# Can a model extract agreed actions without inventing owners or deadlines?

Current status: three model trials recorded; see [provisional results](../reports/meeting-actions.md) and the [dated amendments](../API-RUN-PLAN-2026-09-11.md). The protocol below is the predeclared design. Human baseline/review remain outstanding. All transcripts, people and tasks are synthetic.

## Question and reference

Extract outstanding agreed actions from 12 short meeting cases. The cases cover suggestions, cancellation, reassignment, deadline correction, missing owners, relative dates, explicitly ambiguous dates, completed work, repeated mentions and a quoted instruction. The authored reference contains ten outstanding tasks. Bracketed task identifiers make exact scoring possible; they are a deliberate simplification, not a feature of typical transcripts.

Microsoft documents [meeting recap and follow-up tasks](https://support.microsoft.com/en-us/teams/meetings/recap-in-microsoft-teams). This motivates testing extraction reliability; it does not establish demand or imply this is a test of Teams Copilot. The model under test is the recorded Grok configuration through OpenRouter, using supplied text only.

## Fixed trial

1. Independently review expected.json against inputs.json before any model trial. Record corrections and regenerate the manifest if necessary. Then freeze the materials.
2. Have a human produce the results without seeing expected.json. Preserve their answers and total time, including verification.
3. Use Reality Lab's Experiment trial mode and Meeting actions template. Send the exact prompt.txt in three fresh requests with unchanged settings. Never attach expected.json.
4. Download every run and score it with `node scripts/score-workflow-trial.mjs meeting-actions downloaded-run.json`. Keep failures, missing output and truncated output in the trial record.
5. Time human verification/correction separately. Report model-request latency, total workflow time and cost separately; no time-saving verdict without a measured baseline.

## Predeclared success

All 12 case records must match the reference in all three trials. Action ordering is irrelevant; duplicate actions, invented owners/dates, missing tasks, cancelled/completed tasks presented as open, and extra fields fail exact scoring. Explicitly unknown owners and ambiguous dates must stay null. Report each invented owner or deadline, rather than hiding it inside an aggregate percentage.

Report three-trial case accuracy over 36 attempted case records, plus the list of omission and invention errors. Failed requests count as failed trials. The scorer's protocolEligible field must be true before a syntactically valid response can be treated as a passing fixed-protocol result.

## Limits and commercial hypothesis

This is text extraction from short controlled notes, not speech recognition, meeting participation, speaker identification or a vendor comparison. It says nothing about long recordings, interruptions, accents, time zones or private documents. Even perfect output requires owner confirmation before creating tasks. A possible paid pilot is a permissioned review of one team's action-extraction process with a documented check step; no buyer interest, revenue or production safety is claimed.