iomgaa
144b88480b
- ⑬–⑱：listCells 从采纳决定和冻结协议现算（不写记录）；explanation 有消融才开 阶段；selectCells 照 CellRuleInput 拼输入（余额当时现算、decisionSeed 版本 2）， 写种类 cells 的决定，没选中的格不另写事件；消融实现/核验按第几次编号 （r<轮>-ablationImplementation-1、verification-1、-2……从账本数），重做从上一次 这一格代码接着改，次数用完写 cellNotRunEvent 不跑不扣次；cannotRun 传 null 不核验 不运行不扣次；alsoOff 用 sharedSettings 沿用同设置的第一格；消融运行一个接一个跑， 用途 ablation，登记/扣次/恢复与正向同一入口，恢复时指标缺失策略覆盖为只记缺失 （方法.md 第五部分）；settle 拼 SettleInput 调 settleAblation，结算与 chooseBet 的 bet 决定写在同一事务；wrapUp 阶段关闭后，本轮选中的每个候选在同一个事务里解封 作者自评（取最后一版提案的封存）。 - 一轮接一轮：每轮开头用 runBalance 现算余额，余额为 0 不开新轮；走到 export 停下， 原因"导出还没写"。去掉 roundsNext 里 step.id !== "listCells" 的临时写法。 - 消融中发现由 materials 的 Board.baselineOf 从结算输出读出，flow 不另写采纳记录 （已核实 materials 已实现）。 - 测试 +14：下一步分支 8 条（含收尾关闭后才解封、余额 0 走到导出停、关掉消融时 解释照方法停下）、消融步骤 6 条（选格输入逐字段手拼对账、结算四种结局、同设置 沿用、解封取最后一版封存）；npm run check 通过，npm test 472 条通过。 - 全新上下文审阅一轮（只看消融重做循环、登记扣次恢复、选格/结算/押的因子输入、 解封时机、下一轮与结束条件）：没有会污染账本的问题；采纳一条小修（删掉未用的 导入），其余笔记照列不改。 flow 1302 → 1662 行（上限 3500），materials 未动（1930 行）。
208 lines
10 KiB
TypeScript
208 lines
10 KiB
TypeScript
import assert from "node:assert/strict";
|
||
import { readFileSync } 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 { selectCells, settle, unsealChosen } from "../src/flow/steps/ablation.ts";
|
||
import type { Study } from "../src/flow/study.ts";
|
||
import { Ledger } from "../src/ledger/index.ts";
|
||
import { Board } from "../src/materials/index.ts";
|
||
import { chooseBet, decisionSeed, runCellRule, settleAblation, type AblationModelInput, type CellRuleInput, type SettleInput } from "../src/rules/index.ts";
|
||
import { checkTaskFacts } from "../src/tasks/index.ts";
|
||
import { fullConfig } from "./config-helpers.ts";
|
||
import { definition, exited, proposal, tempDir } from "./ledger-helpers.ts";
|
||
import { judged, r, Writer } from "./materials-helpers.ts";
|
||
|
||
// 选格、结算与押的因子、轮末解封（任务书 17 第 2.2 节）：在真实账本里写到上一步做完，跑这一步，看写下的记录。
|
||
// 预期值按条文手拼：选格输入照 rules 的 CellRuleInput 逐项拼，结算的各格结局照方法.md 6.6 的情形拼；
|
||
// 输出用 rules 的函数当对照（同选批测试的写法，rules 自己有独立预期值的测试）。
|
||
|
||
const config = fullConfig();
|
||
const facts = checkTaskFacts(JSON.parse(readFileSync(join(import.meta.dirname, "..", "tasks", "circle_packing", "task.json"), "utf8")));
|
||
|
||
function study(t: TestContext): { s: Study; w: Writer } {
|
||
const ledger = Ledger.open(join(tempDir(t), "ledger.db"));
|
||
t.after(() => ledger.close());
|
||
const frozen = { config: checkedFull(config), steps: stepList(config), facts, prompts: null, image: "", seed: 1 };
|
||
return { s: { ledger, frozen } as unknown as Study, w: new Writer(ledger) };
|
||
}
|
||
|
||
/** 基线 0.96、被采纳候选 r1-c1（两个因子）完整代码 1.25。 */
|
||
function adoptedBase(w: Writer): void {
|
||
w.l.append("code", { object: "base", version: "v-base", locations: null });
|
||
w.run("base-1", "baseline", 0, null, judged(0.96, "main"));
|
||
w.l.append("code", { object: "r1-c1", version: "v-full", locations: null });
|
||
w.run("r1-forward-r1-c1-1", "forward", 1, "r1-c1", judged(1.25, "main"));
|
||
w.protocol("r1-c1", { factors: [{ name: "因子一" }, { name: "因子二" }], safeguards: [], metrics: { defined: [], reused: [] }, body: "正文", clauses: [], round: 1, baseline: "base", proposals: ["r1-a1-1"] } as never);
|
||
w.decision("adopt", 1, null, { adopted: "r1-c1" });
|
||
}
|
||
|
||
const EXPLANATIONS = [
|
||
{ id: "r1-a1-x1", author: "a1", cells: [{ factor: "p1", range: r(1.0, 1.1, 1.2) }, { factor: "p2", range: r(1.15, 1.2, 1.24) }] },
|
||
{ id: "r1-a2-x1", author: "a2", cells: [{ factor: "p1", range: r(0.9, 1.05, 1.2) }, { factor: "p2", range: r(1.1, 1.18, 1.25) }] },
|
||
] as const;
|
||
|
||
/** 解释阶段关闭：a1、a2 各交一条，对两格都给了三个数。 */
|
||
function explain(w: Writer): void {
|
||
w.open("r1-explanation", "explanation", 1, ["a1", "a2"]);
|
||
for (const e of EXPLANATIONS) {
|
||
w.sub("r1-explanation", e.author, "explanation", null, null, { explanations: [{ body: `机制（${e.author}）`, cells: e.cells.map((c) => ({ factor: c.factor, range: c.range })) }] });
|
||
}
|
||
w.close("r1-explanation");
|
||
}
|
||
|
||
/** 成分模型的输入：选格、结算和押的因子共用的那一份。 */
|
||
function modelInput(): AblationModelInput {
|
||
return { baseline: 0.96, full: 1.25, direction: "maximize", knownBest: facts.knownBest, epsilon: facts.epsilon, lambda: 0.1, factors: ["p1", "p2"], explanations: EXPLANATIONS.map((e) => ({ id: e.id, author: e.author, cells: e.cells.map((c) => ({ factor: c.factor, range: c.range })) })) };
|
||
}
|
||
|
||
test("选格：输入照 CellRuleInput 拼（基线、完整分数、各条解释、余额从运行记录算、种子用选格的版本 2），规则取步骤清单", (t) => {
|
||
const { s, w } = study(t);
|
||
adoptedBase(w);
|
||
explain(w);
|
||
selectCells(s, 1);
|
||
const record = s.ledger.records().find((x) => x.kind === "decision" && x.content.decision === "cells")!;
|
||
const seed = decisionSeed({ studyId: "circle-full-1", round: 1, version: 2 });
|
||
const input: CellRuleInput = {
|
||
...modelInput(),
|
||
// 13 次里基线和正向各确认启动 1 次，余额 11。
|
||
balance: 11,
|
||
batchCap: 2,
|
||
u: 0.5,
|
||
decay: 0.5,
|
||
seed,
|
||
};
|
||
const { rule: used, ...output } = runCellRule("valueMinusCost", input);
|
||
assert.deepEqual(record.content, { decision: "cells", round: 1, rule: used.name, ruleVersion: used.version, params: output.params, seed, input, output });
|
||
});
|
||
|
||
/** 一次消融实现阶段：subs 是 [格, cannotRun, alsoOff]；cannotRun 为 null 表示交了代码。 */
|
||
function impl(w: Writer, n: number, subs: Array<[string, string | null, string[]?]>): void {
|
||
w.open(`r1-ablationImplementation-${n}`, "ablationImplementation", 1, subs.map(([cell]) => `experimenter-${cell}`));
|
||
for (const [cell, cannotRun, alsoOff] of subs) {
|
||
w.sub(`r1-ablationImplementation-${n}`, `experimenter-${cell}`, "ablationImplementation", cell, null, { alsoOff: alsoOff ?? [], cannotRun, concerns: "无" });
|
||
}
|
||
w.close(`r1-ablationImplementation-${n}`);
|
||
}
|
||
|
||
/** 结算一次：拼出 SettleInput，跑 settle，核对 settlement 和 bet 两条记录。 */
|
||
function settled(t: TestContext, prepare: (w: Writer) => void, cells: SettleInput["cells"]): void {
|
||
const { s, w } = study(t);
|
||
adoptedBase(w);
|
||
explain(w);
|
||
w.decision("cells", 1, null, { chosen: ["p1", "p2"] });
|
||
prepare(w);
|
||
settle(s, 1);
|
||
const input: SettleInput = { ...modelInput(), cells };
|
||
const betInput = { ...modelInput(), cells };
|
||
const settlement = s.ledger.records().find((x) => x.kind === "settlement")!;
|
||
assert.deepEqual(settlement.content, { round: 1, result: { candidate: "r1-c1", input, output: settleAblation(input) } });
|
||
const bet = s.ledger.records().find((x) => x.kind === "decision" && x.content.decision === "bet")!;
|
||
assert.deepEqual(bet.content, {
|
||
decision: "bet",
|
||
round: 1,
|
||
rule: "chooseBet",
|
||
ruleVersion: 1,
|
||
params: { direction: "maximize", epsilon: facts.epsilon, knownBest: facts.knownBest, lambda: 0.1 },
|
||
seed: null,
|
||
input: betInput,
|
||
output: chooseBet(betInput),
|
||
});
|
||
}
|
||
|
||
test("结算：有效分数、没测、跑不起来、超时记证据不足，各按 6.6 的结局拼", (t) => {
|
||
settled(
|
||
t,
|
||
(w) => {
|
||
impl(w, 1, [["r1-c1-p1", null], ["r1-c1-p2", null]]);
|
||
w.run("r1-ablation-r1-c1-p1-1", "ablation", 1, "r1-c1-p1", judged(1.1, "main"));
|
||
// p2 启动了但跑崩： judged 为 null，记证据不足（扣了次数）。
|
||
w.l.append("runPlanned", { run: "r1-ablation-r1-c1-p2-1", purpose: "ablation", round: 1, object: "r1-c1-p2", definition });
|
||
w.l.append("runStarted", { run: "r1-ablation-r1-c1-p2-1" });
|
||
w.l.append("runResult", { run: "r1-ablation-r1-c1-p2-1", process: exited, judged: null });
|
||
},
|
||
[
|
||
{ factor: "p1", outcome: { kind: "valid", score: 1.1 } },
|
||
{ factor: "p2", outcome: { kind: "insufficient" } },
|
||
],
|
||
);
|
||
});
|
||
|
||
test("结算：没选中的格记没测；交了跑不起来的记跑不起来", (t) => {
|
||
const { s, w } = study(t);
|
||
adoptedBase(w);
|
||
explain(w);
|
||
w.decision("cells", 1, null, { chosen: ["p1"] });
|
||
impl(w, 1, [["r1-c1-p1", "离了 p1 跑不起来"]]);
|
||
settle(s, 1);
|
||
const input: SettleInput = {
|
||
...modelInput(),
|
||
cells: [
|
||
{ factor: "p1", outcome: { kind: "cannotRun" } },
|
||
{ factor: "p2", outcome: { kind: "untested" } },
|
||
],
|
||
};
|
||
const settlement = s.ledger.records().find((x) => x.kind === "settlement")!;
|
||
assert.deepEqual(settlement.content, { round: 1, result: { candidate: "r1-c1", input, output: settleAblation(input) } });
|
||
});
|
||
|
||
test("结算：核验不通过、次数用完有事件的记证据不足；设置相同的两格都记一起关，沿用第一格的分数", (t) => {
|
||
settled(
|
||
t,
|
||
(w) => {
|
||
impl(w, 1, [["r1-c1-p1", null, ["p2"]]]);
|
||
w.run("r1-ablation-r1-c1-p1-1", "ablation", 1, "r1-c1-p1", judged(1.05, "main"));
|
||
impl(w, 2, [["r1-c1-p2", null]]);
|
||
w.l.append("event", { event: "cellNotRun", reason: "第 2 次核验仍不通过，修复次数上限 1 已用完", objects: ["r1-c1-p2"] });
|
||
},
|
||
[
|
||
{ factor: "p1", outcome: { kind: "joint", off: ["p1", "p2"], score: 1.05 } },
|
||
{ factor: "p2", outcome: { kind: "insufficient" } },
|
||
],
|
||
);
|
||
});
|
||
|
||
test("结算：设置相同的两格都记一起关，沿用第一格的分数", (t) => {
|
||
settled(
|
||
t,
|
||
(w) => {
|
||
impl(w, 1, [["r1-c1-p1", null, ["p2"]]]);
|
||
w.run("r1-ablation-r1-c1-p1-1", "ablation", 1, "r1-c1-p1", judged(1.05, "main"));
|
||
impl(w, 2, [["r1-c1-p2", null, ["p1"]]]);
|
||
},
|
||
[
|
||
{ factor: "p1", outcome: { kind: "joint", off: ["p1", "p2"], score: 1.05 } },
|
||
{ factor: "p2", outcome: { kind: "joint", off: ["p2", "p1"], score: 1.05 } },
|
||
],
|
||
);
|
||
});
|
||
|
||
test("轮末解封：解封本轮选中的候选的全部作者自评（取最后一版的封存），没选中的继续封存", (t) => {
|
||
const { s, w } = study(t);
|
||
w.open("r1-proposal", "proposal", 1, ["a1", "a2"]);
|
||
w.sub("r1-proposal", "a1", "proposal", null, "r1-a1-1", proposal(r(1.0, 1.1, 1.2), "第一版"));
|
||
const v2 = s.ledger.append("submission", {
|
||
stageId: "r1-proposal",
|
||
author: "a1",
|
||
kind: "proposal",
|
||
target: null,
|
||
item: "r1-a1-1",
|
||
version: 2,
|
||
content: proposal(r(1.05, 1.15, 1.25), "第二版"),
|
||
jev: null,
|
||
} as never);
|
||
w.sub("r1-proposal", "a2", "proposal", null, "r1-a2-1", proposal(r(0.9, 1.0, 1.1)));
|
||
w.close("r1-proposal");
|
||
w.decision("dedup", 1, null, { fresh: [{ candidate: "r1-c1", proposals: ["r1-a1-1"] }, { candidate: "r1-c2", proposals: ["r1-a2-1"] }], carried: [] });
|
||
w.decision("batch", 1, null, { chosen: ["r1-c1"] });
|
||
unsealChosen(s, 1);
|
||
const unseals = s.ledger.records().filter((x) => x.kind === "unseal");
|
||
assert.equal(unseals.length, 1);
|
||
assert.equal(unseals[0]!.content.candidate, "r1-c1");
|
||
// 解封的是最后一版（v2）的封存。
|
||
const seal = s.ledger.records().find((x) => x.kind === "seal" && x.content.proposal === v2.seq)!;
|
||
assert.deepEqual(unseals[0]!.content.seals, [seal.seq]);
|
||
assert.deepEqual(new Board(s.ledger.records()).unsealed(), [{ candidate: "r1-c1", author: "a1", range: r(1.05, 1.15, 1.25) }]);
|
||
});
|