Send us a transcript where backcheck got it wrong

Inizia con il pattern pertinente in src/claims.rs, in particolare is_hedged() e is_negated(), quindi riproduci il verdetto falso con una fixture JSONL minima e sanificata in tests/fixtures/. Esegui backcheck -f your-fixture.jsonl --json e aggiungi un test di regressione che mostri che l’affermazione segnalata viene classificata come previsto senza causare un eccesso di corrispondenze.

backcheck is only worth installing if you trust its verdicts. A hook that fires on honest work
gets uninstalled within a day, and then it protects nobody.

So the most valuable thing you can contribute is a case where it was wrong.

Two kinds of wrong

False positive — it flagged work that was fine. This is the expensive kind. Examples we
already fixed this way:

• runners invoked through a virtualenv path (.venv/bin/python -m pytest) were invisible, so
  genuine runs looked like no run at all
• the shell builtin test -f was counted as a test run, which could hide a missing suite
• "Ruff passes with no warnings" was read as a negated sentence and skipped

False negative — an agent claimed something it had not done and backcheck stayed quiet.
Usually an unrecognised runner (#3) or a claim phrasing the patterns miss.

How to report one

Please do not attach a raw transcript. They contain your source, your paths, and sometimes
your credentials.

Send the smallest JSONL that reproduces it, with everything sensitive replaced. Three lines is
usually enough, and it can go straight into tests/fixtures/ as a regression test:

{"type":"assistant","message":{"content":[{"type":"tool_use","id":"t1","name":"Bash","input":{"command":"<command>"}}]}}
{"type":"user","toolUseResult":{"stdout":"<output>","stderr":"","interrupted":false},"message":{"content":[{"type":"tool_result","tool_use_id":"t1","content":"<output>"}]}}
{"type":"assistant","message":{"content":[{"type":"text","text":"<what the agent claimed>"}]}}

Then run backcheck -f your-fixture.jsonl --json and paste the output along with what you
expected instead.

There is an issue template for this: 🎯 Wrong verdict.

Claim phrasings we know are missed

Patterns live in src/claims.rs. Known gaps:

• non-English summaries
• emoji-only status (✅ tests with no verb)
• markdown tables reporting per-check status
• "everything is green", "CI is happy", "all clear"
• claims about coverage thresholds
• claims split across sentences ("Ran the suite. Everything passed.")

Each of those is a pattern plus a test, and the guard against over-matching is the interesting
part — see is_hedged() and is_negated().