Repository navigation
evals: local trial worksheets with separate activation and procedure reports - #188
Merged
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #176
Summary
Adds
evals/trial.py(stdlib only, no network, no agent launch) withinit(unscored worksheet to a new file, never overwrites, refusesevals/so it stays out of the scenario glob) andreport(validate, then text/JSON). The worksheet snapshots scenarios (file and content hashes) and records trial id/date, catalog commit, client, model and reviewer. Reports read only the snapshot, so historical worksheets stay readable after scenarios change (--check-currentnotes drift). No scenario schema change, no new dependency, no transcripts committed.Semantics: activation passes on expected-skill presence (positive) or expected skill present and pilot skill absent (negative); extra activated skills are preserved. Unobserved/pending are never passes and are excluded from the denominator with counts shown. Procedure is reported separately (scored rate, rate over all applicable with unscored as not passed, N/A visibly excluded; N/A needs a reason). Zero denominators print
undefined.Verification
uv run --no-project --python 3.12 --with 'pytest>=9' python -m pytest tests: 93 passed;python -m unittest discover -s tests: OK. No network/model access.tests/test_eval_trials.pyruns the real helper on a mixed fixture and asserts exact per-case outcomes and separate counts (extra skill, pilot wrongly activating on a negative, N/A, incomplete, all-unscored), plus diagnostics for unknown/duplicate cases, bad statuses, missing provenance, mismatched snapshots.initto same path refused.Screenshots: n/a — no user-visible change.