Repository navigation
skill-placebo v0.1.1: a correction to the v1 pass rates, and packaging fixes.
Correction (METHOD.md, amendment 19). Two agent timeouts whose tests passed afterwards had been counted as passes;
METHOD.md sections 5 and 7 say a timeout is a failed trial, and they now count as failed. planning-with-files
moves from "no better than placebo" to "worse than placebo" (pass-rate difference −20 pp [−37, −7], Holm-adjusted
p = 0.043). Costs and every cost ratio R are unchanged. Headline: 2 of 9 skills beat a same-length placebo, both
at Holm-adjusted p = 0.049 (1 worse, 6 no better).
Also:
- README: skills run with the scripted approval turn are marked, as METHOD.md section 5.1 requires.
- PyPI: the project page links to the repository and the issue tracker, with keywords and classifiers.
skill-placebo listandrunnow use this release's tag (v0.1.1).
The per-trial records in results/ are the same as in v0.1.0; only how two of them are counted changed. The
agents' full logs stay attached to release v0.1.0:
https://github.com/simonether/skill-placebo/releases/tag/v0.1.0 (sha256 in results/*/*/AGENT_LOGS.json).
This release has no assets.
Install: uvx skill-placebo run DietrichGebert/ponytail