research(rectification): add reported-offset evaluation and frozen rerun integrity
Independent Staging Quality Gate / validate (push) Successful in 12m7s
Independent Staging Quality Gate / publish (push) Successful in 3m46s

Preserve closed confirmation gates and previously-exposed dataset boundaries. Add auditable 900-trial sensitivity results, current scorer freshness checks, and the v5 collection protocol.

Co-Authored-By: Claude Code <noreply@anthropic.com>
This commit is contained in:
jesse-ux
2026-09-20 12:03:52 +08:00
co-authored by Claude Code
parent 5049e7847b
commit 932f2fffba
22 changed files with 14134 additions and 7 deletions
@@ -81,6 +81,13 @@ const diagnosticReport = JSON.parse(readFileSync(
verified_minute_claim_allowed: boolean;
};
};
const currentTreeRerun = JSON.parse(readFileSync(
new URL("../../docs/research/sealed_holdout_rerun_2026_09_20.json", import.meta.url),
"utf8",
)) as {
frozen_record: { implementation_sha256: string };
official_valid_independent_blind: boolean;
};
type ProductHoldoutContract = typeof SEALED_MINUTE_HOLDOUT & {
metrics_produced_by: {
implementation_sha256: string;
@@ -190,10 +197,16 @@ test("sealed holdout aggregates match the v3 report and still block confirmation
productHoldout.metrics_produced_by.source_report,
"references/real_case_calibration/minute_rectification_holdout_v3_report.json",
);
// 原值: 历史 sidecar 哈希 / 新值: 当前冻结重跑哈希 / 原因: BUG-978 身份新鲜度,不改变确认门。
assert.equal(
productHoldout.current_tree_scorer.implementation_sha256,
diagnosticReport.frozen_scoring.actual_sha256,
currentTreeRerun.frozen_record.implementation_sha256,
);
assert.equal(
productHoldout.current_tree_scorer.source_report,
"docs/research/sealed_holdout_rerun_2026_09_20.json",
);
assert.equal(currentTreeRerun.official_valid_independent_blind, false);
assert.equal(productHoldout.current_tree_scorer.matches_metrics_scorer, false);
assert.equal(productHoldout.current_tree_scorer.official_eval_implementation_hash_matches, false);
assert.equal(productHoldout.current_tree_scorer.official_eval_trial_count, diagnosticReport.trial_count);