md 4.3 ⑫）") },
|
||||
CLOSED,
|
||||
),
|
||||
Type.Object(
|
||||
@@ -83,6 +84,19 @@ const Ablation = Type.Union([
|
||||
),
|
||||
]);
|
||||
|
||||
/** 诊断（方法.md 4.3 诊 1–诊 4、第十一部分"步骤开关"）：关，或开并写两个方法参数（6.11）。 */
|
||||
const Diagnosis = Type.Union([
|
||||
Type.Object({ enabled: method(Type.Literal(false), "诊断关：每轮提案之前不做诊断实验") }, CLOSED),
|
||||
Type.Object(
|
||||
{
|
||||
enabled: method(Type.Literal(true), "诊断开"),
|
||||
perResearcher: method(PositiveInteger, "每人每轮诊断实验上限 A；现行值 2（方法.md 6.11）"),
|
||||
maxExperiments: method(PositiveInteger, "一次诊断最多做几个实验 D；现行值 4（方法.md 6.11）"),
|
||||
},
|
||||
CLOSED,
|
||||
),
|
||||
]);
|
||||
|
||||
/**
|
||||
* 两种流程（方法.md 第十一部分）。完整流程写开关、规则和它才用得到的方法参数；
|
||||
* 朴素循环整套照搬，不能写任何开关、规则，也不写它用不到的 λ、u、衰减（它不预测、不消融）。
|
||||
@@ -96,6 +110,7 @@ const FullFlow = Type.Object(
|
||||
Type.Boolean(),
|
||||
"挑刺与回应开不开（方法.md 4.3 ⑤ ⑥）；关掉时不挑刺、不回应，协议照样冻结；现行值 true",
|
||||
),
|
||||
diagnosis: method(Diagnosis, "诊断（方法.md 4.3 诊 1–诊 4）"),
|
||||
ablation: method(Ablation, "消融（方法.md 4.3 ⑬–⑰）"),
|
||||
batchRule: method(
|
||||
Type.Enum(BATCH_RULE_NAMES),
|
||||
@@ -121,14 +136,12 @@ const NaiveFlow = Type.Object(
|
||||
{
|
||||
preset: method(
|
||||
Type.Literal("naive"),
|
||||
"朴素循环：方法.md 第十一部分的对照流程整套照搬，不挑刺、不回应、不预测、均匀随机选批、不消融",
|
||||
"朴素循环：方法.md 第十一部分的对照流程整套照搬，不挑刺、不回应、不预测、均匀随机选批、不诊断、不消融",
|
||||
),
|
||||
},
|
||||
CLOSED,
|
||||
);
|
||||
|
||||
const PositiveInteger = Type.Integer({ minimum: 1 });
|
||||
|
||||
export const ResearchConfig = Type.Object(
|
||||
{
|
||||
studyId: user(
|
||||
|
||||
+41
-8
@@ -1,4 +1,11 @@
|
||||
import { BATCH_RULES, CELL_RULES, type BatchRuleName, type CellRuleName } from "../rules/index.ts";
|
||||
import {
|
||||
BATCH_RULES,
|
||||
CELL_RULES,
|
||||
DIAGNOSIS_RULES,
|
||||
type BatchRuleName,
|
||||
type CellRuleName,
|
||||
type DiagnosisRuleName,
|
||||
} from "../rules/index.ts";
|
||||
import type { ResearchConfig } from "./schema.ts";
|
||||
|
||||
// 步骤清单（docs/结构.md 第三部分，方法.md 第十一部分）：把配置里的流程、开关、规则翻译成这份研究要走的步骤。
|
||||
@@ -23,6 +30,13 @@ export type StudyStepId = (typeof STUDY_STEPS)[number];
|
||||
export const ROUND_STEPS = [
|
||||
"openRound", // 1. 开轮：冻结本轮名单和提案上限
|
||||
"carryOverPool", // 1. 开轮：处理留池候选
|
||||
"diagnosisProposal", // 诊 1. 诊断解释与诊断实验
|
||||
"diagnosisDedupe", // 诊 2. 判重合并
|
||||
"diagnosisPrediction", // 诊 2. 诊断预测
|
||||
"selectDiagnosis", // 诊 3. 选诊断实验（程序）
|
||||
"diagnosisImplementation", // 诊 4. 实现、核验、修复
|
||||
"diagnosisRun", // 诊 4. 诊断运行
|
||||
"diagnosisSettle", // 诊 4. 诊断结算
|
||||
"proposal", // 2. 提案
|
||||
"dedupe", // 3. 判重合并
|
||||
"firstPrediction", // 4. 第一次预测（看原始提案）
|
||||
@@ -50,11 +64,11 @@ export const ROUND_STEPS = [
|
||||
export type RoundStepId = (typeof ROUND_STEPS)[number];
|
||||
|
||||
/** 决策点用哪条规则：规则名和版本，从 rules 的登记表取。 */
|
||||
export type RuleRef = { name: BatchRuleName | CellRuleName; version: number };
|
||||
export type RuleRef = { name: BatchRuleName | CellRuleName | DiagnosisRuleName; version: number };
|
||||
|
||||
type Status = { status: "execute" } | { status: "skip"; reason: string };
|
||||
|
||||
/** 一步：执行，或按配置跳过（带原因）；只有选批、选格两个决策点带规则，其余为 null。 */
|
||||
/** 一步：执行，或按配置跳过（带原因）；只有选批、选格、选诊断实验三个决策点带规则，其余为 null。 */
|
||||
export type Step<S extends string> = { id: S; rule: RuleRef | null } & Status;
|
||||
|
||||
export interface StepList {
|
||||
@@ -82,7 +96,7 @@ export function stepList(config: ResearchConfig): StepList {
|
||||
const flow = config.flow;
|
||||
if (flow.preset === "naive") {
|
||||
const naive = skip("朴素循环不做这一步（方法.md 第十一部分）");
|
||||
return assemble("naive", (id) => (NAIVE_STEPS.has(id) ? EXECUTE : naive), batchRule("uniformRandom"), null);
|
||||
return assemble("naive", (id) => (NAIVE_STEPS.has(id) ? EXECUTE : naive), batchRule("uniformRandom"), null, null);
|
||||
}
|
||||
|
||||
// 完整流程：只有开关关掉的步骤跳过。
|
||||
@@ -102,15 +116,27 @@ export function stepList(config: ResearchConfig): StepList {
|
||||
if (!flow.firstPrediction) off(["firstPrediction"], "配置关掉了第一次预测");
|
||||
// 挑刺与回应关掉时，协议照样冻结（freezeProtocol 不跟着关）。
|
||||
if (!flow.critique) off(["critique", "response"], "配置关掉了挑刺与回应");
|
||||
// 消融关掉时解释照样写文字，只是不给每格的三个数（方法.md 第十一部分"步骤开关"），所以 explanation 不跟着关。
|
||||
// 消融关掉时，解释也不做：解释就是对消融格的预测，⑬–⑰ 整段都不做（方法.md 4.3 ⑫）。
|
||||
if (!flow.ablation.enabled) {
|
||||
off(
|
||||
["listCells", "selectCells", "ablationImplementation", "ablationVerification", "ablationRun", "settle"],
|
||||
[
|
||||
"listCells", "explanation", "selectCells", "ablationImplementation",
|
||||
"ablationVerification", "ablationRun", "settle",
|
||||
],
|
||||
"配置关掉了消融",
|
||||
);
|
||||
}
|
||||
if (!flow.diagnosis.enabled) {
|
||||
off(
|
||||
[
|
||||
"diagnosisProposal", "diagnosisDedupe", "diagnosisPrediction", "selectDiagnosis",
|
||||
"diagnosisImplementation", "diagnosisRun", "diagnosisSettle",
|
||||
],
|
||||
"配置关掉了诊断",
|
||||
);
|
||||
}
|
||||
const cells = flow.ablation.enabled ? cellRule(flow.ablation.cellRule) : null;
|
||||
return assemble("full", (id) => skipped[id] ?? EXECUTE, batchRule(flow.batchRule), cells);
|
||||
return assemble("full", (id) => skipped[id] ?? EXECUTE, batchRule(flow.batchRule), cells, flow.diagnosis.enabled ? diagnosisRule() : null);
|
||||
}
|
||||
|
||||
function batchRule(name: BatchRuleName): RuleRef {
|
||||
@@ -121,13 +147,20 @@ function cellRule(name: CellRuleName): RuleRef {
|
||||
return { name, version: CELL_RULES[name].version };
|
||||
}
|
||||
|
||||
/** 选诊断实验只有一条规则：信息价值减成本（方法.md 6.12），配置里不选。 */
|
||||
function diagnosisRule(): RuleRef {
|
||||
return { name: "valueMinusCost", version: DIAGNOSIS_RULES.valueMinusCost.version };
|
||||
}
|
||||
|
||||
function assemble(
|
||||
preset: StepList["preset"],
|
||||
status: (id: StudyStepId | RoundStepId) => Status,
|
||||
batch: RuleRef,
|
||||
cells: RuleRef | null,
|
||||
diagnosis: RuleRef | null,
|
||||
): StepList {
|
||||
const ruleOf = (id: RoundStepId) => (id === "selectBatch" ? batch : id === "selectCells" ? cells : null);
|
||||
const ruleOf = (id: RoundStepId) =>
|
||||
id === "selectBatch" ? batch : id === "selectCells" ? cells : id === "selectDiagnosis" ? diagnosis : null;
|
||||
return {
|
||||
preset,
|
||||
study: STUDY_STEPS.map((id) => ({ id, rule: null, ...status(id) })),
|
||||
|
||||
+56
-6
@@ -1,7 +1,18 @@
|
||||
import { rmSync } from "node:fs";
|
||||
import type { Content } from "../ledger/index.ts";
|
||||
import { formalRun as runDefinition, judge, prepareJudging, prepareRun } from "../tasks/index.ts";
|
||||
import type { HypothesisMetric, JudgedResult, ProcessResult } from "../vocab/index.ts";
|
||||
import {
|
||||
diagnosisRun as diagnosisDefinition,
|
||||
formalRun as runDefinition,
|
||||
judge,
|
||||
judgeDiagnosis,
|
||||
prepareDiagnosisJudging,
|
||||
prepareDiagnosisRun,
|
||||
prepareJudging,
|
||||
prepareRun,
|
||||
type DiagnosisExperiment,
|
||||
type DiagnosisJudged,
|
||||
} from "../tasks/index.ts";
|
||||
import type { HypothesisMetric, JudgedResult, ProcessResult, RunDefinition } from "../vocab/index.ts";
|
||||
import { checkoutVersion } from "../workspace/index.ts";
|
||||
import { containerName, freshDir, judgeContainerName, runPaths, workspaceRepo } from "./paths.ts";
|
||||
import type { Study } from "./study.ts";
|
||||
@@ -9,6 +20,8 @@ import type { Study } from "./study.ts";
|
||||
// 共用动作二：跑一次正式运行（任务书 12 第 2.7 节）。先登记、再启动、再写结果，前后顺序由这个入口保证
|
||||
// （方法.md 第七、九部分）：登记占名额，确认启动才扣次，确认没启动的写结果"没启动"、退还名额。
|
||||
// 判分不扣次数，不登记。恢复时接着等一次已经启动的运行，走同一个 finishRun。
|
||||
// 诊断运行（一次运行跑几个实验，用途 diagnosis）走同一个登记、启动、等待的入口，只是准备目录和判分不同：
|
||||
// 运行记录不带判分（judged 为 null），逐个实验的判分在诊断结算前由 judgeDiagnosisRun 从运行目录做。
|
||||
|
||||
export type RunPlan = Omit<Content<"runPlanned">, "definition">;
|
||||
|
||||
@@ -22,11 +35,30 @@ export type RunOutcome = "done" | "stopped" | { notStarted: string };
|
||||
/** 跑一次正式运行：取出 version 的代码，在任务镜像里跑，判分，写结果。用户停下时不等容器，交回 "stopped"。 */
|
||||
export async function formalRun(s: Study, plan: RunPlan, version: string, metrics: HypothesisMetric[]): Promise<RunOutcome> {
|
||||
const definition = runDefinition(s.task, s.frozen.image, String(s.frozen.seed));
|
||||
return launch(s, plan, definition, metrics, (paths) => {
|
||||
checkoutVersion(workspaceRepo(s.dir), version, freshDir(paths.code));
|
||||
prepareRun(s.task, paths.code, freshDir(paths.run), metrics);
|
||||
});
|
||||
}
|
||||
|
||||
/**
|
||||
* 跑一次诊断运行：一个容器、一次登记、一次扣次，里面按顺序跑 experiments 里的实验（总时限平分）。
|
||||
* version 是诊断程序，baselineRunDir 是产出当前基线的那次运行的运行目录（原始产物只读地交给每个实验）。
|
||||
*/
|
||||
export async function diagnosisFormalRun(s: Study, plan: RunPlan, version: string, baselineRunDir: string, experiments: DiagnosisExperiment[]): Promise<RunOutcome> {
|
||||
const definition = diagnosisDefinition(s.task, s.frozen.image, String(s.frozen.seed), experiments.map((e) => e.id));
|
||||
return launch(s, plan, definition, null, (paths) => {
|
||||
checkoutVersion(workspaceRepo(s.dir), version, freshDir(paths.code));
|
||||
prepareDiagnosisRun(s.task, paths.code, baselineRunDir, freshDir(paths.run), experiments);
|
||||
});
|
||||
}
|
||||
|
||||
/** 先登记、准备运行目录、启动，确认启动才写"确认启动"；metrics 为 null 的是诊断运行（运行记录不带判分）。 */
|
||||
async function launch(s: Study, plan: RunPlan, definition: RunDefinition, metrics: HypothesisMetric[] | null, prepare: (paths: ReturnType<typeof runPaths>) => void): Promise<RunOutcome> {
|
||||
s.ledger.append("runPlanned", { ...plan, definition });
|
||||
const paths = runPaths(s.dir, plan.run);
|
||||
freshDir(paths.base);
|
||||
checkoutVersion(workspaceRepo(s.dir), version, freshDir(paths.code));
|
||||
prepareRun(s.task, paths.code, freshDir(paths.run), metrics);
|
||||
prepare(paths);
|
||||
const started = await s.runner.start(containerName(s.frozen.config.studyId, plan.run), definition, paths.run);
|
||||
if (!started.started) {
|
||||
s.ledger.append("runResult", { run: plan.run, process: { ending: "notStarted", reason: started.reason }, judged: null });
|
||||
@@ -39,13 +71,14 @@ export async function formalRun(s: Study, plan: RunPlan, version: string, metric
|
||||
/**
|
||||
* 等一次已经确认启动的运行结束，判分，写结果，删掉容器。日志写进这次运行自己的日志目录。
|
||||
* 收到"停下"信号就不等了，直接交回 "stopped"；容器不管它，下次启动时恢复入口再来接着等。
|
||||
* metrics 为 null 的是诊断运行：这里不判分，结果的 judged 写 null。
|
||||
*/
|
||||
export async function finishRun(s: Study, run: string, metrics: HypothesisMetric[]): Promise<"done" | "stopped"> {
|
||||
export async function finishRun(s: Study, run: string, metrics: HypothesisMetric[] | null): Promise<"done" | "stopped"> {
|
||||
const name = containerName(s.frozen.config.studyId, run);
|
||||
const paths = runPaths(s.dir, run);
|
||||
const waited = await untilStopped(s.stop, s.runner.wait(name, emptyDir(paths.logs)));
|
||||
if (waited === "stopped") return "stopped";
|
||||
const judged = await judgeRun(s, run, waited, metrics);
|
||||
const judged = metrics === null ? null : await judgeRun(s, run, waited, metrics);
|
||||
s.ledger.append("runResult", { run, process: waited, judged });
|
||||
await s.runner.remove(name);
|
||||
return "done";
|
||||
@@ -68,6 +101,23 @@ async function judgeRun(s: Study, run: string, formal: ProcessResult, metrics: H
|
||||
return judge(s.task, paths.run, paths.judge, formal, validator, metrics);
|
||||
}
|
||||
|
||||
/**
|
||||
* 诊断运行的判分（诊断结算前做）：每个交出答案的实验单独判分。同 judgeRun：判分目录每次重新建空的，
|
||||
* 判分容器还在就先删掉再用；恢复时从运行目录重新判，不存半截结果。experiments 是这次运行跑的那几个实验，次序同运行。
|
||||
*/
|
||||
export async function judgeDiagnosisRun(s: Study, run: string, process: ProcessResult, experiments: DiagnosisExperiment[]): Promise<DiagnosisJudged> {
|
||||
const paths = runPaths(s.dir, run);
|
||||
const step = prepareDiagnosisJudging(s.task, paths.run, emptyDir(paths.judge), s.frozen.image, process, experiments);
|
||||
if (step.status === "judged") return step.result;
|
||||
const name = judgeContainerName(s.frozen.config.studyId, run);
|
||||
if ((await s.runner.status(name)) !== "absent") await s.runner.remove(name);
|
||||
const started = await s.runner.start(name, step.definition, paths.judge);
|
||||
if (!started.started) throw new Error(`诊断运行 ${run} 的校验器容器没有启动：${started.reason}`);
|
||||
const validator = await s.runner.wait(name, emptyDir(paths.judgeLogs));
|
||||
await s.runner.remove(name);
|
||||
return judgeDiagnosis(s.task, paths.run, paths.judge, process, validator, experiments);
|
||||
}
|
||||
|
||||
/** 清空重建一个目录：恢复时会再等一次、再判一次，上一次没写完的日志和判分目录不留。 */
|
||||
function emptyDir(dir: string): string {
|
||||
rmSync(dir, { recursive: true, force: true });
|
||||
|
||||
+17
-2
@@ -1,7 +1,7 @@
|
||||
import { mkdirSync } from "node:fs";
|
||||
import { resolve } from "node:path";
|
||||
import type { ResearchConfig } from "../config/index.ts";
|
||||
import { forwardNotRunEvent, cellNotRunEvent } from "../materials/index.ts";
|
||||
import { cellNotRunEvent, diagnosisNotRunEvent, forwardNotRunEvent } from "../materials/index.ts";
|
||||
import { Ledger } from "../ledger/index.ts";
|
||||
import { Runner } from "../runner/index.ts";
|
||||
import { fullConfig, readFrozen } from "./frozen.ts";
|
||||
@@ -10,6 +10,7 @@ import { ledgerFile } from "./paths.ts";
|
||||
import { recover } from "./recover.ts";
|
||||
import { baselineRun, saveBaseline } from "./steps/baseline.ts";
|
||||
import { ablationRun, selectCells, settle, unsealChosen } from "./steps/ablation.ts";
|
||||
import { diagnosisDedupe, diagnosisRun, diagnosisSettle, selectDiagnosisExperiments } from "./steps/diagnosis.ts";
|
||||
import { forwardRun, adoptBatch } from "./steps/forward.ts";
|
||||
import { dispatchStage } from "./steps/dispatch.ts";
|
||||
import { dedupe, freezeProtocols, selectBatch } from "./steps/round.ts";
|
||||
@@ -80,10 +81,12 @@ async function doStep(s: Study, next: Exclude<Next, { do: "freezeConfig" }>): Pr
|
||||
return null;
|
||||
case "baselineRun":
|
||||
case "forwardRun":
|
||||
case "ablationRun": {
|
||||
case "ablationRun":
|
||||
case "diagnosisRun": {
|
||||
const ran =
|
||||
next.do === "baselineRun" ? await baselineRun(s, next.run)
|
||||
: next.do === "forwardRun" ? await forwardRun(s, next.round, next.candidate, next.run)
|
||||
: next.do === "diagnosisRun" ? await diagnosisRun(s, next.round, next.run)
|
||||
: await ablationRun(s, next.round, next.cell, next.run);
|
||||
if (ran === "stopped") return USER_STOPPED;
|
||||
if (ran === "done") return null;
|
||||
@@ -103,6 +106,18 @@ async function doStep(s: Study, next: Exclude<Next, { do: "freezeConfig" }>): Pr
|
||||
case "assignEntries":
|
||||
assignEntries(s);
|
||||
return null;
|
||||
case "diagnosisDedupe":
|
||||
await diagnosisDedupe(s, next.round);
|
||||
return null;
|
||||
case "selectDiagnosis":
|
||||
selectDiagnosisExperiments(s, next.round);
|
||||
return null;
|
||||
case "diagnosisNotRun":
|
||||
s.ledger.append("event", diagnosisNotRunEvent(next.round, next.reason));
|
||||
return null;
|
||||
case "diagnosisSettle":
|
||||
await diagnosisSettle(s, next.round);
|
||||
return null;
|
||||
case "dedupe":
|
||||
await dedupe(s, next.round);
|
||||
return null;
|
||||
|
||||
+24
-9
@@ -4,8 +4,9 @@ import { runSession, type SessionTask, type SessionEnding, type SubmitAttempt, t
|
||||
import { Board, buildMaterial, callId, queryMaterial, researcherId, type TaskMaterial } from "../materials/index.ts";
|
||||
import { receive, type Incoming } from "../submissions/index.ts";
|
||||
import { SUBMISSION_KINDS, type JevJudgment, type Json, type SubmissionKind } from "../vocab/index.ts";
|
||||
import { placeBaselineArtifact } from "../tasks/index.ts";
|
||||
import { checkoutVersion } from "../workspace/index.ts";
|
||||
import { prepareSubmission, taskCode, lightChecker } from "./task-code.ts";
|
||||
import { prepareSubmission, taskArtifact, taskCode, lightChecker } from "./task-code.ts";
|
||||
import { callPaths, freshDir, runPaths, workspaceRepo } from "./paths.ts";
|
||||
import type { Study } from "./study.ts";
|
||||
|
||||
@@ -43,7 +44,7 @@ export async function runModelTask<K extends SubmissionKind>(s: Study, t: ModelT
|
||||
prompts: s.prompts,
|
||||
codeRoot: paths.code,
|
||||
changes,
|
||||
artifacts: t.kind === "explanation" ? runArtifacts(s, t.stageId) : null,
|
||||
artifacts: t.kind === "explanation" ? roundArtifacts(s, t.stageId) : t.kind === "diagnosisProposal" ? baselineArtifacts(s, t.stageId) : null,
|
||||
});
|
||||
const model = role === "researcher" ? researcherModel(s, t.actor) : s.frozen.config.models[role];
|
||||
// 先登记、后建目录：调用编号只在登记之后才算用掉，被杀在两者之间时下次的编号不会撞上旧目录。
|
||||
@@ -60,6 +61,8 @@ export async function runModelTask<K extends SubmissionKind>(s: Study, t: ModelT
|
||||
for (const r of material.readable) checkoutVersion(workspaceRepo(s.dir), r.version, freshDir(r.dir));
|
||||
const workDir = freshDir(paths.work);
|
||||
if (startVersion !== null) checkoutVersion(workspaceRepo(s.dir), startVersion, workDir);
|
||||
const artifactFrom = taskArtifact(s, t);
|
||||
if (artifactFrom !== null) placeBaselineArtifact(s.task, artifactFrom, workDir);
|
||||
let reply = 0;
|
||||
const common = {
|
||||
model,
|
||||
@@ -90,22 +93,34 @@ export async function runModelTask<K extends SubmissionKind>(s: Study, t: ModelT
|
||||
* 解释任务要的产物（materials 的 TaskRequest.artifacts）：本轮每次已公开正向运行的 metrics.json、solution.json
|
||||
* 和标准输出、标准错误日志（运行编号 → 文件名 → 文字）。完整代码那次的答案出自哪里，就靠这些产物判断。
|
||||
*/
|
||||
function runArtifacts(s: Study, stageId: string): Record<string, Record<string, string>> {
|
||||
function roundArtifacts(s: Study, stageId: string): Record<string, Record<string, string>> {
|
||||
const board = new Board(s.ledger.records());
|
||||
const round = board.stage(stageId).round;
|
||||
const artifacts: Record<string, Record<string, string>> = {};
|
||||
for (const r of board.runs()) {
|
||||
if (r.plan.purpose !== "forward" || r.plan.round !== round || !board.runPublic(r)) continue;
|
||||
const paths = runPaths(s.dir, r.plan.run);
|
||||
const files: Record<string, string> = {};
|
||||
for (const [dir, name] of [[paths.run, "metrics.json"], [paths.run, "solution.json"], [paths.logs, "stdout.log"], [paths.logs, "stderr.log"]] as const) {
|
||||
if (existsSync(join(dir, name))) files[name] = readFileSync(join(dir, name), "utf8");
|
||||
}
|
||||
artifacts[r.plan.run] = files;
|
||||
artifacts[r.plan.run] = runFiles(s, r.plan.run);
|
||||
}
|
||||
return artifacts;
|
||||
}
|
||||
|
||||
/** 诊 1 要的产物：产出本轮当前基线的那次运行，同样的几个文件（运行编号 → 文件名 → 文字）。 */
|
||||
function baselineArtifacts(s: Study, stageId: string): Record<string, Record<string, string>> {
|
||||
const board = new Board(s.ledger.records());
|
||||
const run = board.runOf(board.baselineOf(board.stage(stageId).round))!.plan.run;
|
||||
return { [run]: runFiles(s, run) };
|
||||
}
|
||||
|
||||
/** 一次运行留下的产物：metrics.json、solution.json 和两份日志，有哪个读哪个。 */
|
||||
function runFiles(s: Study, run: string): Record<string, string> {
|
||||
const paths = runPaths(s.dir, run);
|
||||
const files: Record<string, string> = {};
|
||||
for (const [dir, name] of [[paths.run, "metrics.json"], [paths.run, "solution.json"], [paths.logs, "stdout.log"], [paths.logs, "stderr.log"]] as const) {
|
||||
if (existsSync(join(dir, name))) files[name] = readFileSync(join(dir, name), "utf8");
|
||||
}
|
||||
return files;
|
||||
}
|
||||
|
||||
/** 研究者用的模型：第 i 位研究者（researcherId(i)）用配置里第 i 个。 */
|
||||
function researcherModel(s: Study, actor: string) {
|
||||
const models = s.frozen.config.models.researchers;
|
||||
|
||||
+111
-7
@@ -1,8 +1,8 @@
|
||||
import type { RoundStepId, StudyStepId } from "../config/index.ts";
|
||||
import type { Content, LedgerRecord } from "../ledger/index.ts";
|
||||
import { Board, CELL_NOT_RUN, cellId, DECISION_KINDS, experimenterId, factorKey, FORWARD_NOT_RUN, ORIGINAL_BASELINE, researcherId, splitCell, verifierId } from "../materials/index.ts";
|
||||
import { Board, CELL_NOT_RUN, cellId, DECISION_KINDS, DIAGNOSIS_NOT_RUN, diagnosisId, experimenterId, factorKey, FORWARD_NOT_RUN, ORIGINAL_BASELINE, researcherId, splitCell, verifierId } from "../materials/index.ts";
|
||||
import { compareIds, listCells, runBalance, sharedSettings } from "../rules/index.ts";
|
||||
import { ablationOutcome, verificationOutcome } from "../submissions/index.ts";
|
||||
import { ablationOutcome, diagnosisOutcome, verificationOutcome } from "../submissions/index.ts";
|
||||
import { SUBMISSION_KINDS, type Stage, type SubmissionKind } from "../vocab/index.ts";
|
||||
import { readFrozen, type Frozen } from "./frozen.ts";
|
||||
|
||||
@@ -23,7 +23,7 @@ const ROUND_ZERO_STAGES = {
|
||||
type RoundZeroStep = keyof typeof ROUND_ZERO_STAGES;
|
||||
|
||||
/** 第 r 轮的研究者阶段：步骤名就是阶段名（vocab STAGES）。 */
|
||||
type RoundStageStep = Extract<RoundStepId, "proposal" | "firstPrediction" | "critique" | "response" | "secondPrediction">;
|
||||
type RoundStageStep = Extract<RoundStepId, "diagnosisProposal" | "diagnosisPrediction" | "proposal" | "firstPrediction" | "critique" | "response" | "secondPrediction">;
|
||||
|
||||
/** 阶段编号由 flow 起：r<轮>-<步骤>，一个步骤一个阶段。第 r 轮按配置跳过的步骤也用这个编号记"按配置跳过"。 */
|
||||
export const stageIdOf = (round: number, step: string): string => `r${round}-${step}`;
|
||||
@@ -52,6 +52,11 @@ export type Next =
|
||||
| { do: "closeStage"; stageId: string }
|
||||
| { do: "dispatch"; stageId: string; tasks: TaskRef[] }
|
||||
| { do: "assignEntries" }
|
||||
| { do: "diagnosisDedupe"; round: number }
|
||||
| { do: "selectDiagnosis"; round: number }
|
||||
| { do: "diagnosisNotRun"; round: number; reason: string }
|
||||
| { do: "diagnosisRun"; round: number; run: string }
|
||||
| { do: "diagnosisSettle"; round: number }
|
||||
| { do: "dedupe"; round: number }
|
||||
| { do: "freezeProtocols"; round: number }
|
||||
| { do: "selectBatch"; round: number }
|
||||
@@ -139,6 +144,28 @@ function roundStepNext(step: RoundStepId, round: number, view: View, frozen: Fro
|
||||
case "carryOverPool":
|
||||
// 不写记录：本轮名单和提案上限就是冻结的配置（方法.md 4.3 ①），留池候选由 materials 从记录算出（Board.carried）。
|
||||
return null;
|
||||
case "diagnosisProposal":
|
||||
case "diagnosisPrediction":
|
||||
// 诊 1、诊 2 → 这个阶段的关闭记录。余额不到 2 次（算到本轮开始前）本轮不诊断：不开阶段、不写记录。
|
||||
return diagnosing(view, round, frozen) ? stageNext(stageIdOf(round, step), step, round, (roster) => roundTasks(step, round, roster, view.board), view, frozen, researchers(frozen)) : null;
|
||||
case "diagnosisDedupe":
|
||||
// 诊 2 判重合并 → 本轮的诊断实验判重合并决定（decision，种类 diagnosisDedup）。
|
||||
if (!diagnosing(view, round, frozen)) return null;
|
||||
return view.board.decision("diagnosisDedup", round) === undefined ? { do: "diagnosisDedupe", round } : null;
|
||||
case "selectDiagnosis":
|
||||
// 诊 3 → 本轮的选诊断实验决定（decision，种类 selectDiagnosis），定了不诊断也写。
|
||||
if (!diagnosing(view, round, frozen)) return null;
|
||||
return view.board.decision("selectDiagnosis", round) === undefined ? { do: "selectDiagnosis", round } : null;
|
||||
case "diagnosisImplementation":
|
||||
// 诊 4 实现与核验 → 最近一次核验里至少有一个实验通过，或修复次数用完而一个都不剩、有"不跑"事件。
|
||||
return diagnosisCodeNext(round, view, frozen);
|
||||
case "diagnosisRun":
|
||||
// 诊 4 运行 → 本轮诊断运行最近一次有结果，而且不是"没启动"。
|
||||
return diagnosisProgramOf(view.board, round, frozen) === null ? null : diagnosisRunNext(round, view);
|
||||
case "diagnosisSettle":
|
||||
// 诊 4 结算 → 本轮的诊断结算记录（diagnosisSettlement，按轮次唯一）。
|
||||
if (diagnosisProgramOf(view.board, round, frozen) === null) return null;
|
||||
return view.board.diagnosisSettlement(round) === undefined ? { do: "diagnosisSettle", round } : null;
|
||||
case "proposal":
|
||||
case "firstPrediction":
|
||||
case "critique":
|
||||
@@ -183,9 +210,6 @@ function roundStepNext(step: RoundStepId, round: number, view: View, frozen: Fro
|
||||
// 解释 → 这一轮解释阶段的关闭记录。
|
||||
const adopted = view.board.decision("adopt", round)?.output.adopted ?? null;
|
||||
if (adopted === null || view.board.protocol(adopted).factors.length < 2) return null;
|
||||
if (frozen.steps.round.some((s) => s.id === "listCells" && s.status === "skip")) {
|
||||
return { do: "halt", reason: "配置关掉了消融、又采纳了多因子改法：不交三个数的解释还没写（cells 只按完整流程全开实现，做步骤开关时再补）", objects: [] };
|
||||
}
|
||||
return stageNext(stageIdOf(round, "explanation"), "explanation", round, (roster) => roster.map((actor) => ({ actor, kind: "explanation" as const, object: null })), view, frozen, researchers(frozen));
|
||||
}
|
||||
case "selectCells":
|
||||
@@ -228,6 +252,7 @@ export function poolOf(board: Board, round: number): string[] {
|
||||
|
||||
/**
|
||||
* 第 r 轮一个研究者阶段的任务（任务书 14 第 2.3 节）。同一个人的几个任务按这里的先后跑。
|
||||
* - 诊断解释与实验、诊断预测：每人一个任务，没有对象；
|
||||
* - 提案：每人一个任务，没有对象；
|
||||
* - 第一次预测：每人 × 本轮每个新候选（含自己的；留池候选不做，方法.md 4.3 ④）；
|
||||
* - 挑刺：每人 × 本轮每条不是自己写的原提案（监督者定，照参考项目）；
|
||||
@@ -238,6 +263,9 @@ function roundTasks(step: RoundStageStep, round: number, roster: readonly string
|
||||
const each = (kind: TaskRef["kind"], objects: readonly string[]): TaskRef[] =>
|
||||
roster.flatMap((actor) => objects.map((object) => ({ actor, kind, object })));
|
||||
switch (step) {
|
||||
case "diagnosisProposal":
|
||||
case "diagnosisPrediction":
|
||||
return roster.map((actor) => ({ actor, kind: step, object: null }));
|
||||
case "proposal":
|
||||
return roster.map((actor) => ({ actor, kind: "proposal", object: null }));
|
||||
case "firstPrediction":
|
||||
@@ -318,7 +346,7 @@ const researchers = (frozen: Frozen): string[] => frozen.config.models.researche
|
||||
function assignedStage(
|
||||
step: string,
|
||||
round: number,
|
||||
kind: "implementation" | "repair" | "verification" | "ablationImplementation" | "ablationVerification",
|
||||
kind: "implementation" | "repair" | "verification" | "ablationImplementation" | "ablationVerification" | "diagnosisImplementation" | "diagnosisRepair" | "diagnosisVerification",
|
||||
candidates: readonly string[],
|
||||
view: View,
|
||||
frozen: Frozen,
|
||||
@@ -355,6 +383,82 @@ function verificationNext(round: number, view: View, frozen: Frozen): Next | nul
|
||||
return null;
|
||||
}
|
||||
|
||||
/**
|
||||
* 第 round 轮诊不诊断的前提（方法.md 4.3 诊断一段）：余额至少 2 次，保证本轮至少还能跑 1 次正向。
|
||||
* 余额只算本轮开始前的运行：本轮诊断运行自己扣了次数以后，诊断这一段的后几步不会反过来算成"不诊断"。
|
||||
* 诊断关着的步骤清单把这几步都标了跳过，不会走到这里。
|
||||
*/
|
||||
function diagnosing(view: View, round: number, frozen: Frozen): boolean {
|
||||
return runBalance(frozen.config.runs, view.board.runStatesBefore(round)).balance >= 2;
|
||||
}
|
||||
|
||||
/**
|
||||
* 诊 4 实现、核验与修复（方法.md 4.3 诊 4）：一位实验员把选中的全部实验写进一份诊断程序，一位核验员逐个实验核对。
|
||||
* 阶段编号按"第几次"排：诊断实现-1、核验-1，不通过的退回修复-1、再核验-2……第几次从已开关的阶段数出来，不另存；
|
||||
* 每次核验都覆盖全部选中的实验。最近一次核验里还有实验不通过、修复次数又用完时，不通过的实验拿掉（diagnosisProgramOf）；
|
||||
* 一个都不剩才写"不跑"事件（方法.md 4.3 诊 4）。
|
||||
*/
|
||||
function diagnosisCodeNext(round: number, view: View, frozen: Frozen): Next | null {
|
||||
if (chosenExperiments(view.board, round) === null) return null;
|
||||
const object = diagnosisId(round);
|
||||
for (let n = 1; ; n++) {
|
||||
const work = n === 1
|
||||
? assignedStage("diagnosisImplementation-1", round, "diagnosisImplementation", [object], view, frozen)
|
||||
: assignedStage(`diagnosisRepair-${n - 1}`, round, "diagnosisRepair", [object], view, frozen);
|
||||
if (work !== null) return work;
|
||||
const verify = assignedStage(`diagnosisVerification-${n}`, round, "diagnosisVerification", [object], view, frozen);
|
||||
if (verify !== null) return verify;
|
||||
const verification = view.board.published("diagnosisVerification").filter((v) => v.round === round).at(-1)!;
|
||||
if (diagnosisOutcome(verification.content).passed) return null;
|
||||
if (n > frozen.config.maxRepairs) {
|
||||
const { experiments, dropped } = diagnosisProgram(view.board, round, frozen.config.maxRepairs);
|
||||
if (experiments.length > 0 || view.events(DIAGNOSIS_NOT_RUN).some((e) => e.objects.includes(object))) return null;
|
||||
return { do: "diagnosisNotRun", round, reason: `选中的实验第 ${n} 次核验都不通过，修复次数上限 ${frozen.config.maxRepairs} 已用完：${dropped.map((d) => `${d.experiment}：${d.reason}`).join("；")}` };
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/** 本轮诊断选中的实验（诊 3 的决定，按实验编号排）；本轮没有诊断（余额不到 2 次、定了不诊断）为 null。 */
|
||||
function chosenExperiments(board: Board, round: number): string[] | null {
|
||||
const decision = board.decision("selectDiagnosis", round);
|
||||
return decision === undefined || !decision.output.diagnose ? null : decision.output.chosen;
|
||||
}
|
||||
|
||||
/**
|
||||
* 诊 4 核验做完之后本轮诊断实际要跑的实验：选中的实验里最近一次核验通过的；不通过的实验拿掉，各带原因
|
||||
* （记"证据不足：核验不通过"）。拿掉实验后诊断程序的代码不动：核验员核过的就是这一份，只是运行时不调不通过的实验的编号。
|
||||
* 只在诊断实现与核验这一步做完以后调用。
|
||||
*/
|
||||
export function diagnosisProgram(board: Board, round: number, maxRepairs: number): { experiments: string[]; dropped: Array<{ experiment: string; reason: string }> } {
|
||||
const verifications = board.published("diagnosisVerification").filter((s) => s.round === round);
|
||||
const last = verifications.at(-1)!.content;
|
||||
const { failing } = diagnosisOutcome(last);
|
||||
return {
|
||||
experiments: board.decision("selectDiagnosis", round)!.output.chosen.filter((e) => !failing.includes(e)),
|
||||
dropped: failing.map((experiment) => {
|
||||
const x = last.experiments.find((y) => y.experiment === experiment)!;
|
||||
const what = `${x.followed ? "" : "没照写法做；"}${x.nothingElse ? "" : "动了写法以外的地方；"}${x.where}`;
|
||||
return { experiment, reason: `第 ${verifications.length} 次核验仍不通过，修复次数上限 ${maxRepairs} 已用完：${what}` };
|
||||
}),
|
||||
};
|
||||
}
|
||||
|
||||
/** 本轮要跑的诊断实验：本轮没有诊断、或一个都没剩下为 null。 */
|
||||
export function diagnosisProgramOf(board: Board, round: number, frozen: Frozen): string[] | null {
|
||||
if (chosenExperiments(board, round) === null) return null;
|
||||
const { experiments } = diagnosisProgram(board, round, frozen.config.maxRepairs);
|
||||
return experiments.length === 0 ? null : experiments;
|
||||
}
|
||||
|
||||
/** 诊 4 运行：和正向同一套，没启动的下次启动起新编号，已登记没结果的必须先由恢复入口核对。 */
|
||||
function diagnosisRunNext(round: number, view: View): Next | null {
|
||||
const runs = view.board.runs().filter((r) => r.plan.purpose === "diagnosis" && r.plan.round === round);
|
||||
const last = runs.at(-1);
|
||||
if (last !== undefined && last.result === null) throw new Error(`诊断运行 ${last.plan.run} 登记了还没有结果：恢复入口应当已经核对过`);
|
||||
if (last !== undefined && last.result!.process.ending !== "notStarted") return null;
|
||||
return { do: "diagnosisRun", round, run: `r${round}-diagnosis-${runs.length + 1}` };
|
||||
}
|
||||
|
||||
/** ⑩：按选批顺序一次只交回一个运行；没启动的下次启动起新编号，不会把半批提前公开。 */
|
||||
function forwardNext(round: number, view: View): Next | null {
|
||||
for (const candidate of view.board.decision("batch", round)!.output.chosen) {
|
||||
|
||||
+6
-2
@@ -43,9 +43,13 @@ export async function recover(s: Study): Promise<{ halt: string } | "stopped" |
|
||||
return null;
|
||||
}
|
||||
|
||||
/** 恢复和第一次运行共用原登记指标定义（引用指标也保留原来的缺失规则）；消融格缺失一律只记缺失（方法.md 第五部分"指标"）。 */
|
||||
function metricsOf(s: Study, plan: Content<"runPlanned">): HypothesisMetric[] {
|
||||
/**
|
||||
* 恢复和第一次运行共用原登记指标定义（引用指标也保留原来的缺失规则）；消融格缺失一律只记缺失（方法.md 第五部分"指标"）。
|
||||
* 诊断运行为 null：运行记录不带判分，逐个实验的判分在诊断结算前做。
|
||||
*/
|
||||
function metricsOf(s: Study, plan: Content<"runPlanned">): HypothesisMetric[] | null {
|
||||
if (plan.purpose === "baseline") return [];
|
||||
if (plan.purpose === "diagnosis") return null;
|
||||
const board = new Board(s.ledger.records());
|
||||
if (plan.purpose === "forward") return protocolMetrics(board, board.protocol(plan.object!));
|
||||
return ablationMetrics(board, splitCell(plan.object!).candidate);
|
||||
|
||||
@@ -0,0 +1,165 @@
|
||||
import { judgeSameExperiment, type DedupExperiment } from "../../jev/index.ts";
|
||||
import { Board, diagnosisId, diagnosisKey, type DiagnosisDedupOutput } from "../../materials/index.ts";
|
||||
import {
|
||||
compareIds,
|
||||
runBalance,
|
||||
runDiagnosisRule,
|
||||
settleDiagnosis,
|
||||
type DiagnosisExplanation,
|
||||
type DiagnosisRuleInput,
|
||||
type DiagnosisRuleName,
|
||||
type ExperimentOutcome,
|
||||
type SettleDiagnosisInput,
|
||||
} from "../../rules/index.ts";
|
||||
import type { DiagnosisExperiment, DiagnosisJudged } from "../../tasks/index.ts";
|
||||
import type { Json } from "../../vocab/index.ts";
|
||||
import { diagnosisFormalRun, judgeDiagnosisRun, type RunOutcome } from "../formal-run.ts";
|
||||
import { diagnosisProgram, diagnosisProgramOf } from "../next-step.ts";
|
||||
import { baselineRunDir } from "../task-code.ts";
|
||||
import type { Study } from "../study.ts";
|
||||
import type { SamePair } from "./round.ts";
|
||||
|
||||
// 诊断这一段里程序自己做的四步（方法.md 4.3 诊 2–诊 4）：判重合并、选诊断实验、诊断运行、诊断结算。
|
||||
// 判重合并和选实验各写一条决定，结算写一条诊断结算，写下之前被打断就从账本重新算一遍；写下之后恢复时照原样用，不重算。
|
||||
|
||||
/** 一个作者交的一个诊断实验：标签 "<作者>.e<第几个>"，只在判重合并的输入里认实验用。 */
|
||||
type Item = { author: string; index: number; label: string; experiment: DedupExperiment };
|
||||
|
||||
/**
|
||||
* 诊 2 判重合并（方法.md 4.3）：本轮已公开的诊断实验两两问一次 Jev，按作者编号、再按第几个排好顺序发，
|
||||
* 全部问完才写决定：中途被打断就整轮重问，判重不扣运行次数。
|
||||
*/
|
||||
export async function diagnosisDedupe(s: Study, round: number): Promise<void> {
|
||||
const board = new Board(s.ledger.records());
|
||||
const model = s.frozen.config.models.jev;
|
||||
const items: Item[] = board
|
||||
.published("diagnosisProposal")
|
||||
.filter((p) => p.round === round)
|
||||
.sort((a, b) => compareIds(a.author, b.author))
|
||||
.flatMap((p) => p.content.experiments.map((experiment, index) => ({ author: p.author, index, label: `${p.author}.e${index + 1}`, experiment })));
|
||||
const pairs: SamePair[] = [];
|
||||
for (const [i, a] of items.entries()) {
|
||||
for (const b of items.slice(i + 1)) pairs.push({ a: a.label, b: b.label, judgment: await judgeSameExperiment(model, a.experiment, b.experiment) });
|
||||
}
|
||||
// SAFETY: JevJudgment 和 DiagnosisDedupOutput 都是纯 Json 值；账本在写入时校验 Json。
|
||||
s.ledger.append("decision", {
|
||||
decision: "diagnosisDedup",
|
||||
round,
|
||||
rule: "jevSameExperiment",
|
||||
ruleVersion: 1,
|
||||
params: { jevModel: model },
|
||||
seed: null,
|
||||
input: { pairs } as unknown as Json,
|
||||
output: groupSameExperiment(items, pairs) as unknown as Json,
|
||||
});
|
||||
}
|
||||
|
||||
/**
|
||||
* 判为相同的连成一组（同判重合并：A 和 B 相同、B 和 C 相同，A、B、C 就是一个实验）。每组合成一个实验，
|
||||
* 按组里排最前的那一个排序，编号用 materials 的 diagnosisKey（d1…）；组内按作者编号、再按第几个排，第一个的正文就是合并后的正文。
|
||||
* items 要先按作者编号、第几个排好。
|
||||
*/
|
||||
export function groupSameExperiment(items: readonly Pick<Item, "author" | "index" | "label">[], pairs: readonly SamePair[]): DiagnosisDedupOutput {
|
||||
const parent = new Map(items.map((x) => [x.label, x.label]));
|
||||
const root = (x: string): string => {
|
||||
while (parent.get(x) !== x) x = parent.get(x)!;
|
||||
return x;
|
||||
};
|
||||
for (const p of pairs) if (p.judgment.yes) parent.set(root(p.a), root(p.b));
|
||||
const groups = new Map<string, Array<{ author: string; index: number }>>();
|
||||
for (const x of items) groups.set(root(x.label), [...(groups.get(root(x.label)) ?? []), { author: x.author, index: x.index }]);
|
||||
return { experiments: [...groups.values()].map((sources, i) => ({ experiment: diagnosisKey(i), sources })) };
|
||||
}
|
||||
|
||||
/**
|
||||
* 诊 3 选诊断实验（方法.md 6.12）：基线、方向、λ、合并后的全部实验、每条解释对每个实验的三个数，加余额（本轮开始前）、
|
||||
* 每批上限、一次诊断最多几个实验 D、u、衰减，交给步骤清单里 selectDiagnosis 那一步的规则，写一条决定。
|
||||
* 规则没有随机，种子写 null；定了不诊断也写。
|
||||
*/
|
||||
export function selectDiagnosisExperiments(s: Study, round: number): void {
|
||||
const board = new Board(s.ledger.records());
|
||||
const { config, facts, steps } = s.frozen;
|
||||
const { diagnosis } = config.flow;
|
||||
if (!diagnosis.enabled) throw new Error("配置关掉了诊断，步骤清单里选诊断实验一步应当标了跳过");
|
||||
const input: DiagnosisRuleInput = {
|
||||
baseline: board.baselineScore(board.baselineOf(round)),
|
||||
direction: facts.finalMetric.direction,
|
||||
lambda: config.flow.openReserve,
|
||||
experiments: board.diagnosisExperiments(round).map((e) => e.id),
|
||||
explanations: predictionsOf(board, round),
|
||||
balance: runBalance(config.runs, board.runStatesBefore(round)).balance,
|
||||
batchCap: config.batchLimit,
|
||||
maxExperiments: diagnosis.maxExperiments,
|
||||
u: config.flow.knowledgeUse,
|
||||
decay: config.flow.decay,
|
||||
};
|
||||
const rule = steps.round.find((x) => x.id === "selectDiagnosis")!.rule!.name as DiagnosisRuleName;
|
||||
const { rule: used, ...output } = runDiagnosisRule(rule, input);
|
||||
// SAFETY: rules 是纯计算，输入输出都是纯 Json 值；账本在写入时校验 Json。
|
||||
s.ledger.append("decision", {
|
||||
decision: "selectDiagnosis",
|
||||
round,
|
||||
rule: used.name,
|
||||
ruleVersion: used.version,
|
||||
params: output.params,
|
||||
seed: null,
|
||||
input: input as unknown as Json,
|
||||
output: output as unknown as Json,
|
||||
});
|
||||
}
|
||||
|
||||
/** 这一轮已冻结的各条诊断解释：作者的第 i 条（materials 的 diagnosisExplanationId），每条对每个合并后的实验给的三个数。选实验和结算共用。 */
|
||||
function predictionsOf(board: Board, round: number): DiagnosisExplanation[] {
|
||||
const predictions = board.published("diagnosisPrediction").filter((p) => p.round === round);
|
||||
return board.diagnosisExplanations(round).map((x) => ({
|
||||
id: x.id,
|
||||
author: x.author,
|
||||
experiments: predictions.find((p) => p.author === x.author)!.content.predictions.filter((y) => y.explanation === x.id).map((y) => ({ experiment: y.experiment, range: y.range })),
|
||||
}));
|
||||
}
|
||||
|
||||
/** 要跑的实验的运行输入：编号加它登记的假设专用指标（全名），次序同 ids，也就是运行的次序。 */
|
||||
function experimentsOf(board: Board, round: number, ids: readonly string[]): DiagnosisExperiment[] {
|
||||
const merged = board.diagnosisExperiments(round);
|
||||
return ids.map((id) => ({ id, metrics: merged.find((e) => e.id === id)!.metrics }));
|
||||
}
|
||||
|
||||
/**
|
||||
* 诊 4 诊断运行：用核验过的诊断程序，复用正式运行的登记、扣次入口，对象是本轮诊断。跑的只是核验里通过的实验：
|
||||
* 不通过而拿掉的实验，代码留在程序里不动（核验员核过的就是这一份），运行时不调它的编号；每个实验分到总时限 ÷ 跑的实验数。
|
||||
*/
|
||||
export function diagnosisRun(s: Study, round: number, run: string): Promise<RunOutcome> {
|
||||
const board = new Board(s.ledger.records());
|
||||
const ran = diagnosisProgramOf(board, round, s.frozen)!;
|
||||
const version = board.code(diagnosisId(round)).version;
|
||||
return diagnosisFormalRun(s, { run, purpose: "diagnosis", round, object: diagnosisId(round) }, version, baselineRunDir(s, board, round), experimentsOf(board, round, ran));
|
||||
}
|
||||
|
||||
/**
|
||||
* 诊 4 结算（方法.md 4.3 诊 4、6.12）：先从运行目录给每个交出答案的实验单独判分，再照 SettleDiagnosisInput 拼输入，
|
||||
* 写一条诊断结算。每个合并后的实验的结局：跑了且有效分数是有效；跑了没有效分数、或核验不通过拿掉的是证据不足；没选中的是没做。
|
||||
* 运行记录不带判分，所以被打断在判分和写结算之间时，恢复后从运行目录重新判。
|
||||
*/
|
||||
export async function diagnosisSettle(s: Study, round: number): Promise<void> {
|
||||
const board = new Board(s.ledger.records());
|
||||
const { experiments: ran, dropped } = diagnosisProgram(board, round, s.frozen.config.maxRepairs);
|
||||
const run = board.runs().filter((r) => r.plan.purpose === "diagnosis" && r.plan.round === round).at(-1)!;
|
||||
const judged = await judgeDiagnosisRun(s, run.plan.run, run.result!.process, experimentsOf(board, round, ran));
|
||||
const input: SettleDiagnosisInput = {
|
||||
baseline: board.baselineScore(board.baselineOf(round)),
|
||||
direction: s.frozen.facts.finalMetric.direction,
|
||||
epsilon: s.frozen.facts.epsilon,
|
||||
lambda: s.frozen.config.flow.openReserve,
|
||||
experiments: board.diagnosisExperiments(round).map((e) => ({ experiment: e.id, outcome: outcomeOf(e.id, ran, dropped.map((d) => d.experiment), judged) })),
|
||||
explanations: predictionsOf(board, round),
|
||||
};
|
||||
// SAFETY: 判分结果、结算的输入输出都是纯 Json 值；账本在写入时校验 Json。
|
||||
s.ledger.append("diagnosisSettlement", { round, result: { run: run.plan.run, judged, dropped, input, output: settleDiagnosis(input) } as unknown as Json });
|
||||
}
|
||||
|
||||
function outcomeOf(experiment: string, ran: readonly string[], dropped: readonly string[], judged: DiagnosisJudged): ExperimentOutcome {
|
||||
if (dropped.includes(experiment)) return { kind: "insufficient" };
|
||||
if (!ran.includes(experiment)) return { kind: "untested" };
|
||||
const j = judged.experiments.find((x) => x.id === experiment)!;
|
||||
return j.verdict === "valid" && j.score.status === "measured" ? { kind: "valid", score: j.score.value } : { kind: "insufficient" };
|
||||
}
|
||||
@@ -94,6 +94,7 @@ function studyFacts(frozen: Frozen, task: Task): StudyFacts {
|
||||
lambda: config.flow.openReserve,
|
||||
totalRuns: config.runs,
|
||||
proposalLimit: config.proposalLimit,
|
||||
diagnosisExperimentLimit: config.flow.diagnosis.enabled ? config.flow.diagnosis.perResearcher : null,
|
||||
factorLimit: config.factorLimit,
|
||||
metrics: metricDefinitions(task),
|
||||
taskText: task.guide,
|
||||
|
||||
+41
-3
@@ -1,19 +1,20 @@
|
||||
import { join } from "node:path";
|
||||
import type { Content } from "../ledger/index.ts";
|
||||
import { codeHunks } from "../jev/index.ts";
|
||||
import { Board, protocolMetrics, splitCell, type CodeChanges } from "../materials/index.ts";
|
||||
import { Board, diagnosisId, protocolMetrics, splitCell, type CodeChanges } from "../materials/index.ts";
|
||||
import type { KnownOf } from "../submissions/index.ts";
|
||||
import { lightCheck, protectedChanges, readLightCheck } from "../tasks/index.ts";
|
||||
import { diagnosisUnversionedNames, lightCheck, protectedChanges, readLightCheck } from "../tasks/index.ts";
|
||||
import type { Json, Submission, SubmissionKind } from "../vocab/index.ts";
|
||||
import { compareVersions, saveVersion } from "../workspace/index.ts";
|
||||
import type { TaskRef } from "./next-step.ts";
|
||||
import { callPaths, containerName, freshDir, workspaceRepo } from "./paths.ts";
|
||||
import { callPaths, containerName, freshDir, runPaths, workspaceRepo } from "./paths.ts";
|
||||
import type { Study } from "./study.ts";
|
||||
|
||||
/**
|
||||
* 实现取本批基线；修复取上次提交。核验、修复的材料差异始终相对本批基线。
|
||||
* 消融实现第一次取被采纳的完整代码，重做取上一次交的这一格代码（materials 的消融实现材料就是这样写的）；
|
||||
* 消融核验的材料差异是完整代码到这一格代码（方法.md 4.3 ⑯）。
|
||||
* 诊断实现取当前基线；诊断修复取上次提交的诊断程序；诊断核验、修复的材料差异是当前基线到诊断程序（方法.md 4.3 诊 4）。
|
||||
*/
|
||||
export function taskCode(s: Study, t: TaskRef & { stageId: string }): { changes: CodeChanges | null; startVersion: string | null } {
|
||||
const board = new Board(s.ledger.records());
|
||||
@@ -26,6 +27,13 @@ export function taskCode(s: Study, t: TaskRef & { stageId: string }): { changes:
|
||||
const { candidate } = splitCell(t.object!);
|
||||
return { changes: compareVersions(workspaceRepo(s.dir), board.code(candidate).version, board.code(t.object!).version), startVersion: null };
|
||||
}
|
||||
if (t.kind === "diagnosisImplementation" || t.kind === "diagnosisRepair" || t.kind === "diagnosisVerification") {
|
||||
const round = board.stage(t.stageId).round;
|
||||
const baseline = board.code(board.baselineOf(round)).version;
|
||||
if (t.kind === "diagnosisImplementation") return { changes: null, startVersion: baseline };
|
||||
const program = board.code(t.object!).version;
|
||||
return { changes: compareVersions(workspaceRepo(s.dir), baseline, program), startVersion: t.kind === "diagnosisRepair" ? program : null };
|
||||
}
|
||||
if (t.kind !== "implementation" && t.kind !== "repair" && t.kind !== "verification") return { changes: null, startVersion: null };
|
||||
const baseline = board.code(board.baselineOf(board.stage(t.stageId).round)).version;
|
||||
if (t.kind === "implementation") return { changes: null, startVersion: baseline };
|
||||
@@ -33,6 +41,18 @@ export function taskCode(s: Study, t: TaskRef & { stageId: string }): { changes:
|
||||
return { changes: compareVersions(workspaceRepo(s.dir), baseline, version), startVersion: t.kind === "repair" ? version : null };
|
||||
}
|
||||
|
||||
/** 产出第 round 轮当前基线的那次运行的运行目录：诊断实验要接着算的原始产物就在里面。 */
|
||||
export function baselineRunDir(s: Study, board: Board, round: number): string {
|
||||
return runPaths(s.dir, board.runOf(board.baselineOf(round))!.plan.run).run;
|
||||
}
|
||||
|
||||
/** 诊断实现、修复的工作目录要放进的原始产物所在的运行目录（取出代码之后，tasks 的 placeBaselineArtifact 放）；别的任务为 null。 */
|
||||
export function taskArtifact(s: Study, t: TaskRef & { stageId: string }): string | null {
|
||||
if (t.kind !== "diagnosisImplementation" && t.kind !== "diagnosisRepair") return null;
|
||||
const board = new Board(s.ledger.records());
|
||||
return baselineRunDir(s, board, board.stage(t.stageId).round);
|
||||
}
|
||||
|
||||
/**
|
||||
* 提交时才知道的东西：提案份数、回应原提案，或实验员当前工作目录的快照和相对本批基线的改动。
|
||||
* 快照写入 workspace 后才问 Jev；收下才把 code 和 submission 在同一个事务写进账本。
|
||||
@@ -61,6 +81,24 @@ export function prepareSubmission(
|
||||
code: { object: t.object!, version: saved.version, locations: null },
|
||||
};
|
||||
}
|
||||
if (t.kind === "diagnosisImplementation" || t.kind === "diagnosisRepair") {
|
||||
// 诊断程序：存的名单多跳过宿主放进来的原始产物目录；Jev 对照的是选中实验的原文和要采集的指标。
|
||||
const saved = saveVersion(workspaceRepo(s.dir), workDir, diagnosisUnversionedNames(s.task));
|
||||
if (!saved.saved) return { known: { code: saved }, code: null };
|
||||
const round = board.stage(t.stageId).round;
|
||||
const changes = compareVersions(workspaceRepo(s.dir), board.code(board.baselineOf(round)).version, saved.version);
|
||||
const chosen = board.diagnosisExperiments(round).filter((e) => board.decision("selectDiagnosis", round)!.output.chosen.includes(e.id));
|
||||
const impl = content as Submission<"diagnosisImplementation">;
|
||||
return {
|
||||
known: { code: {
|
||||
saved: true,
|
||||
protectedFiles: protectedChanges(s.task, changes.files.map((f) => f.path)),
|
||||
hunks: codeHunks(changes),
|
||||
protocol: { body: chosen.map((e) => `### ${e.id} ${e.name}\n\n动哪一处：${e.spot}\n\n${e.body}`).join("\n\n"), metrics: chosen.flatMap((e) => e.metrics) },
|
||||
} },
|
||||
code: { object: t.object!, version: saved.version, locations: impl.locations as Json },
|
||||
};
|
||||
}
|
||||
if (t.kind !== "implementation" && t.kind !== "repair") return { known: null, code: null };
|
||||
const saved = saveVersion(workspaceRepo(s.dir), workDir, s.frozen.facts.unversionedNames);
|
||||
if (!saved.saved) return { known: { code: saved }, code: null };
|
||||
|
||||
@@ -1,11 +1,15 @@
|
||||
// jev：问 Jev 的问法和阈值（docs/结构.md 第二部分）。判断的类型 JevJudgment 在 vocab。
|
||||
export {
|
||||
checkExperiment,
|
||||
checkHunk,
|
||||
checkProtocol,
|
||||
codeHunks,
|
||||
judgeSameChange,
|
||||
judgeSameExperiment,
|
||||
type CodeHunk,
|
||||
type DedupExperiment,
|
||||
type DedupFactor,
|
||||
type DiagnosisExperiment,
|
||||
type FrozenProtocol,
|
||||
type JevCheck,
|
||||
type ProtocolMetric,
|
||||
|
||||
+37
-1
@@ -1,9 +1,10 @@
|
||||
import type { HypothesisMetric, JevJudgment, Json, Submission } from "../vocab/index.ts";
|
||||
import { questionsOf, taskOf, type Question } from "./questions.ts";
|
||||
|
||||
// 问 Jev：从带类型的输入拼材料、发请求、读回答（docs/方法.md 4.3 ③⑥⑨）。
|
||||
// 问 Jev：从带类型的输入拼材料、发请求、读回答（docs/方法.md 4.3 ③⑥⑨ 和诊 1、诊 2）。
|
||||
// 材料里放什么、字段叫什么只在这里定。分请求照参考项目 src/judge/judge.ts：
|
||||
// 判重一对一次请求；冻结条款三问每问单独一次请求，材料相同；代码改动一个差异块一次请求，五问同在。
|
||||
// 诊断实验的判重、冻结检查照判重、冻结条款的写法；诊断程序的代码改动直接用 checkHunk。
|
||||
|
||||
type Factor = Submission<"proposal">["factors"][number];
|
||||
|
||||
@@ -13,6 +14,12 @@ type Factor = Submission<"proposal">["factors"][number];
|
||||
*/
|
||||
export type DedupFactor = Pick<Factor, "name" | "method" | "keySettings">;
|
||||
|
||||
/** 判重看的诊断实验：名称、动哪一处、怎么动。能分开哪几条解释写在正文里，不影响"是不是同一个实验"。 */
|
||||
export type DedupExperiment = Pick<Submission<"diagnosisProposal">["experiments"][number], "name" | "spot" | "body">;
|
||||
|
||||
/** 冻结检查看的诊断实验：名称、动哪一处、怎么动，加登记的假设专用指标（编号是短名）。 */
|
||||
export type DiagnosisExperiment = DedupExperiment & Pick<Submission<"diagnosisProposal">["experiments"][number], "metrics">;
|
||||
|
||||
/** 一份冻结协议的条款（docs/方法.md 4.3 ⑥"冻结协议包括"），字段与统一提交相同。 */
|
||||
export type FrozenProtocol = Pick<Submission<"unification">, "factors" | "safeguards" | "metrics" | "body" | "clauses">;
|
||||
|
||||
@@ -49,6 +56,31 @@ export async function judgeSameChange(model: string, a: readonly DedupFactor[],
|
||||
return judgment as JevJudgment;
|
||||
}
|
||||
|
||||
/** 诊 2 判重：两个诊断实验是不是同一个实验。 */
|
||||
export async function judgeSameExperiment(model: string, a: DedupExperiment, b: DedupExperiment): Promise<JevJudgment> {
|
||||
const state = {
|
||||
task: taskOf("diagnosisDedup"),
|
||||
experiment_a: dedupExperiment(a),
|
||||
experiment_b: dedupExperiment(b),
|
||||
};
|
||||
const [judgment] = await ask(model, state, questionsOf("diagnosisDedup"));
|
||||
return judgment as JevJudgment;
|
||||
}
|
||||
|
||||
/** 诊 1 查一个诊断实验：改动了评测、自相矛盾、写得太含糊。三问逐个发，材料相同。 */
|
||||
export async function checkExperiment(model: string, experiment: DiagnosisExperiment): Promise<JevCheck> {
|
||||
const state = {
|
||||
task: taskOf("diagnosisFreeze"),
|
||||
experiment: {
|
||||
...dedupExperiment(experiment),
|
||||
metrics: experiment.metrics.map((m) => ({ id: m.id, howMeasured: m.howMeasured, missing: m.missing })),
|
||||
},
|
||||
};
|
||||
const judgments: JevJudgment[] = [];
|
||||
for (const question of questionsOf("diagnosisFreeze")) judgments.push(...(await ask(model, state, [question])));
|
||||
return check(judgments);
|
||||
}
|
||||
|
||||
/** ⑥ 查一份冻结协议的条款：改动了评测、自相矛盾、写得太含糊。三问逐个发，材料相同。 */
|
||||
export async function checkProtocol(model: string, protocol: FrozenProtocol): Promise<JevCheck> {
|
||||
const state = {
|
||||
@@ -83,6 +115,10 @@ function dedupFactor(f: DedupFactor): { [key: string]: Json } {
|
||||
return { name: f.name, method: f.method, keySettings: f.keySettings };
|
||||
}
|
||||
|
||||
function dedupExperiment(e: DedupExperiment): { [key: string]: Json } {
|
||||
return { name: e.name, spot: e.spot, body: e.body };
|
||||
}
|
||||
|
||||
function check(judgments: JevJudgment[]): JevCheck {
|
||||
return { flagged: judgments.some((j) => j.yes), judgments };
|
||||
}
|
||||
|
||||
+136
-3
@@ -1,13 +1,17 @@
|
||||
import { hash, type Json } from "../vocab/index.ts";
|
||||
|
||||
// Jev 的九问（docs/方法.md 4.3 ③⑥⑨）：每一处的材料说明、问题原文、阈值。
|
||||
// Jev 的十三问（docs/方法.md 4.3 ③⑥⑨ 和诊 1、诊 2）：每一处的材料说明、问题原文、阈值。
|
||||
// 措辞、判"是"判"否"的说明和示例句取自参考项目 src/judge/questions.ts（填圆那一套），
|
||||
// 只改了必须改的：点 → 因子，字段名对上 vocab 的提案和冻结协议，改动对上 workspace 的比较结果。
|
||||
// 写法照参考项目 research-wiki/参考-Jev使用指南.md：问题用英文、一问只问一件事、
|
||||
// 统一问坏的一面（概率高 = 有问题或"是同一个"）、判据带 not_for 和例子、问题里用反引号点到材料路径。
|
||||
|
||||
/** 用到 Jev 的三处：判重、冻结条款、一段代码改动。 */
|
||||
export type Site = "dedup" | "freeze" | "diff";
|
||||
/**
|
||||
* 用到 Jev 的五处：判重、冻结条款、一段代码改动，以及诊断实验的判重和冻结检查。
|
||||
* 诊断的两处材料是"一个诊断实验"（动当前基线的一处），不是候选提案和冻结协议，所以各有自己的材料说明和问法
|
||||
* （方法.md 4.3 诊 1、诊 2）；一段代码改动的五问直接沿用 diff（诊 4）。
|
||||
*/
|
||||
export type Site = "dedup" | "freeze" | "diff" | "diagnosisDedup" | "diagnosisFreeze";
|
||||
|
||||
/** 一条是非题（Jev 的 noul）。 */
|
||||
export interface Noul {
|
||||
@@ -24,6 +28,9 @@ const THRESHOLDS: Readonly<Record<Site, Readonly<Record<string, number>>>> = {
|
||||
dedup: { same_change: 0.3 },
|
||||
freeze: { self_contradictory: 0.5, changes_measurement: 0.21, too_vague: 0.78 },
|
||||
diff: { unseeded_randomness: 0.55, reads_clock: 0.55, touches_measurement: 0.65, early_exit: 0.63, zero_for_unmeasured: 0.62 },
|
||||
// 诊断实验的两处阈值暂用 dedup、freeze 的值，没有在诊断实验的材料上校准过，等真实数据（任务书 18 第三块）再校准。
|
||||
diagnosisDedup: { same_experiment: 0.3 },
|
||||
diagnosisFreeze: { self_contradictory: 0.5, changes_measurement: 0.21, too_vague: 0.78 },
|
||||
};
|
||||
|
||||
/** 放进材料 `task` 字段的说明：这是什么材料、各部分是什么。 */
|
||||
@@ -44,6 +51,18 @@ const TASKS: Readonly<Record<Site, string>> = {
|
||||
"`protocol.metrics` lists the metrics the protocol requires the implementer " +
|
||||
"to compute in the candidate code and write to metrics.json (each with its `howMeasured` and `missing` rule; an empty list means none). " +
|
||||
"Only this hunk is shown; judge only what the hunk itself adds or changes.",
|
||||
diagnosisDedup:
|
||||
"Two diagnosis experiments from the same round of a study that improves a program by changing its code. " +
|
||||
"A diagnosis experiment takes the current baseline program, changes ONE spot of it (`spot`) as `body` describes " +
|
||||
"(replace a step with a clearly stronger version, replace it with another method, remove it, or relax a restriction), " +
|
||||
"runs the changed program once and scores the answer it returns. " +
|
||||
"The experiments were written by different researchers, in Chinese or English.",
|
||||
diagnosisFreeze:
|
||||
"One diagnosis experiment in a study that improves a program by changing its code. " +
|
||||
"An implementer will change the baseline program at `experiment.spot` exactly as `experiment.body` says and nothing else; " +
|
||||
"the unchanged task harness will then run the changed program once, and the task's validator will score the answer it returns. " +
|
||||
"The implementer may not edit the task's harness or validator files. " +
|
||||
"`experiment.metrics` lists extra diagnostic metrics the changed program must record (an empty list means none).",
|
||||
};
|
||||
|
||||
const QUESTIONS: Readonly<Record<Site, Readonly<Record<string, Noul>>>> = {
|
||||
@@ -292,6 +311,120 @@ const QUESTIONS: Readonly<Record<Site, Readonly<Record<string, Noul>>>> = {
|
||||
},
|
||||
},
|
||||
},
|
||||
// 诊 2 诊断实验判重：一对实验一次请求，写法照 dedup。
|
||||
diagnosisDedup: {
|
||||
same_experiment: {
|
||||
type: "noul",
|
||||
instructions: {
|
||||
question:
|
||||
"Do `experiment_a` and `experiment_b` change the same spot of the baseline in the same way, so that one run would answer both?",
|
||||
focus:
|
||||
"Compare where each experiment changes the baseline (`spot`) and how (`body`): the kind of change (stronger version, other method, removal, relaxed restriction) and its settings. " +
|
||||
"Ignore wording, language, level of detail, the experiment's name, and which explanation it says it can separate from which. " +
|
||||
"The same spot changed in a different way, or the same change with different settings or sizes, is NOT the same experiment: each would need its own run.",
|
||||
},
|
||||
criteria: {
|
||||
true: {
|
||||
what: "Both change the same spot of the baseline in the same way with the same settings; they differ only in wording, language or detail",
|
||||
examples: [
|
||||
"'replace the optimizer step with an exact solver, no time limit' vs 'swap the optimizer for an exact solve without any time cap'",
|
||||
"'remove the final repair pass' vs 'delete the repair pass at the end of the pipeline'",
|
||||
],
|
||||
},
|
||||
false: {
|
||||
what: "They change different spots, change the same spot in a different way, or use the same change with different settings",
|
||||
not_for: "Two experiments that merely test the same explanation, or merely touch the same function at different spots",
|
||||
examples: [
|
||||
"'remove the repair pass' vs 'replace the repair pass with an exact solver'",
|
||||
"'run the search for 10x more steps' vs 'run the search for 100x more steps'",
|
||||
"'relax the iteration limit of the search' vs 'relax the time limit of the whole program'",
|
||||
],
|
||||
},
|
||||
},
|
||||
},
|
||||
},
|
||||
|
||||
// 诊 1 诊断实验冻结检查：每问单独一次请求，材料相同；问法照 freeze，材料换成"一个诊断实验"。
|
||||
diagnosisFreeze: {
|
||||
self_contradictory: {
|
||||
type: "noul",
|
||||
instructions: {
|
||||
question:
|
||||
"Does `experiment` ask for something that cannot be done as written: do two of its statements exclude each other, or does `experiment.body` change places other than the one `experiment.spot` names?",
|
||||
focus:
|
||||
"Take the change described in `experiment.body` and check it against `experiment.spot`: the experiment must change that one spot and keep everything else of the baseline as it is. " +
|
||||
"A change that cannot be carried out without also changing something else, or statements that require things that cannot both hold, make it contradictory. " +
|
||||
"A statement that is merely unusual, ambitious or likely to give a worse score is not a contradiction.",
|
||||
},
|
||||
criteria: {
|
||||
true: {
|
||||
what: "Carrying out the experiment necessarily changes more than the named spot, or two statements require things that cannot both hold",
|
||||
examples: [
|
||||
"`spot` names the radius computation while `body` also asks to replace the starting layout",
|
||||
"`body` says remove the step and also tune its step size",
|
||||
"`body` says change only the solver and also keep the solver's code unchanged",
|
||||
],
|
||||
},
|
||||
false: {
|
||||
what: "All statements can hold at the same time in one implementation that changes only the named spot",
|
||||
not_for: "Experiments that may not improve the score, or statements that are vague but not conflicting",
|
||||
examples: ["replace the solver with an exact LP and keep everything else as in the baseline", "remove the final repair pass"],
|
||||
},
|
||||
},
|
||||
},
|
||||
changes_measurement: {
|
||||
type: "noul",
|
||||
instructions: {
|
||||
question:
|
||||
"Does `experiment` require changing how the answer is scored or checked, rather than changing the solution that gets scored?",
|
||||
focus:
|
||||
"The answer of the changed program is scored by the task's validator and harness. " +
|
||||
"Changing the algorithm that produces the answer is allowed, including relaxing a limit or a check inside the program itself, as long as the returned answer is still scored as before. " +
|
||||
"Changing how the returned value is summed, scored, validated, rounded or reported, or the validator or harness themselves, is a change of measurement. " +
|
||||
"Rules for the experiment's own diagnostic metrics (what they count, how they are aggregated) do not change the score.",
|
||||
},
|
||||
criteria: {
|
||||
true: {
|
||||
what: "The experiment asks to alter the scoring or validation: the formula of the score, what gets counted, tolerances, the harness or validator, or the value reported instead of the real one",
|
||||
examples: [
|
||||
"report the sum of diameters instead of the sum of radii",
|
||||
"count only items above a size threshold when summing",
|
||||
"raise the overlap tolerance in the validator",
|
||||
],
|
||||
},
|
||||
false: {
|
||||
what: "The experiment only changes how the answer is built or searched; the score is computed from its output as before",
|
||||
not_for: "Adding diagnostic metrics that are recorded next to the score without replacing it",
|
||||
examples: ["replace the greedy step with an exact solve", "remove the repair pass and see what the raw answer scores"],
|
||||
},
|
||||
},
|
||||
},
|
||||
too_vague: {
|
||||
type: "noul",
|
||||
instructions: {
|
||||
question:
|
||||
"Does `experiment` leave the implementer to guess what to change, for example which spot of the baseline to edit, what to replace it with, or the settings of the stronger version?",
|
||||
focus:
|
||||
"Read `experiment.spot` and `experiment.body` as an implementer who must follow them without asking questions. " +
|
||||
"Brief text is fine when it names the spot, the kind of change and its settings. " +
|
||||
"Judge only whether each statement is specific; whether two statements conflict is checked separately.",
|
||||
},
|
||||
criteria: {
|
||||
true: {
|
||||
what: "An implementer would have to invent a key part of the experiment: the spot to change, what to put there, or its main settings",
|
||||
examples: ["make the search better somewhere", "replace the optimizer with a stronger one", "relax some of the limits"],
|
||||
},
|
||||
false: {
|
||||
what: "The spot, the kind of change and its settings are stated concretely enough to implement directly",
|
||||
not_for: "Experiments that leave only minor coding details, such as variable names, to the implementer",
|
||||
examples: [
|
||||
"in the radius step, replace the heuristic with an exact LP solved by scipy linprog; no time limit",
|
||||
"remove the final repair pass and return the raw answer",
|
||||
],
|
||||
},
|
||||
},
|
||||
},
|
||||
},
|
||||
};
|
||||
|
||||
/** 一问的全部内容：问题原文、阈值，以及材料说明、问题原文、阈值三者合起来的哈希。 */
|
||||
|
||||
+7
-5
@@ -17,7 +17,7 @@ import {
|
||||
type SubmissionKind,
|
||||
} from "../vocab/index.ts";
|
||||
|
||||
// 账本记录的种类（docs/结构.md 第四部分的 11 种记录，细分成下面 19 种），每种一份内容定义（TypeBox）。
|
||||
// 账本记录的种类（docs/结构.md 第四部分的 11 种记录，细分成下面 20 种），每种一份内容定义（TypeBox）。
|
||||
// 下面写"只有一条""只一次"的，由 schema.ts 的唯一约束保证。
|
||||
// 配置、冻结协议、决定、结算的内容由别的模块产生，这里只检查最少的标识字段，其余只检查是 Json。
|
||||
|
||||
@@ -131,7 +131,7 @@ export const KINDS = {
|
||||
submission: Type.Unsafe<StoredSubmission>({}),
|
||||
/**
|
||||
* 3 退回：被退回的那份内容原样存（它可能正是因为不合格式才被退回），连同退回理由。
|
||||
* target 是这次任务的对象，与提交的 target 同义（候选、原提案、消融格，没有对象的写 null）：
|
||||
* target 是这次任务的对象，与提交的 target 同义（候选、原提案、消融格、本轮诊断，没有对象的写 null）：
|
||||
* 同一人在同一阶段可能对几个对象各做一次同种任务，重新发起时按它找本人这个对象的上一次退回。
|
||||
* jev 是这次退回用到的 Jev 判断（见 JevUsed），判"否"的几问也一起存；没用到 Jev 的写 null。
|
||||
*/
|
||||
@@ -161,11 +161,11 @@ export const KINDS = {
|
||||
/** 6 冻结代码：候选或消融格的代码版本号（由 workspace 给出）和因子位置表。位置表先存 Json。 */
|
||||
code: Type.Object({ object: Id, version: Text, locations: AnyJson }, CLOSED),
|
||||
|
||||
/** 7 运行·登记要跑（占名额）：运行编号由 flow 起，一个编号只登记一次；用途、轮次、对象（候选或格；基线没有）、运行定义。 */
|
||||
/** 7 运行·登记要跑（占名额）：运行编号由 flow 起，一个编号只登记一次；用途、轮次、对象（候选、格或本轮诊断；基线没有）、运行定义。 */
|
||||
runPlanned: Type.Object(
|
||||
{
|
||||
run: Id,
|
||||
purpose: Type.Enum(["baseline", "forward", "ablation"], { type: "string" }),
|
||||
purpose: Type.Enum(["baseline", "forward", "ablation", "diagnosis"], { type: "string" }),
|
||||
round: Round,
|
||||
object: OptionalId,
|
||||
definition: RunDefinition,
|
||||
@@ -178,7 +178,7 @@ export const KINDS = {
|
||||
runResult: Type.Object({ run: Id, process: ProcessResult, judged: Type.Union([JudgedResult, Type.Null()]) }, CLOSED),
|
||||
|
||||
/**
|
||||
* 8 决定：选批、采纳、选格、押的因子、入口分配、判重合并。decision 是决定种类。
|
||||
* 8 决定：选诊断实验、选批、采纳、选格、押的因子、入口分配、判重合并（提案和诊断实验）。decision 是决定种类。
|
||||
* 输入输出存 Json：ledger 不认识 rules。判重合并的输入里带 Jev 的原始概率。
|
||||
*/
|
||||
decision: Type.Object(
|
||||
@@ -197,6 +197,8 @@ export const KINDS = {
|
||||
|
||||
/** 9 结算：每轮最多采纳一个改法，所以最多一次消融，结算按轮次一轮一条。各格结局、配合关系、解释的确认或否定等先存 Json。 */
|
||||
settlement: Type.Object({ round: Round, result: AnyJson }, CLOSED),
|
||||
/** 9 诊断结算：每轮最多一次诊断，结算也按轮次一轮一条（方法.md 6.12）。每个实验的结局与实测、解释的确认或否定等先存 Json。 */
|
||||
diagnosisSettlement: Type.Object({ round: Round, result: AnyJson }, CLOSED),
|
||||
|
||||
/** 10 模型调用开始：先登记再调用（方法.md 第九部分）。附这次任务的全部材料，用来重现模型当时看到了什么。一次调用只有一条。 */
|
||||
callStarted: Type.Object({ call: Id, actor: Id, stageId: OptionalId, model: Text, materials: AnyJson }, CLOSED),
|
||||
|
||||
+13
-5
@@ -83,7 +83,7 @@ export class Ledger {
|
||||
}
|
||||
|
||||
/**
|
||||
* 追加一条记录，返回存下的那一条。先按种类检查内容，再做写入时的四种拒收检查（rejections.ts）。
|
||||
* 追加一条记录，返回存下的那一条。先按种类检查内容，再做写入时的五种拒收检查（rejections.ts）。
|
||||
* 写提案时，作者自评从提案里拆出来，在同一个事务里另存成一条封存，返回的是提案那一条。
|
||||
*/
|
||||
append<K extends AppendKind>(kind: K, content: Input<K>): LedgerRecord<K> {
|
||||
@@ -117,16 +117,24 @@ export class Ledger {
|
||||
this.#first("stageOpen", "json_extract(content, '$.stageId') = ?", stageId) as LedgerRecord<"stageOpen"> | undefined,
|
||||
latest: (slot) => this.latestSubmission(slot),
|
||||
runStarted: (run) => this.#first("runStarted", "json_extract(content, '$.run') = ?", run) !== undefined,
|
||||
ablationStarted: (round) =>
|
||||
ablationStarted: (round) => this.#purposeStarted("ablation", round),
|
||||
diagnosisStarted: (round) => this.#purposeStarted("diagnosis", round),
|
||||
};
|
||||
|
||||
/** 这一轮有没有某种用途的运行已经确认启动。 */
|
||||
#purposeStarted(purpose: "ablation" | "diagnosis", round: number): boolean {
|
||||
return (
|
||||
this.#one(
|
||||
`SELECT s.seq FROM records s JOIN records p
|
||||
ON p.kind = 'runPlanned' AND json_extract(p.content, '$.run') = json_extract(s.content, '$.run')
|
||||
WHERE s.kind = 'runStarted'
|
||||
AND json_extract(p.content, '$.purpose') = 'ablation'
|
||||
AND json_extract(p.content, '$.purpose') = ?
|
||||
AND json_extract(p.content, '$.round') = ?`,
|
||||
purpose,
|
||||
round,
|
||||
) !== undefined,
|
||||
};
|
||||
) !== undefined
|
||||
);
|
||||
}
|
||||
|
||||
// -------------------------------------------------------------------------
|
||||
// 读取
|
||||
|
||||
@@ -1,8 +1,8 @@
|
||||
import { SUBMISSION_KINDS, type Stage } from "../vocab/index.ts";
|
||||
import type { Content, Kind, LedgerRecord, SubmissionSlot } from "./kinds.ts";
|
||||
|
||||
// 写入时由 ledger 拒绝的四种情况（docs/结构.md 第四部分"写入时由 ledger 拒绝的几种情况"），在写入的同一个事务里查。
|
||||
// 只有这四种。记录之间别的先后（例如运行结果前要先有登记）和提交种类与阶段对不对得上都不查：
|
||||
// 写入时由 ledger 拒绝的五种情况（docs/结构.md 第四部分"写入时由 ledger 拒绝的几种情况"），在写入的同一个事务里查。
|
||||
// 只有这五种。记录之间别的先后（例如运行结果前要先有登记）和提交种类与阶段对不对得上都不查：
|
||||
// flow 调 ledger 是模块之间的调用，不是系统边界。"只发生一次"由 schema.ts 的唯一约束管。
|
||||
// 提交内容的字段归 submissions，该不该开下一个阶段归 flow。
|
||||
|
||||
@@ -16,14 +16,18 @@ export interface WriteState {
|
||||
runStarted(run: string): boolean;
|
||||
/** 这一轮有没有消融格的运行已经确认启动。 */
|
||||
ablationStarted(round: number): boolean;
|
||||
/** 这一轮的诊断运行有没有已经确认启动。 */
|
||||
diagnosisStarted(round: number): boolean;
|
||||
}
|
||||
|
||||
/** 实现和核验阶段的提交不能替换（方法.md 第五部分总原则）：实现、修复、核验、消融实现、消融核验。 */
|
||||
/** 实现和核验阶段的提交不能替换（方法.md 第五部分总原则）：实现、修复、核验、消融实现、消融核验、诊断实现（含修复）、诊断核验。 */
|
||||
const NOT_REPLACEABLE: ReadonlySet<Stage> = new Set([
|
||||
"implementation",
|
||||
"verification",
|
||||
"ablationImplementation",
|
||||
"ablationVerification",
|
||||
"diagnosisImplementation",
|
||||
"diagnosisVerification",
|
||||
]);
|
||||
|
||||
export function checkWrite(kind: Kind, content: unknown, state: WriteState): void {
|
||||
@@ -35,7 +39,7 @@ export function checkWrite(kind: Kind, content: unknown, state: WriteState): voi
|
||||
/**
|
||||
* 第 1 条：只收开着的阶段的提交，即有开阶段记录、没有关阶段记录。
|
||||
* 第 2 条：新的一版必须正好接在本人在这个阶段、对同一个对象的最新一版后面（第一版是 1）；
|
||||
* 实现、修复、核验、消融实现、消融核验只能交第一版。
|
||||
* 实现、修复、核验、消融实现、消融核验、诊断实现、诊断修复、诊断核验只能交第一版。
|
||||
*/
|
||||
function checkSubmission(s: Content<"submission">, state: WriteState): void {
|
||||
const where = `提交 ${s.kind}（阶段 ${s.stageId}，作者 ${s.author}，对象 ${s.target ?? "无"}，编号 ${s.item ?? "无"}，第 ${s.version} 版）`;
|
||||
@@ -63,7 +67,10 @@ function checkRunResult(r: Content<"runResult">, state: WriteState): void {
|
||||
}
|
||||
}
|
||||
|
||||
/** 第 4 条：这一轮已经有消融格确认启动了，就不能再关解释阶段（方法.md 4.3 ⑭：解释冻结必须早于任何一格启动）。 */
|
||||
/**
|
||||
* 第 4 条：这一轮已经有消融格确认启动了，就不能再关解释阶段（方法.md 4.3 ⑭：解释冻结必须早于任何一格启动）。
|
||||
* 第 5 条：这一轮的诊断运行已经确认启动了，就不能再关诊断预测阶段（方法.md 4.3 诊 2：冻结必须早于诊断运行启动，范围才算事前预测）。
|
||||
*/
|
||||
function checkStageClose(c: Content<"stageClose">, state: WriteState): void {
|
||||
const open = state.stageOpen(c.stageId);
|
||||
if (open?.content.stage === "explanation" && state.ablationStarted(open.content.round)) {
|
||||
@@ -71,4 +78,9 @@ function checkStageClose(c: Content<"stageClose">, state: WriteState): void {
|
||||
`不能关解释阶段 ${c.stageId}：第 ${open.content.round} 轮已经有消融格确认启动，解释冻结必须早于任何一格启动（方法.md 4.3 ⑭）`,
|
||||
);
|
||||
}
|
||||
if (open?.content.stage === "diagnosisPrediction" && state.diagnosisStarted(open.content.round)) {
|
||||
throw new Error(
|
||||
`不能关诊断预测阶段 ${c.stageId}：第 ${open.content.round} 轮的诊断运行已经确认启动，诊断预测冻结必须早于诊断运行启动（方法.md 4.3 诊 2）`,
|
||||
);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -3,7 +3,7 @@ import type { DatabaseSync } from "node:sqlite";
|
||||
// 账本的表结构（docs/结构.md 第四部分"全部只追加"）：只有一张表，每一行是序号、种类、内容、内容的哈希。
|
||||
// 表结构一改就把版本号加 1；旧文件版本号对不上时打开报错，不迁移，也不逐字核对建表语句。
|
||||
|
||||
export const SCHEMA_VERSION = 4;
|
||||
export const SCHEMA_VERSION = 5;
|
||||
|
||||
/** 新文件要执行的建表语句，由 Ledger 在一个事务里执行，最后写入版本号。 */
|
||||
export const CREATE_SCHEMA = [
|
||||
@@ -33,6 +33,8 @@ export const CREATE_SCHEMA = [
|
||||
`CREATE UNIQUE INDEX call_ended_once ON records (json_extract(content, '$.call')) WHERE kind = 'callEnded'`,
|
||||
// 结算以轮次为编号：每轮最多采纳一个改法，所以最多一次消融。
|
||||
`CREATE UNIQUE INDEX settlement_once ON records (json_extract(content, '$.round')) WHERE kind = 'settlement'`,
|
||||
// 诊断结算同样以轮次为编号：每轮最多一次诊断。
|
||||
`CREATE UNIQUE INDEX diagnosis_settlement_once ON records (json_extract(content, '$.round')) WHERE kind = 'diagnosisSettlement'`,
|
||||
`PRAGMA user_version=${SCHEMA_VERSION}`,
|
||||
];
|
||||
|
||||
|
||||
@@ -3,7 +3,14 @@ import { type Draft, factorLines, KIND_TEXT, offText, registeredMetrics, taskPoi
|
||||
import type { Board } from "./read.ts";
|
||||
import { fenced, json } from "./format.ts";
|
||||
import { unassignedChanges } from "./locations.ts";
|
||||
import { clauseKey, factorKey, splitCell, type FrozenProtocol, type LocationTable } from "./shapes.ts";
|
||||
import {
|
||||
clauseKey,
|
||||
diagnosisId,
|
||||
factorKey,
|
||||
splitCell,
|
||||
type FrozenProtocol,
|
||||
type LocationTable,
|
||||
} from "./shapes.ts";
|
||||
|
||||
// 实验员、核验员的 5 种任务（方法.md 4.3 ⑨ ⑯）：只看分给他的那个候选或那一格——冻结协议、起点代码、
|
||||
// 要实现或核验的东西。不给信念本（用户 2026-10-03 决定 3）、不给任何运行成绩、不给别的候选。
|
||||
@@ -36,8 +43,7 @@ export function protocolMetrics(board: Board, p: FrozenProtocol): HypothesisMetr
|
||||
})];
|
||||
}
|
||||
|
||||
function metricSection(d: Draft, p: FrozenProtocol): void {
|
||||
const all = protocolMetrics(d.board, p);
|
||||
function metricSection(d: Draft, all: HypothesisMetric[]): void {
|
||||
d.section("要实现的指标定义", [
|
||||
all.length === 0 ? "（这份协议没有要实现的假设专用指标）" : json(all),
|
||||
"",
|
||||
@@ -58,15 +64,22 @@ function codeLines(d: Draft): string[] {
|
||||
return [...d.readableLines(), "读这些目录用 read_source（核验员）或 shell（实验员）。"];
|
||||
}
|
||||
|
||||
/** 本轮做过诊断时，诊断程序的代码也能读（只读），冻结协议可以写"照诊断实验 dN 的做法"（方法.md 4.3 ②）；只给代码，不给成绩和实验原文。 */
|
||||
function diagnosisReference(d: Draft): string[] {
|
||||
if (d.board.diagnosisSettlement(d.round) === undefined) return [];
|
||||
d.read(diagnosisId(d.round), "本轮诊断程序（只读，冻结协议引用诊断实验时作参考）");
|
||||
return ["上面还列了本轮诊断程序的代码：冻结协议引用诊断实验时作参考，别把它当成这个候选的起点。"];
|
||||
}
|
||||
|
||||
/** ⑨ 实现：冻结协议、指标定义、要交的位置表、起点代码（本批基线）。 */
|
||||
export function implementation(d: Draft): ChecklistOf<"implementation"> {
|
||||
const candidate = d.req.object!;
|
||||
const { p, items } = frozen(d, candidate);
|
||||
taskSection(d, "实验员", `候选 ${candidate}`);
|
||||
d.read(p.baseline, "起点代码（本批基线）");
|
||||
d.section("基线代码", [...codeLines(d), "你的工作目录是起点代码的副本，改动都落在工作目录里。"]);
|
||||
d.section("基线代码", [...codeLines(d), "你的工作目录是起点代码的副本，改动都落在工作目录里。", ...diagnosisReference(d)]);
|
||||
d.section("本任务要处理的东西", [`### 候选 ${candidate} 的冻结协议`, "", json(p)]);
|
||||
metricSection(d, p);
|
||||
metricSection(d, protocolMetrics(d.board, p));
|
||||
locationTableSection(d, p);
|
||||
return { kind: "implementation", ...items };
|
||||
}
|
||||
@@ -78,7 +91,7 @@ export function repair(d: Draft): ChecklistOf<"repair"> {
|
||||
taskSection(d, "实验员", `候选 ${candidate}`);
|
||||
d.read(p.baseline, "起点代码（本批基线）");
|
||||
d.read(candidate, "你上一次提交的代码");
|
||||
d.section("基线代码", [...codeLines(d), "你的工作目录是上一次提交的代码的副本。"]);
|
||||
d.section("基线代码", [...codeLines(d), "你的工作目录是上一次提交的代码的副本。", ...diagnosisReference(d)]);
|
||||
const v = d.board.published("verification").filter((s) => s.target === candidate).at(-1)!.content;
|
||||
d.section("本任务要处理的东西", [
|
||||
"### 退回意见",
|
||||
@@ -102,7 +115,7 @@ export function repair(d: Draft): ChecklistOf<"repair"> {
|
||||
"",
|
||||
json(p),
|
||||
]);
|
||||
metricSection(d, p);
|
||||
metricSection(d, protocolMetrics(d.board, p));
|
||||
locationTableSection(d, p);
|
||||
return { kind: "repair", ...items };
|
||||
}
|
||||
@@ -148,7 +161,7 @@ export function verification(d: Draft): ChecklistOf<"verification"> {
|
||||
"",
|
||||
fenced(d.req.changes!.diff, "diff"),
|
||||
]);
|
||||
metricSection(d, p);
|
||||
metricSection(d, protocolMetrics(d.board, p));
|
||||
return { kind: "verification", unassigned: found.unassigned.map((u) => u.location), ...items };
|
||||
}
|
||||
|
||||
@@ -212,3 +225,60 @@ export function ablationVerification(d: Draft): ChecklistOf<"ablationVerificatio
|
||||
]);
|
||||
return { kind: "ablationVerification", cell: a.cell, factors: a.factors };
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// 诊断的三种任务（方法.md 4.3 诊 4）：对象是本轮诊断，一位实验员把选中的全部实验写进一份诊断程序，一位核验员逐个核对。
|
||||
// 只看选中的实验原文和当前基线代码，不给信念本、不给任何成绩。
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
/** 三种诊断任务共用：任务与本轮、起点代码（当前基线）、选中的实验原文和指标定义；program 是另外要能读的诊断程序的说明。交回选中的实验编号。 */
|
||||
function diagnosisTask(d: Draft, role: string, program: string | null): string[] {
|
||||
d.group("分给你的材料", "下面是分给你的本轮诊断的全部材料。");
|
||||
const chosen = d.board.decision("selectDiagnosis", d.round)!.output.chosen;
|
||||
const experiments = d.board.diagnosisExperiments(d.round).filter((e) => chosen.includes(e.id));
|
||||
taskSection(d, role, `本轮诊断（对象 ${diagnosisId(d.round)}）`);
|
||||
d.read(d.board.baselineOf(d.round), "起点代码（当前基线）");
|
||||
if (program !== null) d.read(diagnosisId(d.round), program);
|
||||
d.section("基线代码", codeLines(d));
|
||||
d.section("选中的诊断实验", [
|
||||
"每个实验都从当前基线的代码出发，只动它写的那一处，其余照基线；每个实验交出自己的答案。",
|
||||
"",
|
||||
...experiments.flatMap((e) => [`### ${e.id} ${e.name}`, "", `动哪一处：${e.spot}`, "", e.body, ""]),
|
||||
]);
|
||||
metricSection(d, experiments.flatMap((e) => e.metrics));
|
||||
return chosen;
|
||||
}
|
||||
|
||||
/** 诊断实现，加上原始产物的位置；修复（repair 为 true）另放退回意见（没通过的实验和依据、上一次的位置表和改动），通过的实验不要动。 */
|
||||
function diagnosisCode(d: Draft, repair: boolean): string[] {
|
||||
const chosen = diagnosisTask(d, "实验员", repair ? "你上一次提交的诊断程序" : null);
|
||||
d.section("原始产物", [
|
||||
`产出当前基线的那次正式运行（${d.board.runOf(d.board.baselineOf(d.round))!.plan.run}）的原始产物已经只读地放进你的工作目录，实验要接着算的答案就在里面；位置和怎么读写在任务说明里（research_query 查 task）。`,
|
||||
"你的工作目录是当前基线代码的副本，加上这份原始产物，改动都落在工作目录里。",
|
||||
]);
|
||||
if (!repair) return chosen;
|
||||
const v = d.board.published("diagnosisVerification").filter((s) => s.round === d.round).at(-1)!.content;
|
||||
const failed = v.experiments.filter((x) => !x.followed || !x.nothingElse);
|
||||
d.section("退回意见", [
|
||||
...failed.map((x) => `- 实验 ${x.experiment}：${x.followed ? "" : "没照写法做；"}${x.nothingElse ? "" : "动了写法以外的地方；"}${x.where}`),
|
||||
"",
|
||||
`核验员的说明：${v.body}`,
|
||||
"只修上面没通过的实验，通过的实验的代码不要动。",
|
||||
"",
|
||||
...["### 上一次交的位置表", "", json(d.board.code(diagnosisId(d.round)).locations), "", "### 上一次提交相对起点代码的改动", "", fenced(d.req.changes!.diff, "diff")],
|
||||
]);
|
||||
return chosen;
|
||||
}
|
||||
|
||||
export const diagnosisImplementation = (d: Draft): ChecklistOf<"diagnosisImplementation"> => ({ kind: "diagnosisImplementation", experiments: diagnosisCode(d, false) });
|
||||
export const diagnosisRepair = (d: Draft): ChecklistOf<"diagnosisRepair"> => ({ kind: "diagnosisRepair", experiments: diagnosisCode(d, true) });
|
||||
|
||||
/** 诊断核验：选中的实验原文、当前基线到诊断程序的改动、位置表。Jev 的判断不给核验员（同正向核验）。 */
|
||||
export function diagnosisVerification(d: Draft): ChecklistOf<"diagnosisVerification"> {
|
||||
const chosen = diagnosisTask(d, "核验员", "要核验的诊断程序");
|
||||
d.section("本任务要处理的东西", [
|
||||
...["### 实验员交的位置表", "", json(d.board.code(diagnosisId(d.round)).locations), ""],
|
||||
...["### 改动全文（诊断程序相对当前基线，每段前后各 3 行上下文）", "", fenced(d.req.changes!.diff, "diff")],
|
||||
]);
|
||||
return { kind: "diagnosisVerification", experiments: chosen };
|
||||
}
|
||||
|
||||
@@ -12,6 +12,7 @@ import {
|
||||
type Phase,
|
||||
} from "../rules/index.ts";
|
||||
import type { ProcessResult, Range } from "../vocab/index.ts";
|
||||
import { diagnosisHistory } from "./diagnosis.ts";
|
||||
import { measured, num, pct, PLACEMENT_TEXT, rangeText, resourcesText, score, SYNERGY_TEXT, UNTESTED_TEXT } from "./format.ts";
|
||||
import type { StudyFacts } from "./inputs.ts";
|
||||
import { answerPathLines } from "./locations.ts";
|
||||
@@ -21,6 +22,8 @@ import { CELL_NOT_RUN, cellId, factorId, factorKey, ORIGINAL_BASELINE, type Sett
|
||||
// 信念本（方法.md 6.10）：程序从已公开的记录现算，不存。数字全由 rules 的函数算，这里只挑出来写成字。
|
||||
// 只读这几样：已关阶段的提交、按 Board.runPublic 公开的运行、决定、结算、解封；封存、第一次预测、退回一律不读。
|
||||
|
||||
const PHASE_TEXT: Readonly<Record<Phase, string>> = { forward: "正向", diagnosis: "诊断", ablation: "消融" };
|
||||
|
||||
/** 范围怎么换算成把握（方法.md 6.2、6.3），写在开头。只说程序怎么用这三个数，不要求研究者照着算。 */
|
||||
function rangeToBelief(lambda: number): string {
|
||||
return (
|
||||
@@ -118,6 +121,9 @@ export function beliefBookOf(board: Board, study: StudyFacts, round: number | nu
|
||||
}
|
||||
for (const s of settlements) lines.push(...ablationSummary(board, s));
|
||||
|
||||
// 3 诊断历史
|
||||
lines.push("", "**诊断历史**（每轮的诊断实验与实测、每条解释确认或否定、没做的实验和原因，以及程序推出的方向）", ...diagnosisHistory(board, study, latestRound));
|
||||
|
||||
// 4 战绩
|
||||
lines.push("", "**战绩**（只显示，不改任何权重；偏高偏低都按分数原值判）");
|
||||
const samples: Array<{ researcher: string; phase: Phase; range: Range; value: number }> = batches.flatMap((d) =>
|
||||
@@ -127,6 +133,9 @@ export function beliefBookOf(board: Board, study: StudyFacts, round: number | nu
|
||||
return v === null ? [] : c.ranges.map((r) => ({ researcher: r.researcher, phase: "forward" as const, range: r.range, value: v }));
|
||||
}),
|
||||
);
|
||||
for (const s of board.diagnosisSettlements())
|
||||
for (const e of s.input.experiments)
|
||||
if (e.outcome.kind === "valid") for (const x of s.input.explanations) samples.push({ researcher: x.author, phase: "diagnosis", range: x.experiments.find((y) => y.experiment === e.experiment)!.range, value: e.outcome.score });
|
||||
for (const s of settlements)
|
||||
for (const c of s.input.cells) {
|
||||
if (c.outcome.kind !== "valid") continue;
|
||||
@@ -139,7 +148,7 @@ export function beliefBookOf(board: Board, study: StudyFacts, round: number | nu
|
||||
if (tallies.length === 0) lines.push("- 还没有实测可以核对");
|
||||
for (const { researcher, phase, tally: t } of tallies) {
|
||||
lines.push(
|
||||
`- ${researcher}，${phase === "forward" ? "正向" : "消融"} ${t.total} 次：中位数偏高 ${t.medianHigh}、偏低 ${t.medianLow}；` +
|
||||
`- ${researcher}，${PHASE_TEXT[phase]} ${t.total} 次：中位数偏高 ${t.medianHigh}、偏低 ${t.medianLow}；` +
|
||||
`穿过上沿 ${t.aboveTop}、下沿 ${t.belowBottom}；包住 ${t.covered}；落点 ` +
|
||||
(Object.keys(PLACEMENT_TEXT) as Array<keyof typeof PLACEMENT_TEXT>).map((k) => `${PLACEMENT_TEXT[k]} ${t.placements[k]}`).join("，"),
|
||||
);
|
||||
|
||||
+46
-4
@@ -2,11 +2,12 @@ import type { Ledger } from "../ledger/index.ts";
|
||||
import { adoptionLine, compareIds, runBalance } from "../rules/index.ts";
|
||||
import type { Checklist, SubmissionKind } from "../vocab/index.ts";
|
||||
import { beliefBookOf } from "./belief-book.ts";
|
||||
import { directionText, json, measured, score } from "./format.ts";
|
||||
import { diagnosisLines } from "./diagnosis.ts";
|
||||
import { directionText, json, measured, num, resourcesText, score } from "./format.ts";
|
||||
import type { TaskRequest } from "./inputs.ts";
|
||||
import { answerPathLines } from "./locations.ts";
|
||||
import { Board, type Run } from "./read.ts";
|
||||
import { entryId, explanationId, factorKey, splitCell, type FrozenProtocol } from "./shapes.ts";
|
||||
import { entryId, explanationId, factorKey, ORIGINAL_BASELINE, splitCell, type FrozenProtocol } from "./shapes.ts";
|
||||
|
||||
// 任务材料的形状，和研究者各任务共用的几块（方法.md 第三部分"给研究者的材料顺序"）。
|
||||
|
||||
@@ -38,6 +39,11 @@ export const KIND_TEXT: Readonly<Record<SubmissionKind, string>> = {
|
||||
entry: "第零轮闭门写入口",
|
||||
discussion: "第零轮讨论",
|
||||
claim: "第零轮认领入口",
|
||||
diagnosisProposal: "写诊断解释与诊断实验",
|
||||
diagnosisPrediction: "诊断预测",
|
||||
diagnosisImplementation: "实现诊断实验",
|
||||
diagnosisRepair: "按核验意见修复诊断程序",
|
||||
diagnosisVerification: "核验诊断程序",
|
||||
proposal: "提案",
|
||||
firstPrediction: "第一次预测",
|
||||
critique: "挑刺",
|
||||
@@ -142,12 +148,42 @@ export function codeSection(d: Draft): void {
|
||||
d.read(d.board.baselineOf(d.round), "本批基线");
|
||||
for (const r of d.board.runs()) {
|
||||
if (r.plan.object !== null && r.plan.purpose !== "baseline" && d.board.runPublic(r)) {
|
||||
d.read(r.plan.object, `${r.plan.purpose === "forward" ? "候选" : "消融格"} ${r.plan.object} 的代码`);
|
||||
d.read(r.plan.object, r.plan.purpose === "diagnosis" ? `第 ${r.plan.round} 轮诊断程序的代码` : `${r.plan.purpose === "forward" ? "候选" : "消融格"} ${r.plan.object} 的代码`);
|
||||
}
|
||||
}
|
||||
d.section("基线", [...baselineLines(d), "", "能读的代码（用 read_source 读）：", ...d.readableLines()]);
|
||||
}
|
||||
|
||||
/** 本轮诊断结果（方法.md 第三部分材料顺序第 3 条）：从提案起的全部研究者任务都有，没诊断写原因；诊断任务自己读时这一段还没走完，不放。 */
|
||||
export function diagnosisSection(d: Draft): void {
|
||||
if (d.board.diagnosisDone(d.round)) d.section("本轮诊断结果", diagnosisLines(d.board, d.req.study, d.round, true));
|
||||
}
|
||||
|
||||
/** 运行的产物只放进查询，这里列编号。 */
|
||||
export function artifactSection(d: Draft, heading: string): void {
|
||||
const artifacts = d.req.artifacts!;
|
||||
const listed = Object.keys(artifacts)
|
||||
.sort(compareIds)
|
||||
.flatMap((run) => Object.entries(artifacts[run]!).map(([file, text]) => `- ${d.record(`${run}/${file}`, `运行 ${run} 的产物 ${file}`, text)}`));
|
||||
d.section(heading, ["用 research_query 按编号查全文：", ...(listed.length === 0 ? ["（没有）"] : listed)]);
|
||||
}
|
||||
|
||||
/** 本批基线那次运行的结果：分数、判分、其余指标、资源、任务的附加记录；原始基线没有位置表，不写答案路径。 */
|
||||
export function baselineRunSection(d: Draft): void {
|
||||
const base = d.board.baselineOf(d.round);
|
||||
const run = d.board.publicRunOf(base)!;
|
||||
const j = run.result!.judged!;
|
||||
const others = Object.keys(j.metrics).filter((k) => k !== d.req.study.finalMetric).sort(compareIds);
|
||||
d.section("本批基线那次运行的结果", [
|
||||
...(base === ORIGINAL_BASELINE
|
||||
? [`运行 ${run.plan.run}：任务给的原始代码；${j.verdict === "valid" ? "算数" : "不算数"}（${j.verdictReason}）；分数 ${measured(j.score, score)}`]
|
||||
: runLines(d, run)),
|
||||
resourcesText(run.result!.process, d.board.publicRunOf(ORIGINAL_BASELINE)!.result!.process, d.req.study.wallSeconds),
|
||||
...(others.length === 0 ? [] : [`其余指标：${others.map((k) => `${k} ${measured(j.metrics[k]!, num)}`).join("；")}`]),
|
||||
`任务的附加记录（含合法化，原样）：${JSON.stringify(j.extra)}`,
|
||||
]);
|
||||
}
|
||||
|
||||
export function beliefSection(d: Draft): void {
|
||||
d.section("信念本", [beliefBookOf(d.board, d.req.study, d.round)]);
|
||||
}
|
||||
@@ -194,7 +230,13 @@ export function publicRecords(d: Draft): void {
|
||||
e.content.explanations.forEach((x, i) => d.record(explanationId(e.round, e.author, i + 1), `解释 ${explanationId(e.round, e.author, i + 1)}（${e.author}）`, x));
|
||||
}
|
||||
for (const w of b.published("wrapUp")) d.record(`${w.stageId}.${w.author}`, `${w.author} 第 ${w.round} 轮的收尾`, w.content.body);
|
||||
for (const r of b.runs()) if (r.plan.purpose !== "baseline" && b.runPublic(r)) d.record(r.plan.run, `运行 ${r.plan.run}`, [...runLines(d, r), "", json(r.result!.judged)].join("\n"));
|
||||
for (const r of b.runs()) {
|
||||
if (r.plan.purpose !== "baseline" && r.plan.purpose !== "diagnosis" && b.runPublic(r)) d.record(r.plan.run, `运行 ${r.plan.run}`, [...runLines(d, r), "", json(r.result!.judged)].join("\n"));
|
||||
}
|
||||
// 诊断的完整结果：各人给的三个数（结算输入）不放，只有合并后的范围和逐个实验的判分。
|
||||
for (const { round, input: _input, ...settled } of b.diagnosisSettlements()) {
|
||||
d.record(`diagnosis-r${round}`, `第 ${round} 轮诊断的完整结果`, json({ explanations: b.diagnosisExplanations(round), experiments: b.diagnosisExperiments(round), ...settled }));
|
||||
}
|
||||
d.section("研究记录", [
|
||||
"已经公开的记录，一行一条；全文用 research_query 按编号查。",
|
||||
...d.records.map((r) => `- ${r.id}：${r.title}`),
|
||||
|
||||
@@ -12,6 +12,11 @@ const BUILDERS: { [K in SubmissionKind]: (d: Draft) => Extract<Checklist, { kind
|
||||
entry: researcher.entry,
|
||||
discussion: researcher.discussion,
|
||||
claim: researcher.claim,
|
||||
diagnosisProposal: researcher.diagnosisProposal,
|
||||
diagnosisPrediction: researcher.diagnosisPrediction,
|
||||
diagnosisImplementation: assigned.diagnosisImplementation,
|
||||
diagnosisRepair: assigned.diagnosisRepair,
|
||||
diagnosisVerification: assigned.diagnosisVerification,
|
||||
proposal: researcher.proposal,
|
||||
firstPrediction: researcher.firstPrediction,
|
||||
critique: researcher.critique,
|
||||
|
||||
@@ -0,0 +1,78 @@
|
||||
import { runBalance } from "../rules/index.ts";
|
||||
import { measured, num, rangeText, resourcesText, score } from "./format.ts";
|
||||
import type { StudyFacts } from "./inputs.ts";
|
||||
import type { Board, MergedExperiment } from "./read.ts";
|
||||
import { DIAGNOSIS_NOT_RUN, diagnosisId, ORIGINAL_BASELINE, type DiagnosisSettlement } from "./shapes.ts";
|
||||
|
||||
// 一轮诊断的结果写成文字（方法.md 4.3 诊 4、6.10 第 3 条），公共材料的"本轮诊断结果"和信念本的"诊断历史"共用。
|
||||
// 各人对诊断实验给的三个数不放进来，只有合并后的范围。
|
||||
|
||||
const ENDING = { finished: "跑完", crashed: "跑崩了", timeout: "超过分到的时间被杀", notRun: "整次运行先结束，没轮到" } as const;
|
||||
const SKIPPED = { budget: "一次诊断最多做 D 个实验，放不下", value: "信息价值不足以抵过成本" } as const;
|
||||
|
||||
/** 第 round 轮没有诊断的原因（方法.md 4.3 诊 3、诊 4）。 */
|
||||
function noDiagnosisReason(board: Board, study: StudyFacts, round: number): string {
|
||||
const decision = board.decision("selectDiagnosis", round);
|
||||
if (decision === undefined) {
|
||||
const balance = runBalance(study.totalRuns, board.runStatesBefore(round)).balance;
|
||||
return balance < 2 ? `余额只剩 ${balance} 次，不到 2 次，要留 1 次给本轮正向` : "配置关掉了诊断";
|
||||
}
|
||||
if (!decision.output.diagnose) return `信息价值减成本都不大于 0，最好的组合净值 ${num(Math.max(...decision.output.combos.map((c) => c.net)))}`;
|
||||
return `选中的实验核验都没通过、修复次数用完，一个都没跑：${board.eventReason(DIAGNOSIS_NOT_RUN, diagnosisId(round))}`;
|
||||
}
|
||||
|
||||
const signed = (x: number): string => (x > 0 ? `+${score(x)}` : score(x));
|
||||
|
||||
/** 一个合并后的实验：做了什么、实测、结局、测前合并范围、比特、惊讶、用时。 */
|
||||
function experimentLine(board: Board, study: StudyFacts, round: number, s: DiagnosisSettlement, e: MergedExperiment): string {
|
||||
const row = s.output.experiments.find((x) => x.experiment === e.id)!;
|
||||
const judged = s.judged.experiments.find((x) => x.id === e.id);
|
||||
const dropped = s.dropped.find((x) => x.experiment === e.id);
|
||||
const skipped = board.decision("selectDiagnosis", round)!.output.experiments.find((y) => y.experiment === e.id)!;
|
||||
const outcome = {
|
||||
valid: () => `实测 ${measured(judged!.score, score)}，改进量 ${signed(row.gain!)}`,
|
||||
untested: () => `没做：${SKIPPED[skipped.reason!]}${skipped.bestCombo === null ? "" : `（含它的最好组合净值 ${num(skipped.bestCombo.net)}）`}`,
|
||||
insufficient: () => `证据不足：${dropped === undefined ? `${ENDING[judged!.ending]}；${judged!.verdictReason}` : `核验不通过，${dropped.reason}`}`,
|
||||
}[row.outcome]();
|
||||
const parts = [outcome, `测前合并范围 ${rangeText(row.merged)}`];
|
||||
if (row.worldBits !== null) parts.push(`世界比特 ${num(row.worldBits)}`);
|
||||
if (row.surprise !== null) parts.push(`惊讶 ${measured(row.surprise.bits, (x) => `${num(x)} 比特`)}`);
|
||||
if (judged !== undefined) {
|
||||
parts.push(`用时 ${measured(judged.wallSeconds, (x) => `${num(x)} 秒`)}`);
|
||||
const others = Object.keys(judged.metrics).filter((k) => k !== study.finalMetric).sort();
|
||||
if (others.length > 0) parts.push(`其余指标：${others.map((k) => `${k} ${measured(judged.metrics[k]!, num)}`).join("；")}`);
|
||||
}
|
||||
return `- ${e.id} ${e.name}（作者 ${e.authors.join("、")}；动 ${e.spot}）：${parts.join("；")}`;
|
||||
}
|
||||
|
||||
/**
|
||||
* 第 round 轮诊断的结果：每个实验做了什么、实测、用时、结局，每条诊断解释确认或否定，程序推出的方向；没诊断时写原因。
|
||||
* detail 为 true 时每条解释带正文（公共材料），否则只列确认与否定（信念本的诊断历史）。
|
||||
*/
|
||||
export function diagnosisLines(board: Board, study: StudyFacts, round: number, detail: boolean): string[] {
|
||||
const s = board.diagnosisSettlement(round);
|
||||
if (s === undefined) return [`本轮没有诊断：${noDiagnosisReason(board, study, round)}`];
|
||||
const process = board.runs().find((r) => r.plan.run === s.run)!.result!.process;
|
||||
const lines = [
|
||||
`诊断运行 ${s.run}：在基线 ${board.baselineOf(round)}（最终指标 ${score(s.input.baseline)}）上，一次运行做了 ${s.judged.experiments.length} 个实验，只扣 1 次。` +
|
||||
"诊断实验的答案不采纳为基线，只用来看这个方向还剩多少空间。",
|
||||
resourcesText(process, board.publicRunOf(ORIGINAL_BASELINE)!.result!.process, study.wallSeconds),
|
||||
...["", "诊断实验：", ...board.diagnosisExperiments(round).map((e) => experimentLine(board, study, round, s, e)), "", "诊断解释："],
|
||||
];
|
||||
for (const x of board.diagnosisExplanations(round)) {
|
||||
const verdicts = s.output.judgments.find((j) => j.id === x.id)!.verdicts;
|
||||
const of = (v: string) => verdicts.filter((y) => y.verdict === v).map((y) => y.experiment).join("、") || "无";
|
||||
lines.push(`- ${x.id}（作者 ${x.author}）：${verdicts.length === 0 ? "没有有效实验可判" : `确认 ${of("confirmed")}；否定 ${of("denied")}`}`);
|
||||
if (detail) lines.push(...x.body.split("\n").map((l) => ` ${l}`));
|
||||
}
|
||||
const d = s.output.direction;
|
||||
const gain = d.gains.find((g) => g.experiment === d.experiment)!;
|
||||
const way = `程序按诊断推出的方向：${d.experiment}（${gain.measured ? "实测" : "按更新后的权重估计"} ${signed(d.gain)}）。这只是按公式推出来的参考，你可以不跟。`;
|
||||
return [...lines, "", d.found ? way : "诊断没有指出能涨分的方向：所有实验的收益都不为正。"];
|
||||
}
|
||||
|
||||
/** 信念本的"诊断历史"（方法.md 6.10 第 3 条）：每个诊断这一段已走完的轮次一段。 */
|
||||
export function diagnosisHistory(board: Board, study: StudyFacts, latestRound: number): string[] {
|
||||
const rounds = Array.from({ length: latestRound }, (_, i) => i + 1).filter((r) => board.diagnosisDone(r));
|
||||
return rounds.length === 0 ? ["- 还没有"] : rounds.flatMap((r) => [`- 第 ${r} 轮：`, ...diagnosisLines(board, study, r, false).map((l) => (l === "" ? l : ` ${l}`))]);
|
||||
}
|
||||
@@ -7,7 +7,7 @@ export { buildMaterial } from "./build.ts";
|
||||
export type { CodeChanges, PromptFile, Prompts, StudyFacts, TaskRequest } from "./inputs.ts";
|
||||
export { queryMaterial, readPrompts } from "./prompts.ts";
|
||||
// flow 推进时读账本也用这一份（本批基线、基线分数、留池候选、各候选合并了哪些原提案），不另写一份。
|
||||
export { Board, validScore } from "./read.ts";
|
||||
export { Board, validScore, type MergedExperiment } from "./read.ts";
|
||||
export {
|
||||
callId,
|
||||
candidateId,
|
||||
@@ -17,6 +17,7 @@ export {
|
||||
clauseKey,
|
||||
critiqueId,
|
||||
DECISION_KINDS,
|
||||
DIAGNOSIS_NOT_RUN, diagnosisExplanationId, diagnosisId, diagnosisKey, diagnosisNotRunEvent,
|
||||
entryId,
|
||||
experimenterId,
|
||||
verifierId,
|
||||
@@ -33,6 +34,7 @@ export {
|
||||
type DecisionKind,
|
||||
type DecisionShapes,
|
||||
type DedupOutput,
|
||||
type DiagnosisDedupOutput, type DiagnosisLocationTable, type DiagnosisSettlement,
|
||||
type FrozenProtocol,
|
||||
type LocationTable,
|
||||
type Settlement,
|
||||
|
||||
@@ -17,8 +17,10 @@ export type StudyFacts = {
|
||||
lambda: number;
|
||||
/** 正式运行总次数。 */
|
||||
totalRuns: number;
|
||||
/** 每人每轮提案上限 M。 */
|
||||
/** 每人每轮提案上限 M，诊断解释的上限也是它。 */
|
||||
proposalLimit: number;
|
||||
/** 每人每轮诊断实验上限 A；诊断关掉时为 null。 */
|
||||
diagnosisExperimentLimit: number | null;
|
||||
factorLimit: number;
|
||||
/** 最终指标和常规辅助指标的定义（tasks 的 metricDefinitions）。 */
|
||||
metrics: MetricDefinition[];
|
||||
@@ -50,15 +52,15 @@ export type Prompts = { roles: Record<Role, PromptFile>; tasks: Record<Submissio
|
||||
|
||||
/**
|
||||
* 生成一份任务材料要的全部输入。全部必填，用不到的写 null：
|
||||
* - changes：核验是起点到候选代码，修复是起点到上一次提交的代码，消融核验是完整代码到这一格的代码；其余为 null；
|
||||
* - artifacts：解释任务要的本轮各次运行的产物（运行编号 → 文件名 → 文字）；其余为 null；
|
||||
* - changes：核验是起点到候选代码，修复是起点到上一次提交的代码，消融核验是完整代码到这一格的代码；诊断核验是当前基线到诊断程序，诊断修复是当前基线到上一次提交的诊断程序；其余为 null；
|
||||
* - artifacts：解释任务要的本轮各次运行的产物、诊 1 要的本批基线那次运行的产物（运行编号 → 文件名 → 文字）；其余为 null；
|
||||
* - codeRoot：能读的代码版本取出到 `<codeRoot>/<版本号>`，flow 照材料里的目录取出。
|
||||
*/
|
||||
export type TaskRequest = {
|
||||
kind: SubmissionKind;
|
||||
actor: string;
|
||||
stageId: string;
|
||||
/** 一个任务一个对象：候选、原提案（挑刺）或消融格；入口、讨论、志愿、提案、解释、收尾为 null。 */
|
||||
/** 一个任务一个对象：候选、原提案（挑刺）、消融格或本轮诊断（diagnosisId）；入口、讨论、志愿、提案、诊断解释与预测、解释、收尾为 null。 */
|
||||
object: string | null;
|
||||
ledger: Ledger;
|
||||
study: StudyFacts;
|
||||
|
||||
+49
-3
@@ -1,11 +1,13 @@
|
||||
import type { Content, LedgerRecord, StoredSubmission } from "../ledger/index.ts";
|
||||
import { compareIds, type RunState } from "../rules/index.ts";
|
||||
import type { JudgedResult, Range, Stage, SubmissionKind } from "../vocab/index.ts";
|
||||
import type { HypothesisMetric, JudgedResult, Range, Stage, SubmissionKind } from "../vocab/index.ts";
|
||||
import {
|
||||
cellId,
|
||||
diagnosisExplanationId,
|
||||
ORIGINAL_BASELINE,
|
||||
type DecisionKind,
|
||||
type DecisionShapes,
|
||||
type DiagnosisSettlement,
|
||||
type FrozenProtocol,
|
||||
type LocationTable,
|
||||
type Settlement,
|
||||
@@ -22,6 +24,9 @@ export type Run = {
|
||||
result: Content<"runResult"> | null;
|
||||
};
|
||||
|
||||
/** 本轮合并后的一个诊断实验：编号、全部作者、正文（取作者里编号最小的那一份）、登记的指标（编号已是全名"实验编号.短名"）。 */
|
||||
export type MergedExperiment = { id: string; authors: string[]; name: string; spot: string; body: string; metrics: HypothesisMetric[] };
|
||||
|
||||
type StageInfo = { stage: Stage; round: number; roster: string[]; closed: boolean };
|
||||
|
||||
export class Board {
|
||||
@@ -116,12 +121,14 @@ export class Board {
|
||||
|
||||
/**
|
||||
* 运行什么时候公开（任务书 09 设计稿）：基线运行有结果即公开；第 r 轮正向运行在第 r 轮的采纳决定写入后
|
||||
* 整批公开（方法.md 4.3 ⑪）；第 r 轮消融运行在第 r 轮结算写入后公开（4.3 ⑰ ⑱）。
|
||||
* 整批公开（方法.md 4.3 ⑪）；第 r 轮消融运行在第 r 轮结算写入后公开（4.3 ⑰ ⑱）；
|
||||
* 第 r 轮诊断运行在第 r 轮诊断结算写入后公开（4.3 诊 4）。
|
||||
*/
|
||||
runPublic(run: Run): boolean {
|
||||
if (run.result === null) return false;
|
||||
if (run.plan.purpose === "baseline") return true;
|
||||
if (run.plan.purpose === "forward") return this.decision("adopt", run.plan.round) !== undefined;
|
||||
if (run.plan.purpose === "diagnosis") return this.diagnosisSettlement(run.plan.round) !== undefined;
|
||||
return this.settlement(run.plan.round) !== undefined;
|
||||
}
|
||||
|
||||
@@ -140,7 +147,12 @@ export class Board {
|
||||
|
||||
/** 每次运行在账上的状态，算余额用（方法.md 第七部分）。 */
|
||||
runStates(): RunState[] {
|
||||
return this.runs().map((r) => (r.result?.process.ending === "notStarted" ? "notStarted" : r.started ? "started" : "reserved"));
|
||||
return this.runs().map(stateOf);
|
||||
}
|
||||
|
||||
/** 第 round 轮开始前各次运行在账上的状态（算诊断开始时的余额）。 */
|
||||
runStatesBefore(round: number): RunState[] {
|
||||
return this.runs().filter((r) => r.plan.round < round).map(stateOf);
|
||||
}
|
||||
|
||||
/** 某个候选最新的实现或修复提交（已关阶段）。 */
|
||||
@@ -169,6 +181,38 @@ export class Board {
|
||||
return this.#of("settlement").map((r) => ({ round: r.content.round, ...(r.content.result as unknown as Settlement) }));
|
||||
}
|
||||
|
||||
/** 第 round 轮的诊断结算（还没写入为 undefined）。 */
|
||||
diagnosisSettlement(round: number): DiagnosisSettlement | undefined {
|
||||
return this.diagnosisSettlements().find((s) => s.round === round);
|
||||
}
|
||||
|
||||
diagnosisSettlements(): Array<{ round: number } & DiagnosisSettlement> {
|
||||
return this.#of("diagnosisSettlement").map((r) => ({ round: r.content.round, ...(r.content.result as unknown as DiagnosisSettlement) }));
|
||||
}
|
||||
|
||||
/** 第 round 轮诊断这一段走完了没有：提案阶段开了，就是结算写了或本轮不诊断。 */
|
||||
diagnosisDone(round: number): boolean {
|
||||
return [...this.stages.values()].some((s) => s.stage === "proposal" && s.round === round);
|
||||
}
|
||||
|
||||
/** 第 round 轮已公开的诊断解释，按作者、条序排。 */
|
||||
diagnosisExplanations(round: number): Array<{ id: string; author: string; body: string }> {
|
||||
return this.published("diagnosisProposal")
|
||||
.filter((s) => s.round === round)
|
||||
.sort((a, b) => compareIds(a.author, b.author))
|
||||
.flatMap((s) => s.content.explanations.map((x, i) => ({ id: diagnosisExplanationId(round, s.author, i + 1), author: s.author, body: x.body })));
|
||||
}
|
||||
|
||||
/** 第 round 轮合并后的诊断实验：判重合并的决定指明合成自哪几个作者的哪几个实验，正文取第一个。 */
|
||||
diagnosisExperiments(round: number): MergedExperiment[] {
|
||||
const proposals = this.published("diagnosisProposal").filter((s) => s.round === round);
|
||||
const experimentOf = (src: { author: string; index: number }) => proposals.find((s) => s.author === src.author)!.content.experiments[src.index]!;
|
||||
return this.decision("diagnosisDedup", round)!.output.experiments.map(({ experiment, sources }) => {
|
||||
const { name, spot, body, metrics } = experimentOf(sources[0]!);
|
||||
return { id: experiment, authors: [...new Set(sources.map((x) => x.author))], name, spot, body, metrics: metrics.map((m) => ({ ...m, id: `${experiment}.${m.id}` })) };
|
||||
});
|
||||
}
|
||||
|
||||
/** 本人最新的私人笔记（方法.md 第三部分"私人笔记"），没写过为 null。 */
|
||||
note(author: string): string | null {
|
||||
return this.#of("note").filter((r) => r.content.author === author).at(-1)?.content.body ?? null;
|
||||
@@ -249,6 +293,8 @@ export class Board {
|
||||
}
|
||||
}
|
||||
|
||||
const stateOf = (r: Run): RunState => (r.result?.process.ending === "notStarted" ? "notStarted" : r.started ? "started" : "reserved");
|
||||
|
||||
/** 判分结果里的有效分数：算数且测到了分数才有，否则 null。 */
|
||||
export function validScore(judged: JudgedResult | null): number | null {
|
||||
return judged !== null && judged.verdict === "valid" && judged.score.status === "measured" ? judged.score.value : null;
|
||||
|
||||
+38
-12
@@ -1,8 +1,11 @@
|
||||
import { compareIds } from "../rules/index.ts";
|
||||
import { ANSWER_PATH_KEY, SAFEGUARD_TRIGGERS_KEY, type ChecklistOf } from "../vocab/index.ts";
|
||||
import {
|
||||
artifactSection,
|
||||
baselineRunSection,
|
||||
beliefSection,
|
||||
codeSection,
|
||||
diagnosisSection,
|
||||
Draft,
|
||||
entryVersions,
|
||||
factorLines,
|
||||
@@ -33,6 +36,7 @@ function frame(d: Draft, own: Own, middle: () => void): void {
|
||||
researcherTaskSection(d);
|
||||
codeSection(d);
|
||||
if (d.round >= 1) {
|
||||
diagnosisSection(d);
|
||||
publicRecords(d);
|
||||
beliefSection(d);
|
||||
}
|
||||
@@ -126,16 +130,42 @@ function concernLines(d: Draft, round: number): string[] {
|
||||
return lines.length === 0 ? ["（没有）"] : lines;
|
||||
}
|
||||
|
||||
/** 上一轮全部研究者的收尾（提案、诊 1 共用）。 */
|
||||
function previousWrapSection(d: Draft): void {
|
||||
const prev = d.round - 1;
|
||||
const wraps = d.board.published("wrapUp").filter((w) => w.round === prev);
|
||||
d.section("上一轮的收尾", prev < 1 ? ["（这是第 1 轮，没有上一轮）"] : wraps.flatMap((w) => [`### ${w.author} 的收尾`, "", w.content.body, ""]));
|
||||
}
|
||||
|
||||
/** 诊 1 诊断解释与诊断实验：当前基线的代码，产出它的那次运行的答案、分数、指标，信念本，上一轮收尾；第 1 轮的第零轮全部记录在研究记录里。 */
|
||||
export function diagnosisProposal(d: Draft): ChecklistOf<"diagnosisProposal"> {
|
||||
frame(d, [], () => {
|
||||
baselineRunSection(d);
|
||||
artifactSection(d, "本批基线那次运行的答案");
|
||||
previousWrapSection(d);
|
||||
});
|
||||
return { kind: "diagnosisProposal", maxExplanations: d.req.study.proposalLimit, maxExperiments: d.req.study.diagnosisExperimentLimit!, takenMetricIds: factorRules(d).takenMetricIds };
|
||||
}
|
||||
|
||||
/** 诊 2 诊断预测：合并后的诊断实验清单（标明作者）、全部诊断解释；自己的解释另列在"你自己的"里。别人在这一步交的数一概不读（只经 Board.mine 读本人的）。 */
|
||||
export function diagnosisPrediction(d: Draft): ChecklistOf<"diagnosisPrediction"> {
|
||||
const all = d.board.diagnosisExplanations(d.round);
|
||||
const mine = all.filter((x) => x.author === d.req.actor);
|
||||
const experiments = d.board.diagnosisExperiments(d.round);
|
||||
const text = (x: { id: string; author: string; body: string }) => [`#### ${x.id}（作者 ${x.author}）`, "", x.body, ""];
|
||||
frame(d, [{ heading: "你的诊断解释", lines: ["下面每一条都要对每个合并后的诊断实验各给三个数。", "", ...mine.flatMap(text)] }], () => {
|
||||
baselineRunSection(d);
|
||||
const merged = experiments.flatMap((e) => [`#### ${e.id} ${e.name}（作者 ${e.authors.join("、")}）`, "", `动哪一处：${e.spot}`, "", e.body, ""]);
|
||||
d.section("本任务要处理的东西", ["### 合并后的诊断实验", "", ...merged, "### 全部诊断解释（含你自己的）", "", ...all.flatMap(text)]);
|
||||
});
|
||||
return { kind: "diagnosisPrediction", explanations: mine.map((x) => x.id), experiments: experiments.map((e) => e.id) };
|
||||
}
|
||||
|
||||
/** ② 提案：公共材料，加上一轮全部收尾、上一轮实验员的提醒、可以写进"延续"的因子。 */
|
||||
export function proposal(d: Draft): ChecklistOf<"proposal"> {
|
||||
frame(d, [], () => {
|
||||
const prev = d.round - 1;
|
||||
const wraps = d.board.published("wrapUp").filter((w) => w.round === prev);
|
||||
d.section(
|
||||
"上一轮的收尾",
|
||||
prev < 1 ? ["（这是第 1 轮，没有上一轮）"] : wraps.flatMap((w) => [`### ${w.author} 的收尾`, "", w.content.body, ""]),
|
||||
);
|
||||
d.section("上一轮实验员实现时的提醒", prev < 1 ? ["（这是第 1 轮，没有上一轮）"] : concernLines(d, prev));
|
||||
previousWrapSection(d);
|
||||
d.section("上一轮实验员实现时的提醒", d.round < 2 ? ["（这是第 1 轮，没有上一轮）"] : concernLines(d, d.round - 1));
|
||||
const lines = d.board.protocols().flatMap((p) =>
|
||||
p.protocol.factors.map((f, i) => `- ${factorId(p.candidate, factorKey(i))} ${f.name}（关键设定：${f.keySettings}；候选 ${p.candidate}，基线 ${p.protocol.baseline}${f.continues === null ? "" : `；它自己延续 ${f.continues}`}）`),
|
||||
);
|
||||
@@ -258,11 +288,7 @@ export function explanation(d: Draft): ChecklistOf<"explanation"> {
|
||||
"",
|
||||
...factorLines(p),
|
||||
]);
|
||||
const artifacts = d.req.artifacts!;
|
||||
const listed = Object.keys(artifacts)
|
||||
.sort(compareIds)
|
||||
.flatMap((run) => Object.entries(artifacts[run]!).map(([file, text]) => `- ${d.record(`${run}/${file}`, `运行 ${run} 的产物 ${file}`, text)}`));
|
||||
d.section("本轮运行的产物", ["用 research_query 按编号查全文：", ...(listed.length === 0 ? ["（没有）"] : listed)]);
|
||||
artifactSection(d, "本轮运行的产物");
|
||||
});
|
||||
return { kind: "explanation", cells: p.factors.map((_, i) => factorKey(i)), maxExplanations: d.req.study.proposalLimit };
|
||||
}
|
||||
|
||||
+41
-2
@@ -5,13 +5,17 @@ import type {
|
||||
BetResult,
|
||||
CellDecision,
|
||||
CellRuleInput,
|
||||
DiagnosisDecision,
|
||||
DiagnosisRuleInput,
|
||||
EntryAssignment,
|
||||
SettleDiagnosisInput,
|
||||
SettleInput,
|
||||
adopt,
|
||||
settleAblation,
|
||||
settleDiagnosis,
|
||||
} from "../rules/index.ts";
|
||||
import type { Content } from "../ledger/index.ts";
|
||||
import type { Json, Role, Submission, SubmissionKind } from "../vocab/index.ts";
|
||||
import type { Json, Measured, Role, Submission, SubmissionKind, Verdict } from "../vocab/index.ts";
|
||||
|
||||
// 账本里只存成 Json 的几样内容长什么样、编号怎么起（用户 2026-10-03 决定 2）：materials 读它们，flow 照着写。
|
||||
// 决定和结算的输出直接用 rules 的返回类型；ledger 不认识这些类型，只存 Json。
|
||||
@@ -51,6 +55,12 @@ export function splitCell(cell: string): { candidate: string; key: string } {
|
||||
}
|
||||
/** 解释：第 round 轮作者交的第 i 条（从 1 数）。结算里判确认、否定用它。 */
|
||||
export const explanationId = (round: number, author: string, i: number): string => `r${round}-${author}-x${i}`;
|
||||
/** 本轮诊断：诊断程序的冻结代码、诊断运行、诊断实现与核验的对象都是它。 */
|
||||
export const diagnosisId = (round: number): string => `r${round}-diagnosis`;
|
||||
/** 诊断实验在本轮合并后清单里的编号：第 i 个（从 0 数）是 d(i+1)；它登记的指标全名是"实验编号.短名"。 */
|
||||
export const diagnosisKey = (i: number): string => `d${i + 1}`;
|
||||
/** 诊断解释：第 round 轮作者交的第 i 条（从 1 数）。 */
|
||||
export const diagnosisExplanationId = (round: number, author: string, i: number): string => `r${round}-${author}-dx${i}`;
|
||||
/** 模型调用：同一个任务（阶段、谁、哪种提交、对象）第 attempt 次发起。 */
|
||||
export const callId = (stageId: string, actor: string, kind: string, object: string | null, attempt: number): string =>
|
||||
`${stageId}.${actor}.${kind}.${object ?? "none"}.${attempt}`;
|
||||
@@ -104,11 +114,19 @@ export function freezeProtocol(body: ProtocolBody, candidate: string, round: num
|
||||
|
||||
/** 候选代码的位置表，就是实验员交的那三项；基线和消融格的代码记 null（消融格用完整代码的位置表）。 */
|
||||
export type LocationTable = Pick<Submission<"implementation">, "factorLocations" | "safeguardLocations" | "unassigned">;
|
||||
/** 诊断程序的位置表：每个选中的实验一项（诊断实现交的 locations）。 */
|
||||
export type DiagnosisLocationTable = Submission<"diagnosisImplementation">["locations"];
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// 决定（账本记录 decision：decision 是种类，input、output 如下）
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
/**
|
||||
* 诊断实验判重合并的输出（方法.md 4.3 诊 2）：合并后的每个实验由哪几个作者的哪几个实验（index 是它在作者那份诊断提案里的位置，从 0 数）合成；
|
||||
* sources 按作者编号排，第一个就是"作者里编号最小的那一份"，合并后的正文取它的。
|
||||
*/
|
||||
export type DiagnosisDedupOutput = { experiments: Array<{ experiment: string; sources: Array<{ author: string; index: number }> }> };
|
||||
|
||||
/** 判重合并的输出（方法.md 4.3 ③）：本轮新候选各合并了哪几条提案；留池候选各并进了哪几条新提案（可以没有）。 */
|
||||
export type DedupOutput = {
|
||||
fresh: Array<{ candidate: string; proposals: string[] }>;
|
||||
@@ -127,9 +145,11 @@ export type DecisionShapes = {
|
||||
adopt: { input: Parameters<typeof adopt>[0]; output: ReturnType<typeof adopt> };
|
||||
cells: { input: CellRuleInput; output: CellDecision };
|
||||
bet: { input: AblationModelInput & { cells: SettleInput["cells"] }; output: BetResult };
|
||||
diagnosisDedup: { output: DiagnosisDedupOutput };
|
||||
selectDiagnosis: { input: DiagnosisRuleInput; output: DiagnosisDecision };
|
||||
};
|
||||
export type DecisionKind = keyof DecisionShapes;
|
||||
export const DECISION_KINDS: readonly DecisionKind[] = ["entries", "dedup", "batch", "adopt", "cells", "bet"];
|
||||
export const DECISION_KINDS: readonly DecisionKind[] = ["entries", "dedup", "batch", "adopt", "cells", "bet", "diagnosisDedup", "selectDiagnosis"];
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// 结算（账本记录 settlement.result，方法.md 4.3 ⑰）
|
||||
@@ -142,6 +162,25 @@ export type Settlement = {
|
||||
output: ReturnType<typeof settleAblation>;
|
||||
};
|
||||
|
||||
/**
|
||||
* 一次诊断的结算（账本记录 diagnosisSettlement.result，方法.md 6.12）。judged 是每个跑过的实验单独判分的结果，与 tasks 的
|
||||
* DiagnosisJudged 同形（materials 不引用 tasks）；诊断运行记录的 judged 为 null，逐个实验的结果只在这里。
|
||||
* dropped 是核验不通过、修复用完而没跑的实验和原因；input、output 是 settleDiagnosis 的输入输出，没做的实验在里面记 untested。
|
||||
*/
|
||||
export type DiagnosisSettlement = {
|
||||
run: string;
|
||||
judged: {
|
||||
experiments: Array<{ id: string; ending: "finished" | "crashed" | "timeout" | "notRun"; wallSeconds: Measured; verdict: Verdict; verdictReason: string; score: Measured; metrics: Record<string, Measured>; extra: Json }>;
|
||||
};
|
||||
dropped: Array<{ experiment: string; reason: string }>;
|
||||
input: SettleDiagnosisInput;
|
||||
output: ReturnType<typeof settleDiagnosis>;
|
||||
};
|
||||
|
||||
/** 事件（账本记录 event）：选中的诊断实验核验都没通过、修复用完，一个都不剩，不跑、不扣次（方法.md 4.3 诊 4）。objects 只有本轮诊断的编号。 */
|
||||
export const DIAGNOSIS_NOT_RUN = "diagnosisNotRun";
|
||||
export const diagnosisNotRunEvent = (round: number, reason: string): Content<"event"> => ({ event: DIAGNOSIS_NOT_RUN, reason, objects: [diagnosisId(round)] });
|
||||
|
||||
/**
|
||||
* 事件（账本记录 event）：一格消融核验不通过、重做次数用完，这一格不跑、记"证据不足"（方法.md 4.3 ⑯、6.6）。
|
||||
* 这一格没有运行，信念本从这条事件读原因。objects 只有这一格的编号。
|
||||
|
||||
+20
-15
@@ -1,9 +1,17 @@
|
||||
import type { AblationCell } from "./ablation.ts";
|
||||
import { compareIds, improvement, NET_TOLERANCE } from "./basics.ts";
|
||||
import { ablationCellLimit } from "./budget.ts";
|
||||
import { density } from "./distribution.ts";
|
||||
import { seededSample } from "./random.ts";
|
||||
import { ablationModel, evaluateCombos, type AblationModelInput, type ComboParams, type ComboValue } from "./value.ts";
|
||||
import {
|
||||
ablationModel,
|
||||
ablationView,
|
||||
evaluateCombos,
|
||||
itemGain,
|
||||
updatedWeights,
|
||||
type AblationModelInput,
|
||||
type ComboParams,
|
||||
type ComboValue,
|
||||
} from "./value.ts";
|
||||
|
||||
// 方法.md 6.7"选格"、第十一部分的四条选格规则，以及 6.8 押的因子。
|
||||
|
||||
@@ -121,10 +129,15 @@ function netOf(combos: readonly ComboValue[], key: string): number {
|
||||
* "含这一格的净值最大组合"直接按这条挑（只是记录，不要求净值 > 1e-12）。
|
||||
*/
|
||||
export function topCombo<C extends { factors: readonly string[]; net: number }>(combos: readonly C[]): C | null {
|
||||
return topByNet(combos, (c) => c.factors);
|
||||
}
|
||||
|
||||
/** topCombo 的比较规则，组合里的项（因子或实验）怎么取由 items 给；选格和选诊断实验共用（6.7、6.12）。 */
|
||||
export function topByNet<C extends { net: number }>(combos: readonly C[], items: (c: C) => readonly string[]): C | null {
|
||||
if (combos.length === 0) return null;
|
||||
const top = Math.max(...combos.map((c) => c.net));
|
||||
const near = combos.filter((c) => c.net >= top - NET_TOLERANCE);
|
||||
return near.reduce((best, c) => (compareCombos(c.factors, best.factors) < 0 ? c : best));
|
||||
return near.reduce((best, c) => (compareCombos(items(c), items(best)) < 0 ? c : best));
|
||||
}
|
||||
|
||||
/** 格数少的在前；格数相同时按因子编号逐个比。 */
|
||||
@@ -207,19 +220,11 @@ export function chooseBet(input: AblationModelInput & { cells: readonly Ablation
|
||||
const o = input.cells.find((c) => c.factor === f)?.outcome;
|
||||
return o?.kind === "valid" ? o.score : null;
|
||||
});
|
||||
const raw = model.weights.map((w, j) =>
|
||||
scores.reduce<number>((v, x, c) => (x === null ? v : v * density(model.rawKnots[c]![j]!, x)), w),
|
||||
);
|
||||
const sum = raw.reduce((t, v) => t + v, 0);
|
||||
const weightsKept = sum === 0;
|
||||
const weights = weightsKept ? [...model.weights] : raw.map((v) => v / sum);
|
||||
const view = ablationView(model);
|
||||
const { weights, kept: weightsKept } = updatedWeights(view, scores);
|
||||
const gains = model.factors.map((factor, c) => {
|
||||
const x = scores[c];
|
||||
const gain =
|
||||
x !== null && x !== undefined
|
||||
? model.space * (improvement(x, input.full, input.direction) / model.gain)
|
||||
: model.space * model.expected[c]!.reduce((t, e, j) => t + weights[j]! * e, 0);
|
||||
return { factor, gain, measured: x !== null };
|
||||
const x = scores[c] ?? null;
|
||||
return { factor, gain: itemGain(view, c, weights, x), measured: x !== null };
|
||||
});
|
||||
let best = 0;
|
||||
for (let c = 1; c < gains.length; c++) if (gains[c]!.gain > gains[best]!.gain) best = c;
|
||||
|
||||
@@ -0,0 +1,304 @@
|
||||
import type { Direction, Measured, Range } from "../vocab/index.ts";
|
||||
import { judgeExplanations, type AblationExplanation, type CellOutcome, type ExplanationVerdict } from "./ablation.ts";
|
||||
import { combinations, compareIds, improvement, NET_TOLERANCE } from "./basics.ts";
|
||||
import { bitWidth, disagreement, surprise, worldBits } from "./bits.ts";
|
||||
import { topByNet } from "./cells.ts";
|
||||
import { ablationCellMixture, knotsMean, mapKnots, mixtureRange } from "./distribution.ts";
|
||||
import {
|
||||
itemGain,
|
||||
OPEN_RESERVE,
|
||||
runCost,
|
||||
updatedWeights,
|
||||
usePotential,
|
||||
valueOf,
|
||||
type ComboParams,
|
||||
type ValueModel,
|
||||
} from "./value.ts";
|
||||
|
||||
// 方法.md 6.12 诊断：成分模型、方向收益、信息价值、价值减成本、选实验、结算。
|
||||
// 成分、权重更新、精确积分、部分格网格、价值减成本的公式都在 value.ts，这里只给出诊断的"实测 → 收益"换算：
|
||||
// 收益 g(x) = 改进量 x − b（越小越好的任务按 6.1 换成 b − x），另有收益为 0 的"都不走"。
|
||||
|
||||
/** 一条诊断解释：作者，以及对每个合并后的诊断实验给的三个数（实验做完后的分数，原值）。 */
|
||||
export interface DiagnosisExplanation {
|
||||
id: string;
|
||||
author: string;
|
||||
experiments: ReadonlyArray<{ experiment: string; range: Range }>;
|
||||
}
|
||||
|
||||
export interface DiagnosisModelInput {
|
||||
/** 本批基线 b。 */
|
||||
baseline: number;
|
||||
direction: Direction;
|
||||
lambda: number;
|
||||
/** 合并后的全部诊断实验编号（包括没选进组合的）。 */
|
||||
experiments: readonly string[];
|
||||
/** 每条解释对每个实验都给了三个数。 */
|
||||
explanations: readonly DiagnosisExplanation[];
|
||||
}
|
||||
|
||||
export interface DiagnosisModel extends ValueModel {
|
||||
/** 实验编号，按编号排；ValueModel 里"项的下标"按这个次序。 */
|
||||
experiments: string[];
|
||||
/** 成分：各条解释的编号（按编号排），最后是开放预留。 */
|
||||
components: string[];
|
||||
}
|
||||
|
||||
/**
|
||||
* 成分模型（6.12"成分与权重"）：每条诊断解释一个成分，开放预留一个成分；每个成分对每个实验有一个分布，
|
||||
* 解释用它自己的三个数补尾巴，开放预留用这个实验自己的均匀分布（6.3"诊断"，和消融一格同一个合并函数）。
|
||||
*/
|
||||
export function diagnosisModel(input: DiagnosisModelInput): DiagnosisModel {
|
||||
const experiments = [...input.experiments].sort(compareIds);
|
||||
const explanations = [...input.explanations].sort((a, b) => compareIds(a.id, b.id));
|
||||
const mixtures = experiments.map((d) =>
|
||||
ablationCellMixture(
|
||||
explanations.map((e) => ({ id: e.id, author: e.author, range: e.experiments.find((x) => x.experiment === d)!.range })),
|
||||
input.lambda,
|
||||
),
|
||||
);
|
||||
const scale = (x: number): number => improvement(input.baseline, x, input.direction);
|
||||
const rawKnots = mixtures.map((m) => m.map((c) => c.knots));
|
||||
const scaledKnots = rawKnots.map((row) => row.map((k) => mapKnots(k, scale, input.direction === "minimize")));
|
||||
const weights = mixtures[0]!.map((c) => c.weight);
|
||||
const expected = scaledKnots.map((row) => row.map(knotsMean));
|
||||
return {
|
||||
experiments,
|
||||
components: [...explanations.map((e) => e.id), OPEN_RESERVE],
|
||||
weights,
|
||||
rawKnots,
|
||||
scaledKnots,
|
||||
expected,
|
||||
prior: expected.map((row) => row.reduce((s, e, j) => s + weights[j]! * e, 0)),
|
||||
space: 1,
|
||||
floor: 0,
|
||||
scale,
|
||||
};
|
||||
}
|
||||
|
||||
/** 一个诊断实验组合的价值与成本（6.12"价值减成本"）。 */
|
||||
export interface DiagnosisCombo {
|
||||
/** 组合里的实验，按编号排。 */
|
||||
experiments: string[];
|
||||
voi: number;
|
||||
/** R̂ = ⌈(余额 − 1) ÷ 每批上限⌉。 */
|
||||
rHat: number;
|
||||
mPrime: number;
|
||||
/** 价值 = u × m′ × VOI。 */
|
||||
value: number;
|
||||
/** 成本：诊断只占一次运行，任何组合都一样。 */
|
||||
cost: number;
|
||||
net: number;
|
||||
}
|
||||
|
||||
export interface DiagnosisParams extends ComboParams {
|
||||
/** 一次诊断最多做几个实验 D（6.11）。 */
|
||||
maxExperiments: number;
|
||||
}
|
||||
|
||||
/**
|
||||
* 列出可选组合并算价值、成本、净值（6.12"选实验"第 1、2 条）：余额不到 2 次一个都没有；实验数不超过 D 时只有"全做"，
|
||||
* 超过 D 时是所有恰好 D 个实验的组合，按实验编号字典序列出。ĝ₀ = max(0, max_d E(g_d))。
|
||||
*/
|
||||
export function evaluateDiagnosisCombos(
|
||||
model: DiagnosisModel,
|
||||
params: DiagnosisParams,
|
||||
): { g0: number; combos: DiagnosisCombo[] } {
|
||||
const g0 = Math.max(0, ...model.prior);
|
||||
if (params.balance < 2) return { g0, combos: [] };
|
||||
const rHat = Math.ceil((params.balance - 1) / params.batchCap);
|
||||
const mPrime = usePotential(rHat, params.decay);
|
||||
const cost = runCost(g0, params.batchCap, 1);
|
||||
const size = Math.min(model.experiments.length, params.maxExperiments);
|
||||
const combos = combinations(model.experiments.length, size).map((S) => {
|
||||
const voi = valueOf(model, S);
|
||||
const value = params.u * mPrime * voi;
|
||||
return { experiments: S.map((d) => model.experiments[d]!), voi, rHat, mPrime, value, cost, net: value - cost };
|
||||
});
|
||||
return { g0, combos };
|
||||
}
|
||||
|
||||
export interface DiagnosisRuleInput extends DiagnosisModelInput, DiagnosisParams {}
|
||||
|
||||
/**
|
||||
* 没选中的实验为什么没做（6.12"选实验"第 4 条）：budget 放不下（已选的组合装满了 D 个）；
|
||||
* value 信息价值不足以抵过成本。
|
||||
*/
|
||||
export type SkippedReason = "budget" | "value";
|
||||
|
||||
/** 选诊断实验的结果，含 6.12 要求附上的全部数字；规则名和版本由登记表补上。 */
|
||||
export interface DiagnosisDecision {
|
||||
params: { balance: number; batchCap: number; maxExperiments: number; u: number; decay: number; lambda: number };
|
||||
components: string[];
|
||||
/** 起始权重，与 components 对齐。 */
|
||||
weights: number[];
|
||||
/** 各实验的事前期望收益 E(g_d)，按实验编号排。 */
|
||||
priorGains: Array<{ experiment: string; gain: number }>;
|
||||
/** ĝ₀ = max(0, max_d E(g_d))。 */
|
||||
g0: number;
|
||||
/** 可选的全部组合及其价值、成本、净值（余额不到 2 次时为空）。 */
|
||||
combos: DiagnosisCombo[];
|
||||
/** 本轮做不做诊断：选中了净值 > 1e-12 的组合才做。 */
|
||||
diagnose: boolean;
|
||||
/** 选中要做的实验，按编号排；不诊断时为空。 */
|
||||
chosen: string[];
|
||||
/** 选中组合的净值；不诊断时是 0。 */
|
||||
chosenNet: number;
|
||||
/**
|
||||
* 每个合并后的实验：做不做；没做的写原因，并附"含这个实验的组合里净值最大的那个"；
|
||||
* 没有任何含它的可选组合（余额不到 2 次，或没被任何组合包含）时 bestCombo 为 null。
|
||||
*/
|
||||
experiments: Array<{
|
||||
experiment: string;
|
||||
done: boolean;
|
||||
reason: SkippedReason | null;
|
||||
bestCombo: { experiments: string[]; net: number } | null;
|
||||
}>;
|
||||
}
|
||||
|
||||
/**
|
||||
* 选诊断实验，规则"信息价值减成本"（6.12"选实验"）：取净值最大、且净值 > 1e-12 的组合，净值相差不到 1e-12 的
|
||||
* 视为相同、按实验编号从小到大比（和选格同一条挑法，见 topByNet）；都不大于 1e-12 时本轮不诊断。
|
||||
*/
|
||||
export function selectDiagnosis(input: DiagnosisRuleInput): DiagnosisDecision {
|
||||
const model = diagnosisModel(input);
|
||||
const { g0, combos } = evaluateDiagnosisCombos(model, input);
|
||||
const best = topByNet(combos.filter((c) => c.net > NET_TOLERANCE), (c) => c.experiments);
|
||||
const chosen = best === null ? [] : [...best.experiments];
|
||||
return {
|
||||
params: {
|
||||
balance: input.balance,
|
||||
batchCap: input.batchCap,
|
||||
maxExperiments: input.maxExperiments,
|
||||
u: input.u,
|
||||
decay: input.decay,
|
||||
lambda: input.lambda,
|
||||
},
|
||||
components: model.components,
|
||||
weights: model.weights,
|
||||
priorGains: model.experiments.map((experiment, d) => ({ experiment, gain: model.prior[d]! })),
|
||||
g0,
|
||||
combos,
|
||||
diagnose: best !== null,
|
||||
chosen,
|
||||
chosenNet: best === null ? 0 : best.net,
|
||||
experiments: model.experiments.map((experiment) => {
|
||||
const done = chosen.includes(experiment);
|
||||
const top = topByNet(combos.filter((c) => c.experiments.includes(experiment)), (c) => c.experiments);
|
||||
return {
|
||||
experiment,
|
||||
done,
|
||||
reason: done ? null : chosen.length > 0 ? "budget" : "value",
|
||||
bestCombo: done || top === null ? null : { experiments: [...top.experiments], net: top.net },
|
||||
};
|
||||
}),
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* 一个实验的结局（6.12"结算"）：valid 有效分数；insufficient 证据不足（跑崩、超时、没有产物、修不了、核验不通过）；
|
||||
* untested 没做。和消融一格的三种同形，直接用 ablation.ts 的类型。
|
||||
*/
|
||||
export type ExperimentOutcome = Extract<CellOutcome, { kind: "valid" | "insufficient" | "untested" }>;
|
||||
|
||||
export interface SettleDiagnosisInput {
|
||||
/** 本批基线 b。 */
|
||||
baseline: number;
|
||||
direction: Direction;
|
||||
epsilon: number;
|
||||
lambda: number;
|
||||
/** 合并后的全部诊断实验及结局，没做的写 untested。 */
|
||||
experiments: ReadonlyArray<{ experiment: string; outcome: ExperimentOutcome }>;
|
||||
explanations: readonly DiagnosisExplanation[];
|
||||
}
|
||||
|
||||
/** 诊断推出的方向（6.12"结算"）。 */
|
||||
export interface DiagnosisDirection {
|
||||
/** 更新后的成分权重（与模型的成分对齐）；无从更新时等于起始权重。 */
|
||||
weights: number[];
|
||||
/** 实测落在所有成分的范围之外、权重之和为 0 时为 true，此时保持起始权重。 */
|
||||
weightsKept: boolean;
|
||||
/** 各实验的收益（按编号排）：measured 为 true 的是有效实验实测换算的，其余是按更新后的权重估的。 */
|
||||
gains: Array<{ experiment: string; gain: number; measured: boolean }>;
|
||||
/** 收益最大的实验（严格更大才替换，平局取编号小的）及其收益。 */
|
||||
experiment: string;
|
||||
gain: number;
|
||||
/** 最大收益为正才算指出了方向；否则记"诊断没有指出能涨分的方向"。 */
|
||||
found: boolean;
|
||||
}
|
||||
|
||||
/**
|
||||
* 一次诊断的结算（6.12"结算"、6.9）：每个实验的收益、合并范围，有效实验的世界比特、惊讶、分歧量
|
||||
* （格宽 ε/100，从 b 起切）；每条诊断解释在每个有效实验上的确认或否定（和消融同一个判法）；诊断推出的方向。
|
||||
* 证据不足和没做的实验不判、不记账。
|
||||
*/
|
||||
export function settleDiagnosis(input: SettleDiagnosisInput): {
|
||||
experiments: Array<{
|
||||
experiment: string;
|
||||
outcome: ExperimentOutcome["kind"];
|
||||
gain: number | null;
|
||||
merged: Range;
|
||||
worldBits: number | null;
|
||||
surprise: { cell: number; prior: number; bits: Measured } | null;
|
||||
disagreement: number | null;
|
||||
}>;
|
||||
judgments: Array<{ id: string; verdicts: Array<{ experiment: string; verdict: ExplanationVerdict }> }>;
|
||||
direction: DiagnosisDirection;
|
||||
} {
|
||||
const width = bitWidth(input.epsilon);
|
||||
const experiments = [...input.experiments].sort((a, b) => compareIds(a.experiment, b.experiment));
|
||||
const rows = experiments.map(({ experiment, outcome }) => {
|
||||
const entries = input.explanations.map((e) => ({
|
||||
id: e.id,
|
||||
author: e.author,
|
||||
range: e.experiments.find((x) => x.experiment === experiment)!.range,
|
||||
}));
|
||||
const mixture = ablationCellMixture(entries, input.lambda);
|
||||
const valid = outcome.kind === "valid";
|
||||
return {
|
||||
experiment,
|
||||
outcome: outcome.kind,
|
||||
gain: valid ? improvement(input.baseline, outcome.score, input.direction) : null,
|
||||
merged: mixtureRange(mixture),
|
||||
worldBits: valid ? worldBits(mixture, input.baseline, width) : null,
|
||||
surprise: valid ? surprise(mixture, input.baseline, width, outcome.score) : null,
|
||||
disagreement: valid ? disagreement(entries, input.baseline, width) : null,
|
||||
};
|
||||
});
|
||||
// 确认与否定沿用消融的判法：把"实验"当成"格"交给 judgeExplanations，结果再换回实验的叫法。
|
||||
const asCells: AblationExplanation[] = input.explanations.map((e) => ({
|
||||
id: e.id,
|
||||
author: e.author,
|
||||
cells: e.experiments.map((x) => ({ factor: x.experiment, range: x.range })),
|
||||
}));
|
||||
const judgments = judgeExplanations(asCells, experiments.map((e) => ({ factor: e.experiment, outcome: e.outcome }))).map(
|
||||
(j) => ({ id: j.id, verdicts: j.verdicts.map((v) => ({ experiment: v.factor, verdict: v.verdict })) }),
|
||||
);
|
||||
return {
|
||||
experiments: rows,
|
||||
judgments,
|
||||
direction: diagnosisDirection(
|
||||
diagnosisModel({ ...input, experiments: experiments.map((e) => e.experiment) }),
|
||||
experiments.map((e) => (e.outcome.kind === "valid" ? e.outcome.score : null)),
|
||||
),
|
||||
};
|
||||
}
|
||||
|
||||
/** 诊断推出的方向：按有效实验的实测更新成分权重，有效实验用实测换算收益，其余实验用新权重估，取收益最大的（6.12"结算"）。 */
|
||||
function diagnosisDirection(model: DiagnosisModel, scores: ReadonlyArray<number | null>): DiagnosisDirection {
|
||||
const { weights, kept } = updatedWeights(model, scores);
|
||||
const gains = model.experiments.map((experiment, d) => {
|
||||
const x = scores[d] ?? null;
|
||||
return { experiment, gain: itemGain(model, d, weights, x), measured: x !== null };
|
||||
});
|
||||
let best = 0;
|
||||
for (let d = 1; d < gains.length; d++) if (gains[d]!.gain > gains[best]!.gain) best = d;
|
||||
return {
|
||||
weights,
|
||||
weightsKept: kept,
|
||||
gains,
|
||||
experiment: gains[best]!.experiment,
|
||||
gain: gains[best]!.gain,
|
||||
found: gains[best]!.gain > 0,
|
||||
};
|
||||
}
|
||||
@@ -103,6 +103,23 @@ export {
|
||||
worldBits,
|
||||
type Placement,
|
||||
} from "./bits.ts";
|
||||
export {
|
||||
diagnosisModel,
|
||||
evaluateDiagnosisCombos,
|
||||
selectDiagnosis,
|
||||
settleDiagnosis,
|
||||
type DiagnosisCombo,
|
||||
type DiagnosisDecision,
|
||||
type DiagnosisDirection,
|
||||
type DiagnosisExplanation,
|
||||
type DiagnosisModel,
|
||||
type DiagnosisModelInput,
|
||||
type DiagnosisParams,
|
||||
type DiagnosisRuleInput,
|
||||
type ExperimentOutcome,
|
||||
type SettleDiagnosisInput,
|
||||
type SkippedReason,
|
||||
} from "./diagnosis.ts";
|
||||
export { ablationCellLimit, forwardBatchSize, runBalance, type RunState } from "./budget.ts";
|
||||
export { authorVsMerged, trackRecord, type Phase, type Tally } from "./record.ts";
|
||||
export { assignEntries, type EntryAssignment } from "./entries.ts";
|
||||
@@ -110,8 +127,11 @@ export { decisionSeed, seededRandom, seededSample } from "./random.ts";
|
||||
export {
|
||||
BATCH_RULES,
|
||||
CELL_RULES,
|
||||
DIAGNOSIS_RULES,
|
||||
runBatchRule,
|
||||
runCellRule,
|
||||
runDiagnosisRule,
|
||||
type BatchRuleName,
|
||||
type CellRuleName,
|
||||
type DiagnosisRuleName,
|
||||
} from "./registry.ts";
|
||||
|
||||
+7
-5
@@ -2,10 +2,12 @@ import type { Direction, Range } from "../vocab/index.ts";
|
||||
import { compareIds, improvement } from "./basics.ts";
|
||||
import { covers, placement, type Placement } from "./bits.ts";
|
||||
|
||||
// 方法.md 6.10 第 4 条"战绩"、6.9"按人分开、按正向与消融分开"。只用来显示，不改任何权重。
|
||||
// 方法.md 6.10 第 4 条"战绩"、6.9"按人分开、按正向、诊断与消融分开"。只用来显示，不改任何权重。
|
||||
// 一律按分数原值比，不按方向折算。
|
||||
|
||||
export type Phase = "forward" | "ablation";
|
||||
export type Phase = "forward" | "diagnosis" | "ablation";
|
||||
|
||||
const PHASE_ORDER: readonly Phase[] = ["forward", "diagnosis", "ablation"];
|
||||
|
||||
/** 一个人的战绩。 */
|
||||
export interface Tally {
|
||||
@@ -25,8 +27,8 @@ export interface Tally {
|
||||
}
|
||||
|
||||
/**
|
||||
* 战绩：按研究者、按正向与消融分开数（6.9、6.10）。samples 是每人每次的范围（第二次预测或解释的每格）和有效实测。
|
||||
* 结果按研究者编号、再按 forward、ablation 排，只列有样本的组合。
|
||||
* 战绩：按研究者、按正向、诊断与消融分开数（6.9、6.10）。samples 是每人每次的范围（第二次预测、诊断预测或解释的每格）和有效实测。
|
||||
* 结果按研究者编号、再按 forward、diagnosis、ablation 排，只列有样本的组合。
|
||||
*/
|
||||
export function trackRecord(
|
||||
samples: ReadonlyArray<{ researcher: string; phase: Phase; range: Range; value: number }>,
|
||||
@@ -62,7 +64,7 @@ export function trackRecord(
|
||||
t.placements[placement(r, v)] += 1;
|
||||
}
|
||||
return [...groups.values()].sort(
|
||||
(a, b) => compareIds(a.researcher, b.researcher) || (a.phase === b.phase ? 0 : a.phase === "forward" ? -1 : 1),
|
||||
(a, b) => compareIds(a.researcher, b.researcher) || PHASE_ORDER.indexOf(a.phase) - PHASE_ORDER.indexOf(b.phase),
|
||||
);
|
||||
}
|
||||
|
||||
|
||||
@@ -1,4 +1,5 @@
|
||||
import { selectAll, selectByValue, selectNone, selectRandomSameCount, type CellDecision, type CellRuleInput } from "./cells.ts";
|
||||
import { selectDiagnosis, type DiagnosisDecision, type DiagnosisRuleInput } from "./diagnosis.ts";
|
||||
import { selectBestExpected, selectUniformRandom, type BatchDecision, type BatchInput } from "./forward.ts";
|
||||
|
||||
// 规则登记表（结构.md 第五部分"决策规则"、方法.md 第十一部分）：名字 → 函数，带版本号。
|
||||
@@ -22,8 +23,14 @@ export const CELL_RULES = {
|
||||
none: { version: 1, run: selectNone },
|
||||
} as const satisfies Record<string, { version: number; run: (input: CellRuleInput) => CellDecision }>;
|
||||
|
||||
export const DIAGNOSIS_RULES = {
|
||||
/** 信息价值减成本（6.12）；选诊断实验只有这一条规则，配置里不选。 */
|
||||
valueMinusCost: { version: 1, run: selectDiagnosis },
|
||||
} as const satisfies Record<string, { version: number; run: (input: DiagnosisRuleInput) => DiagnosisDecision }>;
|
||||
|
||||
export type BatchRuleName = keyof typeof BATCH_RULES;
|
||||
export type CellRuleName = keyof typeof CELL_RULES;
|
||||
export type DiagnosisRuleName = keyof typeof DIAGNOSIS_RULES;
|
||||
|
||||
/** 按名字跑一条选批规则，结果带上规则名和版本。 */
|
||||
export function runBatchRule(
|
||||
@@ -42,3 +49,12 @@ export function runCellRule(
|
||||
const entry = CELL_RULES[name];
|
||||
return { rule: { name, version: entry.version }, ...entry.run(input) };
|
||||
}
|
||||
|
||||
/** 按名字跑一条选诊断实验规则，结果带上规则名和版本。 */
|
||||
export function runDiagnosisRule(
|
||||
name: DiagnosisRuleName,
|
||||
input: DiagnosisRuleInput,
|
||||
): DiagnosisDecision & { rule: { name: DiagnosisRuleName; version: number } } {
|
||||
const entry = DIAGNOSIS_RULES[name];
|
||||
return { rule: { name, version: entry.version }, ...entry.run(input) };
|
||||
}
|
||||
|
||||
+139
-31
@@ -9,13 +9,44 @@ import {
|
||||
GRID_SUBDIVISIONS,
|
||||
improvement,
|
||||
} from "./basics.ts";
|
||||
import { ablationCellMixture, below, integrateUpperTail, knotsMean, mapKnots, type Knots } from "./distribution.ts";
|
||||
import {
|
||||
ablationCellMixture,
|
||||
below,
|
||||
density,
|
||||
integrateUpperTail,
|
||||
knotsMean,
|
||||
mapKnots,
|
||||
type Knots,
|
||||
} from "./distribution.ts";
|
||||
|
||||
// 方法.md 6.7：方向收益、成分与权重、信息价值、价值减成本。选格规则和押的因子在 cells.ts。
|
||||
//
|
||||
// 一律先在"撑几成"刻度 s(x) = [f(完整) − x] ÷ G（按方向折算）上算，最后乘剩余空间 h：
|
||||
// 收益_p = h × s_p，h 对所有因子相同且为正，所以信息价值、期望收益、ĝ₀ 都等于 h × (撑几成刻度上的同一个量)。
|
||||
// 例外是部分格网格的切段：在分数原值上取折点、切小段，再把中点换成撑几成（见 measuredGrid）。
|
||||
// 成分、权重更新、精确积分、部分格网格、价值减成本这套引擎消融（6.7）和诊断（6.12）共用，只差"实测 → 收益"的换算：
|
||||
// - 消融：一律先在"撑几成"刻度 s(x) = [f(完整) − x] ÷ G（按方向折算）上算，最后乘剩余空间 h：
|
||||
// 收益_p = h × s_p，h 对所有因子相同且为正，所以信息价值、期望收益、ĝ₀ 都等于 h × (撑几成刻度上的同一个量)；
|
||||
// 不设"都不走"的选项。
|
||||
// - 诊断（diagnosis.ts）：收益 g(x) = 改进量 x − b，h 取 1，另有收益为 0 的"都不走"，即 floor = 0。
|
||||
// 例外是部分格网格的切段：在分数原值上取折点、切小段，再把中点换成收益刻度（见 gridOf）。
|
||||
|
||||
/** 引擎要的成分模型（消融和诊断各自建出来）。下标"项"是因子（消融）或实验（诊断）。 */
|
||||
export interface ValueModel {
|
||||
/** 成分起始权重。 */
|
||||
weights: number[];
|
||||
/** rawKnots[项][成分]：分数原值上的折线。 */
|
||||
rawKnots: Knots[][];
|
||||
/** scaledKnots[项][成分]：收益刻度上的折线。 */
|
||||
scaledKnots: Knots[][];
|
||||
/** expected[项][成分]：这个成分下这一项在收益刻度上的期望。 */
|
||||
expected: number[][];
|
||||
/** prior[项]：按起始权重平均的期望。 */
|
||||
prior: number[];
|
||||
/** 收益 = space × 收益刻度上的值：消融是剩余空间 h，诊断是 1。 */
|
||||
space: number;
|
||||
/** "都不走"的收益：诊断是 0，消融没有这个选项，写 null。 */
|
||||
floor: number | null;
|
||||
/** 分数原值 → 收益刻度。 */
|
||||
scale: (x: number) => number;
|
||||
}
|
||||
|
||||
/** 开放预留成分的名字（成分名单里排在最后）。 */
|
||||
export const OPEN_RESERVE = "openReserve";
|
||||
@@ -70,6 +101,20 @@ function supportScale(full: number, direction: Direction, gain: number): (x: num
|
||||
return (x) => improvement(x, full, direction) / gain;
|
||||
}
|
||||
|
||||
/** 消融模型在引擎眼里的样子：收益刻度是撑几成，没有"都不走"。 */
|
||||
export function ablationView(model: AblationModel): ValueModel {
|
||||
return {
|
||||
weights: model.weights,
|
||||
rawKnots: model.rawKnots,
|
||||
scaledKnots: model.supportKnots,
|
||||
expected: model.expected,
|
||||
prior: model.prior,
|
||||
space: model.space,
|
||||
floor: null,
|
||||
scale: supportScale(model.full, model.direction, model.gain),
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* 成分模型（6.7"成分与权重"）：每条解释一个成分，开放预留一个成分；每个成分对每一格有一个分布，
|
||||
* 解释用它自己的三个数补尾巴，开放预留用这一格的均匀分布（6.3"消融"）。同一成分内各格互相独立。
|
||||
@@ -144,26 +189,33 @@ export function cellGrid(knots: readonly Knots[], k: number): { mids: number[];
|
||||
}
|
||||
|
||||
/**
|
||||
* 模型里第 c 格的网格（6.7"只测一部分格"），测 k 格时用。在分数原值上取折点、切小段：方法说"每格在自己的开放预留范围内
|
||||
* 模型里第 c 项的网格（6.7"只测一部分格"），测 k 项时用。在分数原值上取折点、切小段：方法说"每格在自己的开放预留范围内
|
||||
* 按折点切段"，开放预留和折点都是在分数原值上定的（尾巴按 q1 − (q5 − q1)、q9 + (q9 − q5)，预留按 L − W/2、U + W/2
|
||||
* 的浮点原式），去重只合并浮点上完全相等的值。先换到撑几成刻度再去重，相差一个末位的两个原值折点可能换算后变成同一个数，
|
||||
* 段数和 s 就会跟原值上的不一样。切好后只把小段中点换算成撑几成；把握与刻度无关。
|
||||
* 的浮点原式），去重只合并浮点上完全相等的值。先换到收益刻度再去重，相差一个末位的两个原值折点可能换算后变成同一个数，
|
||||
* 段数和 s 就会跟原值上的不一样。切好后只把小段中点换算成收益刻度；把握与刻度无关。
|
||||
*/
|
||||
export function measuredGrid(model: AblationModel, c: number, k: number): { mids: number[]; mass: number[][] } {
|
||||
export function gridOf(model: ValueModel, c: number, k: number): { mids: number[]; mass: number[][] } {
|
||||
const { mids, mass } = cellGrid(model.rawKnots[c]!, k);
|
||||
return { mids: mids.map(supportScale(model.full, model.direction, model.gain)), mass };
|
||||
return { mids: mids.map(model.scale), mass };
|
||||
}
|
||||
|
||||
/** 消融模型里第 c 格的网格，中点换算成撑几成（6.7"只测一部分格"）。 */
|
||||
export function measuredGrid(model: AblationModel, c: number, k: number): { mids: number[]; mass: number[][] } {
|
||||
return gridOf(ablationView(model), c, k);
|
||||
}
|
||||
|
||||
/**
|
||||
* 只测 S 里的格时的 E_{S 的结果}[max_p E(撑_p | S 的结果)]（6.7"信息价值"），在撑几成刻度上。
|
||||
* 对测到的格的全部小段组合求和：成分 j 的联合把握 = 起始权重 × 各测到格在所选小段的把握；它们的和 marg 是这个组合的把握，
|
||||
* 按它归一就是更新后的权重。网格见 measuredGrid。测到的格取小段中点（已换成撑几成），没测的格用更新后的权重平均各成分的期望，取各因子里最大的。
|
||||
* 只测 S 里的项时的 E_{S 的结果}[max(floor, max_p E(收益_p | S 的结果))]（6.7"信息价值"、6.12），在收益刻度上，
|
||||
* 没有"都不走"（floor 为 null）时不和 floor 取大。
|
||||
* 对测到的项的全部小段组合求和：成分 j 的联合把握 = 起始权重 × 各测到项在所选小段的把握；它们的和 marg 是这个组合的把握，
|
||||
* 按它归一就是更新后的权重。网格见 gridOf。测到的项取小段中点（已换成收益刻度），没测的项用更新后的权重平均各成分的期望，取各项里最大的。
|
||||
*/
|
||||
function expectedBestGrid(model: AblationModel, S: readonly number[]): number {
|
||||
function expectedBestGrid(model: ValueModel, S: readonly number[]): number {
|
||||
const k = S.length;
|
||||
const J = model.weights.length;
|
||||
const grids = S.map((c) => measuredGrid(model, c, k));
|
||||
const rest = model.factors.map((_, c) => c).filter((c) => !S.includes(c));
|
||||
const floor = model.floor ?? -Infinity;
|
||||
const grids = S.map((c) => gridOf(model, c, k));
|
||||
const rest = model.prior.map((_, c) => c).filter((c) => !S.includes(c));
|
||||
const partial = Array.from({ length: k + 1 }, () => new Float64Array(J));
|
||||
for (let j = 0; j < J; j++) partial[0]![j] = model.weights[j]!;
|
||||
let total = 0;
|
||||
@@ -173,7 +225,7 @@ function expectedBestGrid(model: AblationModel, S: readonly number[]): number {
|
||||
let marg = 0;
|
||||
for (let j = 0; j < J; j++) marg += cur[j]!;
|
||||
if (marg === 0) return;
|
||||
let best = seen;
|
||||
let best = Math.max(seen, floor);
|
||||
for (const c of rest) {
|
||||
const e = model.expected[c]!;
|
||||
let num = 0;
|
||||
@@ -195,36 +247,80 @@ function expectedBestGrid(model: AblationModel, S: readonly number[]): number {
|
||||
}
|
||||
|
||||
/**
|
||||
* 全部格都测时的 E[max_p 撑_p]（6.7"信息价值"）：同一成分内"各格撑几成的最大值低于 t"的把握 = 各格把握连乘，
|
||||
* 期望 = 下端 + ∫(1 − 连乘) dt，逐段多项式精确积分；再按起始权重相加。
|
||||
* 全部项都测时的 E[max(floor, max_p 收益_p)]（6.7"信息价值"、6.12）：同一成分内"各项最大值低于 t"的把握 = 各项把握连乘，
|
||||
* 期望 = 下端 + ∫(1 − 连乘) dt，逐段多项式精确积分；再按起始权重相加。下端：有"都不走"时是 floor（t 低于它时
|
||||
* max(floor, ·) 一定不低于 t，被积函数是 1），没有时是各项折线的最低点。
|
||||
*/
|
||||
function expectedBestAll(model: AblationModel): number {
|
||||
function expectedBestAll(model: ValueModel): number {
|
||||
let v = 0;
|
||||
for (let j = 0; j < model.weights.length; j++) {
|
||||
const lines = model.supportKnots.map((cell) => cell[j]!);
|
||||
const lo = Math.min(...lines.map((k) => k[0]![0]));
|
||||
const lines = model.scaledKnots.map((cell) => cell[j]!);
|
||||
const lo = model.floor ?? Math.min(...lines.map((k) => k[0]![0]));
|
||||
v += model.weights[j]! * (lo + integrateUpperTail(lines.map((knots) => [{ weight: 1, knots }]), lo));
|
||||
}
|
||||
return v;
|
||||
}
|
||||
|
||||
/** max(floor, max_p E(收益_p))：不测时能拿到的最好期望，在收益刻度上。 */
|
||||
function bestPrior(model: ValueModel): number {
|
||||
return Math.max(model.floor ?? -Infinity, ...model.prior);
|
||||
}
|
||||
|
||||
/** 用网格算的信息价值（不论测几项都走网格），S 是项的下标。 */
|
||||
export function valueByGrid(model: ValueModel, S: readonly number[]): number {
|
||||
return model.space * (expectedBestGrid(model, [...S].sort((a, b) => a - b)) - bestPrior(model));
|
||||
}
|
||||
|
||||
/** 全部项都测时的信息价值（一维精确积分）。 */
|
||||
export function valueExact(model: ValueModel): number {
|
||||
return model.space * (expectedBestAll(model) - bestPrior(model));
|
||||
}
|
||||
|
||||
/**
|
||||
* 信息价值 VOI(S) = E_{S 的结果}[max(floor, max_p E(收益_p | S 的结果))] − max(floor, max_p E(收益_p))（6.7、6.12），
|
||||
* S 是项的下标；全部项都测走一维精确积分，只测一部分走网格，空集是 0。不截断。
|
||||
*/
|
||||
export function valueOf(model: ValueModel, S: readonly number[]): number {
|
||||
if (S.length === 0) return 0;
|
||||
return S.length === model.prior.length ? valueExact(model) : valueByGrid(model, S);
|
||||
}
|
||||
|
||||
/** 网格算法算的信息价值，S 是格的下标（不论测几格都走网格；用来和精确算法互相对照）。 */
|
||||
export function informationValueGrid(model: AblationModel, S: readonly number[]): number {
|
||||
return model.space * (expectedBestGrid(model, [...S].sort((a, b) => a - b)) - Math.max(...model.prior));
|
||||
return valueByGrid(ablationView(model), S);
|
||||
}
|
||||
|
||||
/** 全部格都测时的信息价值（一维精确积分）。 */
|
||||
export function informationValueExact(model: AblationModel): number {
|
||||
return model.space * (expectedBestAll(model) - Math.max(...model.prior));
|
||||
return valueExact(ablationView(model));
|
||||
}
|
||||
|
||||
/** 消融的信息价值 VOI(S)（6.7），S 是格的下标。 */
|
||||
export function informationValue(model: AblationModel, S: readonly number[]): number {
|
||||
return valueOf(ablationView(model), S);
|
||||
}
|
||||
|
||||
/**
|
||||
* 信息价值 VOI(S) = E_{S 的结果}[max_p E(收益_p | S 的结果)] − max_p E(收益_p)（6.7），S 是格的下标；
|
||||
* 全部格都测走一维精确积分，只测一部分走网格，空集是 0。不截断。
|
||||
* 成分新权重（6.8、6.12）：起始权重 × 各有效项（scores 里不是 null 的）上这个成分在实测处的密度（分数原值上，连乘），
|
||||
* 再归一；实测落在所有成分之外、权重之和为 0 时保持起始权重。scores 与项的下标对齐，没有有效实测的项为 null。
|
||||
*/
|
||||
export function informationValue(model: AblationModel, S: readonly number[]): number {
|
||||
if (S.length === 0) return 0;
|
||||
return S.length === model.factors.length ? informationValueExact(model) : informationValueGrid(model, S);
|
||||
export function updatedWeights(
|
||||
model: ValueModel,
|
||||
scores: ReadonlyArray<number | null>,
|
||||
): { weights: number[]; kept: boolean } {
|
||||
const raw = model.weights.map((w, j) =>
|
||||
scores.reduce<number>((v, x, c) => (x === null ? v : v * density(model.rawKnots[c]![j]!, x)), w),
|
||||
);
|
||||
const sum = raw.reduce((t, v) => t + v, 0);
|
||||
const kept = sum === 0;
|
||||
return { weights: kept ? [...model.weights] : raw.map((v) => v / sum), kept };
|
||||
}
|
||||
|
||||
/** 第 c 项的方向收益（6.8、6.12）：有有效实测 x 就用实测换算，否则用 weights 对各成分的期望加权平均。 */
|
||||
export function itemGain(model: ValueModel, c: number, weights: readonly number[], x: number | null): number {
|
||||
return x !== null
|
||||
? model.space * model.scale(x)
|
||||
: model.space * model.expected[c]!.reduce((t, e, j) => t + weights[j]! * e, 0);
|
||||
}
|
||||
|
||||
/** 一个测格组合的价值与成本（6.7"价值减成本"）。 */
|
||||
@@ -253,6 +349,19 @@ export interface ComboParams {
|
||||
decay: number;
|
||||
}
|
||||
|
||||
/** m′ = 1 + r + … + r^(R̂−1)，R̂ = 0 时是 0（6.7"价值减成本"）；消融和诊断共用。 */
|
||||
export function usePotential(rHat: number, decay: number): number {
|
||||
let mPrime = 0;
|
||||
for (let j = 0; j < rHat; j++) mPrime += decay ** j;
|
||||
return mPrime;
|
||||
}
|
||||
|
||||
/** 成本 = 2 max(0, ĝ₀) × [(K + n)/(K + n + 1) − K/(K + 1)]，n 是这次占用的运行次数（6.7、6.12）；消融和诊断共用。 */
|
||||
export function runCost(g0: number, batchCap: number, runs: number): number {
|
||||
const K = batchCap;
|
||||
return 2 * Math.max(0, g0) * ((K + runs) / (K + runs + 1) - K / (K + 1));
|
||||
}
|
||||
|
||||
/**
|
||||
* 列出格数 1..min(格数, 余额 − 1) 的全部非空组合并算价值、成本、净值（6.7"选格"第 1 条）。余额为 0 或 1 时一个都没有。
|
||||
* 组合按格数从少到多、同格数按因子编号字典序列出。ĝ₀ = h × max_p 期望撑几成，原值另给，成本里取 max(0, ĝ₀)。
|
||||
@@ -265,10 +374,9 @@ export function evaluateCombos(model: AblationModel, params: ComboParams): { g0:
|
||||
for (const S of combinations(model.factors.length, size)) {
|
||||
const voi = informationValue(model, S);
|
||||
const rHat = Math.ceil((params.balance - size) / K);
|
||||
let mPrime = 0;
|
||||
for (let j = 0; j < rHat; j++) mPrime += params.decay ** j;
|
||||
const mPrime = usePotential(rHat, params.decay);
|
||||
const value = params.u * mPrime * voi;
|
||||
const cost = 2 * Math.max(0, g0) * ((K + size) / (K + size + 1) - K / (K + 1));
|
||||
const cost = runCost(g0, K, size);
|
||||
combos.push({ factors: S.map((c) => model.factors[c]!), voi, rHat, mPrime, value, cost, net: value - cost });
|
||||
}
|
||||
}
|
||||
|
||||
@@ -23,6 +23,15 @@ export function programProblems(inc: Incoming, protocol: ProtocolBody | null): s
|
||||
...factorProblems(inc.content, inc.checklist),
|
||||
...increasing("authorEstimate", inc.content.authorEstimate),
|
||||
];
|
||||
case "diagnosisProposal":
|
||||
return diagnosisProposalProblems(inc.content, inc.checklist);
|
||||
case "diagnosisPrediction":
|
||||
return diagnosisPredictionProblems(inc.content, inc.checklist);
|
||||
case "diagnosisImplementation":
|
||||
case "diagnosisRepair":
|
||||
return codeProblems(inc.known.code);
|
||||
case "diagnosisVerification":
|
||||
return covers("experiments 的实验", inc.content.experiments.map((x) => x.experiment), inc.checklist.experiments);
|
||||
case "firstPrediction":
|
||||
case "secondPrediction":
|
||||
return increasing("range", inc.content.range);
|
||||
@@ -63,6 +72,45 @@ function claimProblems(c: Submission<"claim">, k: ChecklistOf<"claim">): string[
|
||||
];
|
||||
}
|
||||
|
||||
/**
|
||||
* 诊断解释与诊断实验（方法.md 4.3 诊 1、第五部分"指标"）：解释 1 至 M 条、实验 1 至 A 个（下限由字段定义管）；
|
||||
* 诊断实验登记的指标只能选"缺了只记缺失"；同一个实验里指标短名不重名、不占用已有编号
|
||||
* （不同实验的同名短名冻结时各带自己实验的前缀，不会撞）。
|
||||
*/
|
||||
function diagnosisProposalProblems(c: Submission<"diagnosisProposal">, k: ChecklistOf<"diagnosisProposal">): string[] {
|
||||
return [
|
||||
...(c.explanations.length > k.maxExplanations ? [`交了 ${c.explanations.length} 条诊断解释，最多 ${k.maxExplanations} 条`] : []),
|
||||
...(c.experiments.length > k.maxExperiments ? [`交了 ${c.experiments.length} 个诊断实验，最多 ${k.maxExperiments} 个`] : []),
|
||||
...c.experiments.flatMap((e, i) => {
|
||||
const ids = e.metrics.map((m) => m.id);
|
||||
return [
|
||||
...e.metrics
|
||||
.filter((m) => m.missing !== "recordMissing")
|
||||
.map((m) => `第 ${i + 1} 个实验的指标 ${m.id} 选了 ${m.missing}：诊断实验里的指标只能选 recordMissing（缺了只记缺失）`),
|
||||
...repeated(ids).map((x) => `第 ${i + 1} 个实验的 metrics 里编号 ${x} 写了不止一次`),
|
||||
...unique(ids)
|
||||
.filter((x) => k.takenMetricIds.includes(x))
|
||||
.map((x) => `第 ${i + 1} 个实验的指标编号 ${x} 已被占用（已占用的是 ${k.takenMetricIds.join("、")}）：换一个编号`),
|
||||
];
|
||||
}),
|
||||
];
|
||||
}
|
||||
|
||||
/**
|
||||
* 诊断预测（方法.md 4.3 诊 2）：覆盖自己每条解释 × 每个合并后的实验，不多不少、每一对一次；每个范围三个数严格递增。
|
||||
* 一对写成"解释编号:实验编号"（编号里没有冒号）。
|
||||
*/
|
||||
function diagnosisPredictionProblems(c: Submission<"diagnosisPrediction">, k: ChecklistOf<"diagnosisPrediction">): string[] {
|
||||
return [
|
||||
...covers(
|
||||
"predictions 的（解释:实验）",
|
||||
c.predictions.map((p) => `${p.explanation}:${p.experiment}`),
|
||||
k.explanations.flatMap((e) => k.experiments.map((d) => `${e}:${d}`)),
|
||||
),
|
||||
...c.predictions.flatMap((p) => increasing(`${p.explanation}:${p.experiment} 的 range`, p.range)),
|
||||
];
|
||||
}
|
||||
|
||||
/**
|
||||
* 提案、回应、统一共用（方法.md 第五部分"延续""指标"、4.3 ②⑥）：
|
||||
* 规则 2 因子数上限；3 两个因子不延续同一个、延续的编号看得到；4 保障短名；10 指标的引用与新登记。
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
// submissions：一份提交收下还是退回，先程序检查、再问 Jev（docs/结构.md 第二部分）；
|
||||
// 另有冻结协议的拼法，和核验、消融核验收下之后通过没有。不读也不写账本。
|
||||
export { ablationOutcome, verificationOutcome } from "./outcomes.ts";
|
||||
export { ablationOutcome, diagnosisOutcome, verificationOutcome } from "./outcomes.ts";
|
||||
export { protocolBody } from "./protocol.ts";
|
||||
export { receive } from "./receive.ts";
|
||||
export {
|
||||
|
||||
@@ -34,3 +34,20 @@ export function ablationOutcome(
|
||||
const wronglyOff = v.factors.filter((f) => !shouldBeOff.has(f.factor) && f.off).map((f) => f.factor);
|
||||
return { passed: leftOn.length + wronglyOff.length === 0, leftOn, wronglyOff };
|
||||
}
|
||||
|
||||
/**
|
||||
* 诊断核验通过没有（方法.md 4.3 诊 4）：逐个实验判，任何一项写了 false 的实验不通过。
|
||||
* notFollowed 是没照写法做的实验，touchedElse 是动了写法以外的地方的实验，failing 是两者的并集，按核验里的次序。
|
||||
* 一个实验不通过不影响别的实验：通过的实验代码不再改动。
|
||||
*/
|
||||
export function diagnosisOutcome(v: Submission<"diagnosisVerification">): {
|
||||
passed: boolean;
|
||||
failing: string[];
|
||||
notFollowed: string[];
|
||||
touchedElse: string[];
|
||||
} {
|
||||
const notFollowed = v.experiments.filter((x) => !x.followed).map((x) => x.experiment);
|
||||
const touchedElse = v.experiments.filter((x) => !x.nothingElse).map((x) => x.experiment);
|
||||
const failing = v.experiments.filter((x) => !x.followed || !x.nothingElse).map((x) => x.experiment);
|
||||
return { passed: failing.length === 0, failing, notFollowed, touchedElse };
|
||||
}
|
||||
|
||||
@@ -1,4 +1,4 @@
|
||||
import { checkHunk, checkProtocol, type CodeHunk } from "../jev/index.ts";
|
||||
import { checkExperiment, checkHunk, checkProtocol, type CodeHunk } from "../jev/index.ts";
|
||||
import type { JevJudgment } from "../vocab/index.ts";
|
||||
import { programProblems } from "./checks.ts";
|
||||
import { protocolBody } from "./protocol.ts";
|
||||
@@ -27,7 +27,8 @@ export async function receive(inc: Incoming): Promise<Verdict> {
|
||||
}
|
||||
|
||||
/**
|
||||
* 问 Jev（方法.md 4.3 ⑥⑨）：回应、统一对拼好的协议问冻结条款三问；实现、修复每个差异块问代码改动五问，顺序发。
|
||||
* 问 Jev（方法.md 4.3 ⑥⑨、诊 4）：回应、统一对拼好的协议问冻结条款三问；实现、修复（含诊断程序）每个差异块问代码改动五问，顺序发。
|
||||
* 诊断解释与实验（诊 1）对每个实验问冻结条款三问（改动了评测、自相矛盾、太含糊，用诊断的变体问法），顺序发。
|
||||
* 没有差异块的改动（二进制、只改可执行位、空文件增删）Jev 看不到内容，不问；一块都没有就等于没问，交回 null。
|
||||
* 消融实现不问：冻结协议里没有登记"关掉以后"的随机性和停止条件，五问没有对照（照参考项目）。
|
||||
* flagged 是退回原因里每一处的文字：只列判"是"的几问，问题名用 jev 的英文名（用户 2026-10-03 定）；判"是"只看 jev 交回的 yes。
|
||||
@@ -37,7 +38,19 @@ async function askJev(inc: Incoming, protocol: ProtocolBody | null): Promise<{ j
|
||||
const c = await checkProtocol(inc.jevModel, protocol!);
|
||||
return { judgments: c.judgments, flagged: c.flagged ? [`- 冻结协议：${flaggedText(c.judgments)}`] : [] };
|
||||
}
|
||||
if (inc.kind !== "implementation" && inc.kind !== "repair") return null;
|
||||
if (inc.kind === "diagnosisProposal") {
|
||||
const judgments: JevJudgment[] = [];
|
||||
const flagged: string[] = [];
|
||||
for (const [i, experiment] of inc.content.experiments.entries()) {
|
||||
const c = await checkExperiment(inc.jevModel, experiment);
|
||||
judgments.push(...c.judgments);
|
||||
if (c.flagged) flagged.push(`- 第 ${i + 1} 个诊断实验（${experiment.name}）：${flaggedText(c.judgments)}`);
|
||||
}
|
||||
return { judgments, flagged };
|
||||
}
|
||||
if (inc.kind !== "implementation" && inc.kind !== "repair" && inc.kind !== "diagnosisImplementation" && inc.kind !== "diagnosisRepair") {
|
||||
return null;
|
||||
}
|
||||
// 走到这里程序检查已过，代码一定存成了版本。
|
||||
const code = inc.known.code as Extract<typeof inc.known.code, { saved: true }>;
|
||||
if (code.hunks.length === 0) return null;
|
||||
|
||||
@@ -17,7 +17,7 @@ export type OriginalProposal = Pick<Submission<"proposal">, "factors" | "safegua
|
||||
type NotSaved = { saved: false; path: string; reason: string };
|
||||
|
||||
/**
|
||||
* 实现、修复的代码：存成了版本就带改了的不许改的文件（tasks 的 protectedChanges）、
|
||||
* 实现、修复（候选和诊断程序同样）的代码：存成了版本就带改了的不许改的文件（tasks 的 protectedChanges）、
|
||||
* 和本批基线比出的差异块（jev 的 codeHunks 切好的），以及给 Jev 对照的协议正文和要采集的指标。
|
||||
*/
|
||||
export type ForwardCode = NotSaved | { saved: true; protectedFiles: string[]; hunks: CodeHunk[]; protocol: ProtocolRules };
|
||||
@@ -33,6 +33,9 @@ type KnownByKind = {
|
||||
response: { original: OriginalProposal };
|
||||
implementation: { code: ForwardCode };
|
||||
repair: { code: ForwardCode };
|
||||
/** 诊断程序和候选代码同形：和当前基线比出的差异块，给 Jev 对照的是选中实验的原文和要采集的指标。 */
|
||||
diagnosisImplementation: { code: ForwardCode };
|
||||
diagnosisRepair: { code: ForwardCode };
|
||||
/** 交了"跑不起来"（cannotRun 不是 null）时 flow 不存代码，code 为 null。 */
|
||||
ablationImplementation: { code: AblationCode | null };
|
||||
};
|
||||
|
||||
+227
-13
@@ -1,4 +1,4 @@
|
||||
import { copyFileSync, cpSync, existsSync, readFileSync, writeFileSync } from "node:fs";
|
||||
import { chmodSync, copyFileSync, cpSync, existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
|
||||
import { join } from "node:path";
|
||||
import { Type, type Static } from "typebox";
|
||||
import { Value } from "typebox/value";
|
||||
@@ -13,7 +13,16 @@ import type {
|
||||
} from "../vocab/index.ts";
|
||||
import type { TaskFacts } from "./facts.ts";
|
||||
import { measured, readMetricsFile, unmeasured } from "./metrics-file.ts";
|
||||
import type { LightCheckOutcome, Task, TaskImplementation } from "./types.ts";
|
||||
import type {
|
||||
DiagnosisEnding,
|
||||
DiagnosisExperiment,
|
||||
DiagnosisExperimentJudged,
|
||||
DiagnosisJudged,
|
||||
DiagnosisJudgingStep,
|
||||
LightCheckOutcome,
|
||||
Task,
|
||||
TaskImplementation,
|
||||
} from "./types.ts";
|
||||
|
||||
// 填圆：在单位正方形里放 26 个圆，最大化半径之和。任务说明见 tasks/circle_packing/TASK.md。
|
||||
// 凡是要执行任务代码的事（正式运行、校验器、轻量检查）都只交出运行定义，由调用方交给 runner 在任务镜像里跑；
|
||||
@@ -22,6 +31,7 @@ import type { LightCheckOutcome, Task, TaskImplementation } from "./types.ts";
|
||||
const CIRCLES = 26;
|
||||
const ENTRY = "initial.py";
|
||||
const HARNESS = "harness.py";
|
||||
const DIAGNOSIS_HARNESS = "harness-diagnosis.py";
|
||||
const VALIDATOR = "validate.py";
|
||||
const CHECK_SCRIPT = "check-candidate.py";
|
||||
const METRICS_SPEC = "metrics-spec.json";
|
||||
@@ -29,6 +39,12 @@ const METRICS = "metrics.json";
|
||||
const SOLUTION = "solution.json";
|
||||
const VERDICT = "verdict.json";
|
||||
const CHECK_RESULT = "check.json";
|
||||
/** 诊断运行：原始产物目录、每个实验的答案目录、harness 写的每实验用时报告（路径都相对运行目录）。 */
|
||||
const BASELINE_ARTIFACT = "baseline-artifact";
|
||||
const ANSWERS = "answers";
|
||||
const DIAGNOSIS_REPORT = "diagnosis.json";
|
||||
/** 判分目录里每个实验写的校验器退出码文件（同一个目录下还有这个实验的 solution.json、verdict.json）。 */
|
||||
const EXIT_FILE = "exit";
|
||||
/** 轻量检查目录里放候选代码的子目录；检查脚本放在它的上一层，免得扫到自己。 */
|
||||
const CHECK_CANDIDATE = "candidate";
|
||||
/** 运行目录挂进容器的位置。 */
|
||||
@@ -277,16 +293,16 @@ function readVerdict(judgeDir: string, exitCode: number): ValidatorVerdict {
|
||||
return verdict;
|
||||
}
|
||||
|
||||
function judge(
|
||||
_task: Task,
|
||||
runDir: string,
|
||||
judgeDir: string,
|
||||
formal: ProcessResult,
|
||||
validator: ProcessResult,
|
||||
hypothesisMetrics: HypothesisMetric[],
|
||||
): JudgedResult {
|
||||
const run = started(formal);
|
||||
const verdict = readVerdict(judgeDir, requireExit("校验器", validator, [0, 1]));
|
||||
/** 校验器判定读成的结论：算不算数、理由、最终分数、校验器给的辅助指标、合法化明细。 */
|
||||
interface Scored {
|
||||
valid: boolean;
|
||||
reason: string;
|
||||
score: Measured;
|
||||
fromValidator: Record<string, Measured>;
|
||||
extra: Json;
|
||||
}
|
||||
|
||||
function scoreVerdict(verdict: ValidatorVerdict): Scored {
|
||||
const legal = verdict.legalization;
|
||||
let deduction: Measured;
|
||||
let score: Measured;
|
||||
@@ -328,7 +344,199 @@ function judge(
|
||||
deduction,
|
||||
},
|
||||
};
|
||||
return assemble(runDir, run, hypothesisMetrics, { valid: verdict.valid, reason }, score, fromValidator, extra);
|
||||
return { valid: verdict.valid, reason, score, fromValidator, extra };
|
||||
}
|
||||
|
||||
function judge(
|
||||
_task: Task,
|
||||
runDir: string,
|
||||
judgeDir: string,
|
||||
formal: ProcessResult,
|
||||
validator: ProcessResult,
|
||||
hypothesisMetrics: HypothesisMetric[],
|
||||
): JudgedResult {
|
||||
const run = started(formal);
|
||||
const scored = scoreVerdict(readVerdict(judgeDir, requireExit("校验器", validator, [0, 1])));
|
||||
return assemble(runDir, run, hypothesisMetrics, { valid: scored.valid, reason: scored.reason }, scored.score, scored.fromValidator, scored.extra);
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// 诊断运行（方法.md 4.3"诊断"一段）：一个容器、一次登记、一次扣次，里面按顺序跑几个实验。
|
||||
// 全部是新增的一套函数，用自己的 harness（harness-diagnosis.py），不碰上面正式运行的任何一步。
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
/** 诊断程序存版本时跳过的名字：正式运行那份加上原始产物目录（它是宿主放进工作目录的，不是实验员的代码）。 */
|
||||
function diagnosisUnversionedNames(facts: TaskFacts): string[] {
|
||||
return [...facts.unversionedNames, BASELINE_ARTIFACT];
|
||||
}
|
||||
|
||||
/**
|
||||
* 把产出当前基线的那次正式运行的原始产物（solution.json）放进 dir 下的 baseline-artifact/，设成只读。
|
||||
* 实验员的工作目录和诊断运行的运行目录都用它：实验员能看到实验要接着算的那份答案，运行时实验读的是同一份。
|
||||
* 只读靠文件权限，防的是实验顺手改了它，不是安全边界。
|
||||
*/
|
||||
function placeBaselineArtifact(_task: Task, baselineRunDir: string, dir: string): void {
|
||||
const target = join(dir, BASELINE_ARTIFACT);
|
||||
mkdirSync(target);
|
||||
const file = join(target, SOLUTION);
|
||||
copyFileSync(join(baselineRunDir, SOLUTION), file);
|
||||
chmodSync(file, 0o444);
|
||||
}
|
||||
|
||||
function prepareDiagnosisRun(task: Task, programDir: string, baselineRunDir: string, runDir: string, experiments: DiagnosisExperiment[]): void {
|
||||
cpSync(programDir, runDir, { recursive: true });
|
||||
// 两份 harness 以宿主的为准；指标说明是全部实验登记的指标合在一起，每个实验只写自己的那几条。
|
||||
copyFileSync(join(task.dir, HARNESS), join(runDir, HARNESS));
|
||||
copyFileSync(join(task.dir, DIAGNOSIS_HARNESS), join(runDir, DIAGNOSIS_HARNESS));
|
||||
writeFileSync(join(runDir, METRICS_SPEC), `${JSON.stringify(experiments.flatMap((e) => e.metrics), null, 2)}\n`);
|
||||
placeBaselineArtifact(task, baselineRunDir, runDir);
|
||||
mkdirSync(join(runDir, ANSWERS));
|
||||
}
|
||||
|
||||
/** 总时限、核数照任务规定，实验按顺序跑，每个实验分到总时限 ÷ 实验数（由 harness 执行）。 */
|
||||
function diagnosisRun(task: Task, imageDigest: string, seed: string, experimentIds: string[]): RunDefinition {
|
||||
const eachSeconds = task.facts.run.wallSeconds / experimentIds.length;
|
||||
const command = [
|
||||
"python",
|
||||
"-B",
|
||||
`${WORKDIR}/${DIAGNOSIS_HARNESS}`,
|
||||
"--program",
|
||||
ENTRY,
|
||||
"--baseline",
|
||||
BASELINE_ARTIFACT,
|
||||
"--answers",
|
||||
ANSWERS,
|
||||
"--seconds-each",
|
||||
String(eachSeconds),
|
||||
"--report",
|
||||
DIAGNOSIS_REPORT,
|
||||
"--experiments",
|
||||
...experimentIds,
|
||||
];
|
||||
return definition(task, imageDigest, command, task.facts.run.wallSeconds, { INFORSI_SEED: seed });
|
||||
}
|
||||
|
||||
/** 一个实验在运行目录里的结局：harness 报告里的一项，或者整次运行先结束、没轮到它。 */
|
||||
type ExperimentOutcome = { ending: "finished" | "crashed" | "timeout"; wallSeconds: number } | { ending: "notRun"; reason: string };
|
||||
|
||||
/** harness 写的报告（我们自己的程序，按约定的格式直接读）。 */
|
||||
interface ReportEntry {
|
||||
id: string;
|
||||
status: "finished" | "crashed" | "timeout";
|
||||
wall_seconds: number;
|
||||
}
|
||||
|
||||
function experimentOutcomes(
|
||||
task: Task,
|
||||
runDir: string,
|
||||
process: Exclude<ProcessResult, { ending: "notStarted" }>,
|
||||
experiments: DiagnosisExperiment[],
|
||||
): ExperimentOutcome[] {
|
||||
const path = join(runDir, DIAGNOSIS_REPORT);
|
||||
const report = existsSync(path) ? (readJson(path) as ReportEntry[]) : [];
|
||||
return experiments.map((e) => {
|
||||
const entry = report.find((r) => r.id === e.id);
|
||||
if (entry !== undefined) return { ending: entry.status, wallSeconds: entry.wall_seconds };
|
||||
const failure = processFailure(task, process);
|
||||
if (failure === null) throw new Error(`诊断运行正常退出，但 ${DIAGNOSIS_REPORT} 里没有实验 ${e.id}`);
|
||||
return { ending: "notRun", reason: `没轮到：整次诊断运行${failure}` };
|
||||
});
|
||||
}
|
||||
|
||||
/** 没有答案可判的实验：不算数，辅助指标和假设专用指标都记测不到并附原因。 */
|
||||
function unjudgedExperiment(
|
||||
id: string,
|
||||
ending: Exclude<DiagnosisEnding, "finished">,
|
||||
wall: Measured,
|
||||
reason: string,
|
||||
hypothesisMetrics: HypothesisMetric[],
|
||||
): DiagnosisExperimentJudged {
|
||||
const metrics = {
|
||||
...Object.fromEntries(FROM_VALIDATOR.map((m) => [m, unmeasured(`没有跑校验器：${reason}`)])),
|
||||
...Object.fromEntries(hypothesisMetrics.map((m) => [m.id, unmeasured(`没有结果：${reason}`)])),
|
||||
};
|
||||
return { id, ending, wallSeconds: wall, verdict: "invalid", verdictReason: reason, score: unmeasured(reason), metrics, extra: { legalization: null } };
|
||||
}
|
||||
|
||||
function judgedExperiments(
|
||||
task: Task,
|
||||
runDir: string,
|
||||
judgeDir: string,
|
||||
experiments: DiagnosisExperiment[],
|
||||
outcomes: ExperimentOutcome[],
|
||||
): DiagnosisJudged {
|
||||
const eachSeconds = task.facts.run.wallSeconds / experiments.length;
|
||||
return {
|
||||
experiments: experiments.map((e, i): DiagnosisExperimentJudged => {
|
||||
const outcome = outcomes[i]!;
|
||||
if (outcome.ending === "notRun") return unjudgedExperiment(e.id, "notRun", unmeasured(outcome.reason), outcome.reason, e.metrics);
|
||||
const wall = measured(outcome.wallSeconds);
|
||||
if (outcome.ending === "timeout") {
|
||||
return unjudgedExperiment(e.id, "timeout", wall, `超过分到的时间 ${eachSeconds} 秒，被结束`, e.metrics);
|
||||
}
|
||||
if (outcome.ending === "crashed") {
|
||||
return unjudgedExperiment(e.id, "crashed", wall, `实验的进程以非零退出码结束（报错见这次运行的 stderr.log）`, e.metrics);
|
||||
}
|
||||
const dir = join(judgeDir, e.id);
|
||||
const exitCode = Number(readFileSync(join(dir, EXIT_FILE), "utf8").trim());
|
||||
const scored = scoreVerdict(readVerdict(dir, exitCode));
|
||||
const hypothesis = readMetricsFile(join(runDir, ANSWERS, e.id, METRICS), `${ANSWERS}/${e.id}/${METRICS}`, e.metrics).hypothesis;
|
||||
return {
|
||||
id: e.id,
|
||||
ending: "finished",
|
||||
wallSeconds: wall,
|
||||
verdict: scored.valid ? "valid" : "invalid",
|
||||
verdictReason: scored.reason,
|
||||
score: scored.score,
|
||||
metrics: { ...scored.fromValidator, ...hypothesis },
|
||||
extra: scored.extra,
|
||||
};
|
||||
}),
|
||||
};
|
||||
}
|
||||
|
||||
/** 逐个实验判分。校验器一个容器、按顺序对每个有答案的实验各跑一次，退出码各存一个文件。 */
|
||||
const VALIDATE_EACH = `for id in "$@"; do
|
||||
python -B ${VALIDATOR} --circles ${CIRCLES} --solution "$id/${SOLUTION}" --atol ${TOLERANCE} --legalize --verdict "$id/${VERDICT}"
|
||||
echo $? > "$id/${EXIT_FILE}"
|
||||
done`;
|
||||
|
||||
function prepareDiagnosisJudging(
|
||||
task: Task,
|
||||
runDir: string,
|
||||
judgeDir: string,
|
||||
imageDigest: string,
|
||||
process: ProcessResult,
|
||||
experiments: DiagnosisExperiment[],
|
||||
): DiagnosisJudgingStep {
|
||||
const run = started(process);
|
||||
const outcomes = experimentOutcomes(task, runDir, run, experiments);
|
||||
const finished = experiments.filter((_, i) => outcomes[i]!.ending === "finished");
|
||||
if (finished.length === 0) return { status: "judged", result: judgedExperiments(task, runDir, judgeDir, experiments, outcomes) };
|
||||
for (const e of finished) {
|
||||
mkdirSync(join(judgeDir, e.id));
|
||||
copyFileSync(join(runDir, ANSWERS, e.id, SOLUTION), join(judgeDir, e.id, SOLUTION));
|
||||
}
|
||||
copyFileSync(join(task.dir, VALIDATOR), join(judgeDir, VALIDATOR));
|
||||
const command = ["sh", "-c", VALIDATE_EACH, "inforsi-validate-each", ...finished.map((e) => e.id)];
|
||||
return { status: "validate", definition: definition(task, imageDigest, command, VALIDATOR_SECONDS * finished.length, {}) };
|
||||
}
|
||||
|
||||
function judgeDiagnosis(
|
||||
task: Task,
|
||||
runDir: string,
|
||||
judgeDir: string,
|
||||
process: ProcessResult,
|
||||
validator: ProcessResult | null,
|
||||
experiments: DiagnosisExperiment[],
|
||||
): DiagnosisJudged {
|
||||
const run = started(process);
|
||||
const outcomes = experimentOutcomes(task, runDir, run, experiments);
|
||||
if (outcomes.some((o) => o.ending === "finished")) {
|
||||
if (validator === null) throw new Error("有实验交出了答案，却没有校验器的运行结果");
|
||||
requireExit("校验器", validator, [0]);
|
||||
}
|
||||
return judgedExperiments(task, runDir, judgeDir, experiments, outcomes);
|
||||
}
|
||||
|
||||
export const circlePacking: TaskImplementation = {
|
||||
@@ -339,4 +547,10 @@ export const circlePacking: TaskImplementation = {
|
||||
readLightCheck,
|
||||
prepareJudging,
|
||||
judge,
|
||||
diagnosisUnversionedNames,
|
||||
placeBaselineArtifact,
|
||||
prepareDiagnosisRun,
|
||||
diagnosisRun,
|
||||
prepareDiagnosisJudging,
|
||||
judgeDiagnosis,
|
||||
};
|
||||
|
||||
+16
-1
@@ -2,14 +2,29 @@
|
||||
// 不起子进程：要执行任务代码的事都只交出运行定义，由调用方交给 runner 在任务镜像里跑。
|
||||
export { checkTaskFacts, TASK_KINDS, TaskFacts, type TaskKind } from "./facts.ts";
|
||||
export {
|
||||
diagnosisRun,
|
||||
diagnosisUnversionedNames,
|
||||
formalRun,
|
||||
judge,
|
||||
judgeDiagnosis,
|
||||
lightCheck,
|
||||
loadTask,
|
||||
metricDefinitions,
|
||||
placeBaselineArtifact,
|
||||
prepareDiagnosisJudging,
|
||||
prepareDiagnosisRun,
|
||||
prepareJudging,
|
||||
prepareRun,
|
||||
protectedChanges,
|
||||
readLightCheck,
|
||||
} from "./task.ts";
|
||||
export type { JudgingStep, LightCheckOutcome, Task } from "./types.ts";
|
||||
export type {
|
||||
DiagnosisEnding,
|
||||
DiagnosisExperiment,
|
||||
DiagnosisExperimentJudged,
|
||||
DiagnosisJudged,
|
||||
DiagnosisJudgingStep,
|
||||
JudgingStep,
|
||||
LightCheckOutcome,
|
||||
Task,
|
||||
} from "./types.ts";
|
||||
|
||||
+80
-1
@@ -3,7 +3,15 @@ import { join, resolve } from "node:path";
|
||||
import type { HypothesisMetric, JudgedResult, MetricDefinition, ProcessResult, RunDefinition } from "../vocab/index.ts";
|
||||
import { circlePacking } from "./circle-packing.ts";
|
||||
import { checkTaskFacts, type TaskKind } from "./facts.ts";
|
||||
import type { JudgingStep, LightCheckOutcome, Task, TaskImplementation } from "./types.ts";
|
||||
import type {
|
||||
DiagnosisExperiment,
|
||||
DiagnosisJudged,
|
||||
DiagnosisJudgingStep,
|
||||
JudgingStep,
|
||||
LightCheckOutcome,
|
||||
Task,
|
||||
TaskImplementation,
|
||||
} from "./types.ts";
|
||||
|
||||
// tasks 对外的函数。任务目录的约定：task.json（任务事实）、TASK.md（给实验员的说明）、
|
||||
// Dockerfile 与依赖清单（建镜像）、baseline/（基线代码），以及各任务自己的 Python 工具。
|
||||
@@ -105,3 +113,74 @@ export function judge(
|
||||
export function protectedChanges(task: Task, changedFiles: string[]): string[] {
|
||||
return changedFiles.filter((file) => task.facts.protectedFiles.includes(file));
|
||||
}
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// 诊断运行（方法.md 4.3"诊断"一段）：一个容器、一次登记、一次扣次，里面按顺序跑几个实验。
|
||||
// 是和上面正式运行并列的另一套函数，上面的函数不认识它。各目录同样由调用方建好、传空目录进来。
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
/** 存诊断程序版本时跳过的名字：task.json 的清单，加上宿主放进工作目录的原始产物目录。 */
|
||||
export function diagnosisUnversionedNames(task: Task): string[] {
|
||||
return implementation(task).diagnosisUnversionedNames(task.facts);
|
||||
}
|
||||
|
||||
/**
|
||||
* 把产出当前基线的那次正式运行的原始产物，只读地放进 dir（实验员工作目录，调用方已用 checkoutVersion 取出基线代码）。
|
||||
* baselineRunDir 是那次运行的运行目录。放在哪、叫什么，由任务定，写在 TASK.md 里给实验员看。
|
||||
*/
|
||||
export function placeBaselineArtifact(task: Task, baselineRunDir: string, dir: string): void {
|
||||
implementation(task).placeBaselineArtifact(task, baselineRunDir, dir);
|
||||
}
|
||||
|
||||
/**
|
||||
* 准备一次诊断运行的运行目录：诊断程序、两份 harness、全部实验登记的指标说明（metrics-spec.json）、
|
||||
* 只读的原始产物。runDir 由调用方建好，应为空目录；实验的先后顺序就是 experiments 的顺序。
|
||||
*/
|
||||
export function prepareDiagnosisRun(
|
||||
task: Task,
|
||||
programDir: string,
|
||||
baselineRunDir: string,
|
||||
runDir: string,
|
||||
experiments: DiagnosisExperiment[],
|
||||
): void {
|
||||
implementation(task).prepareDiagnosisRun(task, programDir, baselineRunDir, runDir, experiments);
|
||||
}
|
||||
|
||||
/**
|
||||
* 诊断运行的运行定义：总时限、核数照任务规定，harness 按顺序跑 experimentIds 里的实验，
|
||||
* 每个实验分到总时限 ÷ 实验数。种子同正式运行（同一份研究共用一个）。
|
||||
*/
|
||||
export function diagnosisRun(task: Task, imageDigest: string, seed: string, experimentIds: string[]): RunDefinition {
|
||||
return implementation(task).diagnosisRun(task, imageDigest, seed, experimentIds);
|
||||
}
|
||||
|
||||
/**
|
||||
* 诊断判分第一步：每个交出答案的实验的答案复制进 judgeDir 下以实验编号命名的子目录，
|
||||
* 交回校验器的运行定义（一个容器逐个实验判分）；没有一个实验交出答案时直接交回判分结果。
|
||||
*/
|
||||
export function prepareDiagnosisJudging(
|
||||
task: Task,
|
||||
runDir: string,
|
||||
judgeDir: string,
|
||||
imageDigest: string,
|
||||
process: ProcessResult,
|
||||
experiments: DiagnosisExperiment[],
|
||||
): DiagnosisJudgingStep {
|
||||
return implementation(task).prepareDiagnosisJudging(task, runDir, judgeDir, imageDigest, process, experiments);
|
||||
}
|
||||
|
||||
/**
|
||||
* 诊断判分第二步：按实验分开交回分数、合法化、登记的指标、用时、结局。validator 是第一步交回过运行定义时跑出的结果，
|
||||
* 第一步直接判完时传 null。资源用量记在整次运行的 ProcessResult 里，不在这里。
|
||||
* 校验器自己没正常跑完、判定文件认不出、修后分数高于原分数时报错停下，同正式运行的 judge。
|
||||
*/
|
||||
export function judgeDiagnosis(
|
||||
task: Task,
|
||||
runDir: string,
|
||||
judgeDir: string,
|
||||
process: ProcessResult,
|
||||
validator: ProcessResult | null,
|
||||
experiments: DiagnosisExperiment[],
|
||||
): DiagnosisJudged {
|
||||
return implementation(task).judgeDiagnosis(task, runDir, judgeDir, process, validator, experiments);
|
||||
}
|
||||
|
||||
+60
-1
@@ -1,4 +1,4 @@
|
||||
import type { HypothesisMetric, JudgedResult, MetricDefinition, ProcessResult, RunDefinition } from "../vocab/index.ts";
|
||||
import type { HypothesisMetric, JudgedResult, Json, Measured, MetricDefinition, ProcessResult, RunDefinition, Verdict } from "../vocab/index.ts";
|
||||
import type { TaskFacts } from "./facts.ts";
|
||||
|
||||
/** 读进来的一个任务。 */
|
||||
@@ -22,6 +22,44 @@ export type LightCheckOutcome = { passed: true } | { passed: false; reasons: str
|
||||
*/
|
||||
export type JudgingStep = { status: "judged"; result: JudgedResult } | { status: "validate"; definition: RunDefinition };
|
||||
|
||||
/** 一个要跑的诊断实验：编号，以及它自己登记的假设专用指标（冻结后的全名；诊断实验只能登记"缺了只记缺失"的指标）。 */
|
||||
export interface DiagnosisExperiment {
|
||||
id: string;
|
||||
metrics: HypothesisMetric[];
|
||||
}
|
||||
|
||||
/**
|
||||
* 一个诊断实验怎么结束：finished 跑完交出了答案；crashed 实验自己崩了；timeout 超过分到的时间被杀；
|
||||
* notRun 整次运行先被总时限或从外面结束，没轮到它。
|
||||
*/
|
||||
export type DiagnosisEnding = "finished" | "crashed" | "timeout" | "notRun";
|
||||
|
||||
/** 一个诊断实验单独判分的结果。资源用量（CPU、峰值内存）记在整次运行的结果里，这里只有它自己的用时。 */
|
||||
export interface DiagnosisExperimentJudged {
|
||||
id: string;
|
||||
ending: DiagnosisEnding;
|
||||
/** 这个实验实际用掉的墙钟秒数；notRun 时测不到。 */
|
||||
wallSeconds: Measured;
|
||||
verdict: Verdict;
|
||||
verdictReason: string;
|
||||
/** 合法化之后的最终指标分数；不算数时测不到并附原因。 */
|
||||
score: Measured;
|
||||
/** 校验器给的辅助指标，加上这个实验登记的假设专用指标，键是指标编号。 */
|
||||
metrics: Record<string, Measured>;
|
||||
/** 同正式运行的 extra：填圆是合法化的明细。 */
|
||||
extra: Json;
|
||||
}
|
||||
|
||||
/** 一次诊断运行的判分结果：每个实验一项，顺序同运行定义里的实验顺序。 */
|
||||
export interface DiagnosisJudged {
|
||||
experiments: DiagnosisExperimentJudged[];
|
||||
}
|
||||
|
||||
/** 诊断判分第一步：不用跑校验器就已判定（没有一个实验交出答案），或交回校验器的运行定义。 */
|
||||
export type DiagnosisJudgingStep =
|
||||
| { status: "judged"; result: DiagnosisJudged }
|
||||
| { status: "validate"; definition: RunDefinition };
|
||||
|
||||
/** 一种任务的实现。对外函数按 task.facts.kind 从对照表里取出它来调。参数的含义见 task.ts 里同名的对外函数。 */
|
||||
export interface TaskImplementation {
|
||||
/** 最终指标（编号和方向取自 task.json）与常规辅助指标的定义。 */
|
||||
@@ -46,4 +84,25 @@ export interface TaskImplementation {
|
||||
validator: ProcessResult,
|
||||
hypothesisMetrics: HypothesisMetric[],
|
||||
): JudgedResult;
|
||||
/** 诊断运行（方法.md 4.3"诊断"一段）：一次运行跑几个实验。上面七个函数的行为与它无关。 */
|
||||
diagnosisUnversionedNames(facts: TaskFacts): string[];
|
||||
placeBaselineArtifact(task: Task, baselineRunDir: string, dir: string): void;
|
||||
prepareDiagnosisRun(task: Task, programDir: string, baselineRunDir: string, runDir: string, experiments: DiagnosisExperiment[]): void;
|
||||
diagnosisRun(task: Task, imageDigest: string, seed: string, experimentIds: string[]): RunDefinition;
|
||||
prepareDiagnosisJudging(
|
||||
task: Task,
|
||||
runDir: string,
|
||||
judgeDir: string,
|
||||
imageDigest: string,
|
||||
process: ProcessResult,
|
||||
experiments: DiagnosisExperiment[],
|
||||
): DiagnosisJudgingStep;
|
||||
judgeDiagnosis(
|
||||
task: Task,
|
||||
runDir: string,
|
||||
judgeDir: string,
|
||||
process: ProcessResult,
|
||||
validator: ProcessResult | null,
|
||||
experiments: DiagnosisExperiment[],
|
||||
): DiagnosisJudged;
|
||||
}
|
||||
|
||||
@@ -22,6 +22,9 @@ type FactorRules = {
|
||||
/** 实现、修复、核验共用：冻结协议的实现条款（c1…）、因子（p1…）、保障短名。 */
|
||||
type FrozenItems = { candidate: string; clauses: string[]; factors: string[]; safeguards: string[] };
|
||||
|
||||
/** 诊断实现、修复、核验共用：本次诊断选中的实验编号（d1…）。 */
|
||||
type DiagnosisItems = { experiments: string[] };
|
||||
|
||||
/** 入口个数的范围（方法.md 4.2：2–3 个）。 */
|
||||
type EntryCount = { min: 2; max: 3 };
|
||||
|
||||
@@ -30,6 +33,16 @@ export type Checklist =
|
||||
| { kind: "discussion"; entries: EntryCount }
|
||||
/** minChoices = min(N, 入口总数)（方法.md 4.2 认领）。 */
|
||||
| { kind: "claim"; entries: string[]; minChoices: number }
|
||||
/**
|
||||
* 诊断解释最多 M 条、诊断实验最多 A 个（方法.md 4.3 诊 1）；takenMetricIds 是新登记的指标短名不能用的编号
|
||||
* （最终指标、常规辅助指标和答案路径的两个保留键）。
|
||||
*/
|
||||
| { kind: "diagnosisProposal"; maxExplanations: number; maxExperiments: number; takenMetricIds: string[] }
|
||||
/** 对自己每条解释（explanations）× 每个合并后的实验（experiments）各给三个数（方法.md 4.3 诊 2）。 */
|
||||
| { kind: "diagnosisPrediction"; explanations: string[]; experiments: string[] }
|
||||
| ({ kind: "diagnosisImplementation" } & DiagnosisItems)
|
||||
| ({ kind: "diagnosisRepair" } & DiagnosisItems)
|
||||
| ({ kind: "diagnosisVerification" } & DiagnosisItems)
|
||||
| ({ kind: "proposal"; maxProposals: number } & FactorRules)
|
||||
| { kind: "firstPrediction"; candidate: string }
|
||||
| { kind: "secondPrediction"; candidate: string }
|
||||
|
||||
@@ -10,6 +10,10 @@ export const STAGES = [
|
||||
"entryDraft", // 0.1 闭门写入口
|
||||
"discussion", // 0.2 讨论，可配置为 1 或 2 次，每次都是这个阶段
|
||||
"claim", // 0.3 认领
|
||||
"diagnosisProposal", // 诊 1 诊断解释与诊断实验
|
||||
"diagnosisPrediction", // 诊 2 诊断预测
|
||||
"diagnosisImplementation", // 诊 4 诊断实现，以及核验不通过后的修复
|
||||
"diagnosisVerification", // 诊 4 诊断核验
|
||||
"proposal", // ② 提案
|
||||
"firstPrediction", // ④ 第一次预测
|
||||
"critique", // ⑤ 挑刺
|
||||
@@ -33,6 +37,11 @@ export const SUBMISSION_KINDS = {
|
||||
entry: { stage: "entryDraft", role: "researcher" },
|
||||
discussion: { stage: "discussion", role: "researcher" },
|
||||
claim: { stage: "claim", role: "researcher" },
|
||||
diagnosisProposal: { stage: "diagnosisProposal", role: "researcher" },
|
||||
diagnosisPrediction: { stage: "diagnosisPrediction", role: "researcher" },
|
||||
diagnosisImplementation: { stage: "diagnosisImplementation", role: "experimenter" },
|
||||
diagnosisRepair: { stage: "diagnosisImplementation", role: "experimenter" },
|
||||
diagnosisVerification: { stage: "diagnosisVerification", role: "verifier" },
|
||||
proposal: { stage: "proposal", role: "researcher" },
|
||||
firstPrediction: { stage: "firstPrediction", role: "researcher" },
|
||||
critique: { stage: "critique", role: "researcher" },
|
||||
|
||||
@@ -503,6 +503,146 @@ export const VerificationSubmission = Type.Object(
|
||||
},
|
||||
);
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// 诊断：解释与实验、预测、实现、核验（方法.md 4.3 诊 1–诊 4、第五部分"诊断"三行）
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
const DiagnosisExplanation = Type.Object(
|
||||
{
|
||||
body: Type.String({
|
||||
...TEXT,
|
||||
description:
|
||||
"Markdown：为什么卡在这里，剩下的空间在哪，什么样的实测会推翻它（写不出推翻它的实测，说明这条解释还不能被检验，先改具体再交）。" +
|
||||
"诊断问的是剩下的空间在哪，不是上一步的功劳怎么分。没有具体判断就直说，写明查了什么、为什么拿不准",
|
||||
}),
|
||||
},
|
||||
CLOSED,
|
||||
);
|
||||
|
||||
const DiagnosisExperiment = Type.Object(
|
||||
{
|
||||
name: Type.String({
|
||||
...TEXT,
|
||||
description: "名称：离开这份清单单独看也读得懂",
|
||||
}),
|
||||
spot: Type.String({
|
||||
...TEXT,
|
||||
description: "动哪一处：在当前基线的哪一处代码或哪一步上动手，只写一处（文件、函数或步骤名），其余照基线保留；想测两处就拆成两个实验",
|
||||
}),
|
||||
body: Type.String({
|
||||
...TEXT,
|
||||
description:
|
||||
"Markdown：在这一处怎么动（换成明显更强的版本、换成另一种做法、拿掉，或放宽某个限制）、" +
|
||||
"做完交出什么能判分的答案、几种可能的结果各排除哪几条解释（有分不开的也写明）。不写自己估的分数，分数在下一步交",
|
||||
}),
|
||||
metrics: Type.Array(HypothesisMetric, {
|
||||
description:
|
||||
"登记假设专用指标：这个实验的代码要额外记录的量，编号只写短名；没有就交空数组。" +
|
||||
"诊断实验里的指标只能选 recordMissing（缺了只记缺失），选 invalidateRun 会被退回",
|
||||
}),
|
||||
},
|
||||
CLOSED,
|
||||
);
|
||||
|
||||
export const DiagnosisProposalSubmission = Type.Object(
|
||||
{
|
||||
explanations: Type.Array(DiagnosisExplanation, {
|
||||
minItems: 1,
|
||||
description: "诊断解释清单，编号由程序分配；条数上限见材料。真有第二个看法再写第二条，各条之间要真的不同；没有具体判断，就写一条直说没有的",
|
||||
}),
|
||||
experiments: Type.Array(DiagnosisExperiment, {
|
||||
minItems: 1,
|
||||
description:
|
||||
"诊断实验清单，个数上限见材料。好的实验是不同解释对它的预测差得远：大家都预测涨不了多少的实验，测了也不改变下一步。" +
|
||||
"每个实验只动一处，用时要放得进分到的那份时间；想不出能把解释分开的实验，就交最能暴露你的解释错在哪的那个",
|
||||
}),
|
||||
},
|
||||
{
|
||||
...CLOSED,
|
||||
description: "对当前基线的诊断：几条解释，和能把它们分开的几个实验。实验正文里不写自己估的分数",
|
||||
},
|
||||
);
|
||||
|
||||
export const DiagnosisPredictionSubmission = Type.Object(
|
||||
{
|
||||
predictions: Type.Array(
|
||||
Type.Object(
|
||||
{
|
||||
explanation: Type.String({ ...Id, description: "你自己的诊断解释编号，照材料里列的写" }),
|
||||
experiment: Type.String({ ...Id, description: "合并后的诊断实验编号，照材料里列的写，别人提的也要给" }),
|
||||
range: Range,
|
||||
},
|
||||
CLOSED,
|
||||
),
|
||||
{
|
||||
description:
|
||||
"你自己的每条诊断解释对每个合并后的实验各给三个数：按这条解释，做了这个实验，分数有一成、一半、九成可能低于多少。" +
|
||||
"不多不少，每一对只写一次；按这条解释本身写，不往折中的范围靠；解释和这个实验没关系就按基线附近写并放宽；拿不准就把 q1 和 q9 拉开",
|
||||
},
|
||||
),
|
||||
body: Type.String({
|
||||
...TEXT,
|
||||
description: "理由，Markdown：这些数是怎么来的，不同解释给的数差在哪里；几条解释给的数差不多，写明它们在这些实验上分不开",
|
||||
}),
|
||||
},
|
||||
CLOSED,
|
||||
);
|
||||
|
||||
/** 诊断实现和修复用同一个定义。代码本身由程序从实验员的工作目录冻结，不作为字段交。 */
|
||||
export const DiagnosisImplementationSubmission = Type.Object(
|
||||
{
|
||||
locations: Type.Array(
|
||||
Type.Object(
|
||||
{
|
||||
experiment: Type.String({ ...Id, description: "诊断实验编号，照材料里列的写" }),
|
||||
locations: LOCATIONS,
|
||||
},
|
||||
CLOSED,
|
||||
),
|
||||
{
|
||||
description: "位置表：选中的每个实验一项，列出本次代码中这个实验实际所在的全部位置，行号以本次提交代码为准",
|
||||
},
|
||||
),
|
||||
concerns: CONCERNS,
|
||||
body: Type.String({
|
||||
...TEXT,
|
||||
description: "实现说明，Markdown：位置表之外需要说明的事；实验原文没规定规模时，写你选了什么、为什么放得进分到的时间。修复时逐条回应核验意见：改在哪里，或以实验原文和代码位置说明为什么不改。没有要补充的就写\"无\"",
|
||||
}),
|
||||
},
|
||||
CLOSED,
|
||||
);
|
||||
|
||||
export const DiagnosisVerificationSubmission = Type.Object(
|
||||
{
|
||||
experiments: Type.Array(
|
||||
Type.Object(
|
||||
{
|
||||
experiment: Type.String({ ...Id, description: "诊断实验编号，照材料里列的写" }),
|
||||
followed: Type.Boolean({
|
||||
description: "照实验写的做法实现了写 true；没找到、做得不一样或材料不足以核清，写 false",
|
||||
}),
|
||||
nothingElse: Type.Boolean({
|
||||
description: "没有动实验写的那一处以外的地方写 true；顺手改了别处（含顺手改进）写 false",
|
||||
}),
|
||||
where: WHERE,
|
||||
},
|
||||
CLOSED,
|
||||
),
|
||||
{
|
||||
description: "选中的每个实验一项，逐个判断，一个实验不通过不影响别的实验",
|
||||
},
|
||||
),
|
||||
body: Type.String({
|
||||
...TEXT,
|
||||
description: "说明，Markdown：逐项结果之外需要说明的事。没有要补充的就写\"无\"",
|
||||
}),
|
||||
},
|
||||
{
|
||||
...CLOSED,
|
||||
description: "只交逐个实验的结果，通过与否由程序推出：任何一项写 false 的实验退回实验员修复，用完修复次数仍不通过的实验从这次诊断里拿掉",
|
||||
},
|
||||
);
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// 解释、消融、收尾（方法.md 4.3 ⑭⑯⑱）
|
||||
// ---------------------------------------------------------------------------
|
||||
@@ -611,6 +751,11 @@ export const SUBMISSION_SCHEMAS = {
|
||||
entry: EntrySubmission,
|
||||
discussion: DiscussionSubmission,
|
||||
claim: ClaimSubmission,
|
||||
diagnosisProposal: DiagnosisProposalSubmission,
|
||||
diagnosisPrediction: DiagnosisPredictionSubmission,
|
||||
diagnosisImplementation: DiagnosisImplementationSubmission,
|
||||
diagnosisRepair: DiagnosisImplementationSubmission,
|
||||
diagnosisVerification: DiagnosisVerificationSubmission,
|
||||
proposal: ProposalSubmission,
|
||||
firstPrediction: PredictionSubmission,
|
||||
critique: CritiqueSubmission,
|
||||
|
||||
@@ -154,6 +154,34 @@ import 之前它还会删掉工作目录里已有的 `solution.json` 和 `metric
|
||||
**不要自己跑正式规模的求解。** 正式运行要先登记、再启动、再扣次数，由程序统一做；
|
||||
正式规模的求解哪怕叫"调试"，也是正式运行，不能借轻量检查或自己的 shell 绕开预算。
|
||||
|
||||
## 诊断实验怎么拿到原始产物
|
||||
|
||||
诊断实验（在当前基线上只动一处、再看分数）不走上面的 `run_packing()`。实验员把本轮选中的全部实验写进**一份诊断程序**，
|
||||
它从当前基线的代码出发，入口仍是 `initial.py`，但要定义的是：
|
||||
|
||||
```python
|
||||
def run_experiment(experiment_id, baseline_dir):
|
||||
...
|
||||
return centers, radii, sum_of_radii # 与 run_packing() 的返回值同形
|
||||
```
|
||||
|
||||
- `experiment_id` 是实验编号原文，函数里按它分派到各个实验；每个实验只动它写的那一处，其余照基线。
|
||||
- `baseline_dir` 是**产出当前基线的那次正式运行的原始产物**所在的目录，相对工作目录，现在就是 `baseline-artifact`。
|
||||
里面有一个文件 `solution.json`：当时那次运行交出的原始解（`centers`、`radii`、`reported_sum`，合法化之前），
|
||||
实验要接着算的答案就在里面。程序把它**只读**地放进你的工作目录（文件权限 0444），运行时每个实验读到的是同一份；
|
||||
不要改它，也不要把它当成自己代码的一部分：存代码版本时 `baseline-artifact` 这个名字被跳过。
|
||||
- 一次诊断运行是**一个容器、一次正式运行的名额**，总时限和核数同上（120 秒、1 核）。容器里的诊断 harness
|
||||
（`harness-diagnosis.py`，宿主保管，由程序放进运行目录）按实验顺序一个一个跑，每个实验开一个单独的进程，
|
||||
分到**总时限 ÷ 实验数**的时间，到时被直接杀掉。numba 的首次编译、数据载入都算在这份时间里。
|
||||
一个实验崩了或超时，只记这一个实验失败，后面的照常跑；整次运行被总时限杀掉时，没轮到的实验记"没轮到"。
|
||||
- harness 记下每个实验实际用了多久、有没有超时。每个实验的答案写在 `answers/<实验编号>/solution.json`，
|
||||
由宿主的校验器按上面"评测会收下什么"的次序**逐个单独判分**（包括只缩不移合法化）。
|
||||
- 实验登记的假设专用指标，写法同下面"指标怎么交给宿主"：在 `run_experiment` 返回之前往工作目录的 `metrics.json` 写，
|
||||
键照抄 `metrics-spec.json` 里**这个实验自己登记**的指标编号（该文件里是本轮全部实验登记的指标，别的实验的不用写）。
|
||||
每个实验开始前 harness 会清掉工作目录里的 `metrics.json`，不会串到下一个实验。诊断实验登记的指标一律是"缺了只记缺失"，
|
||||
`answer-path` 和 `safeguard-triggers` 两个保留键不用写。
|
||||
- 诊断实验的答案**不会被采纳为基线**，只用来看这个方向还剩多少空间。
|
||||
|
||||
## 指标怎么交给宿主（这一节最容易写错，先读完再动手）
|
||||
|
||||
工作目录里有两个指标文件，一进一出。**路径都相对工作目录**，也就是 `harness.py`
|
||||
|
||||
@@ -0,0 +1,125 @@
|
||||
"""填圆任务的诊断运行 harness（宿主的固定工具，候选不得修改）。
|
||||
|
||||
一次诊断运行是一个容器、一次正式运行的名额，里面按顺序跑若干个诊断实验（docs/方法.md 4.3"诊断"一段）。
|
||||
每个实验各开一个子进程：实验崩了、超时了，只影响它自己，后面的实验照常跑。
|
||||
**它不判定解是否合法**：每个实验交出的答案，事后由独立校验器 ``validate.py`` 逐个重算。
|
||||
读候选程序、写原始解、整理 metrics.json 都用 ``harness.py`` 里现成的函数，两份 harness 在这些事上不分叉。
|
||||
|
||||
命令行（由宿主的运行定义给出，实验员不用自己跑）：
|
||||
python -B harness-diagnosis.py --program <程序> --baseline <目录> --answers <目录>
|
||||
--seconds-each <秒> --report <文件> --experiments <实验编号> [<实验编号> ...]
|
||||
|
||||
每个实验：
|
||||
- 调用候选程序里的 ``run_experiment(experiment_id, baseline_dir)``，返回 (centers, radii, reported_sum)，
|
||||
与 ``run_packing()`` 的返回值同形。``experiment_id`` 是实验编号原文，``baseline_dir`` 是只读的原始产物目录
|
||||
（相对工作目录），里面是产出当前基线的那次正式运行的 ``solution.json``。
|
||||
- 分到的时限是 ``--seconds-each``（总时限 ÷ 实验数），到时被直接杀掉。
|
||||
- 随机种子同正式运行：环境变量 INFORSI_SEED，每个实验的子进程各自从头设一次。
|
||||
- 答案写进 ``<answers>/<实验编号>/solution.json``。指标写法同 TASK.md"指标怎么交给宿主"：
|
||||
在 ``run_experiment`` 返回之前往工作目录里的 ``metrics.json`` 写，写完后 harness 把它整理好、
|
||||
搬到 ``<answers>/<实验编号>/metrics.json``，下一个实验开始前工作目录里不留它。
|
||||
|
||||
每个实验结束后，harness 重写一次 ``--report``：每个实验一项
|
||||
``{"id", "status", "wall_seconds"}``，status 是 finished（子进程退出码 0）、crashed（非零退出）、
|
||||
timeout（超过分到的时限，被杀）。跑到哪写到哪：整个容器被总时限杀掉时，已经跑完的实验还留着记录，
|
||||
没跑到的实验没有记录。
|
||||
"""
|
||||
|
||||
import argparse
|
||||
import hashlib
|
||||
import importlib
|
||||
import json
|
||||
import os
|
||||
import random
|
||||
import signal
|
||||
import subprocess
|
||||
import sys
|
||||
import time
|
||||
from typing import Any, cast
|
||||
|
||||
import harness
|
||||
|
||||
SOLUTION_FILE = "solution.json"
|
||||
|
||||
|
||||
def run_experiment_child(args) -> int:
|
||||
"""子进程：设种子、载入候选程序、跑一个实验、落盘答案和指标。"""
|
||||
np = importlib.import_module("numpy")
|
||||
|
||||
seed = int.from_bytes(
|
||||
hashlib.sha256(os.environ["INFORSI_SEED"].encode()).digest()[:4], "big"
|
||||
)
|
||||
random.seed(seed)
|
||||
np.random.seed(seed)
|
||||
|
||||
module = harness.load_program(args.program)
|
||||
run_experiment = getattr(module, "run_experiment", None)
|
||||
if not callable(run_experiment):
|
||||
raise RuntimeError(f"{args.program} 没有定义可调用的 run_experiment(experiment_id, baseline_dir)")
|
||||
centers, radii, reported_sum = cast(Any, run_experiment(args.id, args.baseline))
|
||||
solution = {
|
||||
"centers": [[float(x), float(y)] for x, y in centers],
|
||||
"radii": [float(r) for r in radii],
|
||||
"reported_sum": float(reported_sum),
|
||||
}
|
||||
harness.write_json_atomically(args.output, solution)
|
||||
harness.finalize_metrics()
|
||||
return 0
|
||||
|
||||
|
||||
def run_experiment_with_limit(args, experiment_id: str) -> tuple[str, float]:
|
||||
"""在子进程里跑一个实验，交回结局和用时。"""
|
||||
answer_dir = os.path.join(args.answers, experiment_id)
|
||||
os.makedirs(answer_dir)
|
||||
# 上一个实验或候选目录里遗留的指标文件，不能被当成这个实验的。
|
||||
if os.path.exists(harness.METRICS_FILE):
|
||||
os.remove(harness.METRICS_FILE)
|
||||
command = [
|
||||
sys.executable, "-B", os.path.abspath(__file__), "--child",
|
||||
"--program", args.program, "--baseline", args.baseline,
|
||||
"--id", experiment_id, "--output", os.path.join(answer_dir, SOLUTION_FILE),
|
||||
]
|
||||
start = time.monotonic()
|
||||
# 新开一个进程组：超时时连实验自己开的子进程一起杀掉。
|
||||
child = subprocess.Popen(command, start_new_session=True)
|
||||
try:
|
||||
status = "finished" if child.wait(timeout=args.seconds_each) == 0 else "crashed"
|
||||
except subprocess.TimeoutExpired:
|
||||
os.killpg(child.pid, signal.SIGKILL)
|
||||
child.wait()
|
||||
status = "timeout"
|
||||
wall = time.monotonic() - start
|
||||
if os.path.exists(harness.METRICS_FILE):
|
||||
if status == "finished":
|
||||
os.replace(harness.METRICS_FILE, os.path.join(answer_dir, harness.METRICS_FILE))
|
||||
else:
|
||||
os.remove(harness.METRICS_FILE)
|
||||
return status, wall
|
||||
|
||||
|
||||
def main() -> int:
|
||||
parser = argparse.ArgumentParser(description="按顺序运行若干个诊断实验，各自限时、互不影响。")
|
||||
parser.add_argument("--program", required=True)
|
||||
parser.add_argument("--baseline", required=True)
|
||||
parser.add_argument("--child", action="store_true")
|
||||
parser.add_argument("--id")
|
||||
parser.add_argument("--output")
|
||||
parser.add_argument("--answers")
|
||||
parser.add_argument("--seconds-each", type=float)
|
||||
parser.add_argument("--report")
|
||||
parser.add_argument("--experiments", nargs="+")
|
||||
args = parser.parse_args()
|
||||
if args.child:
|
||||
return run_experiment_child(args)
|
||||
|
||||
report: list[dict] = []
|
||||
for experiment_id in args.experiments:
|
||||
status, wall = run_experiment_with_limit(args, experiment_id)
|
||||
print(f"[诊断实验 {experiment_id}] {status}，用时 {wall:.2f} 秒", file=sys.stderr, flush=True)
|
||||
report.append({"id": experiment_id, "status": status, "wall_seconds": wall})
|
||||
harness.write_json_atomically(args.report, report)
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
sys.exit(main())
|
||||
@@ -177,6 +177,7 @@ test("写不出来的组合：朴素循环里写开关、规则或它用不到
|
||||
critique: false,
|
||||
firstPrediction: false,
|
||||
roundZero: { enabled: false },
|
||||
diagnosis: { enabled: false },
|
||||
ablation: { enabled: false },
|
||||
batchRule: "uniformRandom",
|
||||
openReserve: 0.1,
|
||||
@@ -200,6 +201,35 @@ test("写不出来的组合：关掉的第零轮写讨论次数、关掉的消
|
||||
assert.match(errorOf(ablation), /^\/flow\/ablation\/cellRule：不认识的字段$/m);
|
||||
});
|
||||
|
||||
/** 诊断开着的完整流程配置。 */
|
||||
function diagnosisConfig(): ReturnType<typeof fullConfig> {
|
||||
const config = fullConfig();
|
||||
(config.flow as Obj).diagnosis = { enabled: true, perResearcher: 2, maxExperiments: 4 };
|
||||
return config;
|
||||
}
|
||||
|
||||
test("诊断：开着的配置通过校验；逐个删字段、写 0、写小数、多写字段都报错；关着的不能写两个参数", () => {
|
||||
assert.deepEqual(checkConfig(structuredClone(diagnosisConfig())), diagnosisConfig());
|
||||
|
||||
for (const key of ["enabled", "perResearcher", "maxExperiments"]) {
|
||||
const config = diagnosisConfig();
|
||||
delete ((config.flow as Obj).diagnosis as Obj)[key];
|
||||
assert.match(errorOf(config), new RegExp(`^/flow/diagnosis/${key}：`, "m"));
|
||||
}
|
||||
const cases: Array<[Obj, RegExp]> = [
|
||||
[{ enabled: true, perResearcher: 0, maxExperiments: 4 }, /^\/flow\/diagnosis\/perResearcher：不能小于 1$/m],
|
||||
[{ enabled: true, perResearcher: 2, maxExperiments: 2.5 }, /^\/flow\/diagnosis\/maxExperiments：应为整数$/m],
|
||||
[{ enabled: true, perResearcher: 2, maxExperiments: 4, rule: "x" }, /^\/flow\/diagnosis\/rule：不认识的字段$/m],
|
||||
[{ enabled: false, perResearcher: 2 }, /^\/flow\/diagnosis\/perResearcher：不认识的字段$/m],
|
||||
[{ enabled: "yes" }, /^\/flow\/diagnosis\/enabled：缺少或取值不对，只能是 false、true 之一$/m],
|
||||
];
|
||||
for (const [diagnosis, expected] of cases) {
|
||||
const config = fullConfig();
|
||||
(config.flow as Obj).diagnosis = diagnosis;
|
||||
assert.match(errorOf(config), expected);
|
||||
}
|
||||
});
|
||||
|
||||
test("写不出来的组合：规则名不在 rules 登记表里", () => {
|
||||
const batch = fullConfig();
|
||||
(batch.flow as Obj).batchRule = "expected-best";
|
||||
|
||||
@@ -29,6 +29,7 @@ export function fullConfig(): ResearchConfig {
|
||||
roundZero: { enabled: true, discussions: 2 },
|
||||
firstPrediction: true,
|
||||
critique: true,
|
||||
diagnosis: { enabled: false },
|
||||
ablation: { enabled: true, cellRule: "valueMinusCost" },
|
||||
batchRule: "bestExpected",
|
||||
openReserve: 0.1,
|
||||
|
||||
@@ -28,7 +28,7 @@ function assertReasons(list: StepList): void {
|
||||
}
|
||||
|
||||
test("步骤清单：完整流程全开", () => {
|
||||
const list = stepList(fullConfig());
|
||||
const list = listOf((f) => (f.diagnosis = { enabled: true, perResearcher: 2, maxExperiments: 4 }));
|
||||
assert.equal(list.preset, "full");
|
||||
assert.deepEqual(rows(list.study), [
|
||||
["freezeConfig", E, null],
|
||||
@@ -44,6 +44,13 @@ test("步骤清单：完整流程全开", () => {
|
||||
assert.deepEqual(rows(list.round), [
|
||||
["openRound", E, null],
|
||||
["carryOverPool", E, null],
|
||||
["diagnosisProposal", E, null],
|
||||
["diagnosisDedupe", E, null],
|
||||
["diagnosisPrediction", E, null],
|
||||
["selectDiagnosis", E, { name: "valueMinusCost", version: 1 }],
|
||||
["diagnosisImplementation", E, null],
|
||||
["diagnosisRun", E, null],
|
||||
["diagnosisSettle", E, null],
|
||||
["proposal", E, null],
|
||||
["dedupe", E, null],
|
||||
["firstPrediction", E, null],
|
||||
@@ -95,6 +102,13 @@ test("步骤清单：只关第一次预测", () => {
|
||||
assert.deepEqual(rows(list.round), [
|
||||
["openRound", E, null],
|
||||
["carryOverPool", E, null],
|
||||
["diagnosisProposal", S, null],
|
||||
["diagnosisDedupe", S, null],
|
||||
["diagnosisPrediction", S, null],
|
||||
["selectDiagnosis", S, null],
|
||||
["diagnosisImplementation", S, null],
|
||||
["diagnosisRun", S, null],
|
||||
["diagnosisSettle", S, null],
|
||||
["proposal", E, null],
|
||||
["dedupe", E, null],
|
||||
["firstPrediction", S, null],
|
||||
@@ -128,6 +142,13 @@ test("步骤清单：只关挑刺与回应，协议照样冻结", () => {
|
||||
assert.deepEqual(rows(list.round), [
|
||||
["openRound", E, null],
|
||||
["carryOverPool", E, null],
|
||||
["diagnosisProposal", S, null],
|
||||
["diagnosisDedupe", S, null],
|
||||
["diagnosisPrediction", S, null],
|
||||
["selectDiagnosis", S, null],
|
||||
["diagnosisImplementation", S, null],
|
||||
["diagnosisRun", S, null],
|
||||
["diagnosisSettle", S, null],
|
||||
["proposal", E, null],
|
||||
["dedupe", E, null],
|
||||
["firstPrediction", E, null],
|
||||
@@ -154,13 +175,20 @@ test("步骤清单：只关挑刺与回应，协议照样冻结", () => {
|
||||
]);
|
||||
});
|
||||
|
||||
test("步骤清单：只关消融，解释照样写，选格没有规则", () => {
|
||||
test("步骤清单：只关消融，解释也跳过，选格没有规则", () => {
|
||||
const list = listOf((f) => (f.ablation = { enabled: false }));
|
||||
assertReasons(list);
|
||||
assert.deepEqual(rows(list.study), rows(stepList(fullConfig()).study));
|
||||
assert.deepEqual(rows(list.round), [
|
||||
["openRound", E, null],
|
||||
["carryOverPool", E, null],
|
||||
["diagnosisProposal", S, null],
|
||||
["diagnosisDedupe", S, null],
|
||||
["diagnosisPrediction", S, null],
|
||||
["selectDiagnosis", S, null],
|
||||
["diagnosisImplementation", S, null],
|
||||
["diagnosisRun", S, null],
|
||||
["diagnosisSettle", S, null],
|
||||
["proposal", E, null],
|
||||
["dedupe", E, null],
|
||||
["firstPrediction", E, null],
|
||||
@@ -177,7 +205,7 @@ test("步骤清单：只关消融，解释照样写，选格没有规则", () =>
|
||||
["informationAccount", E, null],
|
||||
["adopt", E, null],
|
||||
["listCells", S, null],
|
||||
["explanation", E, null],
|
||||
["explanation", S, null],
|
||||
["selectCells", S, null],
|
||||
["ablationImplementation", S, null],
|
||||
["ablationVerification", S, null],
|
||||
@@ -205,6 +233,13 @@ test("步骤清单：朴素循环", () => {
|
||||
assert.deepEqual(rows(list.round), [
|
||||
["openRound", E, null],
|
||||
["carryOverPool", S, null], // 候选不留池
|
||||
["diagnosisProposal", S, null], // 不诊断
|
||||
["diagnosisDedupe", S, null], // 不诊断
|
||||
["diagnosisPrediction", S, null], // 不诊断
|
||||
["selectDiagnosis", S, null], // 不诊断
|
||||
["diagnosisImplementation", S, null], // 不诊断
|
||||
["diagnosisRun", S, null], // 不诊断
|
||||
["diagnosisSettle", S, null], // 不诊断
|
||||
["proposal", E, null],
|
||||
["dedupe", E, null], // 判重照常
|
||||
["firstPrediction", S, null], // 不预测
|
||||
@@ -264,3 +299,30 @@ test("步骤清单：换规则后清单里的规则跟着换", () => {
|
||||
assert.deepEqual(rule("selectBatch"), { name: "uniformRandom", version: 1 });
|
||||
assert.deepEqual(rule("selectCells"), { name: "randomSameCount", version: 1 });
|
||||
});
|
||||
|
||||
test("步骤清单：诊断开、消融关（现行配置）：诊断七步执行，消融七步（含解释）跳过", () => {
|
||||
const list = listOf((f) => {
|
||||
f.diagnosis = { enabled: true, perResearcher: 2, maxExperiments: 4 };
|
||||
f.ablation = { enabled: false };
|
||||
});
|
||||
assertReasons(list);
|
||||
const status = (id: string) => list.round.find((s) => s.id === id)!.status;
|
||||
for (const id of [
|
||||
"diagnosisProposal", "diagnosisDedupe", "diagnosisPrediction", "selectDiagnosis",
|
||||
"diagnosisImplementation", "diagnosisRun", "diagnosisSettle",
|
||||
]) {
|
||||
assert.equal(status(id), "execute", id);
|
||||
}
|
||||
for (const id of [
|
||||
"listCells", "explanation", "selectCells", "ablationImplementation", "ablationVerification", "ablationRun", "settle",
|
||||
]) {
|
||||
assert.equal(status(id), "skip", id);
|
||||
}
|
||||
// 诊断七步排在留池之后、提案之前（方法.md 4.3"诊断"一段）。
|
||||
const ids = list.round.map((s) => s.id);
|
||||
assert.deepEqual(ids.slice(ids.indexOf("carryOverPool"), ids.indexOf("proposal") + 1), [
|
||||
"carryOverPool", "diagnosisProposal", "diagnosisDedupe", "diagnosisPrediction", "selectDiagnosis",
|
||||
"diagnosisImplementation", "diagnosisRun", "diagnosisSettle", "proposal",
|
||||
]);
|
||||
assert.deepEqual(list.round.find((s) => s.id === "selectDiagnosis")!.rule, { name: "valueMinusCost", version: 1 });
|
||||
});
|
||||
|
||||
Vendored
+1937
File diff suppressed because it is too large
Load Diff
+318
@@ -0,0 +1,318 @@
|
||||
"""生成诊断（方法.md 6.12）的预期值 JSON（diagnosis.json）。
|
||||
|
||||
用法（在仓库根目录）：
|
||||
python3 tests/fixtures/rules/gen_diagnosis.py
|
||||
只用标准库；依赖同目录的 rules_ref.py 里的基础件（折线、有理数分布、Stieltjes 精确积分、比特、惊讶、分歧量）。
|
||||
重跑结果逐位相同。
|
||||
|
||||
独立性：诊断的成分模型、方向收益、信息价值、价值减成本、选实验、结算都在本文件里只照 docs/方法.md 6.12 和 6.7、6.9
|
||||
的文字写成，没有读过 src/rules/diagnosis.ts 和 src/rules/value.ts。
|
||||
|
||||
算法（有理数精确，浮点输入先精确转成有理数）：
|
||||
- 每个成分每个实验的分布：各条解释的折线加这个实验自己的开放预留（6.3"诊断"），换成收益 g = 改进量（x − b，越小越好时 b − x）；
|
||||
- E(g_d)：各成分期望按起始权重平均；
|
||||
- 全做（k = 实验总数）的 VOI：Σ_j w_j · E_j[max(0, max_d g_d)]（rules_ref.expect_max 的 floor0 模式）− max(0, max_d E g_d)；
|
||||
- 只做一个实验 d0 的 VOI（exactK1）：d0 的各成分在相邻折点之间密度是常数，所以后验权重在每一段里是常数，
|
||||
没做的实验的期望收益 c_e 在每一段里是常数，max(0, v, c_e…) 对 v 的积分有闭式；
|
||||
- 价值 = u·m′·VOI，R̂ = ⌈(余额 − 1) ÷ K⌉，成本 = 2·max(0, ĝ₀)·[(K+1)/(K+2) − K/(K+1)]；
|
||||
- 推出的方向：成分新权重 ∝ 起始权重 × 各有效实验在实测处的密度（所在段斜率，段左闭右开）。
|
||||
|
||||
生成的文件：diagnosis.json，格式：
|
||||
{"generator", "tolerance", "scenarios": [{"name", "section", "input", "expected", "note"}]}
|
||||
- 场景 input：{"b", "direction", "lambda", "experiments": ["d1", ...],
|
||||
"explanations": [{"id", "author", "ranges": {"d1": {"q1","q5","q9"}, ...}}]}，
|
||||
选实验场景另带 "balance"、"batchCap"、"maxExperiments"、"u"、"decay"，结算场景另带 "epsilon" 和 "outcomes"。
|
||||
- 分数一律是原值；收益、VOI 是收益刻度（改进量）。
|
||||
"""
|
||||
|
||||
import json
|
||||
import math
|
||||
import os
|
||||
import sys
|
||||
from fractions import Fraction as Fr
|
||||
from itertools import combinations
|
||||
|
||||
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
|
||||
import rules_ref as R # noqa: E402
|
||||
|
||||
HERE = os.path.dirname(os.path.abspath(__file__))
|
||||
F = float
|
||||
|
||||
|
||||
def tri(t):
|
||||
return {"q1": t[0], "q5": t[1], "q9": t[2]}
|
||||
|
||||
|
||||
def ids_key(s):
|
||||
return R.factor_key(s)
|
||||
|
||||
|
||||
class Model:
|
||||
"""诊断的成分模型（有理数）。explanations 按编号排好；成分 = 各解释 + 开放预留。"""
|
||||
|
||||
def __init__(self, b, direction, lam, experiments, explanations):
|
||||
self.b, self.direction, self.lam = fr(b), direction, fr(lam)
|
||||
self.exps = sorted(experiments, key=ids_key)
|
||||
self.expl = sorted(explanations, key=lambda e: ids_key(e["id"]))
|
||||
authors = [e["author"] for e in self.expl]
|
||||
self.w = R.explanation_weights(authors, lam) + [fr(lam)]
|
||||
# raw[d][j]：分数原值上的点列；gain[d][j]：收益（改进量）上的点列
|
||||
self.raw, self.gain = {}, {}
|
||||
for d in self.exps:
|
||||
pts = [R.polyline(*e["ranges"][d]) for e in self.expl]
|
||||
lo, hi = R.reserve_bounds(pts)
|
||||
pts.append(R.uniform_pts(lo, hi))
|
||||
self.raw[d] = pts
|
||||
self.gain[d] = [R.improvement_pts(p, self.b, direction) for p in pts]
|
||||
self.mu = {d: [R.mean(R.single(p)) for p in self.gain[d]] for d in self.exps}
|
||||
self.prior = {d: sum((wj * m for wj, m in zip(self.w, self.mu[d])), Fr(0)) for d in self.exps}
|
||||
|
||||
def g0(self):
|
||||
return max(Fr(0), max(self.prior.values()))
|
||||
|
||||
def voi_all(self):
|
||||
total = Fr(0)
|
||||
for j, wj in enumerate(self.w):
|
||||
if wj == 0:
|
||||
continue
|
||||
total += wj * R.expect_max([R.single(self.gain[d][j]) for d in self.exps], floor0=True)
|
||||
return total - self.g0()
|
||||
|
||||
def voi_k1(self, d0):
|
||||
"""只做 d0：每一段里后验权重和没做的实验的期望收益都是常数，对 max(0, v, c…) 精确积分。"""
|
||||
bps = sorted(set().union(*[set(x for x, _ in p) for p in self.gain[d0]]))
|
||||
total = Fr(0)
|
||||
for a, b in zip(bps, bps[1:]):
|
||||
dens = [R.pts_left(p, b) - R.pts_left(p, a) for p in self.gain[d0]]
|
||||
dens = [x / (b - a) for x in dens]
|
||||
marg = sum((wj * x for wj, x in zip(self.w, dens)), Fr(0))
|
||||
if marg == 0:
|
||||
continue
|
||||
c0 = Fr(0)
|
||||
for e in self.exps:
|
||||
if e == d0:
|
||||
continue
|
||||
num = sum((wj * x * m for wj, x, m in zip(self.w, dens, self.mu[e])), Fr(0))
|
||||
c0 = max(c0, num / marg)
|
||||
t = min(max(c0, a), b)
|
||||
total += marg * (c0 * (t - a) + (b * b - t * t) / 2)
|
||||
return total - self.g0()
|
||||
|
||||
def mixture(self, d):
|
||||
"""实验 d 的成分混合（分数原值）。"""
|
||||
return R.Dist(list(zip(self.w, self.raw[d])))
|
||||
|
||||
|
||||
def fr(x):
|
||||
return R.fr(x)
|
||||
|
||||
|
||||
def model_of(inp):
|
||||
expl = [{"id": e["id"], "author": e["author"], "ranges": {d: (r["q1"], r["q5"], r["q9"]) for d, r in e["ranges"].items()}} for e in inp["explanations"]]
|
||||
return Model(inp["b"], inp["direction"], inp["lambda"], inp["experiments"], expl)
|
||||
|
||||
|
||||
def explanation(eid, author, ranges):
|
||||
return {"id": eid, "author": author, "ranges": {d: tri(t) for d, t in ranges.items()}}
|
||||
|
||||
|
||||
def m_prime(balance, K, r):
|
||||
R_ = -(-(balance - 1) // K)
|
||||
return R_, sum((fr(r) ** j for j in range(R_)), Fr(0))
|
||||
|
||||
|
||||
def cost_of(g0, K):
|
||||
return 2 * g0 * (Fr(K + 1, K + 2) - Fr(K, K + 1))
|
||||
|
||||
|
||||
S = []
|
||||
|
||||
|
||||
def select_case(name, inp, balance, K, D, u, r, note):
|
||||
inp = dict(inp, balance=balance, batchCap=K, maxExperiments=D, u=u, decay=r)
|
||||
m = model_of(inp)
|
||||
rows = []
|
||||
if balance >= 2:
|
||||
size = min(len(m.exps), D)
|
||||
R_, mp = m_prime(balance, K, r)
|
||||
c = cost_of(m.g0(), K)
|
||||
for combo in combinations(m.exps, size):
|
||||
if size == len(m.exps):
|
||||
voi = m.voi_all()
|
||||
elif size == 1:
|
||||
voi = m.voi_k1(combo[0])
|
||||
else:
|
||||
voi = None # 网格近似，没有精确值，只检查结构
|
||||
if voi is None:
|
||||
rows.append({"experiments": list(combo), "voi": None})
|
||||
continue
|
||||
value = fr(u) * mp * voi
|
||||
rows.append({"experiments": list(combo), "voi": F(voi), "rHat": R_, "mPrime": F(mp), "value": F(value),
|
||||
"cost": F(c), "net": F(value - c)})
|
||||
expected = {
|
||||
"priorGains": {d: F(m.prior[d]) for d in m.exps},
|
||||
"g0": F(m.g0()),
|
||||
"weights": [F(x) for x in m.w],
|
||||
"combos": rows,
|
||||
}
|
||||
if all(r_["voi"] is not None for r_ in rows) and rows:
|
||||
eligible = [r_ for r_ in rows if r_["net"] > 1e-12]
|
||||
if eligible:
|
||||
top = max(r_["net"] for r_ in eligible)
|
||||
near = [r_ for r_ in eligible if r_["net"] >= top - 1e-12]
|
||||
best = min(near, key=lambda r_: [ids_key(x) for x in r_["experiments"]])
|
||||
expected["diagnose"] = True
|
||||
expected["chosen"] = best["experiments"]
|
||||
expected["chosenNet"] = best["net"]
|
||||
else:
|
||||
expected["diagnose"] = False
|
||||
expected["chosen"] = []
|
||||
expected["chosenNet"] = 0.0
|
||||
elif not rows:
|
||||
expected["diagnose"] = False
|
||||
expected["chosen"] = []
|
||||
expected["chosenNet"] = 0.0
|
||||
S.append({"name": name, "section": "6.12", "input": inp, "expected": expected, "note": note})
|
||||
|
||||
|
||||
def settle_case(name, inp, outcomes, note):
|
||||
inp = dict(inp, outcomes=outcomes)
|
||||
m = model_of(inp)
|
||||
eps = fr(inp["epsilon"])
|
||||
delta = F(eps / 100)
|
||||
b = F(m.b)
|
||||
scores = {d: (o["score"] if o["kind"] == "valid" else None) for d, o in outcomes.items()}
|
||||
rows = {}
|
||||
for d in m.exps:
|
||||
dist = m.mixture(d)
|
||||
row = {"outcome": outcomes[d]["kind"]}
|
||||
row["merged"] = {"q1": F(R.quantile(dist, Fr(1, 10))), "q5": F(R.quantile(dist, Fr(1, 2))), "q9": F(R.quantile(dist, Fr(9, 10)))}
|
||||
if scores[d] is not None:
|
||||
x = scores[d]
|
||||
row["gain"] = F(R.improve(x, m.b, m.direction))
|
||||
row["worldBits"] = R.world_bits(dist, b, delta)
|
||||
sp = R.surprise(dist, x, b, delta)
|
||||
row["surprise"] = sp
|
||||
persons = {}
|
||||
for e in m.expl:
|
||||
persons.setdefault(e["author"], []).append(R.polyline(*e["ranges"][d]))
|
||||
person_dists = [R.Dist([(Fr(1, len(ps)), p) for p in ps]) for _, ps in sorted(persons.items(), key=lambda kv: ids_key(kv[0]))]
|
||||
row["disagreement"] = R.disagreement(person_dists, b, delta)[0]
|
||||
rows[d] = row
|
||||
# 确认与否定：只有有效实验判，落在 [q1, q9]（含两端）确认。
|
||||
judgments = {}
|
||||
for e in m.expl:
|
||||
judgments[e["id"]] = {d: ("confirmed" if e["ranges"][d][0] <= scores[d] <= e["ranges"][d][2] else "denied")
|
||||
for d in m.exps if scores[d] is not None}
|
||||
# 诊断推出的方向
|
||||
new = []
|
||||
for j, wj in enumerate(m.w):
|
||||
v = wj
|
||||
for d in m.exps:
|
||||
if scores[d] is not None:
|
||||
v *= R.pts_density(m.raw[d][j], fr(scores[d]))
|
||||
new.append(v)
|
||||
tot = sum(new, Fr(0))
|
||||
kept = tot == 0
|
||||
w2 = m.w if kept else [v / tot for v in new]
|
||||
gains = {}
|
||||
for d in m.exps:
|
||||
if scores[d] is not None:
|
||||
gains[d] = {"gain": F(R.improve(scores[d], m.b, m.direction)), "measured": True}
|
||||
else:
|
||||
gains[d] = {"gain": F(sum((wj * mu for wj, mu in zip(w2, m.mu[d])), Fr(0))), "measured": False}
|
||||
best = m.exps[0]
|
||||
for d in m.exps[1:]:
|
||||
if gains[d]["gain"] > gains[best]["gain"]:
|
||||
best = d
|
||||
expected = {"experiments": rows, "judgments": judgments,
|
||||
"direction": {"weights": [F(x) for x in w2], "weightsKept": kept, "gains": gains,
|
||||
"experiment": best, "gain": gains[best]["gain"], "found": gains[best]["gain"] > 0}}
|
||||
S.append({"name": name, "section": "6.12 结算；6.9", "input": inp, "expected": expected, "note": note})
|
||||
|
||||
|
||||
# ------------------------------------------------------------------ 场景
|
||||
|
||||
def base(explanations, experiments, b=0.0, direction="maximize", lam=0.1):
|
||||
return {"b": b, "direction": direction, "lambda": lam, "experiments": experiments, "explanations": explanations}
|
||||
|
||||
|
||||
# 1. 两条解释对同一个实验预测相反：A 说涨 0.2 左右，B 说跌 0.2 左右。
|
||||
OPP = [explanation("e1", "researcher-1", {"d1": (0.15, 0.2, 0.25)}),
|
||||
explanation("e2", "researcher-2", {"d1": (-0.25, -0.2, -0.15)})]
|
||||
select_case("opposite_lambda0", base(OPP, ["d1"], lam=0.0), 7, 2, 4, 0.5, 0.5,
|
||||
"λ = 0，两个成分的范围不相交：做了就知道是哪一条。手算：做了以后 E[max(0, g)] = 0.5 × 0.2 = 0.1，"
|
||||
"不做时 max(0, E g) = max(0, 0.5×0.2 + 0.5×(−0.2)) = 0，VOI = 0.1。")
|
||||
select_case("opposite_lambda01", base(OPP, ["d1"]), 7, 2, 4, 0.5, 0.5, "同上，λ = 0.1，开放预留铺在两端之外。")
|
||||
|
||||
# 2. 两条解释预测一样：λ = 0 时信息价值为 0；λ = 0.1 时只剩开放预留带来的一点。
|
||||
SAME = [explanation("e1", "researcher-1", {"d1": (0.1, 0.2, 0.3)}),
|
||||
explanation("e2", "researcher-2", {"d1": (0.1, 0.2, 0.3)})]
|
||||
select_case("same_lambda0", base(SAME, ["d1"], lam=0.0), 7, 2, 4, 0.5, 0.5,
|
||||
"λ = 0，两条解释一样，做不做都不改变看法：VOI = 0（手算：g 的取值范围 0 到 0.4 全在 0 以上，E[max(0, g)] = E g = 0.2）。净值 = −成本 < 0，不诊断。")
|
||||
select_case("same_lambda01", base(SAME, ["d1"]), 7, 2, 4, 0.5, 0.5, "λ = 0.1：解释们的范围都在 0 以上，只有开放预留（铺在 −0.2 到 0.6）能给出负的结果，所以 VOI 只剩预留带来的一小部分。")
|
||||
|
||||
# 3. 三个实验、三条解释，全做（3 ≤ D）。
|
||||
MULTI = [explanation("e1", "researcher-1", {"d1": (0.02, 0.05, 0.08), "d2": (-0.05, 0.0, 0.05), "d3": (0.0, 0.1, 0.2)}),
|
||||
explanation("e2", "researcher-2", {"d1": (0.0, 0.03, 0.06), "d2": (0.1, 0.2, 0.3), "d3": (-0.1, -0.05, 0.0)}),
|
||||
explanation("e3", "researcher-3", {"d1": (-0.02, 0.04, 0.1), "d2": (-0.2, -0.1, 0.0), "d3": (0.05, 0.1, 0.15)})]
|
||||
select_case("multi_all", base(MULTI, ["d1", "d2", "d3"]), 9, 2, 4, 0.5, 0.5,
|
||||
"三个实验不超过 D = 4，只有全做一个组合。R̂ = ⌈8 ÷ 2⌉ = 4，m′ = 1.875，成本 = 2ĝ₀ × (3/4 − 2/3)。")
|
||||
|
||||
# 4. 实验数超过 D：只比较装满的组合。D = 1 时是三个单实验组合（有精确值）；D = 2 时只有三个两实验组合（只查结构）。
|
||||
select_case("over_cap_d1", base(MULTI, ["d1", "d2", "d3"]), 9, 2, 1, 0.5, 0.5,
|
||||
"三个实验超过 D = 1：只有三个单实验组合，没有空组合；精确值用逐段闭式积分。")
|
||||
select_case("over_cap_d2", base(MULTI, ["d1", "d2", "d3"]), 9, 2, 2, 0.5, 0.5,
|
||||
"三个实验超过 D = 2：只有三个两实验组合，没有单实验组合；只检查组合列表（网格近似没有精确值）。")
|
||||
|
||||
# 5. 余额不到 2 次不诊断。
|
||||
select_case("balance_one", base(MULTI, ["d1", "d2", "d3"]), 1, 2, 4, 0.5, 0.5, "余额 1：没有可选组合，不诊断。")
|
||||
|
||||
# 6. 平票按实验编号：d2 和 d10 的预测完全一样，D = 1 时两个单实验组合净值相同，取 d2（自然序，d2 排在 d10 前面）。
|
||||
TIE = [explanation("e1", "researcher-1", {"d2": (-0.2, 0.0, 0.2), "d10": (-0.2, 0.0, 0.2)}),
|
||||
explanation("e2", "researcher-2", {"d2": (0.1, 0.3, 0.5), "d10": (0.1, 0.3, 0.5)})]
|
||||
select_case("tie_by_id", base(TIE, ["d10", "d2"]), 7, 2, 1, 0.5, 0.5, "两个实验的预测完全一样，净值相同（差不到 1e-12），取实验编号小的 d2。")
|
||||
|
||||
# 7. 结算。
|
||||
def shifted(expls, b):
|
||||
"""把以基线为 0 的改进量范围搬成分数原值（基线 b）。"""
|
||||
return [dict(e, ranges={d: {k: v + b for k, v in r.items()} for d, r in e["ranges"].items()}) for e in expls]
|
||||
|
||||
|
||||
SETTLE = shifted(MULTI[:2] + [explanation("e3", "researcher-2", {"d1": (-0.02, 0.04, 0.1), "d2": (-0.2, -0.1, 0.0), "d3": (0.05, 0.1, 0.15)})], 1.0)
|
||||
SINP = dict(base(SETTLE, ["d1", "d2", "d3"], b=1.0), epsilon=0.01)
|
||||
settle_case("settle_mixed", SINP,
|
||||
{"d1": {"kind": "valid", "score": 1.045}, "d2": {"kind": "insufficient"}, "d3": {"kind": "untested"}},
|
||||
"d1 有效，d2 证据不足（不判、不记账），d3 没做。研究者 2 交了两条解释，各占他那份的一半。")
|
||||
settle_case("settle_two_valid", SINP,
|
||||
{"d1": {"kind": "valid", "score": 1.07}, "d2": {"kind": "valid", "score": 1.25}, "d3": {"kind": "untested"}},
|
||||
"两个有效实验：权重按两处密度连乘更新；d2 的实测 +0.25 落在 e2 的范围里、其余两条之外。")
|
||||
settle_case("settle_no_direction", dict(SINP),
|
||||
{"d1": {"kind": "valid", "score": 0.97}, "d2": {"kind": "valid", "score": 0.8}, "d3": {"kind": "valid", "score": 0.9}},
|
||||
"全部实验的收益都不为正（−0.03、−0.2、−0.1）：最大收益 −0.03 ≤ 0，记没有指出能涨分的方向。")
|
||||
settle_case("settle_all_insufficient", SINP,
|
||||
{"d1": {"kind": "insufficient"}, "d2": {"kind": "insufficient"}, "d3": {"kind": "insufficient"}},
|
||||
"没有一个有效实验：不判、不记账，权重保持起始权重，方向按起始权重估的期望收益。")
|
||||
settle_case("settle_outside_all", SINP,
|
||||
{"d1": {"kind": "valid", "score": 3.0}, "d2": {"kind": "untested"}, "d3": {"kind": "untested"}},
|
||||
"d1 实测 +2.0 落在所有成分（含开放预留）的范围之外，权重无从更新：保持起始权重（weightsKept）。")
|
||||
|
||||
MIN_EXPL = [explanation("e1", "researcher-1", {"d1": (0.95, 0.9, 0.85)[::-1], "d2": (0.9, 1.0, 1.1)}),
|
||||
explanation("e2", "researcher-2", {"d1": (0.8, 0.85, 0.9), "d2": (0.95, 1.05, 1.15)})]
|
||||
MIN_INP = dict(base(MIN_EXPL, ["d1", "d2"], b=1.0, direction="minimize"), epsilon=0.01)
|
||||
select_case("minimize_all", dict(base(MIN_EXPL, ["d1", "d2"], b=1.0, direction="minimize")), 7, 2, 4, 0.5, 0.5,
|
||||
"越小越好：收益 = b − x。d1 两条解释都预测降到 0.9 左右（收益约 +0.1），d2 预测上升。")
|
||||
settle_case("settle_minimize", MIN_INP, {"d1": {"kind": "valid", "score": 0.88}, "d2": {"kind": "valid", "score": 1.05}},
|
||||
"越小越好：d1 实测 0.88，收益 = 1.0 − 0.88 = 0.12；d2 实测 1.05，收益 −0.05。")
|
||||
|
||||
doc = {"generator": "gen_diagnosis.py",
|
||||
"tolerance": {
|
||||
"exact": "有理数精确算出的量（成分权重、事前期望收益、ĝ₀、全做的 VOI、R̂、m′、成本、价值、净值、收益、更新后的权重）：绝对误差 1e-12",
|
||||
"exactK1": "只做一个实验的 VOI 是逐段闭式积分，实现用部分格网格近似：|差| ≤ 1e-9 + 2e-4·|精确值|",
|
||||
"bits": "世界比特、惊讶、分歧量是浮点：绝对误差 1e-9；合并范围是分位点：绝对误差 1e-9",
|
||||
"discrete": "是否诊断、选中的实验、确认与否定、有没有指出方向、实验的结局：必须完全相同"},
|
||||
"scenarios": S}
|
||||
with open(os.path.join(HERE, "diagnosis.json"), "w", encoding="utf-8") as f:
|
||||
json.dump(doc, f, ensure_ascii=False, indent=1)
|
||||
f.write("\n")
|
||||
print(len(S), "个场景")
|
||||
@@ -0,0 +1,230 @@
|
||||
import assert from "node:assert/strict";
|
||||
import { spawnSync } from "node:child_process";
|
||||
import { mkdirSync, mkdtempSync, rmSync, writeFileSync } from "node:fs";
|
||||
import { tmpdir } from "node:os";
|
||||
import { join } from "node:path";
|
||||
import { test, type TestContext } from "node:test";
|
||||
import { stepList } from "../src/config/index.ts";
|
||||
import { fullConfig as checkedFull } from "../src/flow/frozen.ts";
|
||||
import { judgeDiagnosisRun } from "../src/flow/formal-run.ts";
|
||||
import { containerName, freshDir, runPaths, workspaceRepo } from "../src/flow/paths.ts";
|
||||
import { recover } from "../src/flow/recover.ts";
|
||||
import { diagnosisRun, diagnosisSettle } from "../src/flow/steps/diagnosis.ts";
|
||||
import type { Study } from "../src/flow/study.ts";
|
||||
import { Ledger } from "../src/ledger/index.ts";
|
||||
import { Board } from "../src/materials/index.ts";
|
||||
import { runBalance, settleDiagnosis, type SettleDiagnosisInput } from "../src/rules/index.ts";
|
||||
import { Runner } from "../src/runner/index.ts";
|
||||
import { diagnosisRun as diagnosisDefinition, diagnosisUnversionedNames, loadTask, prepareDiagnosisRun, type DiagnosisJudged } from "../src/tasks/index.ts";
|
||||
import { checkoutVersion, createRepository, saveVersion } from "../src/workspace/index.ts";
|
||||
import { fullConfig } from "./config-helpers.ts";
|
||||
import { judged, r, Writer } from "./materials-helpers.ts";
|
||||
|
||||
// 诊断运行在 flow 里的几处，用真实 Docker 测，不花模型的钱（任务书 18 第 1.6 节）：
|
||||
// 选中的实验拿掉一个后跑的只是剩下的实验、逐个实验判分和结算；诊断运行恢复只核对、不重做、只扣一次。
|
||||
// 镜像一律 inforsi-test-flow-dx 开头的固定名字，只动自己前缀的。判分的预期值是手算的相切布局：26 个半径 1/64 的圆排成一行，
|
||||
// 圆心间距 1/32，半径和 26/64 = 0.40625（同 tests/tasks-diagnosis.test.ts，都是 2 的幂有理数，没有舍入）。
|
||||
|
||||
const TASK_DIR = join(import.meta.dirname, "..", "tasks", "circle_packing");
|
||||
const BASELINE_SUM = 26 / 64;
|
||||
|
||||
function docker(args: string[]): { status: number | null; stderr: string } {
|
||||
return spawnSync("docker", args, { encoding: "utf8" });
|
||||
}
|
||||
|
||||
function removeImage(tag: string): void {
|
||||
const removed = docker(["rmi", tag]);
|
||||
if (removed.status !== 0 && !removed.stderr.includes("No such image")) throw new Error(removed.stderr);
|
||||
}
|
||||
|
||||
function tempDir(t: TestContext): string {
|
||||
const dir = mkdtempSync(join(tmpdir(), "inforsi-flow-dx-"));
|
||||
t.after(() => rmSync(dir, { recursive: true, force: true }));
|
||||
return dir;
|
||||
}
|
||||
|
||||
/** 诊断程序：d1 原样交回基线的原始答案；别的实验编号被调用就崩（用来证明拿掉的实验没被调用）。 */
|
||||
const PROGRAM = `import json
|
||||
|
||||
|
||||
def run_experiment(experiment_id, baseline_dir):
|
||||
if experiment_id == "d1":
|
||||
with open(baseline_dir + "/solution.json") as handle:
|
||||
solution = json.load(handle)
|
||||
return solution["centers"], solution["radii"], solution["reported_sum"]
|
||||
raise RuntimeError("不该被调用的实验 " + experiment_id)
|
||||
`;
|
||||
|
||||
/** 真实任务镜像、账本、裸仓库；第 0 轮基线已跑完（运行目录里放着相切布局的原始答案，半径和 0.40625），诊断程序已存成版本。 */
|
||||
async function setup(t: TestContext, suffix: string, maxRepairs: number): Promise<{ s: Study; w: Writer; version: string }> {
|
||||
const dir = tempDir(t);
|
||||
const id = `test-flow-dx-${suffix}`;
|
||||
removeImage(`inforsi-${id}`);
|
||||
t.after(() => removeImage(`inforsi-${id}`));
|
||||
const task = loadTask(TASK_DIR);
|
||||
const image = await new Runner(null).buildImage(task.dir, `inforsi-${id}`);
|
||||
const ledger = Ledger.open(join(dir, "ledger.sqlite"));
|
||||
t.after(() => ledger.close());
|
||||
const config = { ...fullConfig(), studyId: id, maxRepairs };
|
||||
if (config.flow.preset !== "full") throw new Error("测试配置应是完整流程");
|
||||
config.flow = { ...config.flow, diagnosis: { enabled: true, perResearcher: 2, maxExperiments: 4 }, ablation: { enabled: false } };
|
||||
const s = { dir, task, runner: new Runner(null), ledger, stop: new AbortController().signal,
|
||||
frozen: { config: checkedFull(config), steps: stepList(config), facts: task.facts, prompts: null, image, seed: 1 } } as unknown as Study;
|
||||
const w = new Writer(ledger);
|
||||
ledger.append("code", { object: "base", version: "v-base", locations: null });
|
||||
w.run("base-1", "baseline", 0, null, judged(BASELINE_SUM, "main"));
|
||||
const baseRun = runPaths(dir, "base-1").run;
|
||||
mkdirSync(baseRun, { recursive: true });
|
||||
writeFileSync(join(baseRun, "solution.json"), JSON.stringify({
|
||||
centers: Array.from({ length: 26 }, (_, i) => [7 / 64 + i / 32, 0.5]),
|
||||
radii: Array(26).fill(1 / 64),
|
||||
reported_sum: 99,
|
||||
}));
|
||||
createRepository(workspaceRepo(dir));
|
||||
const program = join(dir, "program");
|
||||
mkdirSync(program);
|
||||
writeFileSync(join(program, "initial.py"), PROGRAM);
|
||||
const saved = saveVersion(workspaceRepo(dir), program, diagnosisUnversionedNames(task));
|
||||
if (!saved.saved) throw new Error(`诊断程序存不成版本：${saved.reason}`);
|
||||
return { s, w, version: saved.version };
|
||||
}
|
||||
|
||||
const RANGES = {
|
||||
a1: [r(0.5, 0.6, 0.7), r(0.3, 0.4, 0.45)],
|
||||
a2: [r(0.2, 0.3, 0.4), r(0.3, 0.4, 0.45)],
|
||||
} as const;
|
||||
|
||||
/** 本轮诊断写到"选中 d1、d2，核验只有 d1 通过"为止：诊断解释与实验、判重合并、诊断预测、选实验、诊断程序、核验。 */
|
||||
function selectedAndVerified({ w, version }: { w: Writer; version: string }): void {
|
||||
const experiment = (name: string) => ({ name, spot: `${name}的位置`, body: `${name}的做法`, metrics: [] });
|
||||
w.open("r1-diagnosisProposal", "diagnosisProposal", 1, ["a1", "a2"]);
|
||||
for (const [author, name] of [["a1", "实验甲"], ["a2", "实验乙"]] as const) {
|
||||
w.sub("r1-diagnosisProposal", author, "diagnosisProposal", null, null, { explanations: [{ body: `${author}的解释` }], experiments: [experiment(name)] });
|
||||
}
|
||||
w.close("r1-diagnosisProposal");
|
||||
w.decision("diagnosisDedup", 1, null, {
|
||||
experiments: [{ experiment: "d1", sources: [{ author: "a1", index: 0 }] }, { experiment: "d2", sources: [{ author: "a2", index: 0 }] }],
|
||||
});
|
||||
w.open("r1-diagnosisPrediction", "diagnosisPrediction", 1, ["a1", "a2"]);
|
||||
for (const author of ["a1", "a2"] as const) {
|
||||
w.sub("r1-diagnosisPrediction", author, "diagnosisPrediction", null, null, {
|
||||
predictions: ["d1", "d2"].map((d, i) => ({ explanation: `r1-${author}-dx1`, experiment: d, range: RANGES[author][i]! })),
|
||||
body: "理由",
|
||||
});
|
||||
}
|
||||
w.close("r1-diagnosisPrediction");
|
||||
w.decision("selectDiagnosis", 1, null, { diagnose: true, chosen: ["d1", "d2"] });
|
||||
w.l.append("code", { object: "r1-diagnosis", version, locations: [] });
|
||||
w.open("r1-diagnosisVerification-1", "diagnosisVerification", 1, ["verifier-r1-diagnosis"]);
|
||||
w.sub("r1-diagnosisVerification-1", "verifier-r1-diagnosis", "diagnosisVerification", "r1-diagnosis", null, {
|
||||
experiments: [
|
||||
{ experiment: "d1", followed: true, nothingElse: true, where: "initial.py:4" },
|
||||
{ experiment: "d2", followed: false, nothingElse: true, where: "没找到" },
|
||||
],
|
||||
body: "无",
|
||||
});
|
||||
w.close("r1-diagnosisVerification-1");
|
||||
}
|
||||
|
||||
test("诊断运行：核验没通过又没有修复次数的实验拿掉，只跑剩下的实验；逐个实验判分，结算里拿掉的记证据不足", { timeout: 180_000 }, async (t) => {
|
||||
const fixture = await setup(t, "settle", 0);
|
||||
const { s } = fixture;
|
||||
selectedAndVerified(fixture);
|
||||
t.after(() => { docker(["rm", "-f", containerName(s.frozen.config.studyId, "r1-diagnosis-1")]); });
|
||||
assert.equal(await diagnosisRun(s, 1, "r1-diagnosis-1"), "done");
|
||||
const records = s.ledger.records();
|
||||
const planned = records.find((x) => x.kind === "runPlanned" && x.content.purpose === "diagnosis")!;
|
||||
if (planned.kind !== "runPlanned") throw new Error("缺登记");
|
||||
assert.deepEqual({ round: planned.content.round, object: planned.content.object }, { round: 1, object: "r1-diagnosis" });
|
||||
// 只调 d1，总时限 120 秒由它一个实验独享。
|
||||
const { command } = planned.content.definition;
|
||||
assert.deepEqual(command.slice(command.indexOf("--experiments")), ["--experiments", "d1"]);
|
||||
assert.equal(command[command.indexOf("--seconds-each") + 1], "120");
|
||||
assert.equal(records.filter((x) => x.kind === "runStarted" && x.content.run === "r1-diagnosis-1").length, 1);
|
||||
const result = records.findLast((x) => x.kind === "runResult")!;
|
||||
if (result.kind !== "runResult") throw new Error("缺结果");
|
||||
assert.equal(result.content.process.ending, "exited");
|
||||
assert.equal(result.content.judged, null);
|
||||
assert.deepEqual(runBalance(13, new Board(records).runStates()), { charged: 2, reserved: 0, balance: 11 });
|
||||
|
||||
await diagnosisSettle(s, 1);
|
||||
const settled = s.ledger.records().find((x) => x.kind === "diagnosisSettlement")!;
|
||||
if (settled.kind !== "diagnosisSettlement") throw new Error("缺结算");
|
||||
const input: SettleDiagnosisInput = {
|
||||
baseline: BASELINE_SUM,
|
||||
direction: "maximize",
|
||||
epsilon: s.frozen.facts.epsilon,
|
||||
lambda: 0.1,
|
||||
experiments: [
|
||||
{ experiment: "d1", outcome: { kind: "valid", score: BASELINE_SUM } },
|
||||
{ experiment: "d2", outcome: { kind: "insufficient" } },
|
||||
],
|
||||
explanations: (["a1", "a2"] as const).map((author) => ({
|
||||
id: `r1-${author}-dx1`,
|
||||
author,
|
||||
experiments: ["d1", "d2"].map((experiment, i) => ({ experiment, range: RANGES[author][i]! })),
|
||||
})),
|
||||
};
|
||||
const { judged: j, dropped, run, ...rest } = settled.content.result as unknown as { judged: DiagnosisJudged; dropped: Array<{ experiment: string; reason: string }>; run: string };
|
||||
assert.equal(run, "r1-diagnosis-1");
|
||||
assert.deepEqual(rest, { input, output: settleDiagnosis(input) });
|
||||
// 判分只有 d1 一项：拿掉的 d2 没跑、没判。
|
||||
assert.equal(j.experiments.length, 1);
|
||||
const d1 = j.experiments[0]!;
|
||||
assert.deepEqual([d1.id, d1.ending, d1.verdict, d1.score], ["d1", "finished", "valid", { status: "measured", value: BASELINE_SUM }]);
|
||||
assert.deepEqual(dropped.map((d) => d.experiment), ["d2"]);
|
||||
assert.match(dropped[0]!.reason, /第 1 次核验仍不通过，修复次数上限 0 已用完：没照写法做；没找到/);
|
||||
// 结算写进去以后诊断运行公开；每轮最多一条结算。
|
||||
assert.ok(new Board(s.ledger.records()).runs().filter((x) => x.plan.purpose === "diagnosis").every((x) => new Board(s.ledger.records()).runPublic(x)));
|
||||
assert.throws(() => s.ledger.append("diagnosisSettlement", { round: 1, result: null }));
|
||||
});
|
||||
|
||||
/** 手工把一次诊断运行做到"登记了"：和 diagnosisRun 同样的目录准备，但不启动（启动由各个测试自己决定）。 */
|
||||
function planned(s: Study, version: string): { run: string; name: string; dir: string } {
|
||||
const run = "r1-diagnosis-1";
|
||||
const paths = runPaths(s.dir, run);
|
||||
const definition = diagnosisDefinition(s.task, s.frozen.image, String(s.frozen.seed), ["d1"]);
|
||||
s.ledger.append("runPlanned", { run, purpose: "diagnosis", round: 1, object: "r1-diagnosis", definition });
|
||||
freshDir(paths.base);
|
||||
checkoutVersion(workspaceRepo(s.dir), version, freshDir(paths.code));
|
||||
prepareDiagnosisRun(s.task, paths.code, runPaths(s.dir, "base-1").run, freshDir(paths.run), [{ id: "d1", metrics: [] }]);
|
||||
return { run, name: containerName(s.frozen.config.studyId, run), dir: paths.run };
|
||||
}
|
||||
|
||||
test("诊断运行恢复：容器已启动、账本没写确认启动和结果，接上原容器不重跑，只扣一次，运行记录不带判分，判分从运行目录重做", { timeout: 180_000 }, async (t) => {
|
||||
const { s, version } = await setup(t, "recover", 0);
|
||||
const { run, name, dir } = planned(s, version);
|
||||
t.after(() => { docker(["rm", "-f", name]); });
|
||||
assert.deepEqual(await s.runner.start(name, diagnosisDefinition(s.task, s.frozen.image, String(s.frozen.seed), ["d1"]), dir), { started: true });
|
||||
// 中断点：容器已启动，账本只有登记。恢复不调用 diagnosisRun，而是找到原容器等它。
|
||||
assert.equal(await recover(s), null);
|
||||
const records = s.ledger.records();
|
||||
assert.deepEqual(["runPlanned", "runStarted", "runResult"].map((kind) => records.filter((x) => x.kind === kind && (x.content as { run?: string }).run === run).length), [1, 1, 1]);
|
||||
const result = records.findLast((x) => x.kind === "runResult")!;
|
||||
if (result.kind !== "runResult") throw new Error("缺结果");
|
||||
assert.equal(result.content.run, run);
|
||||
assert.equal(result.content.process.ending, "exited");
|
||||
assert.equal(result.content.judged, null);
|
||||
assert.deepEqual(runBalance(13, new Board(records).runStates()), { charged: 2, reserved: 0, balance: 11 });
|
||||
assert.equal(await s.runner.status(name), "absent");
|
||||
assert.equal(await recover(s), null);
|
||||
assert.equal(s.ledger.records().length, records.length);
|
||||
// 恢复后的判分：从运行目录重判，逐个实验单独判分。
|
||||
const judgedRun = await judgeDiagnosisRun(s, run, result.content.process, [{ id: "d1", metrics: [] }]);
|
||||
assert.equal(judgedRun.experiments.length, 1);
|
||||
const d1 = judgedRun.experiments[0]!;
|
||||
assert.deepEqual([d1.id, d1.ending, d1.verdict, d1.score], ["d1", "finished", "valid", { status: "measured", value: BASELINE_SUM }]);
|
||||
});
|
||||
|
||||
test("诊断运行恢复：登记了、容器不存在，确认没启动：写结果没启动、退还名额，不补启动", { timeout: 120_000 }, async (t) => {
|
||||
const { s, version } = await setup(t, "notstarted", 0);
|
||||
planned(s, version);
|
||||
assert.equal(await recover(s), null);
|
||||
const records = s.ledger.records();
|
||||
const result = records.findLast((x) => x.kind === "runResult")!;
|
||||
if (result.kind !== "runResult") throw new Error("缺结果");
|
||||
assert.equal(result.content.process.ending, "notStarted");
|
||||
assert.equal(result.content.judged, null);
|
||||
assert.equal(records.filter((x) => x.kind === "runStarted" && x.content.run === "r1-diagnosis-1").length, 0);
|
||||
assert.deepEqual(runBalance(13, new Board(records).runStates()), { charged: 1, reserved: 0, balance: 12 });
|
||||
});
|
||||
@@ -0,0 +1,185 @@
|
||||
import assert from "node:assert/strict";
|
||||
import { appendFileSync, cpSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
|
||||
import { join } from "node:path";
|
||||
import { test, type TestContext } from "node:test";
|
||||
import { stepList } from "../src/config/index.ts";
|
||||
import { fullConfig as checkedFull } from "../src/flow/frozen.ts";
|
||||
import { runPaths, workspaceRepo } from "../src/flow/paths.ts";
|
||||
import { groupSameExperiment, selectDiagnosisExperiments } from "../src/flow/steps/diagnosis.ts";
|
||||
import { prepareSubmission, taskArtifact, taskCode } from "../src/flow/task-code.ts";
|
||||
import type { Study } from "../src/flow/study.ts";
|
||||
import { Ledger } from "../src/ledger/index.ts";
|
||||
import { runDiagnosisRule, type DiagnosisRuleInput } from "../src/rules/index.ts";
|
||||
import type { ForwardCode } from "../src/submissions/index.ts";
|
||||
import { checkTaskFacts, diagnosisUnversionedNames, loadTask, placeBaselineArtifact } from "../src/tasks/index.ts";
|
||||
import type { JevJudgment } from "../src/vocab/index.ts";
|
||||
import { compareVersions, createRepository, saveVersion } from "../src/workspace/index.ts";
|
||||
import { fullConfig } from "./config-helpers.ts";
|
||||
import { tempDir } from "./ledger-helpers.ts";
|
||||
import { judged, r, Writer } from "./materials-helpers.ts";
|
||||
|
||||
// 诊断里程序自己做的两步（任务书 18 第 1.6 节）：判重合并的分组、选诊断实验；加上诊断实现、修复、核验的取材和存版本接线。判重合并的 Jev 问法由 jev 的测试管；
|
||||
// 诊断运行和结算要真实 Docker，在 flow-diagnosis-docker.test.ts。预期值照条文手拼，输出用 rules 的函数当对照
|
||||
// （同选格测试的写法，rules 自己有独立预期值的测试）。
|
||||
|
||||
const facts = checkTaskFacts(JSON.parse(readFileSync(join(import.meta.dirname, "..", "tasks", "circle_packing", "task.json"), "utf8")));
|
||||
|
||||
const yes = (probability: number): JevJudgment => ({ question: "diagnosisDedup", probability, threshold: 0.3, yes: probability > 0.3 }) as unknown as JevJudgment;
|
||||
const item = (author: string, index: number) => ({ author, index, label: `${author}.e${index + 1}` });
|
||||
|
||||
test("诊断实验判重合并：判为相同的连成一组，编号 d1… 按组里排最前的实验排，组内按作者、第几个排", () => {
|
||||
const items = [item("researcher-1", 0), item("researcher-1", 1), item("researcher-2", 0), item("researcher-3", 0)];
|
||||
const pairs = [
|
||||
{ a: "researcher-1.e1", b: "researcher-2.e1", judgment: yes(0.9) },
|
||||
{ a: "researcher-1.e2", b: "researcher-3.e1", judgment: yes(0.1) },
|
||||
{ a: "researcher-2.e1", b: "researcher-3.e1", judgment: yes(0.8) },
|
||||
];
|
||||
// researcher-1.e1、researcher-2.e1、researcher-3.e1 连成一组（A=B，B=C）；researcher-1.e2 自成一个。
|
||||
assert.deepEqual(groupSameExperiment(items, pairs), {
|
||||
experiments: [
|
||||
{ experiment: "d1", sources: [{ author: "researcher-1", index: 0 }, { author: "researcher-2", index: 0 }, { author: "researcher-3", index: 0 }] },
|
||||
{ experiment: "d2", sources: [{ author: "researcher-1", index: 1 }] },
|
||||
],
|
||||
});
|
||||
});
|
||||
|
||||
test("诊断实验判重合并：没有一对判相同时，每个实验各成一个", () => {
|
||||
const items = [item("researcher-1", 0), item("researcher-2", 0)];
|
||||
const pairs = [{ a: "researcher-1.e1", b: "researcher-2.e1", judgment: yes(0.1) }];
|
||||
assert.deepEqual(groupSameExperiment(items, pairs).experiments.map((e) => e.experiment), ["d1", "d2"]);
|
||||
});
|
||||
|
||||
/** 基线 0.96；a1、a2 各一条诊断解释，对合并后的 d1、d2 各给三个数；余额：13 次里基线用了 1 次。 */
|
||||
function study(t: TestContext, ranges: Record<string, ReturnType<typeof r>[]>): { s: Study; input: DiagnosisRuleInput } {
|
||||
const ledger = Ledger.open(join(tempDir(t), "ledger.db"));
|
||||
t.after(() => ledger.close());
|
||||
const w = new Writer(ledger);
|
||||
ledger.append("code", { object: "base", version: "v-base", locations: null });
|
||||
w.run("base-1", "baseline", 0, null, judged(0.96, "main"));
|
||||
w.open("r1-diagnosisProposal", "diagnosisProposal", 1, ["a1", "a2"]);
|
||||
const experiment = (name: string) => ({ name, spot: `${name}的位置`, body: `${name}的做法`, metrics: [] });
|
||||
for (const [author, name] of [["a1", "实验甲"], ["a2", "实验乙"]] as const) {
|
||||
w.sub("r1-diagnosisProposal", author, "diagnosisProposal", null, null, { explanations: [{ body: `${author}的解释` }], experiments: [experiment(name)] });
|
||||
}
|
||||
w.close("r1-diagnosisProposal");
|
||||
w.decision("diagnosisDedup", 1, null, {
|
||||
experiments: [{ experiment: "d1", sources: [{ author: "a1", index: 0 }] }, { experiment: "d2", sources: [{ author: "a2", index: 0 }] }],
|
||||
});
|
||||
w.open("r1-diagnosisPrediction", "diagnosisPrediction", 1, ["a1", "a2"]);
|
||||
for (const author of ["a1", "a2"]) {
|
||||
const mine = ranges[author]!;
|
||||
w.sub("r1-diagnosisPrediction", author, "diagnosisPrediction", null, null, {
|
||||
predictions: ["d1", "d2"].map((d, i) => ({ explanation: `r1-${author}-dx1`, experiment: d, range: mine[i]! })),
|
||||
body: "理由",
|
||||
});
|
||||
}
|
||||
w.close("r1-diagnosisPrediction");
|
||||
const config = fullConfig();
|
||||
if (config.flow.preset !== "full") throw new Error("测试配置应是完整流程");
|
||||
config.flow = { ...config.flow, diagnosis: { enabled: true, perResearcher: 2, maxExperiments: 4 }, ablation: { enabled: false } };
|
||||
const frozen = { config: checkedFull(config), steps: stepList(config), facts, prompts: null, image: "", seed: 1 };
|
||||
const input: DiagnosisRuleInput = {
|
||||
baseline: 0.96,
|
||||
direction: "maximize",
|
||||
lambda: 0.1,
|
||||
experiments: ["d1", "d2"],
|
||||
explanations: ["a1", "a2"].map((author) => ({
|
||||
id: `r1-${author}-dx1`,
|
||||
author,
|
||||
experiments: ["d1", "d2"].map((experiment, i) => ({ experiment, range: ranges[author]![i]! })),
|
||||
})),
|
||||
balance: 12,
|
||||
batchCap: 2,
|
||||
maxExperiments: 4,
|
||||
u: 0.5,
|
||||
decay: 0.5,
|
||||
};
|
||||
return { s: { ledger, frozen } as unknown as Study, input };
|
||||
}
|
||||
|
||||
test("选诊断实验：输入照 DiagnosisRuleInput 拼（基线、每条解释对每个实验的三个数、余额从本轮之前的运行算），规则取步骤清单，种子 null", (t) => {
|
||||
// 两人对 d1 预测相反（一个说涨到 1.3，一个说跌到 0.9），对 d2 都说涨不了多少。
|
||||
const { s, input } = study(t, {
|
||||
a1: [r(1.2, 1.3, 1.4), r(0.95, 0.97, 0.99)],
|
||||
a2: [r(0.8, 0.9, 1.0), r(0.95, 0.97, 0.99)],
|
||||
});
|
||||
selectDiagnosisExperiments(s, 1);
|
||||
const record = s.ledger.records().find((x) => x.kind === "decision" && x.content.decision === "selectDiagnosis")!;
|
||||
const { rule: used, ...output } = runDiagnosisRule("valueMinusCost", input);
|
||||
assert.deepEqual(record.content, { decision: "selectDiagnosis", round: 1, rule: "valueMinusCost", ruleVersion: used.version, params: output.params, seed: null, input, output });
|
||||
// 分歧大的实验信息价值为正：本轮做诊断，d1 在选中的组合里。
|
||||
assert.equal(output.diagnose, true);
|
||||
assert.ok(output.chosen.includes("d1"));
|
||||
});
|
||||
|
||||
/** 诊断实现的代码接线：真实裸仓库里存了基线；工作目录是基线副本加一处改动，再放进只读的原始产物。 */
|
||||
function codeStudy(t: TestContext): { s: Study; work: string; baseline: string } {
|
||||
const dir = tempDir(t);
|
||||
const ledger = Ledger.open(join(dir, "ledger.db"));
|
||||
t.after(() => ledger.close());
|
||||
const task = loadTask(join(import.meta.dirname, "..", "tasks", "circle_packing"));
|
||||
const repo = workspaceRepo(dir);
|
||||
createRepository(repo);
|
||||
const names = diagnosisUnversionedNames(task);
|
||||
const saved = saveVersion(repo, task.baselineDir, names);
|
||||
if (!saved.saved) throw new Error("基线存不成版本");
|
||||
const w = new Writer(ledger);
|
||||
ledger.append("code", { object: "base", version: saved.version, locations: null });
|
||||
w.run("base-1", "baseline", 0, null, judged(0.96, "main"));
|
||||
const baseRun = runPaths(dir, "base-1").run;
|
||||
mkdirSync(baseRun, { recursive: true });
|
||||
writeFileSync(join(baseRun, "solution.json"), "{}");
|
||||
w.open("r1-diagnosisProposal", "diagnosisProposal", 1, ["a1"]);
|
||||
const metric = { id: "note", meaning: "记录", unit: "无", direction: "maximize" as const, howMeasured: "数一数", missing: "recordMissing" as const };
|
||||
w.sub("r1-diagnosisProposal", "a1", "diagnosisProposal", null, null, { explanations: [{ body: "解释" }], experiments: [{ name: "实验甲", spot: "第一步", body: "换个做法", metrics: [metric] }] });
|
||||
w.close("r1-diagnosisProposal");
|
||||
w.decision("diagnosisDedup", 1, null, { experiments: [{ experiment: "d1", sources: [{ author: "a1", index: 0 }] }] });
|
||||
w.decision("selectDiagnosis", 1, null, { diagnose: true, chosen: ["d1"] });
|
||||
const work = join(dir, "work");
|
||||
cpSync(task.baselineDir, work, { recursive: true });
|
||||
appendFileSync(join(work, "initial.py"), "\n\ndef run_experiment(experiment_id, baseline_dir):\n return [], [], 0.0\n");
|
||||
placeBaselineArtifact(task, baseRun, work);
|
||||
const frozen = { config: checkedFull(fullConfig()), steps: stepList(fullConfig()), facts: task.facts, prompts: null, image: "", seed: 1 };
|
||||
return { s: { dir, task, ledger, frozen } as unknown as Study, work, baseline: saved.version };
|
||||
}
|
||||
|
||||
const dxTask = (kind: "diagnosisImplementation" | "diagnosisRepair" | "diagnosisVerification") =>
|
||||
({ actor: "x", kind, object: "r1-diagnosis", stageId: `r1-${kind}-1` });
|
||||
|
||||
test("诊断实现的接线：存版本时跳过只读的原始产物目录；Jev 对照选中实验的原文和全名指标；工作目录放进原始产物", (t) => {
|
||||
const { s, work, baseline } = codeStudy(t);
|
||||
const w = new Writer(s.ledger);
|
||||
w.open("r1-diagnosisImplementation-1", "diagnosisImplementation", 1, ["x"]);
|
||||
const content = { locations: [{ experiment: "d1", locations: ["initial.py:90-92"] }], concerns: "无", body: "无" };
|
||||
const { known, code } = prepareSubmission(s, dxTask("diagnosisImplementation"), content, 0, work);
|
||||
const saved = (known as { code: ForwardCode }).code;
|
||||
if (!saved.saved) throw new Error("应存成版本");
|
||||
assert.deepEqual(saved.protectedFiles, []);
|
||||
assert.ok(saved.hunks.length > 0);
|
||||
assert.ok(saved.hunks.every((h) => h.path === "initial.py"));
|
||||
assert.match(saved.protocol.body, /### d1 实验甲\n\n动哪一处：第一步\n\n换个做法/);
|
||||
assert.deepEqual(saved.protocol.metrics.map((m) => m.id), ["d1.note"]);
|
||||
assert.deepEqual(code, { object: "r1-diagnosis", version: code!.version, locations: content.locations });
|
||||
// 工作目录里放着的原始产物没有进版本：当前基线到诊断程序的改动里只有 initial.py。
|
||||
const changes = compareVersions(workspaceRepo(s.dir), baseline, code!.version);
|
||||
assert.deepEqual(changes.files.map((f) => f.path), ["initial.py"]);
|
||||
});
|
||||
|
||||
test("诊断实现、修复、核验取材：实现从当前基线出发，修复从上次提交的诊断程序出发，核验和修复的差异都是当前基线到诊断程序；原始产物只给实现和修复", (t) => {
|
||||
const { s, work, baseline } = codeStudy(t);
|
||||
const w = new Writer(s.ledger);
|
||||
const content = { locations: [{ experiment: "d1", locations: ["initial.py:90-92"] }], concerns: "无", body: "无" };
|
||||
w.open("r1-diagnosisImplementation-1", "diagnosisImplementation", 1, ["x"]);
|
||||
w.open("r1-diagnosisRepair-1", "diagnosisImplementation", 1, ["x"]);
|
||||
w.open("r1-diagnosisVerification-1", "diagnosisVerification", 1, ["v"]);
|
||||
assert.deepEqual(taskCode(s, dxTask("diagnosisImplementation")), { changes: null, startVersion: baseline });
|
||||
const { code } = prepareSubmission(s, dxTask("diagnosisImplementation"), content, 0, work);
|
||||
s.ledger.append("code", code!);
|
||||
const changes = compareVersions(workspaceRepo(s.dir), baseline, code!.version);
|
||||
assert.deepEqual(taskCode(s, dxTask("diagnosisRepair")), { changes, startVersion: code!.version });
|
||||
assert.deepEqual(taskCode(s, dxTask("diagnosisVerification")), { changes, startVersion: null });
|
||||
const baseRun = runPaths(s.dir, "base-1").run;
|
||||
assert.equal(taskArtifact(s, dxTask("diagnosisImplementation")), baseRun);
|
||||
assert.equal(taskArtifact(s, dxTask("diagnosisRepair")), baseRun);
|
||||
assert.equal(taskArtifact(s, dxTask("diagnosisVerification")), null);
|
||||
});
|
||||
+217
-13
@@ -4,11 +4,11 @@ import { join } from "node:path";
|
||||
import { test } from "node:test";
|
||||
import { stepList, type ResearchConfig } from "../src/config/index.ts";
|
||||
import type { AppendKind, Content, LedgerRecord } from "../src/ledger/index.ts";
|
||||
import { Board, forwardNotRunEvent, type StudyRecord } from "../src/materials/index.ts";
|
||||
import { Board, diagnosisNotRunEvent, forwardNotRunEvent, type StudyRecord } from "../src/materials/index.ts";
|
||||
import { runBalance } from "../src/rules/index.ts";
|
||||
import { definition, exited } from "./ledger-helpers.ts";
|
||||
import { judged } from "./materials-helpers.ts";
|
||||
import { nextStep } from "../src/flow/next-step.ts";
|
||||
import { diagnosisProgram, nextStep } from "../src/flow/next-step.ts";
|
||||
import type { Json, Submission, SubmissionKind } from "../src/vocab/index.ts";
|
||||
import { fullConfig } from "./config-helpers.ts";
|
||||
|
||||
@@ -217,10 +217,30 @@ function assigned(book: Book): Book {
|
||||
return book.add("decision", { decision: "entries", round: 0, rule: "assignEntries", ruleVersion: 1, params: null, seed: null, input: null, output });
|
||||
}
|
||||
|
||||
test("认领关了：分配；分配做完：开第 1 轮的提案阶段，名单是全部研究者", () => {
|
||||
const DIAGNOSIS_STEPS = [
|
||||
"diagnosisProposal", "diagnosisDedupe", "diagnosisPrediction", "selectDiagnosis",
|
||||
"diagnosisImplementation", "diagnosisRun", "diagnosisSettle",
|
||||
];
|
||||
|
||||
/** 诊断关着（fullConfig 的写法）：第 round 轮开头的诊断七步各记一条按配置跳过（方法.md 第十一部分"步骤开关"）。 */
|
||||
function diagnosisSkipped(book: Book, round: number): Book {
|
||||
for (const step of DIAGNOSIS_STEPS) {
|
||||
assert.deepEqual(nextStep(book.records), { do: "skip", step: `r${round}-${step}`, reason: "配置关掉了诊断" });
|
||||
book.add("event", { event: "skippedByConfig", reason: "配置关掉了诊断", objects: [`r${round}-${step}`] });
|
||||
}
|
||||
return book;
|
||||
}
|
||||
|
||||
/** 第零轮做完、入口分配好，第 1 轮的诊断七步也记完跳过。 */
|
||||
function inRoundOne(book: Book): Book {
|
||||
return diagnosisSkipped(assigned(book), 1);
|
||||
}
|
||||
|
||||
test("认领关了：分配；分配做完：诊断关着，七步各记按配置跳过，再开第 1 轮的提案阶段，名单是全部研究者", () => {
|
||||
const book = afterRoundZero(fullConfig());
|
||||
assert.deepEqual(nextStep(book.records), { do: "assignEntries" });
|
||||
assigned(book);
|
||||
diagnosisSkipped(book, 1);
|
||||
assert.deepEqual(nextStep(book.records), { do: "openStage", stageId: "r1-proposal", stage: "proposal", round: 1, roster: R });
|
||||
});
|
||||
|
||||
@@ -246,7 +266,7 @@ const each = (kind: SubmissionKind, objects: string[]) => R.flatMap((actor) => o
|
||||
|
||||
/** 第 1 轮写到 last 这一步做完为止。 */
|
||||
function roundOne(last: "proposal" | "dedupe" | "firstPrediction" | "critique" | "response" | "freezeProtocol" | "secondPrediction" | "selectBatch"): Book {
|
||||
const book = assigned(afterRoundZero(fullConfig()));
|
||||
const book = inRoundOne(afterRoundZero(fullConfig()));
|
||||
const order = ["proposal", "dedupe", "firstPrediction", "critique", "response", "freezeProtocol", "secondPrediction", "selectBatch"];
|
||||
const steps: Record<string, () => void> = {
|
||||
proposal: () => stage(book, "r1-proposal", "proposal", [["researcher-1", "proposal", null, P1], ["researcher-2", "proposal", null, P2], ["researcher-3", "proposal", null, P3]]),
|
||||
@@ -278,7 +298,7 @@ function roundOne(last: "proposal" | "dedupe" | "firstPrediction" | "critique" |
|
||||
}
|
||||
|
||||
test("提案阶段一人交了两条、两人没交：只派没交的两人，提案没有对象", () => {
|
||||
const book = assigned(afterRoundZero(fullConfig())).open("r1-proposal", "proposal", 1, R).sub("r1-proposal", "researcher-1", "proposal", null, P1);
|
||||
const book = inRoundOne(afterRoundZero(fullConfig())).open("r1-proposal", "proposal", 1, R).sub("r1-proposal", "researcher-1", "proposal", null, P1);
|
||||
book.sub("r1-proposal", "researcher-1", "proposal", null, "r1-researcher-1-2");
|
||||
const tasks = ["researcher-2", "researcher-3"].map((actor) => ({ actor, kind: "proposal", object: null }));
|
||||
assert.deepEqual(nextStep(book.records), { do: "dispatch", stageId: "r1-proposal", tasks });
|
||||
@@ -350,7 +370,7 @@ test("关掉第一次预测：第 1 轮写一条按配置跳过（对象 r1-firs
|
||||
const config = fullConfig();
|
||||
if (config.flow.preset !== "full") throw new Error("测试配置应是完整流程");
|
||||
config.flow.firstPrediction = false;
|
||||
const book = assigned(afterRoundZero(config));
|
||||
const book = inRoundOne(afterRoundZero(config));
|
||||
stage(book, "r1-proposal", "proposal", [["researcher-1", "proposal", null, P1], ["researcher-2", "proposal", null, P2], ["researcher-3", "proposal", null, P3]]);
|
||||
const output = { fresh: [{ candidate: "r1-c1", proposals: [P1] }, { candidate: "r1-c2", proposals: [P2, P3] }], carried: [] };
|
||||
book.add("decision", { decision: "dedup", round: 1, rule: "jevSameChange", ruleVersion: 1, params: null, seed: null, input: null, output });
|
||||
@@ -648,7 +668,8 @@ test("结算和押的因子写完：开收尾；收尾关闭后解封本轮选
|
||||
book.add("unseal", { candidate: "r1-c1", seals: [1] });
|
||||
assert.deepEqual(nextStep(book.records), { do: "unsealChosen", round: 1 });
|
||||
book.add("unseal", { candidate: "r1-c2", seals: [2] });
|
||||
// 余额还有（已扣 5 次）：开第 2 轮。
|
||||
// 余额还有（已扣 5 次）：开第 2 轮，先是诊断七步的按配置跳过。
|
||||
diagnosisSkipped(book, 2);
|
||||
assert.deepEqual(nextStep(book.records), { do: "openStage", stageId: "r2-proposal", stage: "proposal", round: 2, roster: R });
|
||||
});
|
||||
|
||||
@@ -679,7 +700,7 @@ test("余额为 0：不开下一轮，走到导出停下，原因写明", () =>
|
||||
assert.deepEqual(nextStep(book.records), { do: "halt", reason: "导出还没写", objects: [] });
|
||||
});
|
||||
|
||||
test("配置关掉消融、又采纳了多因子改法：先记按配置跳过，再在解释停下", () => {
|
||||
test("配置关掉消融、又采纳了多因子改法：⑬–⑰ 整段（含解释）各记按配置跳过，再开收尾", () => {
|
||||
const book = firstVerification(batchTwo(1), true, true);
|
||||
for (const c of C) recordForward(book, c, 1, "valid");
|
||||
book.add("decision", { decision: "adopt", round: 1, rule: "adopt", ruleVersion: 1, params: null, seed: null, input: null, output: { adopted: ADOPTED } });
|
||||
@@ -693,9 +714,192 @@ test("配置关掉消融、又采纳了多因子改法：先记按配置跳过
|
||||
study.config = config as unknown as Json;
|
||||
study.steps = stepList(config) as unknown as Json;
|
||||
const reason = "配置关掉了消融";
|
||||
assert.deepEqual(nextStep(book.records), { do: "skip", step: "r1-listCells", reason });
|
||||
book.add("event", { event: "skippedByConfig", reason, objects: ["r1-listCells"] });
|
||||
const next = nextStep(book.records);
|
||||
assert.equal(next.do, "halt");
|
||||
assert.match((next as { reason: string }).reason, /不交三个数的解释还没写/);
|
||||
// 方法.md 4.3 ⑫：解释是 ⑭，和列格、选格、消融、结算一起不做；没有"解释停下"的情形。
|
||||
for (const step of ["listCells", "explanation", "selectCells", "ablationImplementation", "ablationVerification", "ablationRun", "settle"]) {
|
||||
assert.deepEqual(nextStep(book.records), { do: "skip", step: `r1-${step}`, reason });
|
||||
book.add("event", { event: "skippedByConfig", reason, objects: [`r1-${step}`] });
|
||||
}
|
||||
assert.deepEqual(nextStep(book.records), { do: "openStage", stageId: "r1-wrapUp", stage: "wrapUp", round: 1, roster: R });
|
||||
});
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// 诊断（任务书 18 第 1.6 节，方法.md 4.3 诊 1–诊 4）：预期动作手写，不调模型、Jev 或 Docker。
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
const DX = "r1-diagnosis";
|
||||
const DX_ACTOR = "experimenter-r1-diagnosis";
|
||||
const DX_VERIFIER = "verifier-r1-diagnosis";
|
||||
|
||||
/** 诊断开着的配置：A = 2、D = 4（方法.md 6.11 现行值），消融关；maxRepairs 和运行次数可调。 */
|
||||
function diagnosisConfig(maxRepairs: number, runs: number): ResearchConfig {
|
||||
const config = { ...fullConfig(), maxRepairs, runs };
|
||||
if (config.flow.preset !== "full") throw new Error("测试配置应是完整流程");
|
||||
config.flow = { ...config.flow, diagnosis: { enabled: true, perResearcher: 2, maxExperiments: 4 }, ablation: { enabled: false } };
|
||||
return config;
|
||||
}
|
||||
|
||||
/** 第零轮做完、入口分配好；诊断开着，第 1 轮的诊断从 r1-diagnosisProposal 开始。 */
|
||||
function diagnosisRoundOne(maxRepairs: number, runs: number): Book {
|
||||
return assigned(afterRoundZero(diagnosisConfig(maxRepairs, runs)));
|
||||
}
|
||||
|
||||
const dxTasks = (kind: SubmissionKind) => R.map((actor) => ({ actor, kind, object: null }));
|
||||
|
||||
/** 诊 1、诊 2 阶段关闭，两个实验合并成 d1、d2，诊 3 的决定写下。 */
|
||||
function selected(book: Book, diagnose: boolean, chosen: string[]): Book {
|
||||
for (const [stageId, stage, kind] of [["r1-diagnosisProposal", "diagnosisProposal", "diagnosisProposal"], ["r1-diagnosisPrediction", "diagnosisPrediction", "diagnosisPrediction"]] as const) {
|
||||
book.open(stageId, stage, 1, R);
|
||||
for (const r of R) book.sub(stageId, r, kind, null, null);
|
||||
book.add("stageClose", { stageId });
|
||||
if (kind === "diagnosisProposal") {
|
||||
book.add("decision", { decision: "diagnosisDedup", round: 1, rule: "jevSameExperiment", ruleVersion: 1, params: null, seed: null, input: null, output: { experiments: [] } });
|
||||
}
|
||||
}
|
||||
return book.add("decision", { decision: "selectDiagnosis", round: 1, rule: "valueMinusCost", ruleVersion: 1, params: null, seed: null, input: null, output: { diagnose, chosen } });
|
||||
}
|
||||
|
||||
/** 诊断实现（n = 1）或修复（n ≥ 2）阶段：实验员交一份，关闭。 */
|
||||
function dxWork(book: Book, n: number): Book {
|
||||
const stageId = n === 1 ? "r1-diagnosisImplementation-1" : `r1-diagnosisRepair-${n - 1}`;
|
||||
book.open(stageId, "diagnosisImplementation", 1, [DX_ACTOR]);
|
||||
book.sub(stageId, DX_ACTOR, n === 1 ? "diagnosisImplementation" : "diagnosisRepair", DX, null);
|
||||
return book.add("stageClose", { stageId });
|
||||
}
|
||||
|
||||
/** 第 n 次诊断核验：results 是每个选中实验"是否通过"。 */
|
||||
function dxVerified(book: Book, n: number, results: Record<string, boolean>): Book {
|
||||
const stageId = `r1-diagnosisVerification-${n}`;
|
||||
book.open(stageId, "diagnosisVerification", 1, [DX_VERIFIER]);
|
||||
const content: Submission<"diagnosisVerification"> = {
|
||||
experiments: Object.entries(results).map(([experiment, ok]) => ({ experiment, followed: ok, nothingElse: true, where: ok ? "initial.py:10" : "没找到" })),
|
||||
body: "无",
|
||||
};
|
||||
book.add("submission", { stageId, author: DX_VERIFIER, kind: "diagnosisVerification", target: DX, item: null, version: 1, content, jev: null });
|
||||
return book.add("stageClose", { stageId });
|
||||
}
|
||||
|
||||
/** 一次诊断运行：没启动，或启动了（运行记录不带判分）。 */
|
||||
function dxRun(book: Book, n: number, ending: "valid" | "notStarted"): Book {
|
||||
const run = `r1-diagnosis-${n}`;
|
||||
book.add("runPlanned", { run, purpose: "diagnosis", round: 1, object: DX, definition });
|
||||
if (ending === "notStarted") return book.add("runResult", { run, process: { ending, reason: "Docker 未启动" }, judged: null });
|
||||
book.add("runStarted", { run });
|
||||
return book.add("runResult", { run, process: exited, judged: null });
|
||||
}
|
||||
|
||||
const openProposal = { do: "openStage", stageId: "r1-proposal", stage: "proposal", round: 1, roster: R };
|
||||
|
||||
test("诊断开着：诊 1 开阶段，每人一个任务、没有对象；交齐关闭后判重合并，再开诊断预测，同样每人一个任务", () => {
|
||||
const book = diagnosisRoundOne(1, 13);
|
||||
assert.deepEqual(nextStep(book.records), { do: "openStage", stageId: "r1-diagnosisProposal", stage: "diagnosisProposal", round: 1, roster: R });
|
||||
book.open("r1-diagnosisProposal", "diagnosisProposal", 1, R);
|
||||
assert.deepEqual(nextStep(book.records), { do: "dispatch", stageId: "r1-diagnosisProposal", tasks: dxTasks("diagnosisProposal") });
|
||||
for (const r of R) book.sub("r1-diagnosisProposal", r, "diagnosisProposal", null, null);
|
||||
assert.deepEqual(nextStep(book.records), { do: "closeStage", stageId: "r1-diagnosisProposal" });
|
||||
book.add("stageClose", { stageId: "r1-diagnosisProposal" });
|
||||
assert.deepEqual(nextStep(book.records), { do: "diagnosisDedupe", round: 1 });
|
||||
book.add("decision", { decision: "diagnosisDedup", round: 1, rule: "jevSameExperiment", ruleVersion: 1, params: null, seed: null, input: null, output: { experiments: [] } });
|
||||
assert.deepEqual(nextStep(book.records), { do: "openStage", stageId: "r1-diagnosisPrediction", stage: "diagnosisPrediction", round: 1, roster: R });
|
||||
book.open("r1-diagnosisPrediction", "diagnosisPrediction", 1, R);
|
||||
assert.deepEqual(nextStep(book.records), { do: "dispatch", stageId: "r1-diagnosisPrediction", tasks: dxTasks("diagnosisPrediction") });
|
||||
for (const r of R) book.sub("r1-diagnosisPrediction", r, "diagnosisPrediction", null, null);
|
||||
book.add("stageClose", { stageId: "r1-diagnosisPrediction" });
|
||||
assert.deepEqual(nextStep(book.records), { do: "selectDiagnosis", round: 1 });
|
||||
});
|
||||
|
||||
test("余额不到 2 次：诊 1–诊 4 都不做，不开阶段、不写记录，直接开提案（运行 2 次，基线用掉 1 次，余额 1）", () => {
|
||||
const book = diagnosisRoundOne(1, 2);
|
||||
assert.deepEqual(nextStep(book.records), openProposal);
|
||||
});
|
||||
|
||||
test("诊 3 定了不诊断：诊 4 的几步都不做，直接开提案；决定记录照写", () => {
|
||||
const book = selected(diagnosisRoundOne(1, 13), false, []);
|
||||
assert.deepEqual(nextStep(book.records), openProposal);
|
||||
});
|
||||
|
||||
test("诊断实现与核验：一位实验员对本轮诊断，一位核验员逐个实验交结果；全通过就运行，没启动起新编号，已登记没结果要先恢复", () => {
|
||||
const book = selected(diagnosisRoundOne(1, 13), true, ["d1", "d2"]);
|
||||
assert.deepEqual(nextStep(book.records), { do: "openStage", stageId: "r1-diagnosisImplementation-1", stage: "diagnosisImplementation", round: 1, roster: [DX_ACTOR] });
|
||||
book.open("r1-diagnosisImplementation-1", "diagnosisImplementation", 1, [DX_ACTOR]);
|
||||
assert.deepEqual(nextStep(book.records), { do: "dispatch", stageId: "r1-diagnosisImplementation-1", tasks: [{ actor: DX_ACTOR, kind: "diagnosisImplementation", object: DX }] });
|
||||
book.sub("r1-diagnosisImplementation-1", DX_ACTOR, "diagnosisImplementation", DX, null);
|
||||
book.add("stageClose", { stageId: "r1-diagnosisImplementation-1" });
|
||||
assert.deepEqual(nextStep(book.records), { do: "openStage", stageId: "r1-diagnosisVerification-1", stage: "diagnosisVerification", round: 1, roster: [DX_VERIFIER] });
|
||||
dxVerified(book, 1, { d1: true, d2: true });
|
||||
assert.deepEqual(nextStep(book.records), { do: "diagnosisRun", round: 1, run: "r1-diagnosis-1" });
|
||||
dxRun(book, 1, "notStarted");
|
||||
assert.deepEqual(nextStep(book.records), { do: "diagnosisRun", round: 1, run: "r1-diagnosis-2" });
|
||||
assert.deepEqual(runBalance(13, new Board(book.records).runStates()), { charged: 1, reserved: 0, balance: 12 });
|
||||
book.add("runPlanned", { run: "r1-diagnosis-2", purpose: "diagnosis", round: 1, object: DX, definition });
|
||||
assert.throws(() => nextStep(book.records), /诊断运行 r1-diagnosis-2 登记了还没有结果：恢复入口应当已经核对过/);
|
||||
book.add("runStarted", { run: "r1-diagnosis-2" });
|
||||
book.add("runResult", { run: "r1-diagnosis-2", process: exited, judged: null });
|
||||
assert.deepEqual(nextStep(book.records), { do: "diagnosisSettle", round: 1 });
|
||||
// 没启动的那次退还名额：基线 1 次、诊断 1 次，诊断运行只扣一次。
|
||||
assert.deepEqual(runBalance(13, new Board(book.records).runStates()), { charged: 2, reserved: 0, balance: 11 });
|
||||
book.add("diagnosisSettlement", { round: 1, result: null });
|
||||
assert.deepEqual(nextStep(book.records), openProposal);
|
||||
});
|
||||
|
||||
test("核验有实验不通过：只退回修复，只重验一次；用完修复仍不通过的实验拿掉，跑的只是剩下的实验", () => {
|
||||
const book = selected(diagnosisRoundOne(1, 13), true, ["d1", "d2"]);
|
||||
dxWork(book, 1);
|
||||
dxVerified(book, 1, { d1: true, d2: false });
|
||||
assert.deepEqual(nextStep(book.records), { do: "openStage", stageId: "r1-diagnosisRepair-1", stage: "diagnosisImplementation", round: 1, roster: [DX_ACTOR] });
|
||||
book.open("r1-diagnosisRepair-1", "diagnosisImplementation", 1, [DX_ACTOR]);
|
||||
assert.deepEqual(nextStep(book.records), { do: "dispatch", stageId: "r1-diagnosisRepair-1", tasks: [{ actor: DX_ACTOR, kind: "diagnosisRepair", object: DX }] });
|
||||
book.sub("r1-diagnosisRepair-1", DX_ACTOR, "diagnosisRepair", DX, null);
|
||||
book.add("stageClose", { stageId: "r1-diagnosisRepair-1" });
|
||||
assert.deepEqual(nextStep(book.records), { do: "openStage", stageId: "r1-diagnosisVerification-2", stage: "diagnosisVerification", round: 1, roster: [DX_VERIFIER] });
|
||||
dxVerified(book, 2, { d1: true, d2: false });
|
||||
// 修复次数上限 1 已用完：d2 拿掉，d1 照跑；没有"不跑"事件。
|
||||
assert.deepEqual(nextStep(book.records), { do: "diagnosisRun", round: 1, run: "r1-diagnosis-1" });
|
||||
const dropped = diagnosisProgram(new Board(book.records), 1, 1);
|
||||
assert.deepEqual(dropped.experiments, ["d1"]);
|
||||
assert.deepEqual(dropped.dropped.map((d) => d.experiment), ["d2"]);
|
||||
assert.match(dropped.dropped[0]!.reason, /第 2 次核验仍不通过，修复次数上限 1 已用完：没照写法做；没找到/);
|
||||
});
|
||||
|
||||
test("修复后上次通过的实验又不通过：以最近一次核验为准，修复前通过过也不能跑", () => {
|
||||
const book = selected(diagnosisRoundOne(1, 13), true, ["d1", "d2"]);
|
||||
dxWork(book, 1);
|
||||
dxVerified(book, 1, { d1: true, d2: false });
|
||||
dxWork(book, 2);
|
||||
dxVerified(book, 2, { d1: false, d2: true });
|
||||
assert.deepEqual(diagnosisProgram(new Board(book.records), 1, 1).experiments, ["d2"]);
|
||||
assert.deepEqual(nextStep(book.records), { do: "diagnosisRun", round: 1, run: "r1-diagnosis-1" });
|
||||
});
|
||||
|
||||
test("maxRepairs = 0：第一次核验不通过的实验就拿掉，不开修复阶段", () => {
|
||||
const book = selected(diagnosisRoundOne(0, 13), true, ["d1", "d2"]);
|
||||
dxWork(book, 1);
|
||||
dxVerified(book, 1, { d1: false, d2: true });
|
||||
assert.deepEqual(nextStep(book.records), { do: "diagnosisRun", round: 1, run: "r1-diagnosis-1" });
|
||||
assert.deepEqual(diagnosisProgram(new Board(book.records), 1, 0).experiments, ["d2"]);
|
||||
});
|
||||
|
||||
test("选中的实验一个都不剩：写不跑事件（一次），不登记运行、不扣次，不结算，直接开提案", () => {
|
||||
const book = selected(diagnosisRoundOne(1, 13), true, ["d1", "d2"]);
|
||||
dxWork(book, 1);
|
||||
dxVerified(book, 1, { d1: false, d2: false });
|
||||
dxWork(book, 2);
|
||||
dxVerified(book, 2, { d1: false, d2: false });
|
||||
const next = nextStep(book.records);
|
||||
assert.equal(next.do, "diagnosisNotRun");
|
||||
if (next.do !== "diagnosisNotRun") throw new Error("应写不跑事件");
|
||||
assert.equal(next.round, 1);
|
||||
assert.match(next.reason, /选中的实验第 2 次核验都不通过，修复次数上限 1 已用完：d1：.*；d2：/);
|
||||
book.add("event", diagnosisNotRunEvent(1, next.reason));
|
||||
assert.deepEqual(nextStep(book.records), openProposal);
|
||||
assert.deepEqual(runBalance(13, new Board(book.records).runStates()), { charged: 1, reserved: 0, balance: 12 });
|
||||
});
|
||||
|
||||
test("诊断运行扣次以后：余额只算本轮开始之前的运行，诊断这一段的后几步不会算成不诊断", () => {
|
||||
// 运行 3 次：本轮开始前余额 2，诊断运行扣 1 次后余额 1。诊断这一段的后几步照常走，不会因余额变小算成"不诊断"。
|
||||
const book = selected(diagnosisRoundOne(1, 3), true, ["d1"]);
|
||||
dxWork(book, 1);
|
||||
dxVerified(book, 1, { d1: true });
|
||||
dxRun(book, 1, "valid");
|
||||
assert.equal(runBalance(3, new Board(book.records).runStates()).balance, 1);
|
||||
assert.deepEqual(nextStep(book.records), { do: "diagnosisSettle", round: 1 });
|
||||
});
|
||||
|
||||
+59
-1
@@ -3,7 +3,17 @@ import { mkdirSync, mkdtempSync, rmSync, writeFileSync } from "node:fs";
|
||||
import { tmpdir } from "node:os";
|
||||
import { join } from "node:path";
|
||||
import { test, type TestContext } from "node:test";
|
||||
import { checkHunk, checkProtocol, codeHunks, judgeSameChange, type FrozenProtocol, type ProtocolRules } from "../src/jev/index.ts";
|
||||
import {
|
||||
checkExperiment,
|
||||
checkHunk,
|
||||
checkProtocol,
|
||||
codeHunks,
|
||||
judgeSameChange,
|
||||
judgeSameExperiment,
|
||||
type DiagnosisExperiment,
|
||||
type FrozenProtocol,
|
||||
type ProtocolRules,
|
||||
} from "../src/jev/index.ts";
|
||||
import { hash, type Json } from "../src/vocab/index.ts";
|
||||
import { compareVersions, createRepository, saveVersion } from "../src/workspace/index.ts";
|
||||
|
||||
@@ -307,3 +317,51 @@ test("把 workspace 的比较结果切成差异块，每块从 @@ 行起", (t) =
|
||||
assert.equal(hunks[0]?.text, "@@ -1,5 +1,5 @@\n line 1\n-line 2\n+changed 2\n line 3\n line 4\n line 5");
|
||||
assert.equal(hunks[2]?.text, "@@ -0,0 +1 @@\n+x = 1");
|
||||
});
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// 诊断实验的判重和冻结检查（方法.md 4.3 诊 1、诊 2）
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
const EXPERIMENT: DiagnosisExperiment = {
|
||||
name: "拿掉收尾",
|
||||
spot: "solve.py 的收尾修复",
|
||||
body: "拿掉收尾修复，直接交出原始答案",
|
||||
metrics: [{ id: "raw-score", meaning: "收尾之前的分数", unit: "无", direction: "maximize", howMeasured: "收尾之前的分数", missing: "recordMissing" }],
|
||||
};
|
||||
|
||||
test("诊断实验判重：一对一次请求，只给名称、动哪一处、怎么动；阈值暂用判重的 0.3", async (t) => {
|
||||
const sent = fakeJev(t, answerAll(() => 0.31));
|
||||
const other = { name: "强起点", spot: "solve.py 的起点", body: "换成更强的起点" };
|
||||
const j = await judgeSameExperiment("jev-x", EXPERIMENT, other);
|
||||
assert.equal(sent.length, 1);
|
||||
assert.deepEqual(Object.keys(sent[0]!.body.questions), ["same_experiment"]);
|
||||
assert.deepEqual(sent[0]!.body.state.experiment_a, { name: "拿掉收尾", spot: "solve.py 的收尾修复", body: "拿掉收尾修复，直接交出原始答案" });
|
||||
assert.deepEqual(sent[0]!.body.state.experiment_b, other);
|
||||
assert.deepEqual([j.question, j.threshold, j.yes, j.requestId], ["same_experiment", 0.3, true, "req_1"]);
|
||||
});
|
||||
|
||||
test("诊断实验判重：概率等于阈值判否", async (t) => {
|
||||
fakeJev(t, answerAll(() => 0.3));
|
||||
assert.equal((await judgeSameExperiment("jev-x", EXPERIMENT, EXPERIMENT)).yes, false);
|
||||
});
|
||||
|
||||
test("诊断实验冻结检查：三问各一次请求，材料相同，问法是诊断的变体；有一问判是就标出", async (t) => {
|
||||
const sent = fakeJev(t, answerAll((name) => (name === "changes_measurement" ? 0.22 : 0.1)));
|
||||
const result = await checkExperiment("jev-x", EXPERIMENT);
|
||||
assert.deepEqual(
|
||||
result.judgments.map((j) => [j.question, j.threshold, j.yes]),
|
||||
[["self_contradictory", 0.5, false], ["changes_measurement", 0.21, true], ["too_vague", 0.78, false]],
|
||||
);
|
||||
assert.equal(result.flagged, true);
|
||||
assert.deepEqual(sent.map((s) => Object.keys(s.body.questions)), [["self_contradictory"], ["changes_measurement"], ["too_vague"]]);
|
||||
for (const s of sent) assert.deepEqual(s.body.state, sent[0]!.body.state);
|
||||
assert.deepEqual(sent[0]!.body.state.experiment, {
|
||||
name: "拿掉收尾",
|
||||
spot: "solve.py 的收尾修复",
|
||||
body: "拿掉收尾修复，直接交出原始答案",
|
||||
metrics: [{ id: "raw-score", howMeasured: "收尾之前的分数", missing: "recordMissing" }],
|
||||
});
|
||||
// 材料说明和问法都是诊断的变体，哈希和冻结协议同名三问的不同。
|
||||
const protocol = await checkProtocol("jev-x", PROTOCOL);
|
||||
result.judgments.forEach((j, i) => assert.notEqual(j.questionHash, protocol.judgments[i]!.questionHash, j.question));
|
||||
});
|
||||
|
||||
@@ -80,7 +80,7 @@ export function closeStage(ledger: Ledger, stageId: string): number {
|
||||
}
|
||||
|
||||
/** 登记一次运行。 */
|
||||
export function planRun(ledger: Ledger, run: string, purpose: "baseline" | "forward" | "ablation", round: number): number {
|
||||
export function planRun(ledger: Ledger, run: string, purpose: "baseline" | "forward" | "ablation" | "diagnosis", round: number): number {
|
||||
return ledger.append("runPlanned", { run, purpose, round, object: purpose === "baseline" ? null : `obj-${run}`, definition }).seq;
|
||||
}
|
||||
|
||||
|
||||
@@ -16,12 +16,13 @@ import {
|
||||
wrapUp,
|
||||
} from "./ledger-helpers.ts";
|
||||
|
||||
// 写入时由 ledger 拒绝的四种情况（docs/结构.md 第四部分"写入时由 ledger 拒绝的几种情况"），每种一条拒收、一条对应的正常情况。
|
||||
// 写入时由 ledger 拒绝的五种情况（docs/结构.md 第四部分"写入时由 ledger 拒绝的几种情况"），每种一条拒收、一条对应的正常情况。
|
||||
// 预期值来自文字：
|
||||
// 1. 只收开着的阶段（有开阶段记录、没有关阶段记录）的提交，别的拒收；
|
||||
// 2. 要替换的那一份不是本人在这个阶段、对同一个对象的最新一版 → 拒收；实现、修复、核验、消融实现、消融核验不能替换（方法.md 第五部分）；
|
||||
// 3. 运行结果写在"确认启动"之前 → 拒收；"没启动"的结果不需要启动记录（方法.md 第七部分）；
|
||||
// 4. 这一轮消融已经有格确认启动了，才来关这一轮的解释阶段 → 拒收（方法.md 4.3 ⑭）。
|
||||
// 4. 这一轮消融已经有格确认启动了，才来关这一轮的解释阶段 → 拒收（方法.md 4.3 ⑭）；
|
||||
// 5. 这一轮的诊断运行已经确认启动了，才来关这一轮的诊断预测阶段 → 拒收（方法.md 4.3 诊 2）；诊断实现、修复、核验和实现、核验一样不能替换。
|
||||
|
||||
function submitWrapUp(ledger: Ledger, stageId: string, author: string, version: number, body = "收尾") {
|
||||
return ledger.append("submission", { stageId, author, target: null, item: null, version, jev: null, kind: "wrapUp", content: wrapUp(body) });
|
||||
@@ -156,3 +157,53 @@ test("拒收 4 的正常情况：只登记没启动、别的轮的消融启动
|
||||
closeStage(ledger, "wrap");
|
||||
assert.equal(ledger.stageClosed("wrap"), true);
|
||||
});
|
||||
|
||||
test("拒收 5：这一轮的诊断运行已经确认启动，就不能关这一轮的诊断预测阶段", (t) => {
|
||||
const { ledger } = freshLedger(t);
|
||||
openStage(ledger, "dpred", "diagnosisPrediction", 2);
|
||||
planRun(ledger, "diagrun", "diagnosis", 2);
|
||||
ledger.append("runStarted", { run: "diagrun" });
|
||||
assert.throws(() => closeStage(ledger, "dpred"), /诊断预测冻结必须早于诊断运行启动/);
|
||||
assert.equal(ledger.stageClosed("dpred"), false);
|
||||
});
|
||||
|
||||
test("拒收 5 的正常情况：只登记没启动、别的轮的诊断启动、本轮消融或正向启动，都不挡关诊断预测阶段；消融启动也不被诊断运行连累", (t) => {
|
||||
const { ledger } = freshLedger(t);
|
||||
openStage(ledger, "dpred", "diagnosisPrediction", 2);
|
||||
planRun(ledger, "diag-planned", "diagnosis", 2); // 只登记，还没确认启动
|
||||
planRun(ledger, "diag-round1", "diagnosis", 1);
|
||||
ledger.append("runStarted", { run: "diag-round1" });
|
||||
planRun(ledger, "fwd-round2", "forward", 2);
|
||||
ledger.append("runStarted", { run: "fwd-round2" });
|
||||
closeStage(ledger, "dpred");
|
||||
assert.equal(ledger.stageClosed("dpred"), true);
|
||||
// 诊断运行启动后，解释阶段照旧关得掉：两种用途各管各的。
|
||||
openStage(ledger, "expl", "explanation", 2);
|
||||
planRun(ledger, "diag-round2", "diagnosis", 2);
|
||||
ledger.append("runStarted", { run: "diag-round2" });
|
||||
closeStage(ledger, "expl");
|
||||
assert.equal(ledger.stageClosed("expl"), true);
|
||||
});
|
||||
|
||||
test("拒收 2：诊断实现、诊断修复、诊断核验只收第一版，不能替换；诊断预测可以替换", (t) => {
|
||||
const { ledger } = freshLedger(t);
|
||||
openStage(ledger, "dimpl", "diagnosisImplementation", 1, ["e1"]);
|
||||
openStage(ledger, "dver", "diagnosisVerification", 1, ["v1"]);
|
||||
openStage(ledger, "dpred", "diagnosisPrediction", 1);
|
||||
const base = { author: "e1", target: "r1-diagnosis", item: null };
|
||||
const code = { locations: [{ experiment: "d1", locations: ["solve.py:3-9"] }], concerns: "无", body: "无" };
|
||||
const cases = [
|
||||
{ ...base, stageId: "dimpl", kind: "diagnosisImplementation" as const, content: code },
|
||||
{ ...base, stageId: "dimpl", author: "e2", kind: "diagnosisRepair" as const, content: code },
|
||||
{ ...base, author: "v1", stageId: "dver", kind: "diagnosisVerification" as const, content: { experiments: [], body: "无" } },
|
||||
];
|
||||
for (const c of cases) {
|
||||
ledger.append("submission", { ...c, version: 1, jev: null });
|
||||
assert.throws(() => ledger.append("submission", { ...c, version: 2, jev: null }), /不能替换/, c.kind);
|
||||
}
|
||||
const slot = { stageId: "dpred", author: "r1", target: null, item: null, kind: "diagnosisPrediction" as const };
|
||||
const content = { predictions: [{ explanation: "r1-r1-dx1", experiment: "d1", range: range(1, 2, 3) }], body: "理由" };
|
||||
ledger.append("submission", { ...slot, version: 1, jev: null, content });
|
||||
ledger.append("submission", { ...slot, version: 2, jev: null, content });
|
||||
assert.equal(ledger.latestSubmission(slot)?.content.version, 2);
|
||||
});
|
||||
|
||||
@@ -438,6 +438,16 @@ test("关掉再打开：读出的记录不变，序号接着往下排", (t) => {
|
||||
assert.equal(reopened.append("runStarted", { run: "run1" }).seq, 9);
|
||||
});
|
||||
|
||||
test("唯一约束：诊断结算按轮次，同一轮两次报错；同一轮的消融结算另算，不受它影响", (t) => {
|
||||
const { ledger, path } = freshLedger(t);
|
||||
ledger.append("diagnosisSettlement", { round: 1, result: { experiments: [] } });
|
||||
assert.throws(() => ledger.append("diagnosisSettlement", { round: 1, result: { experiments: [] } }), /diagnosis_settlement_once/);
|
||||
ledger.append("diagnosisSettlement", { round: 2, result: { experiments: [] } });
|
||||
ledger.append("settlement", { round: 1, result: { cells: {} } });
|
||||
assert.throws(() => ledger.append("diagnosisSettlement", { round: 3, run: "x", result: {} } as never), /diagnosisSettlement/);
|
||||
assert.equal(countRows(path), 3);
|
||||
});
|
||||
|
||||
test("打开：表结构版本不对报错；不是空文件也不是账本的库报错", (t) => {
|
||||
const dir = tempDir(t);
|
||||
const wrongVersion = join(dir, "v99.db");
|
||||
@@ -446,12 +456,12 @@ test("打开：表结构版本不对报错；不是空文件也不是账本的
|
||||
db.close();
|
||||
assert.throws(() => Ledger.open(wrongVersion), /版本是 99/);
|
||||
|
||||
// 版本 3 是退回还没有对象字段时的账本，不迁移，打开报错。
|
||||
const previous = join(dir, "v3.db");
|
||||
const v3 = new DatabaseSync(previous);
|
||||
v3.exec("PRAGMA user_version = 3");
|
||||
v3.close();
|
||||
assert.throws(() => Ledger.open(previous), /版本是 3.*只认版本 4/s);
|
||||
// 版本 4 是还没有诊断结算和诊断运行用途时的账本，不迁移，打开报错。
|
||||
const previous = join(dir, "v4.db");
|
||||
const v4 = new DatabaseSync(previous);
|
||||
v4.exec("PRAGMA user_version = 4");
|
||||
v4.close();
|
||||
assert.throws(() => Ledger.open(previous), /版本是 4.*只认版本 5/s);
|
||||
|
||||
const foreign = join(dir, "foreign.db");
|
||||
const other = new DatabaseSync(foreign);
|
||||
|
||||
@@ -0,0 +1,337 @@
|
||||
import assert from "node:assert/strict";
|
||||
import { test, type TestContext } from "node:test";
|
||||
import { Ledger } from "../src/ledger/index.ts";
|
||||
import {
|
||||
Board,
|
||||
buildMaterial,
|
||||
type CodeChanges,
|
||||
diagnosisExplanationId,
|
||||
diagnosisId,
|
||||
diagnosisNotRunEvent,
|
||||
type DiagnosisSettlement,
|
||||
type StudyFacts,
|
||||
type TaskMaterial,
|
||||
} from "../src/materials/index.ts";
|
||||
import { selectDiagnosis, settleDiagnosis, type DiagnosisRuleInput, type SettleDiagnosisInput } from "../src/rules/index.ts";
|
||||
import type { SubmissionKind } from "../src/vocab/index.ts";
|
||||
import { freshLedger } from "./ledger-helpers.ts";
|
||||
import { batch, CHANGES_C1, everything, judged, material, NONE, P_A1, prompts, protocolOf, r, study, TABLE_C1, Writer } from "./materials-helpers.ts";
|
||||
|
||||
// 诊断的材料（任务书 18 第 1.5 节、方法.md 第三部分、4.3 诊 1–诊 4）：在真实账本里写一轮诊断，看每个任务该有的有、不该有的没有。
|
||||
// 研究者 a1、a2，实验员 x1，核验员 v1。基线 0.96。三个合并后的诊断实验：d1（a1、a2 的"强起点"合并）、d2（a1 的）、d3（a2 的第二个）。
|
||||
// a1 的诊断预测数 7.771…、a2 的 8.881…是好认的数：任何别人的材料里都不该出现。
|
||||
// 选实验的决定、结算都用 rules 的函数算（不手写数）；这里只查材料怎么读它们。
|
||||
|
||||
const SEALED = { a1: ["7.771", "7.772", "7.773"], a2: ["8.881", "8.882", "8.883"] };
|
||||
const D = diagnosisId(1);
|
||||
const headings = (m: TaskMaterial) => m.sections.map((s) => s.heading);
|
||||
const has = (m: TaskMaterial, text: string) => assert.ok(everything(m).includes(text), `材料里应有"${text}"`);
|
||||
const lacks = (m: TaskMaterial, text: string) => assert.ok(!everything(m).includes(text), `材料里不该有"${text}"`);
|
||||
|
||||
const experiment = (name: string, spot: string) => ({ name, spot, body: `${name}：${spot}上怎么动`, metrics: [] });
|
||||
const gapExperiment = { ...experiment("拿掉收尾", "solve.py 收尾"), metrics: [{ id: "gap", meaning: "间隙", unit: "无", direction: "minimize" as const, howMeasured: "最后量一次", missing: "recordMissing" as const }] };
|
||||
|
||||
/** 两条解释对三个实验的三个数：a1 说 d1 能涨很多，a2 说 d1 涨不了；d2、d3 两人差不多。 */
|
||||
const ranges = {
|
||||
[diagnosisExplanationId(1, "a1", 1)]: { author: "a1", d1: r(1.1, 1.25, 1.3), d2: r(0.95, 1.0, 1.05), d3: r(0.95, 1.0, 1.05) },
|
||||
[diagnosisExplanationId(1, "a2", 1)]: { author: "a2", d1: r(0.9, 0.95, 0.99), d2: r(0.95, 1.0, 1.05), d3: r(0.95, 1.0, 1.05) },
|
||||
};
|
||||
const ruleInput = (balance: number, same: boolean): DiagnosisRuleInput => ({
|
||||
baseline: 0.96,
|
||||
direction: "maximize",
|
||||
lambda: 0.1,
|
||||
experiments: ["d1", "d2", "d3"],
|
||||
explanations: Object.entries(ranges).map(([id, x]) => ({
|
||||
id,
|
||||
author: x.author,
|
||||
experiments: (["d1", "d2", "d3"] as const).map((e) => ({ experiment: e, range: same ? ranges[diagnosisExplanationId(1, "a1", 1)]![e] : x[e] })),
|
||||
})),
|
||||
balance,
|
||||
batchCap: 1,
|
||||
u: 0.5,
|
||||
decay: 0.5,
|
||||
maxExperiments: 2,
|
||||
});
|
||||
|
||||
type Step = "diagProposalOpen" | "diagProposalClosed" | "diagPredictionOpen" | "diagPredictionClosed" | "selected" | "implemented" | "verifiedFailed" | "repairOpen" | "run" | "settled";
|
||||
const ORDER: Step[] = ["diagProposalOpen", "diagProposalClosed", "diagPredictionOpen", "diagPredictionClosed", "selected", "implemented", "verifiedFailed", "repairOpen", "run", "settled"];
|
||||
|
||||
/** 第 1 轮诊断写到某一步（含）为止。选中的是净值最大的两个实验；d 里验证不通过的那个在结算里记证据不足。 */
|
||||
function diagnosisStudy(t: TestContext, upTo: Step): { l: Ledger; w: Writer; chosen: string[] } {
|
||||
const { ledger: l } = freshLedger(t);
|
||||
const w = new Writer(l);
|
||||
const reached = (s: Step) => ORDER.indexOf(upTo) >= ORDER.indexOf(s);
|
||||
l.append("code", { object: "base", version: "v-base", locations: null });
|
||||
w.run("run-0", "baseline", 0, null, judged(0.96, "main"));
|
||||
w.open("s1-dp", "diagnosisProposal", 1, ["a1", "a2"]);
|
||||
w.sub("s1-dp", "a1", "diagnosisProposal", null, null, {
|
||||
explanations: [{ body: "甲的诊断解释：卡在起点 UNIQUE-A1" }],
|
||||
experiments: [experiment("强起点", "solve.py 起点"), gapExperiment],
|
||||
});
|
||||
if (!reached("diagProposalClosed")) return { l, w, chosen: [] };
|
||||
w.sub("s1-dp", "a2", "diagnosisProposal", null, null, {
|
||||
explanations: [{ body: "乙的诊断解释：卡在收尾 UNIQUE-A2" }],
|
||||
experiments: [experiment("强起点", "solve.py 起点（乙的写法）"), experiment("放宽间隙", "solve.py 间隙限制")],
|
||||
});
|
||||
w.close("s1-dp");
|
||||
w.decision("diagnosisDedup", 1, null, {
|
||||
experiments: [
|
||||
{ experiment: "d1", sources: [{ author: "a1", index: 0 }, { author: "a2", index: 0 }] },
|
||||
{ experiment: "d2", sources: [{ author: "a1", index: 1 }] },
|
||||
{ experiment: "d3", sources: [{ author: "a2", index: 1 }] },
|
||||
],
|
||||
});
|
||||
w.open("s1-dpred", "diagnosisPrediction", 1, ["a1", "a2"]);
|
||||
const pred = (author: "a1" | "a2", numbers: string[]) =>
|
||||
(["d1", "d2", "d3"] as const).map((e) => ({ explanation: diagnosisExplanationId(1, author, 1), experiment: e, range: e === "d1" ? r(...(numbers.map(Number) as [number, number, number])) : ranges[diagnosisExplanationId(1, author, 1)]![e] }));
|
||||
w.sub("s1-dpred", "a1", "diagnosisPrediction", null, null, { predictions: pred("a1", SEALED.a1), body: "甲的预测理由" });
|
||||
if (!reached("diagPredictionClosed")) return { l, w, chosen: [] };
|
||||
w.sub("s1-dpred", "a2", "diagnosisPrediction", null, null, { predictions: pred("a2", SEALED.a2), body: "乙的预测理由" });
|
||||
w.close("s1-dpred");
|
||||
const input = ruleInput(9, false);
|
||||
const decision = selectDiagnosis(input);
|
||||
assert.equal(decision.diagnose, true);
|
||||
const chosen = decision.chosen;
|
||||
assert.equal(chosen.length, 2);
|
||||
if (!reached("selected")) return { l, w, chosen };
|
||||
w.decision("selectDiagnosis", 1, input, decision);
|
||||
w.open("s1-dimpl", "diagnosisImplementation", 1, ["x1"]);
|
||||
const table = chosen.map((experiment, i) => ({ experiment, locations: [`solve.py:${10 + 10 * i}-${19 + 10 * i}`] }));
|
||||
w.sub("s1-dimpl", "x1", "diagnosisImplementation", D, null, { locations: table, concerns: "收尾里有个旧开关", body: "诊断实现说明 UNIQUE-X1" });
|
||||
w.close("s1-dimpl");
|
||||
l.append("code", { object: D, version: "v-diag", locations: table as never });
|
||||
if (!reached("implemented")) return { l, w, chosen };
|
||||
w.open("s1-dver", "diagnosisVerification", 1, ["v1"]);
|
||||
w.sub("s1-dver", "v1", "diagnosisVerification", D, null, {
|
||||
experiments: chosen.map((experiment, i) => ({ experiment, followed: i === 0, nothingElse: true, where: `solve.py:${10 + 10 * i}` })),
|
||||
body: "核验员：第二个实验没照写法做",
|
||||
});
|
||||
w.close("s1-dver");
|
||||
if (!reached("repairOpen")) return { l, w, chosen };
|
||||
w.open("s1-dimpl2", "diagnosisImplementation", 1, ["x1"]);
|
||||
if (!reached("run")) return { l, w, chosen };
|
||||
w.run("run-1", "diagnosis", 1, D, null); // 诊断运行的记录里没有判分：逐个实验的结果在结算里
|
||||
if (!reached("settled")) return { l, w, chosen };
|
||||
const settleInput: SettleDiagnosisInput = {
|
||||
baseline: 0.96,
|
||||
direction: "maximize",
|
||||
epsilon: 0.01,
|
||||
lambda: 0.1,
|
||||
experiments: ["d1", "d2", "d3"].map((experiment) => ({
|
||||
experiment,
|
||||
outcome: experiment === chosen[0] ? { kind: "valid", score: 1.1 } : experiment === chosen[1] ? { kind: "insufficient" } : { kind: "untested" },
|
||||
})),
|
||||
explanations: ruleInput(9, false).explanations,
|
||||
};
|
||||
const settlement: DiagnosisSettlement = {
|
||||
run: "run-1",
|
||||
judged: {
|
||||
experiments: [
|
||||
{
|
||||
id: chosen[0]!, ending: "finished", wallSeconds: { status: "measured", value: 31.5 }, verdict: "valid", verdictReason: "校验器判合法",
|
||||
score: { status: "measured", value: 1.1 }, metrics: { "sum-of-radii": { status: "measured", value: 1.1 }, [`${chosen[0]}.gap`]: { status: "measured", value: 0.02 } }, extra: null,
|
||||
},
|
||||
],
|
||||
},
|
||||
dropped: [{ experiment: chosen[1]!, reason: "第二个实验两次核验都没落实" }],
|
||||
input: settleInput,
|
||||
output: settleDiagnosis(settleInput),
|
||||
};
|
||||
l.append("diagnosisSettlement", { round: 1, result: settlement as never });
|
||||
return { l, w, chosen };
|
||||
}
|
||||
|
||||
const mat = (l: Ledger, kind: SubmissionKind, actor: string, stageId: string, object: string | null, extra: { changes: CodeChanges | null; artifacts: Record<string, Record<string, string>> | null } = NONE, facts: StudyFacts = study) =>
|
||||
buildMaterial({ kind, actor, stageId, object, ledger: l, study: facts, prompts, codeRoot: "/code", ...extra });
|
||||
|
||||
test("诊 1：基线代码、基线那次运行的结果和答案、没有别人这一轮的诊断；上限取自配置（A、M）", (t) => {
|
||||
const { l } = diagnosisStudy(t, "diagProposalOpen");
|
||||
const m = mat(l, "diagnosisProposal", "a2", "s1-dp", null, { changes: null, artifacts: { "run-0": { "solution.json": "{\"centers\": []}" } } });
|
||||
assert.deepEqual(m.checklist, {
|
||||
kind: "diagnosisProposal",
|
||||
maxExplanations: 2,
|
||||
maxExperiments: 2,
|
||||
takenMetricIds: ["sum-of-radii", "wall-seconds", "answer-path", "safeguard-triggers"],
|
||||
});
|
||||
assert.deepEqual(m.readable, [{ version: "v-base", label: "本批基线", dir: "/code/v-base" }]);
|
||||
has(m, "运行 run-0：任务给的原始代码；算数（校验器判合法）；分数 0.96");
|
||||
has(m, "run-0/solution.json");
|
||||
assert.equal(m.records.find((x) => x.id === "run-0/solution.json")!.text, "{\"centers\": []}");
|
||||
assert.ok(headings(m).includes("上一轮的收尾"));
|
||||
assert.ok(!headings(m).includes("本轮诊断结果"), "诊断这一段还没走完，不放本轮诊断结果");
|
||||
lacks(m, "UNIQUE-A1"); // a1 在同一阶段已经交了，阶段没关，别人读不到
|
||||
lacks(m, "强起点");
|
||||
});
|
||||
|
||||
test("诊 2：合并后的实验标明作者、全部诊断解释、自己的解释另列并给对号清单；看不到别人在这一步交的数", (t) => {
|
||||
const { l } = diagnosisStudy(t, "diagPredictionOpen");
|
||||
const m = mat(l, "diagnosisPrediction", "a2", "s1-dpred", null);
|
||||
assert.deepEqual(m.checklist, { kind: "diagnosisPrediction", explanations: ["r1-a2-dx1"], experiments: ["d1", "d2", "d3"] });
|
||||
has(m, "#### d1 强起点（作者 a1、a2）");
|
||||
has(m, "动哪一处：solve.py 起点\n"); // 合并后的正文取作者里编号最小的那一份（a1）
|
||||
lacks(m, "solve.py 起点（乙的写法）");
|
||||
has(m, "#### d3 放宽间隙（作者 a2）");
|
||||
has(m, "UNIQUE-A1");
|
||||
has(m, "UNIQUE-A2");
|
||||
assert.ok(m.sections.some((s) => s.heading === "你的诊断解释" && s.text.includes("r1-a2-dx1")));
|
||||
for (const x of SEALED.a1) lacks(m, x); // a1 已经交了诊断预测，阶段还开着
|
||||
});
|
||||
|
||||
test("诊断预测的数：关闭之后，也不进任何别人的材料（诊 2 本人、实现、提案都没有）", (t) => {
|
||||
const { l } = diagnosisStudy(t, "settled");
|
||||
l.append("stageOpen", { stageId: "s1-prop", stage: "proposal", round: 1, roster: ["a1", "a2"] });
|
||||
const proposals = [mat(l, "proposal", "a1", "s1-prop", null), mat(l, "proposal", "a2", "s1-prop", null)];
|
||||
for (const m of proposals) {
|
||||
for (const x of [...SEALED.a1, ...SEALED.a2]) lacks(m, x);
|
||||
}
|
||||
const own = mat(l, "diagnosisPrediction", "a1", "s1-dpred", null);
|
||||
for (const x of SEALED.a2) lacks(own, x);
|
||||
});
|
||||
|
||||
test("诊断实现：选中的实验原文、基线代码、原始产物的位置；不给信念本和成绩", (t) => {
|
||||
const { l, chosen } = diagnosisStudy(t, "selected");
|
||||
const m = mat(l, "diagnosisImplementation", "x1", "s1-dimpl", D);
|
||||
assert.deepEqual(m.checklist, { kind: "diagnosisImplementation", experiments: chosen });
|
||||
assert.deepEqual(m.readable, [{ version: "v-base", label: "起点代码（当前基线）", dir: "/code/v-base" }]);
|
||||
has(m, "产出当前基线的那次正式运行（run-0）的原始产物已经只读地放进你的工作目录");
|
||||
for (const id of chosen) has(m, `### ${id} `);
|
||||
assert.ok(!headings(m).includes("信念本") && !headings(m).includes("本批基线那次运行的结果"));
|
||||
lacks(m, "UNIQUE-A2"); // 诊断解释不给实验员：只给选中的实验
|
||||
});
|
||||
|
||||
test("诊断修复：实现的材料加退回意见（没通过的实验、依据、上一次位置表和改动）、能读上一次的诊断程序；通过的实验不要动", (t) => {
|
||||
const { l, chosen } = diagnosisStudy(t, "repairOpen");
|
||||
const m = mat(l, "diagnosisRepair", "x1", "s1-dimpl2", D, { changes: CHANGES_C1, artifacts: null });
|
||||
assert.deepEqual(m.checklist, { kind: "diagnosisRepair", experiments: chosen });
|
||||
assert.deepEqual(m.readable.map((x) => x.version), ["v-base", "v-diag"]);
|
||||
const opinion = m.sections.find((s) => s.heading === "退回意见")!.text;
|
||||
assert.ok(opinion.includes(`实验 ${chosen[1]}：没照写法做；solve.py:20`));
|
||||
assert.ok(!opinion.includes(`实验 ${chosen[0]}：`), "通过的实验不列");
|
||||
assert.ok(opinion.includes("核验员：第二个实验没照写法做"));
|
||||
assert.ok(opinion.includes(`"experiment": "${chosen[0]}"`) && opinion.includes("多起点循环"));
|
||||
});
|
||||
|
||||
test("诊断核验：选中的实验原文、基线到诊断程序的改动、位置表；不给信念本和实验员的说明", (t) => {
|
||||
const { l, chosen } = diagnosisStudy(t, "implemented");
|
||||
const m = mat(l, "diagnosisVerification", "v1", "s1-dver", D, { changes: CHANGES_C1, artifacts: null });
|
||||
assert.deepEqual(m.checklist, { kind: "diagnosisVerification", experiments: chosen });
|
||||
assert.deepEqual(m.readable.map((x) => x.version), ["v-base", "v-diag"]);
|
||||
has(m, "多起点循环");
|
||||
has(m, "solve.py:10-19");
|
||||
for (const id of chosen) has(m, `### ${id} `);
|
||||
assert.ok(!headings(m).includes("信念本"));
|
||||
lacks(m, "UNIQUE-X1");
|
||||
});
|
||||
|
||||
test("本轮诊断结果：提案材料里每个实验的做法、实测、用时、结局，每条解释确认或否定，推出的方向；排在基线之后、研究记录之前", (t) => {
|
||||
const { l, chosen } = diagnosisStudy(t, "settled");
|
||||
l.append("stageOpen", { stageId: "s1-prop", stage: "proposal", round: 1, roster: ["a1", "a2"] });
|
||||
const m = mat(l, "proposal", "a1", "s1-prop", null);
|
||||
const h = headings(m);
|
||||
assert.ok(h.indexOf("基线") < h.indexOf("本轮诊断结果") && h.indexOf("本轮诊断结果") < h.indexOf("研究记录"));
|
||||
const section = m.sections.find((s) => s.heading === "本轮诊断结果")!.text;
|
||||
const [first, second] = chosen as [string, string];
|
||||
assert.ok(section.includes(`诊断运行 run-1：在基线 base（最终指标 0.96）上，一次运行做了 1 个实验，只扣 1 次`));
|
||||
const line = (id: string) => section.split("\n").find((x) => x.startsWith(`- ${id} `))!;
|
||||
for (const part of ["实测 1.1，改进量 +0.14", "用时 31.5 秒", "其余指标：d1.gap 0.02", "测前合并范围"]) assert.ok(line(first!).includes(part), part);
|
||||
assert.ok(line(second!).includes("证据不足：核验不通过，第二个实验两次核验都没落实"));
|
||||
assert.match(section, /没做：一次诊断最多做 D 个实验，放不下/);
|
||||
assert.match(section, /- r1-a1-dx1（作者 a1）：确认 .*；否定 /);
|
||||
assert.ok(section.includes("甲的诊断解释：卡在起点 UNIQUE-A1") && section.includes("乙的诊断解释：卡在收尾 UNIQUE-A2"));
|
||||
assert.match(section, /程序按诊断推出的方向：d\d（实测 \+0\.14）。这只是按公式推出来的参考，你可以不跟。/);
|
||||
assert.ok(m.records.some((x) => x.id === "diagnosis-r1" && x.text.includes("run-1") && !x.text.includes("7.771")));
|
||||
assert.ok(m.readable.some((x) => x.version === "v-diag" && x.label === "第 1 轮诊断程序的代码"));
|
||||
});
|
||||
|
||||
/** 只有基线、第 1 轮提案阶段开着的账本：本轮没有诊断记录。 */
|
||||
function bare(t: TestContext): Ledger {
|
||||
const { ledger: l } = freshLedger(t);
|
||||
const w = new Writer(l);
|
||||
l.append("code", { object: "base", version: "v-base", locations: null });
|
||||
w.run("run-0", "baseline", 0, null, judged(0.96, "main"));
|
||||
w.open("s1-prop", "proposal", 1, ["a1", "a2"]);
|
||||
return l;
|
||||
}
|
||||
const diagnosisText = (m: TaskMaterial) => m.sections.find((s) => s.heading === "本轮诊断结果")!.text;
|
||||
|
||||
test("本轮没有诊断：写原因。余额不到 2 次、配置关掉（余额够却没有决定）", (t) => {
|
||||
const l = bare(t);
|
||||
// 基线占 1 次：运行总数 2 时余额 1（不到 2），总数 10 时余额 9。
|
||||
assert.equal(diagnosisText(mat(l, "proposal", "a1", "s1-prop", null, NONE, { ...study, totalRuns: 2 })), "本轮没有诊断：余额只剩 1 次，不到 2 次，要留 1 次给本轮正向");
|
||||
assert.equal(diagnosisText(mat(l, "proposal", "a1", "s1-prop", null)), "本轮没有诊断：配置关掉了诊断");
|
||||
});
|
||||
|
||||
test("本轮没有诊断：信息价值不抵成本时，原因带最好的组合净值；选中的实验一个都没剩下时，原因取事件", (t) => {
|
||||
const l = bare(t);
|
||||
const w = new Writer(l);
|
||||
const input = ruleInput(9, true); // 两条解释对每个实验预测一样：信息价值只剩开放预留那一点，抵不过成本
|
||||
const decision = selectDiagnosis(input);
|
||||
assert.equal(decision.diagnose, false);
|
||||
w.decision("selectDiagnosis", 1, input, decision);
|
||||
const best = Math.max(...decision.combos.map((c) => c.net));
|
||||
const text = diagnosisText(mat(l, "proposal", "a1", "s1-prop", null));
|
||||
assert.equal(text, `本轮没有诊断：信息价值减成本都不大于 0，最好的组合净值 ${Number(best.toPrecision(4))}`);
|
||||
|
||||
// 同一轮另一份账本：决定是做，但核验都没通过、修复用完。
|
||||
const l2 = bare(t);
|
||||
const w2 = new Writer(l2);
|
||||
const done = selectDiagnosis(ruleInput(9, false));
|
||||
w2.decision("selectDiagnosis", 1, ruleInput(9, false), done);
|
||||
l2.append("event", diagnosisNotRunEvent(1, "d1、d2 都没通过核验"));
|
||||
assert.equal(diagnosisText(mat(l2, "proposal", "a1", "s1-prop", null)), "本轮没有诊断：选中的实验核验都没通过、修复次数用完，一个都没跑：d1、d2 都没通过核验");
|
||||
});
|
||||
|
||||
test("诊断运行在诊断结算写入之前不公开，写入之后公开（和正向的采纳、消融的结算同一类规则）", (t) => {
|
||||
const run = (l: Ledger) => new Board(l.records()).runs().find((x) => x.plan.run === "run-1")!;
|
||||
const running = diagnosisStudy(t, "run").l;
|
||||
assert.equal(new Board(running.records()).runPublic(run(running)), false);
|
||||
const settled = diagnosisStudy(t, "settled").l;
|
||||
assert.equal(new Board(settled.records()).runPublic(run(settled)), true);
|
||||
});
|
||||
|
||||
test("信念本的诊断历史：各轮的实验、确认与否定、方向；战绩按人加一行诊断；诊断实验的实测不算手里最好", (t) => {
|
||||
const { l, chosen } = diagnosisStudy(t, "settled");
|
||||
l.append("stageOpen", { stageId: "s1-prop", stage: "proposal", round: 1, roster: ["a1", "a2"] });
|
||||
const book = mat(l, "proposal", "a1", "s1-prop", null).sections.find((s) => s.heading === "信念本")!.text;
|
||||
assert.ok(book.includes("**诊断历史**"));
|
||||
assert.ok(book.includes(`- 第 1 轮：\n 诊断运行 run-1：在基线 base`));
|
||||
assert.ok(book.includes(` - ${chosen[0]} `) && book.includes("确认 d1；否定 无"));
|
||||
assert.ok(book.includes("手里最好 0.96"), "诊断实验实测 1.1，但诊断的答案不采纳，手里最好仍是基线");
|
||||
// 战绩：a1、a2 各一条诊断：只有 d1 一个有效实验，样本各 1 个。a1 的 q5 = 1.25 > 实测 1.1 偏高；a2 的 q5 = 0.95 < 1.1 偏低。
|
||||
assert.ok(book.includes("- a1，诊断 1 次：中位数偏高 1、偏低 0"));
|
||||
assert.ok(book.includes("- a2，诊断 1 次：中位数偏高 0、偏低 1"));
|
||||
});
|
||||
|
||||
test("信念本的诊断历史：第 2 轮诊断关着时，第 1 轮的历史还在，第 2 轮没走完不列；第 2 轮诊断任务读不到第 2 轮的诊断结果", (t) => {
|
||||
const { l } = diagnosisStudy(t, "settled");
|
||||
l.append("stageOpen", { stageId: "s1-prop", stage: "proposal", round: 1, roster: ["a1", "a2"] });
|
||||
l.append("stageClose", { stageId: "s1-prop" });
|
||||
// 第 1 轮采纳 r1-c1 作基线要写一大串记录，这里只看第 2 轮的诊断任务：基线仍是 base，所以直接开第 2 轮的诊断阶段。
|
||||
l.append("stageOpen", { stageId: "s2-dp", stage: "diagnosisProposal", round: 2, roster: ["a1", "a2"] });
|
||||
const m = mat(l, "diagnosisProposal", "a1", "s2-dp", null, { changes: null, artifacts: {} });
|
||||
assert.ok(!headings(m).includes("本轮诊断结果"));
|
||||
const book = m.sections.find((s) => s.heading === "信念本")!.text;
|
||||
assert.ok(book.includes("- 第 1 轮：") && !book.includes("- 第 2 轮："));
|
||||
});
|
||||
|
||||
test("实现正向候选的实验员：本轮有诊断时能读诊断程序（只读，注明冻结协议引用诊断实验时作参考）；没有诊断时没有", (t) => {
|
||||
const { l } = diagnosisStudy(t, "settled");
|
||||
const w = new Writer(l);
|
||||
l.append("stageOpen", { stageId: "s1-prop", stage: "proposal", round: 1, roster: ["a1", "a2"] });
|
||||
w.protocol("r1-c1", protocolOf("r1-c1", P_A1, 1, "base", ["r1-a1-1"], ["加多起点"]));
|
||||
w.open("s1-impl-c1", "implementation", 1, ["x1"]);
|
||||
const m = mat(l, "implementation", "x1", "s1-impl-c1", "r1-c1");
|
||||
assert.deepEqual(m.readable, [
|
||||
{ version: "v-base", label: "起点代码（本批基线）", dir: "/code/v-base" },
|
||||
{ version: "v-diag", label: "本轮诊断程序（只读，冻结协议引用诊断实验时作参考）", dir: "/code/v-diag" },
|
||||
]);
|
||||
has(m, "冻结协议引用诊断实验时作参考");
|
||||
lacks(m, "UNIQUE-A1"); // 只给代码，不给诊断解释和成绩
|
||||
lacks(m, "实测 1.1");
|
||||
|
||||
const l2 = bare(t);
|
||||
const w2 = new Writer(l2);
|
||||
w2.protocol("r1-c1", protocolOf("r1-c1", P_A1, 1, "base", ["r1-a1-1"], ["加多起点"]));
|
||||
w2.open("s1-impl-c1", "implementation", 1, ["x1"]);
|
||||
assert.deepEqual(mat(l2, "implementation", "x1", "s1-impl-c1", "r1-c1").readable.map((x) => x.version), ["v-base"]);
|
||||
});
|
||||
@@ -36,6 +36,7 @@ export const study: StudyFacts = {
|
||||
lambda: 0.1,
|
||||
totalRuns: 10,
|
||||
proposalLimit: 2,
|
||||
diagnosisExperimentLimit: 2,
|
||||
factorLimit: 3,
|
||||
metrics: [
|
||||
{ id: "sum-of-radii", layer: "final", meaning: "半径和", unit: "无", direction: "maximize", howMeasured: "校验器", missing: "invalidateRun" },
|
||||
@@ -114,7 +115,7 @@ export class Writer {
|
||||
decision(decision: string, round: number, input: unknown, output: unknown): void {
|
||||
this.l.append("decision", { decision, round, rule: "test", ruleVersion: 1, params: null, seed: null, input: input as never, output: output as never });
|
||||
}
|
||||
run(run: string, purpose: "baseline" | "forward" | "ablation", round: number, object: string | null, j: JudgedResult): void {
|
||||
run(run: string, purpose: "baseline" | "forward" | "ablation" | "diagnosis", round: number, object: string | null, j: JudgedResult | null): void {
|
||||
this.l.append("runPlanned", { run, purpose, round, object, definition });
|
||||
this.l.append("runStarted", { run });
|
||||
this.l.append("runResult", { run, process: exited, judged: j });
|
||||
|
||||
@@ -0,0 +1,294 @@
|
||||
import assert from "node:assert/strict";
|
||||
import { test } from "node:test";
|
||||
import {
|
||||
diagnosisModel,
|
||||
runDiagnosisRule,
|
||||
selectDiagnosis,
|
||||
settleDiagnosis,
|
||||
trackRecord,
|
||||
type DiagnosisExplanation,
|
||||
type DiagnosisRuleInput,
|
||||
type ExperimentOutcome,
|
||||
} from "../src/rules/index.ts";
|
||||
import { close, loadFixture, scenarioPicker, type Scenario } from "./rules-fixtures-helpers.ts";
|
||||
|
||||
// 预期值来自 tests/fixtures/rules/gen_diagnosis.py：只照方法.md 6.12、6.7、6.9 的文字、用有理数写成的独立参考，
|
||||
// 没有读过 src/rules/diagnosis.ts 和 value.ts（重新生成：python3 tests/fixtures/rules/gen_diagnosis.py）。
|
||||
// 场景里有手算的几个（opposite_lambda0 的 VOI = 0.1，same_lambda0 的 VOI = 0，说明写在场景的 note 里）。
|
||||
|
||||
const fixture = loadFixture("diagnosis.json");
|
||||
const pick = scenarioPicker(fixture);
|
||||
|
||||
const EXACT = 1e-12;
|
||||
const BITS = 1e-9;
|
||||
/** 只做一个实验时的 VOI：实现用网格近似，参考是逐段闭式积分。 */
|
||||
const closeK1 = (actual: number, expected: number, what: string): void =>
|
||||
close(actual, expected, 1e-9 + 2e-4 * Math.abs(expected), what);
|
||||
|
||||
type Ranges = Record<string, { q1: number; q5: number; q9: number }>;
|
||||
|
||||
function explanationsOf(s: Scenario): DiagnosisExplanation[] {
|
||||
return s.input.explanations.map((e: { id: string; author: string; ranges: Ranges }) => ({
|
||||
id: e.id,
|
||||
author: e.author,
|
||||
experiments: Object.entries(e.ranges).map(([experiment, range]) => ({ experiment, range })),
|
||||
}));
|
||||
}
|
||||
|
||||
function ruleInput(s: Scenario): DiagnosisRuleInput {
|
||||
return {
|
||||
baseline: s.input.b,
|
||||
direction: s.input.direction,
|
||||
lambda: s.input.lambda,
|
||||
experiments: s.input.experiments,
|
||||
explanations: explanationsOf(s),
|
||||
balance: s.input.balance,
|
||||
batchCap: s.input.batchCap,
|
||||
maxExperiments: s.input.maxExperiments,
|
||||
u: s.input.u,
|
||||
decay: s.input.decay,
|
||||
};
|
||||
}
|
||||
|
||||
/** 选实验场景：事前数字、组合表、选中的实验都和独立参考一致。 */
|
||||
function checkSelection(name: string): ReturnType<typeof selectDiagnosis> {
|
||||
const s = pick.get(name);
|
||||
const got = selectDiagnosis(ruleInput(s));
|
||||
const e = s.expected;
|
||||
for (const [d, v] of Object.entries(e.priorGains as Record<string, number>)) {
|
||||
close(got.priorGains.find((x) => x.experiment === d)!.gain, v, EXACT, `${name} ${d} 事前期望收益`);
|
||||
}
|
||||
close(got.g0, e.g0, EXACT, `${name} ĝ₀`);
|
||||
e.weights.forEach((w: number, j: number) => close(got.weights[j]!, w, EXACT, `${name} 起始权重 ${j}`));
|
||||
assert.deepEqual(got.combos.map((c) => c.experiments), e.combos.map((c: any) => c.experiments), `${name} 组合列表`);
|
||||
e.combos.forEach((c: any, i: number) => {
|
||||
const g = got.combos[i]!;
|
||||
if (c.voi === null) return;
|
||||
const what = `${name} {${c.experiments}}`;
|
||||
if (c.experiments.length === s.input.experiments.length) close(g.voi, c.voi, EXACT, `${what} VOI`);
|
||||
else closeK1(g.voi, c.voi, `${what} VOI`);
|
||||
assert.equal(g.rHat, c.rHat, `${what} R̂`);
|
||||
close(g.mPrime, c.mPrime, EXACT, `${what} m′`);
|
||||
close(g.cost, c.cost, EXACT, `${what} 成本`);
|
||||
close(g.value, c.value, 1e-9 + 2e-4 * Math.abs(c.value), `${what} 价值`);
|
||||
close(g.net, c.net, 1e-9 + 2e-4 * Math.abs(c.value), `${what} 净值`);
|
||||
});
|
||||
if ("diagnose" in e) {
|
||||
assert.equal(got.diagnose, e.diagnose, `${name} 做不做`);
|
||||
assert.deepEqual(got.chosen, e.chosen, `${name} 选中的实验`);
|
||||
}
|
||||
return got;
|
||||
}
|
||||
|
||||
test("6.12 信息价值：两条解释对实验预测相反时为正（手算 0.1），预测一样时为 0", () => {
|
||||
const opposite = checkSelection("opposite_lambda0");
|
||||
close(opposite.combos[0]!.voi, 0.1, EXACT, "手算：0.5 × 0.2 − 0");
|
||||
assert.equal(opposite.diagnose, true);
|
||||
|
||||
const same = checkSelection("same_lambda0");
|
||||
close(same.combos[0]!.voi, 0, EXACT, "预测一样、范围都在 0 以上：做不做都不改变看法");
|
||||
assert.equal(same.diagnose, false, "净值 = −成本 < 0");
|
||||
assert.equal(same.experiments[0]!.reason, "value");
|
||||
assert.equal(same.chosenNet, 0);
|
||||
|
||||
// λ = 0.1：预测一样时只剩开放预留带来的一小部分，比预测相反小得多。
|
||||
const oppositeReserve = checkSelection("opposite_lambda01");
|
||||
const sameReserve = checkSelection("same_lambda01");
|
||||
close(sameReserve.combos[0]!.voi, 0.0025, EXACT, "手算：λ × (E[max(0, U)] − E U) = 0.1 × 0.025");
|
||||
assert.ok(sameReserve.combos[0]!.voi > 0);
|
||||
assert.ok(oppositeReserve.combos[0]!.voi > 10 * sameReserve.combos[0]!.voi);
|
||||
});
|
||||
|
||||
test("6.12 价值减成本：不管做几个实验成本相同；R̂ = ⌈(余额 − 1) ÷ 每批上限⌉", () => {
|
||||
const got = checkSelection("multi_all");
|
||||
// 手算：余额 9、K = 2：R̂ = ⌈8 ÷ 2⌉ = 4，m′ = 1 + 0.5 + 0.25 + 0.125 = 1.875；
|
||||
// 成本 = 2 × 0.0525 × (3/4 − 2/3) = 0.00875（ĝ₀ = 0.0525 来自参考脚本的事前期望）。
|
||||
assert.equal(got.combos.length, 1, "实验数 3 ≤ D = 4：只有全做一个组合");
|
||||
assert.equal(got.combos[0]!.rHat, 4);
|
||||
close(got.combos[0]!.mPrime, 1.875, EXACT, "m′");
|
||||
close(got.combos[0]!.cost, 0.00875, EXACT, "成本");
|
||||
assert.deepEqual(got.chosen, ["d1", "d2", "d3"]);
|
||||
assert.ok(got.experiments.every((x) => x.done && x.reason === null && x.bestCombo === null));
|
||||
});
|
||||
|
||||
test("6.12 选实验：余额 < 2 时不诊断，没有可选组合", () => {
|
||||
const got = checkSelection("balance_one");
|
||||
assert.deepEqual(got.combos, []);
|
||||
assert.equal(got.diagnose, false);
|
||||
assert.deepEqual(got.chosen, []);
|
||||
assert.ok(got.experiments.every((x) => !x.done && x.bestCombo === null));
|
||||
});
|
||||
|
||||
test("6.12 选实验：实验数超过 D 时只比较装满的组合", () => {
|
||||
const one = checkSelection("over_cap_d1");
|
||||
assert.equal(one.combos.length, 3, "C(3,1) 个单实验组合");
|
||||
assert.ok(one.combos.every((c) => c.experiments.length === 1), "没有空组合，也没有更大的组合");
|
||||
assert.deepEqual(one.chosen, ["d2"]);
|
||||
// 没选中的实验：放不下，并附上含它的组合里净值最大的那个（这里就是它自己的单实验组合）。
|
||||
const d3 = one.experiments.find((x) => x.experiment === "d3")!;
|
||||
assert.equal(d3.done, false);
|
||||
assert.equal(d3.reason, "budget");
|
||||
assert.deepEqual(d3.bestCombo!.experiments, ["d3"]);
|
||||
close(d3.bestCombo!.net, pick.get("over_cap_d1").expected.combos[2].net, 1e-9 + 2e-4 * 0.1, "d3 的净值");
|
||||
|
||||
const two = checkSelection("over_cap_d2");
|
||||
assert.deepEqual(two.combos.map((c) => c.experiments), [["d1", "d2"], ["d1", "d3"], ["d2", "d3"]]);
|
||||
assert.equal(two.chosen.length, 2);
|
||||
// 多做一个实验信息价值只会不变或变大：每个两实验组合的 VOI 不小于它包含的任一单实验（参考脚本的精确值）。
|
||||
const singles = pick.get("over_cap_d1").expected.combos as Array<{ experiments: string[]; voi: number }>;
|
||||
for (const pair of two.combos) {
|
||||
for (const d of pair.experiments) {
|
||||
const single = singles.find((c) => c.experiments[0] === d)!.voi;
|
||||
assert.ok(pair.voi >= single - (1e-9 + 2e-4 * single), `{${pair.experiments}} 的 VOI ${pair.voi} 小于 {${d}} 的 ${single}`);
|
||||
}
|
||||
}
|
||||
});
|
||||
|
||||
test("6.12 选实验：平票按实验编号（自然序，d2 排在 d10 前面）；越小越好按改进量算", () => {
|
||||
const tie = checkSelection("tie_by_id");
|
||||
assert.deepEqual(tie.chosen, ["d2"]);
|
||||
assert.deepEqual(tie.experiments.find((x) => x.experiment === "d10")!.bestCombo!.experiments, ["d10"]);
|
||||
checkSelection("minimize_all");
|
||||
});
|
||||
|
||||
test("登记表：选诊断实验只有一条规则，结果带规则名和版本；同样的输入算两次逐位相同", () => {
|
||||
const s = pick.get("multi_all");
|
||||
const run = runDiagnosisRule("valueMinusCost", ruleInput(s));
|
||||
assert.deepEqual(run.rule, { name: "valueMinusCost", version: 1 });
|
||||
assert.deepEqual(JSON.parse(JSON.stringify(run)), JSON.parse(JSON.stringify({ rule: run.rule, ...selectDiagnosis(ruleInput(s)) })));
|
||||
});
|
||||
|
||||
test("成分模型：每条解释一个成分加开放预留；解释合起来 1 − λ，同一人平分", () => {
|
||||
const s = pick.get("settle_mixed");
|
||||
const m = diagnosisModel({ ...ruleInput({ ...s, input: { ...s.input, balance: 9, batchCap: 2, maxExperiments: 4, u: 0.5, decay: 0.5 } }), });
|
||||
// 手算：researcher-1 一条 0.9 × 1/2，researcher-2 两条各 0.9 × 1/4，预留 0.1。
|
||||
assert.deepEqual(m.components, ["e1", "e2", "e3", "openReserve"]);
|
||||
[0.45, 0.225, 0.225, 0.1].forEach((w, j) => close(m.weights[j]!, w, EXACT, `权重 ${j}`));
|
||||
});
|
||||
|
||||
type Settled = ReturnType<typeof settleDiagnosis>;
|
||||
|
||||
function settle(name: string): { got: Settled; e: any } {
|
||||
const s = pick.get(name);
|
||||
const outcomes = s.input.outcomes as Record<string, { kind: ExperimentOutcome["kind"]; score?: number }>;
|
||||
const got = settleDiagnosis({
|
||||
baseline: s.input.b,
|
||||
direction: s.input.direction,
|
||||
epsilon: s.input.epsilon,
|
||||
lambda: s.input.lambda,
|
||||
experiments: Object.entries(outcomes).map(([experiment, o]) => ({
|
||||
experiment,
|
||||
outcome: o.kind === "valid" ? { kind: "valid", score: o.score! } : ({ kind: o.kind } as ExperimentOutcome),
|
||||
})),
|
||||
explanations: explanationsOf(s),
|
||||
});
|
||||
const e = s.expected;
|
||||
assert.deepEqual(got.experiments.map((x) => x.outcome), Object.keys(e.experiments).map((d) => e.experiments[d].outcome), `${name} 结局`);
|
||||
for (const row of got.experiments) {
|
||||
const x = e.experiments[row.experiment];
|
||||
const what = `${name} ${row.experiment}`;
|
||||
close(row.merged.q1, x.merged.q1, BITS, `${what} 合并一成`);
|
||||
close(row.merged.q5, x.merged.q5, BITS, `${what} 合并一半`);
|
||||
close(row.merged.q9, x.merged.q9, BITS, `${what} 合并九成`);
|
||||
if (x.gain === undefined) {
|
||||
// 证据不足、没做：不判、不记账。
|
||||
assert.equal(row.gain, null, `${what} 收益`);
|
||||
assert.equal(row.worldBits, null, `${what} 世界比特`);
|
||||
assert.equal(row.surprise, null, `${what} 惊讶`);
|
||||
assert.equal(row.disagreement, null, `${what} 分歧量`);
|
||||
} else {
|
||||
close(row.gain!, x.gain, EXACT, `${what} 收益`);
|
||||
close(row.worldBits!, x.worldBits, BITS, `${what} 世界比特`);
|
||||
assert.equal(row.surprise!.cell, x.surprise.bin, `${what} 惊讶所在格`);
|
||||
close(row.surprise!.prior, x.surprise.p, BITS, `${what} 事前把握`);
|
||||
assert.equal(row.surprise!.bits.status, x.surprise.status === "ok" ? "measured" : "unmeasured");
|
||||
if (row.surprise!.bits.status === "measured") close(row.surprise!.bits.value, x.surprise.surprise, BITS, `${what} 惊讶`);
|
||||
close(row.disagreement!, x.disagreement, BITS, `${what} 分歧量`);
|
||||
}
|
||||
}
|
||||
// 确认与否定：只列有效实验。
|
||||
assert.deepEqual(
|
||||
Object.fromEntries(got.judgments.map((j) => [j.id, Object.fromEntries(j.verdicts.map((v) => [v.experiment, v.verdict]))])),
|
||||
e.judgments,
|
||||
`${name} 确认与否定`,
|
||||
);
|
||||
const dir = got.direction;
|
||||
assert.equal(dir.weightsKept, e.direction.weightsKept, `${name} 权重是否保持`);
|
||||
e.direction.weights.forEach((w: number, j: number) => close(dir.weights[j]!, w, EXACT, `${name} 新权重 ${j}`));
|
||||
for (const g of dir.gains) {
|
||||
const x = e.direction.gains[g.experiment];
|
||||
close(g.gain, x.gain, EXACT, `${name} ${g.experiment} 方向收益`);
|
||||
assert.equal(g.measured, x.measured, `${name} ${g.experiment} 是实测还是估计`);
|
||||
}
|
||||
assert.equal(dir.experiment, e.direction.experiment, `${name} 推出的方向`);
|
||||
close(dir.gain, e.direction.gain, EXACT, `${name} 方向收益`);
|
||||
assert.equal(dir.found, e.direction.found, `${name} 有没有指出方向`);
|
||||
return { got, e };
|
||||
}
|
||||
|
||||
test("6.12 结算：证据不足的实验不判、不记账；没做的实验也不判", () => {
|
||||
const { got } = settle("settle_mixed");
|
||||
assert.deepEqual(got.experiments.map((x) => x.outcome), ["valid", "insufficient", "untested"]);
|
||||
// d2、d3 没有有效分数：每条解释只有 d1 一行判定，没有比特。
|
||||
assert.ok(got.judgments.every((j) => j.verdicts.length === 1 && j.verdicts[0]!.experiment === "d1"));
|
||||
assert.equal(got.experiments.filter((x) => x.worldBits !== null).length, 1);
|
||||
// 手算：d1 实测 1.045 = b + 0.045，收益 +0.045；落在 e1 的 [1.02, 1.08] 里确认，e2 的 [1.00, 1.06] 里确认，e3 的 [0.98, 1.10] 里确认。
|
||||
assert.deepEqual(got.judgments.map((j) => j.verdicts[0]!.verdict), ["confirmed", "confirmed", "confirmed"]);
|
||||
});
|
||||
|
||||
test("6.12 结算：两个有效实验、确认与否定、权重更新", () => {
|
||||
const { got } = settle("settle_two_valid");
|
||||
// d2 实测 1.25，e2 的 [1.1, 1.3] 确认；e1 的 [0.95, 1.05]、e3 的 [0.8, 1.0] 否定。
|
||||
assert.deepEqual(got.judgments.find((j) => j.id === "e2")!.verdicts.find((v) => v.experiment === "d2")!.verdict, "confirmed");
|
||||
assert.deepEqual(got.judgments.find((j) => j.id === "e1")!.verdicts.find((v) => v.experiment === "d2")!.verdict, "denied");
|
||||
assert.deepEqual(got.judgments.find((j) => j.id === "e3")!.verdicts.find((v) => v.experiment === "d2")!.verdict, "denied");
|
||||
});
|
||||
|
||||
test("6.12 结算：全部实验收益不为正时，记没有指出能涨分的方向", () => {
|
||||
const { got } = settle("settle_no_direction");
|
||||
assert.equal(got.direction.found, false);
|
||||
assert.ok(got.direction.gains.every((g) => g.gain <= 0));
|
||||
// 手算：收益分别是 0.97 − 1 = −0.03、0.8 − 1 = −0.2、0.9 − 1 = −0.1，最大的是 d1 的 −0.03。
|
||||
assert.equal(got.direction.experiment, "d1");
|
||||
close(got.direction.gain, -0.03, EXACT, "最大收益");
|
||||
});
|
||||
|
||||
test("6.12 结算：没有有效实验、实测落在所有成分之外时保持起始权重", () => {
|
||||
const none = settle("settle_all_insufficient");
|
||||
assert.equal(none.got.direction.weightsKept, false, "没有有效实验时没有密度可乘，权重原样（不算'无从更新'）");
|
||||
assert.ok(none.got.judgments.every((j) => j.verdicts.length === 0));
|
||||
const outside = settle("settle_outside_all");
|
||||
assert.equal(outside.got.direction.weightsKept, true);
|
||||
assert.deepEqual(outside.got.direction.weights, diagnosisModel({
|
||||
baseline: 1, direction: "maximize", lambda: 0.1, experiments: ["d1", "d2", "d3"],
|
||||
explanations: explanationsOf(pick.get("settle_outside_all")),
|
||||
}).weights);
|
||||
});
|
||||
|
||||
test("6.12 结算：越小越好按改进量 b − x", () => {
|
||||
const { got } = settle("settle_minimize");
|
||||
close(got.experiments[0]!.gain!, 0.12, EXACT, "手算：1.0 − 0.88");
|
||||
close(got.experiments[1]!.gain!, -0.05, EXACT, "手算：1.0 − 1.05");
|
||||
assert.equal(got.direction.experiment, "d1");
|
||||
});
|
||||
|
||||
test("战绩：按人、按正向 / 诊断 / 消融分开数，次序是编号、正向、诊断、消融", () => {
|
||||
// 手算：researcher-2 的诊断样本 (q1, q5, q9) = (0.1, 0.2, 0.3)，实测 0.35 在 q9 之上：中位数偏低、被穿过上沿、没包住、落点 aboveQ9。
|
||||
const range = { q1: 0.1, q5: 0.2, q9: 0.3 };
|
||||
const rows = trackRecord([
|
||||
{ researcher: "researcher-2", phase: "ablation", range, value: 0.2 },
|
||||
{ researcher: "researcher-2", phase: "diagnosis", range, value: 0.35 },
|
||||
{ researcher: "researcher-1", phase: "diagnosis", range, value: 0.05 },
|
||||
{ researcher: "researcher-2", phase: "forward", range, value: 0.25 },
|
||||
]);
|
||||
assert.deepEqual(rows.map((r) => `${r.researcher}/${r.phase}`), [
|
||||
"researcher-1/diagnosis", "researcher-2/forward", "researcher-2/diagnosis", "researcher-2/ablation",
|
||||
]);
|
||||
const t = rows[2]!.tally;
|
||||
assert.deepEqual([t.total, t.medianLow, t.medianHigh, t.aboveTop, t.belowBottom, t.covered], [1, 1, 0, 1, 0, 0]);
|
||||
assert.equal(t.placements.aboveQ9, 1);
|
||||
});
|
||||
|
||||
test("diagnosis.json 里的场景都被测到", () => {
|
||||
assert.deepEqual(pick.unused(), []);
|
||||
});
|
||||
@@ -2,6 +2,7 @@ import assert from "node:assert/strict";
|
||||
import { test, type TestContext } from "node:test";
|
||||
import {
|
||||
ablationOutcome,
|
||||
diagnosisOutcome,
|
||||
protocolBody,
|
||||
receive,
|
||||
verificationOutcome,
|
||||
@@ -442,3 +443,138 @@ test("消融实现的代码存不成版本：退回", async () => {
|
||||
});
|
||||
assert.match(reason, /代码存不成版本：bin\/x（符号链接/);
|
||||
});
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// 诊断（方法.md 4.3 诊 1–诊 4）
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
const DX_EXPERIMENT = (name: string, missing: "recordMissing" | "invalidateRun" = "recordMissing", ids: string[] = []) => ({
|
||||
name,
|
||||
spot: "solve.py 的起点生成",
|
||||
body: "换成更强的起点，交出答案",
|
||||
metrics: ids.map((id) => ({ ...metric(id), missing })),
|
||||
});
|
||||
const DX_PROPOSAL = (over: Partial<Submission<"diagnosisProposal">> = {}): Submission<"diagnosisProposal"> => ({
|
||||
explanations: [{ body: "卡在起点" }, { body: "卡在收尾" }],
|
||||
experiments: [DX_EXPERIMENT("强起点"), DX_EXPERIMENT("拿掉收尾", "recordMissing", ["raw-score"])],
|
||||
...over,
|
||||
});
|
||||
const DX_LIST: ChecklistOf<"diagnosisProposal"> = { kind: "diagnosisProposal", maxExplanations: 2, maxExperiments: 2, takenMetricIds: ["sum-of-radii", "answer-path"] };
|
||||
const dxProposal = (content: Submission<"diagnosisProposal">): Incoming => ({ kind: "diagnosisProposal", content, checklist: DX_LIST, known: null, jevModel: MODEL });
|
||||
|
||||
test("诊断解释与实验：解释 1 至 M 条、实验 1 至 A 个；超过就退回", async () => {
|
||||
const many = DX_PROPOSAL({ explanations: [{ body: "a" }, { body: "b" }, { body: "c" }], experiments: [DX_EXPERIMENT("1"), DX_EXPERIMENT("2"), DX_EXPERIMENT("3")] });
|
||||
const reason = await rejected(dxProposal(many));
|
||||
assert.match(reason, /交了 3 条诊断解释，最多 2 条/);
|
||||
assert.match(reason, /交了 3 个诊断实验，最多 2 个/);
|
||||
});
|
||||
|
||||
test("诊断解释与实验：登记的指标只能是 recordMissing；同一个实验里不重名、不占用已有编号", async () => {
|
||||
const reason = await rejected(
|
||||
dxProposal(
|
||||
DX_PROPOSAL({
|
||||
experiments: [
|
||||
DX_EXPERIMENT("甲", "invalidateRun", ["raw-score"]),
|
||||
{ ...DX_EXPERIMENT("乙"), metrics: [metric("a"), metric("a"), metric("answer-path")] },
|
||||
],
|
||||
}),
|
||||
),
|
||||
);
|
||||
assert.match(reason, /第 1 个实验的指标 raw-score 选了 invalidateRun：诊断实验里的指标只能选 recordMissing/);
|
||||
assert.match(reason, /第 2 个实验的 metrics 里编号 a 写了不止一次/);
|
||||
assert.match(reason, /第 2 个实验的指标编号 answer-path 已被占用/);
|
||||
});
|
||||
|
||||
test("诊断解释与实验：对每个实验问冻结三问；两个实验用同一个指标短名没问题（冻结时各带自己实验的前缀）", async (t) => {
|
||||
fakeJev(t, () => false);
|
||||
const v = await receive(dxProposal(DX_PROPOSAL({ experiments: [DX_EXPERIMENT("甲", "recordMissing", ["x"]), DX_EXPERIMENT("乙", "recordMissing", ["x"])] })));
|
||||
assert.equal(v.accepted, true);
|
||||
assert.equal(v.jev!.length, 6, "两个实验各三问");
|
||||
assert.deepEqual(v.jev!.map((j) => j.question), ["self_contradictory", "changes_measurement", "too_vague", "self_contradictory", "changes_measurement", "too_vague"]);
|
||||
});
|
||||
|
||||
test("诊断解释与实验：Jev 判是就退回，原因写明第几个实验和哪一问", async (t) => {
|
||||
fakeJev(t, (q) => q === "too_vague");
|
||||
const v = await receive(dxProposal(DX_PROPOSAL()));
|
||||
assert.equal(v.accepted, false);
|
||||
const reason = (v as Extract<Verdict, { accepted: false }>).reason;
|
||||
assert.match(reason, /Jev 判出下面几处有问题/);
|
||||
assert.match(reason, /- 第 1 个诊断实验（强起点）：too_vague 概率 0\.99/);
|
||||
assert.match(reason, /- 第 2 个诊断实验（拿掉收尾）：too_vague 概率 0\.99/);
|
||||
});
|
||||
|
||||
const DX_PRED_LIST: ChecklistOf<"diagnosisPrediction"> = { kind: "diagnosisPrediction", explanations: ["r1-dx1", "r1-dx2"], experiments: ["d1", "d2"] };
|
||||
const dxPred = (predictions: Array<[string, string, ReturnType<typeof r>]>): Incoming => ({
|
||||
kind: "diagnosisPrediction",
|
||||
content: { predictions: predictions.map(([explanation, experiment, range]) => ({ explanation, experiment, range })), body: "理由" },
|
||||
checklist: DX_PRED_LIST,
|
||||
known: null,
|
||||
jevModel: MODEL,
|
||||
});
|
||||
|
||||
test("诊断预测：自己每条解释 × 每个合并后的实验，不多不少；收下时不问 Jev", async () => {
|
||||
const all: Array<[string, string, ReturnType<typeof r>]> = [
|
||||
["r1-dx1", "d1", r(1, 1.1, 1.2)], ["r1-dx1", "d2", r(1, 1.1, 1.2)], ["r1-dx2", "d1", r(1, 1.1, 1.2)], ["r1-dx2", "d2", r(1, 1.1, 1.2)],
|
||||
];
|
||||
assert.deepEqual(await receive(dxPred(all)), { accepted: true, jev: null, protocol: null });
|
||||
const reason = await rejected(dxPred([all[0]!, all[0]!, all[1]!, ["r1-dx3", "d1", r(1, 1.1, 1.2)]]));
|
||||
assert.match(reason, /predictions 的（解释:实验）应当是 r1-dx1:d1、r1-dx1:d2、r1-dx2:d1、r1-dx2:d2/);
|
||||
assert.match(reason, /你写的 r1-dx3:d1 不在里面/);
|
||||
assert.match(reason, /还缺 r1-dx2:d1、r1-dx2:d2/);
|
||||
assert.match(reason, /r1-dx1:d1 写了不止一次/);
|
||||
});
|
||||
|
||||
test("诊断预测：三个数要严格递增", async () => {
|
||||
const reason = await rejected(dxPred([["r1-dx1", "d1", r(1, 1, 1.2)], ["r1-dx1", "d2", r(1, 1.1, 1.2)], ["r1-dx2", "d1", r(1, 1.1, 1.2)], ["r1-dx2", "d2", r(1.3, 1.2, 1.1)]]));
|
||||
assert.match(reason, /r1-dx1:d1 的 range 的三个数要严格递增 q1 < q5 < q9，你写的是 1 \/ 1 \/ 1\.2/);
|
||||
assert.match(reason, /r1-dx2:d2 的 range 的三个数要严格递增/);
|
||||
});
|
||||
|
||||
const DX_CODE_BODY = { locations: [{ experiment: "d1", locations: ["solve.py:3-9"] }], concerns: "无", body: "无" };
|
||||
type DxCode = Extract<Incoming, { kind: "diagnosisImplementation" }>["known"]["code"];
|
||||
const dxCode = (kind: "diagnosisImplementation" | "diagnosisRepair", code: DxCode): Incoming =>
|
||||
kind === "diagnosisImplementation"
|
||||
? { kind, content: DX_CODE_BODY, checklist: { kind, experiments: ["d1"] }, known: { code }, jevModel: MODEL }
|
||||
: { kind, content: DX_CODE_BODY, checklist: { kind, experiments: ["d1"] }, known: { code }, jevModel: MODEL };
|
||||
|
||||
test("诊断实现、修复：改了不许改的文件、代码存不了就退回；有差异块时每块问一次五问", async (t) => {
|
||||
for (const kind of ["diagnosisImplementation", "diagnosisRepair"] as const) {
|
||||
assert.match(await rejected(dxCode(kind, { ...SAVED, protectedFiles: ["validate.py"] })), /改了不许改的文件：validate\.py/);
|
||||
assert.match(await rejected(dxCode(kind, { saved: false, path: "bin/x", reason: "符号链接" })), /代码存不成版本：bin\/x（符号链接）/);
|
||||
}
|
||||
fakeJev(t, () => false);
|
||||
const hunks = [
|
||||
{ path: "solve.py", change: "modified" as const, text: "@@ -1 +1 @@\n-a\n+b" },
|
||||
{ path: "solve.py", change: "modified" as const, text: "@@ -9 +9 @@\n-c\n+d" },
|
||||
];
|
||||
const v = await receive(dxCode("diagnosisRepair", { ...SAVED, hunks }));
|
||||
assert.equal(v.accepted, true);
|
||||
assert.equal(v.jev!.length, 10, "两块各五问");
|
||||
});
|
||||
|
||||
test("诊断核验：逐个实验的结果覆盖全部选中的实验", async () => {
|
||||
const verification = (ids: string[]): Incoming => ({
|
||||
kind: "diagnosisVerification",
|
||||
content: { experiments: ids.map((experiment) => ({ experiment, followed: true, nothingElse: true, where: "solve.py:3" })), body: "无" },
|
||||
checklist: { kind: "diagnosisVerification", experiments: ["d1", "d2"] },
|
||||
known: null,
|
||||
jevModel: MODEL,
|
||||
});
|
||||
assert.deepEqual(await receive(verification(["d2", "d1"])), { accepted: true, jev: null, protocol: null });
|
||||
const reason = await rejected(verification(["d1", "d3"]));
|
||||
assert.match(reason, /experiments 的实验应当是 d1、d2；你写的 d3 不在里面；还缺 d2/);
|
||||
});
|
||||
|
||||
test("诊断核验通过没有：逐个实验判，一个不通过不影响别的实验", () => {
|
||||
const v: Submission<"diagnosisVerification"> = {
|
||||
experiments: [
|
||||
{ experiment: "d1", followed: true, nothingElse: true, where: "a" },
|
||||
{ experiment: "d2", followed: false, nothingElse: true, where: "没找到" },
|
||||
{ experiment: "d3", followed: true, nothingElse: false, where: "顺手改了别处" },
|
||||
{ experiment: "d4", followed: false, nothingElse: false, where: "都不对" },
|
||||
],
|
||||
body: "无",
|
||||
};
|
||||
assert.deepEqual(diagnosisOutcome(v), { passed: false, failing: ["d2", "d3", "d4"], notFollowed: ["d2", "d4"], touchedElse: ["d3", "d4"] });
|
||||
assert.deepEqual(diagnosisOutcome({ experiments: [v.experiments[0]!], body: "无" }), { passed: true, failing: [], notFollowed: [], touchedElse: [] });
|
||||
});
|
||||
|
||||
@@ -0,0 +1,254 @@
|
||||
import assert from "node:assert/strict";
|
||||
import { spawnSync } from "node:child_process";
|
||||
import { cpSync, existsSync, mkdirSync, mkdtempSync, readFileSync, rmSync, statSync, writeFileSync } from "node:fs";
|
||||
import { tmpdir } from "node:os";
|
||||
import { join } from "node:path";
|
||||
import { after, test, type TestContext } from "node:test";
|
||||
import { Runner } from "../src/runner/index.ts";
|
||||
import {
|
||||
diagnosisRun,
|
||||
diagnosisUnversionedNames,
|
||||
judgeDiagnosis,
|
||||
loadTask,
|
||||
placeBaselineArtifact,
|
||||
prepareDiagnosisJudging,
|
||||
prepareDiagnosisRun,
|
||||
type DiagnosisExperiment,
|
||||
type DiagnosisJudged,
|
||||
} from "../src/tasks/index.ts";
|
||||
import type { HypothesisMetric, Measured, ProcessResult, RunDefinition } from "../src/vocab/index.ts";
|
||||
|
||||
// 诊断运行的测试（任务书 18 第 1.4 节）：一次运行跑几个实验，经真实 Docker 在任务镜像里跑；没有 Docker 时失败，不跳过。
|
||||
// 镜像和容器用 inforsi-test-tasks-diag 开头的固定名字，与别的测试文件错开。
|
||||
// 判分的预期值是手算的 2 的幂有理数布局（同 tests/tasks.test.ts）：宽松布局半径和 26/128 = 0.203125，
|
||||
// 相切布局 26/64 = 0.40625。
|
||||
|
||||
const TASK_DIR = join(import.meta.dirname, "..", "tasks", "circle_packing");
|
||||
const IMAGE_TAG = "inforsi-test-tasks-diag";
|
||||
/** 测试用的总时限：6 个实验各分到 4 秒。 */
|
||||
const TOTAL_SECONDS = 24;
|
||||
|
||||
/** 任务目录抄一份，只把总时限改小，其余（Dockerfile、校验器、harness）原样，镜像编号也就相同。 */
|
||||
function shortTask() {
|
||||
const dir = mkdtempSync(join(tmpdir(), "inforsi-tasks-diag-task-"));
|
||||
cpSync(TASK_DIR, dir, { recursive: true });
|
||||
const facts = JSON.parse(readFileSync(join(dir, "task.json"), "utf8"));
|
||||
facts.run.wallSeconds = TOTAL_SECONDS;
|
||||
writeFileSync(join(dir, "task.json"), JSON.stringify(facts));
|
||||
return loadTask(dir);
|
||||
}
|
||||
|
||||
const task = shortTask();
|
||||
const runner = new Runner(null);
|
||||
const digest = await runner.buildImage(task.dir, IMAGE_TAG);
|
||||
after(() => {
|
||||
rmSync(task.dir, { recursive: true, force: true });
|
||||
docker(["rmi", IMAGE_TAG]);
|
||||
});
|
||||
|
||||
function docker(args: string[]): { status: number | null; stderr: string } {
|
||||
return spawnSync("docker", args, { encoding: "utf8" });
|
||||
}
|
||||
|
||||
function forceRemove(name: string): void {
|
||||
const removed = docker(["rm", "-f", name]);
|
||||
if (removed.status !== 0 && !removed.stderr.includes("No such container")) throw new Error(removed.stderr);
|
||||
}
|
||||
|
||||
function tempDir(t: TestContext, label: string): string {
|
||||
const dir = mkdtempSync(join(tmpdir(), `inforsi-tasks-diag-${label}-`));
|
||||
t.after(() => rmSync(dir, { recursive: true, force: true }));
|
||||
return dir;
|
||||
}
|
||||
|
||||
async function run(t: TestContext, suffix: string, definition: RunDefinition, dir: string): Promise<ProcessResult> {
|
||||
const name = `inforsi-test-tasks-diag-${suffix}`;
|
||||
forceRemove(name);
|
||||
t.after(() => forceRemove(name));
|
||||
const outcome = await runner.start(name, definition, dir);
|
||||
assert.deepEqual(outcome, { started: true });
|
||||
return runner.wait(name, tempDir(t, "log"));
|
||||
}
|
||||
|
||||
const m = (value: number): Measured => ({ status: "measured", value });
|
||||
|
||||
function rowLayout(radius: number, spacing: number, firstX: number): object {
|
||||
const n = 26;
|
||||
return {
|
||||
centers: Array.from({ length: n }, (_, i) => [firstX + i * spacing, 0.5]),
|
||||
radii: Array(n).fill(radius),
|
||||
reported_sum: 99,
|
||||
};
|
||||
}
|
||||
|
||||
const metric = (id: string): HypothesisMetric => ({ id, meaning: "测试用", unit: "无", direction: "maximize", howMeasured: "测试用", missing: "recordMissing" });
|
||||
|
||||
/** 六个实验：分别测"拿到了原始产物""指标不串""崩了不影响后面""超时""原始产物只读""答案不合法"。 */
|
||||
const EXPERIMENTS: DiagnosisExperiment[] = [
|
||||
{ id: "d-loose", metrics: [metric("d-loose.note")] },
|
||||
{ id: "d-same", metrics: [metric("d-same.note")] },
|
||||
{ id: "d-crash", metrics: [] },
|
||||
{ id: "d-slow", metrics: [] },
|
||||
{ id: "d-tamper", metrics: [] },
|
||||
{ id: "d-outside", metrics: [] },
|
||||
];
|
||||
|
||||
const PROGRAM = `import json
|
||||
import time
|
||||
|
||||
|
||||
def layout(radius, spacing, first_x):
|
||||
return ([[first_x + i * spacing, 0.5] for i in range(26)], [radius] * 26, 99.0)
|
||||
|
||||
|
||||
def run_experiment(experiment_id, baseline_dir):
|
||||
if experiment_id == "d-loose":
|
||||
with open("metrics.json", "w") as handle:
|
||||
json.dump({"d-loose.note": 0.5}, handle)
|
||||
return layout(1 / 128, 1 / 32, 7 / 64)
|
||||
if experiment_id == "d-same":
|
||||
with open(baseline_dir + "/solution.json") as handle:
|
||||
solution = json.load(handle)
|
||||
return solution["centers"], solution["radii"], solution["reported_sum"]
|
||||
if experiment_id == "d-crash":
|
||||
raise RuntimeError("故意崩")
|
||||
if experiment_id == "d-slow":
|
||||
time.sleep(60)
|
||||
if experiment_id == "d-tamper":
|
||||
with open(baseline_dir + "/solution.json", "w") as handle:
|
||||
handle.write("{}")
|
||||
if experiment_id == "d-outside":
|
||||
return layout(1 / 128, 1 / 32, 2.0)
|
||||
raise RuntimeError("不认识的实验 " + experiment_id)
|
||||
`;
|
||||
|
||||
/** 基线那次正式运行的运行目录：只要有 solution.json。放的是相切布局，半径和 0.40625。 */
|
||||
function baselineRunDir(t: TestContext): string {
|
||||
const dir = tempDir(t, "baseline-run");
|
||||
writeFileSync(join(dir, "solution.json"), JSON.stringify(rowLayout(1 / 64, 1 / 32, 7 / 64)));
|
||||
return dir;
|
||||
}
|
||||
|
||||
function programDir(t: TestContext): string {
|
||||
const dir = tempDir(t, "program");
|
||||
writeFileSync(join(dir, "initial.py"), PROGRAM);
|
||||
return dir;
|
||||
}
|
||||
|
||||
function value(item: Measured | undefined): number {
|
||||
assert.equal(item?.status, "measured", JSON.stringify(item));
|
||||
return (item as { value: number }).value;
|
||||
}
|
||||
|
||||
function reason(item: Measured | undefined): string {
|
||||
assert.equal(item?.status, "unmeasured", JSON.stringify(item));
|
||||
return (item as { reason: string }).reason;
|
||||
}
|
||||
|
||||
test("运行定义：一个容器，总时限和核数照任务，实验按顺序、每个分到总时限 ÷ 实验数，种子经环境变量交给 harness", () => {
|
||||
const definition = diagnosisRun(task, digest, "seed-x", ["a", "b", "c"]);
|
||||
assert.equal(definition.wallSeconds, TOTAL_SECONDS);
|
||||
assert.equal(definition.cores, 1);
|
||||
assert.equal(definition.env.INFORSI_SEED, "seed-x");
|
||||
assert.deepEqual(definition.command, [
|
||||
"python", "-B", "/work/harness-diagnosis.py", "--program", "initial.py", "--baseline", "baseline-artifact",
|
||||
"--answers", "answers", "--seconds-each", "8", "--report", "diagnosis.json", "--experiments", "a", "b", "c",
|
||||
]);
|
||||
});
|
||||
|
||||
test("原始产物：只读地放进实验员工作目录，存版本时被跳过；运行目录里同样有，并带全部实验登记的指标说明", (t) => {
|
||||
assert.deepEqual(diagnosisUnversionedNames(task), [...task.facts.unversionedNames, "baseline-artifact"]);
|
||||
const workDir = tempDir(t, "work");
|
||||
placeBaselineArtifact(task, baselineRunDir(t), workDir);
|
||||
const placed = join(workDir, "baseline-artifact", "solution.json");
|
||||
assert.equal(statSync(placed).mode & 0o777, 0o444);
|
||||
assert.deepEqual(JSON.parse(readFileSync(placed, "utf8")), rowLayout(1 / 64, 1 / 32, 7 / 64));
|
||||
|
||||
const runDir = tempDir(t, "run");
|
||||
prepareDiagnosisRun(task, programDir(t), baselineRunDir(t), runDir, EXPERIMENTS);
|
||||
for (const name of ["initial.py", "harness.py", "harness-diagnosis.py", "baseline-artifact/solution.json", "metrics-spec.json"]) {
|
||||
assert.ok(existsSync(join(runDir, name)), name);
|
||||
}
|
||||
assert.ok(statSync(join(runDir, "answers")).isDirectory());
|
||||
const spec = JSON.parse(readFileSync(join(runDir, "metrics-spec.json"), "utf8")) as HypothesisMetric[];
|
||||
assert.deepEqual(spec.map((s) => s.id), ["d-loose.note", "d-same.note"]);
|
||||
});
|
||||
|
||||
test("从头到尾：六个实验一个容器按顺序跑，崩了、超时、改只读文件的各记各的，不影响别人；逐个单独判分", async (t) => {
|
||||
const baseline = baselineRunDir(t);
|
||||
const runDir = tempDir(t, "e2e-run");
|
||||
prepareDiagnosisRun(task, programDir(t), baseline, runDir, EXPERIMENTS);
|
||||
const ids = EXPERIMENTS.map((e) => e.id);
|
||||
const process = await run(t, "e2e", diagnosisRun(task, digest, "inforsi-test-seed", ids), runDir);
|
||||
assert.equal(process.ending, "exited", JSON.stringify(process));
|
||||
assert.equal((process as { exitCode: number }).exitCode, 0);
|
||||
|
||||
const judgeDir = tempDir(t, "e2e-judge");
|
||||
const step = prepareDiagnosisJudging(task, runDir, judgeDir, digest, process, EXPERIMENTS);
|
||||
assert.equal(step.status, "validate");
|
||||
if (step.status !== "validate") return;
|
||||
const validator = await run(t, "e2e-validate", step.definition, judgeDir);
|
||||
const judged = judgeDiagnosis(task, runDir, judgeDir, process, validator, EXPERIMENTS);
|
||||
const by = Object.fromEntries(judged.experiments.map((e) => [e.id, e]));
|
||||
assert.deepEqual(judged.experiments.map((e) => e.id), ids);
|
||||
|
||||
assert.equal(by["d-loose"]!.ending, "finished");
|
||||
assert.equal(by["d-loose"]!.verdict, "valid");
|
||||
assert.equal(value(by["d-loose"]!.score), 0.203125);
|
||||
assert.equal(value(by["d-loose"]!.metrics["d-loose.note"]), 0.5);
|
||||
assert.equal(value(by["d-loose"]!.metrics["legalization-deduction"]), 0);
|
||||
assert.ok(value(by["d-loose"]!.wallSeconds) > 0 && value(by["d-loose"]!.wallSeconds) < 4);
|
||||
|
||||
// 读到了原始产物：原样返回，成绩就是那份产物的半径和。指标没写，记测不到，也没有串进上一个实验写的值。
|
||||
assert.equal(by["d-same"]!.verdict, "valid");
|
||||
assert.equal(value(by["d-same"]!.score), 0.40625);
|
||||
assert.match(reason(by["d-same"]!.metrics["d-same.note"]), /metrics\.json 不存在/);
|
||||
|
||||
assert.equal(by["d-crash"]!.ending, "crashed");
|
||||
assert.equal(by["d-crash"]!.verdict, "invalid");
|
||||
assert.match(reason(by["d-crash"]!.score), /非零退出码/);
|
||||
assert.match(reason(by["d-crash"]!.metrics["raw-score"]), /没有跑校验器/);
|
||||
|
||||
// 超时：被杀在分到的 4 秒上，用时不短于 4 秒；后面的实验照常跑完。
|
||||
assert.equal(by["d-slow"]!.ending, "timeout");
|
||||
assert.match(by["d-slow"]!.verdictReason, /超过分到的时间 4 秒/);
|
||||
assert.ok(value(by["d-slow"]!.wallSeconds) >= 4 && value(by["d-slow"]!.wallSeconds) < 8);
|
||||
|
||||
assert.equal(by["d-tamper"]!.ending, "crashed");
|
||||
assert.deepEqual(JSON.parse(readFileSync(join(runDir, "baseline-artifact", "solution.json"), "utf8")), rowLayout(1 / 64, 1 / 32, 7 / 64));
|
||||
|
||||
assert.equal(by["d-outside"]!.ending, "finished");
|
||||
assert.equal(by["d-outside"]!.verdict, "invalid");
|
||||
assert.match(by["d-outside"]!.verdictReason, /^校验器判不合法/);
|
||||
assert.match(reason(by["d-outside"]!.score), /^不算数/);
|
||||
});
|
||||
|
||||
/** 手写一次"只跑了一部分"的运行目录：report 里只有头几项。 */
|
||||
function partialRun(t: TestContext, report: object[]): string {
|
||||
const dir = tempDir(t, "partial");
|
||||
writeFileSync(join(dir, "diagnosis.json"), JSON.stringify(report));
|
||||
return dir;
|
||||
}
|
||||
|
||||
test("整次运行被总时限杀掉：报告里没有的实验记没轮到，写明整次运行的结局；已有记录的照常", (t) => {
|
||||
const timeout: ProcessResult = { ending: "timeout", cpuSeconds: m(23), wallSeconds: m(24), peakMemoryBytes: m(5) };
|
||||
const two = EXPERIMENTS.slice(0, 2);
|
||||
const runDir = partialRun(t, [{ id: "d-loose", status: "timeout", wall_seconds: 4.01 }]);
|
||||
const step = prepareDiagnosisJudging(task, runDir, tempDir(t, "judge"), digest, timeout, two);
|
||||
assert.equal(step.status, "judged");
|
||||
const judged = (step as { result: DiagnosisJudged }).result;
|
||||
assert.equal(judged.experiments[0]!.ending, "timeout");
|
||||
assert.equal(value(judged.experiments[0]!.wallSeconds), 4.01);
|
||||
assert.equal(judged.experiments[1]!.ending, "notRun");
|
||||
assert.match(judged.experiments[1]!.verdictReason, /没轮到：整次诊断运行超过单次时限 24 秒/);
|
||||
assert.match(reason(judged.experiments[1]!.wallSeconds), /没轮到/);
|
||||
// 没有交出答案的实验，judgeDiagnosis 不要校验器的结果。
|
||||
assert.deepEqual(judgeDiagnosis(task, runDir, tempDir(t, "judge2"), timeout, null, two), judged);
|
||||
});
|
||||
|
||||
test("运行正常退出、报告里却缺实验：报错停下，不当成候选的失败", (t) => {
|
||||
const exited: ProcessResult = { ending: "exited", exitCode: 0, cpuSeconds: m(1), wallSeconds: m(2), peakMemoryBytes: m(5) };
|
||||
const runDir = partialRun(t, []);
|
||||
mkdirSync(join(runDir, "answers"));
|
||||
assert.throws(() => prepareDiagnosisJudging(task, runDir, tempDir(t, "judge"), digest, exited, EXPERIMENTS), /没有实验 d-loose/);
|
||||
});
|
||||
@@ -102,6 +102,14 @@ const samples: { [K in SubmissionKind]: Submission<K> } = {
|
||||
],
|
||||
},
|
||||
claim: { choices: ["researcher-2-e1", "researcher-1-e2"] },
|
||||
diagnosisProposal: {
|
||||
explanations: [{ body: "卡在起点：……；若换成明显更强的起点分数不涨，这条解释就不成立" }],
|
||||
experiments: [{ name: "更强的起点", spot: "solve.py 的起点生成", body: "换成……，交出答案", metrics: metrics.defined }],
|
||||
},
|
||||
diagnosisPrediction: { predictions: [{ explanation: "r2-dx1", experiment: "d1", range }], body: "理由" },
|
||||
diagnosisImplementation: { locations: [{ experiment: "d1", locations: ["solve.py:10-40"] }], concerns: "无", body: "说明" },
|
||||
diagnosisRepair: { locations: [{ experiment: "d1", locations: ["solve.py:10-40"] }], concerns: "无", body: "按核验意见补上" },
|
||||
diagnosisVerification: { experiments: [{ experiment: "d1", followed: true, nothingElse: true, where: "solve.py:10-40" }], body: "无" },
|
||||
proposal: { factors, safeguards: ["final-shrink"], metrics, authorEstimate: range, body },
|
||||
firstPrediction: { range, body: "理由" },
|
||||
critique: { target: "p2", body: "（推测）扰动可能和起点数混杂" },
|
||||
|
||||
@@ -260,3 +260,23 @@ test("比较的行号：改一行、加几行、只删不加，新旧两边的
|
||||
// 给模型看的差异原文带 3 行上下文：三处改动相隔不到 6 行，合成一块。
|
||||
assert.match(result.diff, /^@@ -1,10 \+1,12 @@\n a\n-b\n\+B\n c\n d\n e\n\+X\n\+Y\n\+Z\n f\n g\n-h\n i\n j\n$/m);
|
||||
});
|
||||
|
||||
test("诊断程序（任务书 18）：基线代码取出后放进只读的原始产物、改成诊断程序再存，产物被跳过，比较只看程序的改动，取出再存号不变", (t) => {
|
||||
const repo = newRepo(t);
|
||||
const baseline = save(repo, makeDir(t, { "initial.py": "a\nb\n", "evaluate.py": "e\n" }), []);
|
||||
const workDir = checkout(t, repo, baseline);
|
||||
// 原始产物放进实验员工作目录，设成只读（0444），与 tasks 的 placeBaselineArtifact 一样。
|
||||
mkdirSync(join(workDir, "baseline-artifact"));
|
||||
writeFileSync(join(workDir, "baseline-artifact", "solution.json"), "{}");
|
||||
chmodSync(join(workDir, "baseline-artifact", "solution.json"), 0o444);
|
||||
writeFileSync(join(workDir, "initial.py"), "a\nb\nexperiment\n");
|
||||
const excluded = ["__pycache__", "solution.json", "baseline-artifact"];
|
||||
const program = save(repo, workDir, excluded);
|
||||
assert.notEqual(program, baseline);
|
||||
const comparison = compareVersions(repo, baseline, program);
|
||||
assert.deepEqual(comparison.files.map((f) => f.path), ["initial.py"]);
|
||||
assert.deepEqual((comparison.files[0] as Extract<FileChange, { binary: false }>).hunks, [{ from: { after: 2 }, to: { start: 3, end: 3 } }]);
|
||||
const again = checkout(t, repo, program);
|
||||
assert.equal(existsSync(join(again, "baseline-artifact")), false);
|
||||
assert.equal(save(repo, again, excluded), program);
|
||||
});
|
||||
|
||||
Reference in New Issue
Block a user