Validation explorer
Every pytest result Claude ran is logged automatically by a hook in tests/conftest.py. Pick a run, then search and filter. Each test shows its source code, so you can judge what it really checks.
Why a skeptical reviewer can trust the results
- No fabricated quotes:
test_evidence_quotes_are_verbatimchecks that every quote on the dashboard occurs in the downloaded source text. - Independent ground truth: the stance coder is scored against a gold set of course pages whose policies were described independently by web-search summaries (table below).
- Aggregates recomputed: dashboard numbers are recomputed from the raw coded data in
test_aggregates_match_recount. - Tests caught real bugs: rule tests found two coder defects ("may not be used" missed, "human collaborator" missed). See the ad-hoc checks below and the failed runs in the history.
- Human review: a 25% stratified sample is set aside for independent human coding, and Cohen's κ is reported on the dashboard.