跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
其他技能

independence-bookkeeping

Provenance rules that keep checks independent. Use when setting up any comparison, oracle, or reference value — and whenever tempted to certify a result with machinery that helped produce it.

BootLoops-ai · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add BootLoops-ai/skills --skill independence-bookkeeping

技能说明


name: independence-bookkeeping description: Provenance rules that keep checks independent. Use when setting up any comparison, oracle, or reference value — and whenever tempted to certify a result with machinery that helped produce it.

independence-bookkeeping — keeping the second route honest

THE LAW: independence is an auditable property of records, never an assumption. Two routes are independent when you can produce the trail showing what each one depends on — code, inputs, tuning history — and the trails do not meet at any layer where the feared error could live. If you cannot produce the trail, you do not have independence; you have a feeling, and correlated routes agree wrongly with exactly the same warm feeling as independent ones.

The strongest evidence a computed result can carry is agreement with a second route that could not have failed the same way. The entire content of that check is the "could not" — and the "could not" is a fact about records, not about intentions. Nobody sets out to build a circular check. Independence decays silently, through ordinary tidy-minded acts: a stale reference gets "refreshed" with the current code; two evaluators quietly converge on the same library; an oracle's values leak into the tuning of the thing the oracle later certifies. None of these announce themselves, and every one of them leaves the comparison script printing PASS. The only defense is bookkeeping kept from the start, because independence lost cannot be reconstructed afterward — once a value has touched a fit, no amount of later diligence untouches it.

The one-way contamination rule

An oracle that fed a fit may never certify the result. Contamination is one-way and permanent. The moment a reference value influences any choice — a fitted parameter, a model form, a tolerance, a stopping decision, even which of several candidates you kept — that value belongs to the production side of the ledger forever. Certification draws only from the disjoint set: values the fitting process never saw and could not have seen.

This is stricter than it first looks. Influence includes the soft channels: you glanced at the reference while debugging and "fixed" the code until it agreed; you used it to decide how many terms were enough; you discarded a run because it disagreed with it. All of these are fits in everything but name. The test is counterfactual: if this reference had been different, could anything about the result have come out different? If yes, it fed the fit.

The ledger — what to write down and when

Keep provenance as an append-only record, written at the moment of creation, not reconstructed at certification time. Memory of provenance is worthless; the whole point is that six weeks later nobody remembers which script produced the number.

1. A birth record for every reference value. Generator (program plus version or commit), inputs and settings, working precision, the estimated accuracy and where that estimate comes from (internal convergence, an error bound, agreement with something else — name it), and the date. A reference without a recorded accuracy cannot support a digit claim: a comparison is only as good as the weaker side, and an undocumented reference has no known strength. Comparisons have stalled at exactly the digit count of a reference nobody recorded — the agreement was real, and every claimed digit past it was manufactured.

2. A contact log for every oracle. Each time a value enters a fit, a tuning decision, a model selection, or a debugging session, append the event. The log is append-only; entries are never removed, because the rule above is one-way. Certification of claim C requires a reference whose contact log contains neither C nor anything C was built on.

3. A lineage audit for any claimed "two independent routes." Walk each route's dependency stack down to the libraries, the tabulated inputs, and the constants. Two programs in different languages that call the same special-function library share that library's defects; two derivations that read the same published table share its typos. Shared hardware and compilers rarely matter; shared mathematics and shared data almost always do. Write down the deepest shared layer and then argue, in writing, that the failure you are guarding against cannot live there. "Fully disjoint" is often impossible — that is fine; undocumented sharing is what kills you, not sharing itself.

4. One deliberately different engine. Keep an evaluator around precisely because it shares nothing with the rest of your stack — different algorithm, different language, different author if you can get one. It will be slower. Its value is not speed; its value is that its errors are uncorrelated with everyone else's, which makes its agreement worth more per digit than any in-family check.

5. The comparison itself is code — test it before believing it. Before any pass counts as evidence, feed the comparator a deliberately mismatched pair and watch it fail. A comparator that has never been seen to fail proves nothing: loose tolerances, a wrong column, a string comparison of truncated prints — all of these return agreement on everything, forever, and the output looks identical to a real pass. Also write down, before running the real comparison, what outcome would count as failure. A failure criterion chosen after seeing the numbers is a fit.

6. Close the loop in the record. When a comparison certifies a claim, the certificate cites the reference's birth record; the reference's contact log gains the claim. The reference is now burned for anything built on top of that claim. This is the price of using it, and the ledger is how you know the price was paid.

Failure-mode catalogue

The classes this discipline exists to catch. Every one of them prints PASS.

  • The shared-library twins. Two "independent" implementations both call the same underlying expansion; a defect there makes th
阅读完整来源 ↗