roshworldwide/holdout eval gate

Run an LLM eval on this PR and fail only when quality regressed beyond statistical noise. Posts the full comparison (effects, confidence intervals, corrected p-values) as a PR comment.

View on GitHub

Trust Signals

Scorecard Score
not yet scored
Maintenance Recency
Stale
License
None
namedescriptionrequireddefault
evalEval reference: a .jsonl file or Python 'module:attr'yes
targetTarget reference: 'ollama:M' | 'openai:M' | 'anthropic:M' | 'module:attr'yes
baselineBaseline run id (or unambiguous prefix) present in the storeyes
storeRun store directory (commit your baseline artifact inside it)no.holdout
scorerScorer for .jsonl evals: exact | exact-strict | regex:<pattern>noexact
alphaSignificance level applied to corrected p-valuesno0.05
correctionMultiple-comparison correction: benjamini-hochberg | holm | nonenobenjamini-hochberg
seedSeed for generation and resamplingno0
python-versionPython version to set upno3.12
installpip spec for holdout ('.' to install from this repo checkout)noholdout
commentPost the comparison as a PR commentnotrue
github-tokenToken used to post the PR commentno${{ github.token }}
namedescription
run-idRun id of the candidate evaluation
verdictimproved | regressed | no_significant_change | insufficient_data