publish-the-run-that-did-not-work
Report the experiment that failed, preserve the superseded one unchanged, and never re-roll for significance. Selection is the bias, not the number.
カードを読んだレビュアーが与えるものであり、ファイルが自分について主張するものではありません。信頼できない内容を扱った実行中にエージェントが抽出したカードは汚染された状態で生まれ、取得される前にレビュー待ちとして保留されます。
どんなときに思い出すか
- the A/B came out negative
- rerunning the benchmark
- the result contradicts what we claimed
- writing up a measurement
価値はたいてい「回避」と「確認」にあります。「実施」は誰でも書ける節です。
トリガー
You ran a comparison to decide whether a change helps, and the result is inconvenient: negative, not significant, or contradicting something already announced.
It does not apply to a run that was simply broken — a crashed harness measured nothing, and publishing it as a result would be its own kind of noise. Fix and rerun, and say that you did.
実施
- Write down, before running, what outcome would kill the claim. If you cannot name it, the experiment cannot fail, and an experiment that cannot fail is not one.
- Report the effect with an interval, not a point estimate.
+9.8pp [−3.5, +16.7]says something+9.8ppdoes not. - When a later run supersedes an earlier one, keep the earlier one published and unedited, and link the correction. Amending it removes the only evidence that the correction happened.
- If a claim does not survive replication, retract it in the same place it was made, in the same size type.
回避
Rerunning until it looks better and reporting the best run. That is not a bad measurement, it is a selected one, and selection cannot be detected in the number afterwards by anyone — including you.
Avoid quietly dropping the negative result while keeping the positive one from the same series. Avoid pooling a fresh sample with an old one after seeing that pooling helps; if pooling was not pre-registered as primary, it is secondary and must be labelled that way.
And avoid describing a non-significant delta with words that imply significance: "improved", "lifted", "up from". "Not significant on its own" is the sentence.
確認
Ask what a sceptical reader could conclude from the artefacts you published. If the answer requires trusting that you did not discard anything, the artefacts are incomplete — publish the journal, the per-item outcomes, or the raw pairs.
Then check the claim in your own README against the results section. If one says a thing the other retracts, the retraction is the true one and the claim has to move.
リスク
This is slower, and it publishes work that looks like failure to anyone reading a single page out of context. Some of those readers will be evaluating whether to use the thing.
The subtler cost: a project that publishes negatives can be quoted against itself. That is a real price, and it is smaller than the alternative — one discovered overstatement makes every other number you have published worthless.
使い方
カードはデータです。リポジトリをクローンし、パスで取り込んでください。ネットワーク経由で届いたものは汚染扱いとなり、承認されるまで保留されます。それが望ましい挙動であり、ここにワンライナーのインストーラーがない理由です。
git clone https://github.com/brcampidelli/chimera-agent.gitchimera skills-import chimera-agent/skills/publish-the-run-that-did-not-work/SKILL.md完全性
公開された状態のファイルの SHA-256。取り込む側は、受け取ったものがこのページに表示されたものと同じか確認できます。
a0a59845f9bef586b26fd372445c1e2688fa69882121a0f5e2ef8e4848b1218a