Reproducible meeting evidence · MIT + CC BY 4.0

Validate action items before they become tasks

A transcript can support a commitment, contradict it, or say nothing about it. This fixture keeps that distinction visible. It checks each supplied action candidate for a source quote, known owner, real deadline, explicit acceptance, and duplicates before allowing it into the review queue.

The transcript and candidates are synthetic. The run tests validation logic; it does not measure transcription quality, model accuracy, or time saved. Accepted items still require human approval before any external side effect.
Four transcript segments produce eight candidates; the validator accepts two and rejects six with explicit reason codes

Files

Evidence first

A short quote must appear in the declared transcript segment. An invented action cannot pass on polished wording alone.

No silent guesses

Unknown owners, impossible dates, proposals, and duplicate tasks stay in the rejection list with machine-readable reasons.

Review before write

The validator produces a review queue. Creating tracker tasks or sending messages remains a separate, explicitly approved step.

Verified scope

Twenty tests cover the six visible rejection reasons, normalized evidence, canonical participant names, calendar dates, stable ordering, malformed input, saved JSON, generated SVG, and license files. In the reviewed synthetic run, two of eight candidates passed all gates.