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Report the experiment that failed, preserve the superseded one unchanged, and never re-roll for significance.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.
由读过该卡片的评审者赋予，而不是文件自我声明的。智能体在处理不可信内容的运行中提炼出的卡片，生来即被标记为受污染，在被检索之前会一直等待评审。
何时会想到它
- the A/B came out negative
- rerunning the benchmark
- the result contradicts what we claimed
- writing up a measurement
价值通常在「避免」和「检查」里。「该做」是人人都会写的那一节。
触发
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.
该做
- 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.
避免
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.
检查
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.
风险
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.
如何使用
卡片就是数据。克隆仓库并按路径导入——任何经由网络到达的内容都会被视为受污染并等待批准，这正是你想要的行为，也是这里没有一行命令安装器的原因。
git clone https://github.com/brcampidelli/chimera-agent.gitchimera skills-import chimera-agent/skills/publish-the-run-that-did-not-work/SKILL.md完整性
文件发布版本的 SHA-256。导入方可以据此核对收到的内容与本页展示的一致。
a0a59845f9bef586b26fd372445c1e2688fa69882121a0f5e2ef8e4848b1218a