# Answer key — reveal only after extraction

Synthetic demonstration, 10 independent excerpts. Date-only deadlines must not acquire an invented time. Equivalent task wording is acceptable; owners, final dates, and status must match.

| ID | Final action | Owner | Final deadline | Status | Decisive evidence |
|---|---|---|---|---|---|
| M01 | Send revised product brief | Maya | 2026-09-30 | Active | “Agreed. Wednesday it is.” |
| M02 | Upload final icons | Nina | 2026-10-02 | Active | Nina accepts the upload; Priya confirms. |
| M03 | Send customer survey | Omar (before cancellation) | None; former 2026-10-01 revoked | Cancelled | “Cancel the send entirely.” |
| M04 | Submit budget summary | Ben | 2026-10-02 | Active | October 5 proposal rejected; Ben confirms October 2. |
| M05 | Send supplier comparison | Grace | 2026-09-30 | Active | Thursday email explicitly superseded. |
| M06 | Upload demo video | Chen | 2026-09-30 14:00 Europe/London | Active | “Wednesday, September 30, at 2 p.m. London time.” |
| M07 | Circulate final launch checklist | Tess | 2026-10-02 | Active | “Tess owns circulation, due October 2.” |
| M08 | Send analytics report | Lina | Unspecified / TBD | Active | “No replacement date yet.” |
| M09 | Send partner deck | Hugo | 2026-10-02 | Active | “Then keep October 2 as the final deadline.” |
| M10 | Publish revised release announcement | Ava | 2026-10-01 | Active | Old draft cancelled, revised announcement confirmed. |

## Review rules

- Keep the initial model output before revealing this key. Score that output, not a corrected answer produced after seeing the key.
- Expect nine active actions. M03 may be omitted from an active-only list or shown separately as cancelled; listing it as an active obligation is wrong.
- For M03, retaining the former date as explicitly historical is acceptable. Treating it as an effective deadline is wrong.
- For M08, an explicit unknown deadline is correct; omitting the still-active task is not.
- M06 needs the agreed time and London timezone, or an accurately converted equivalent.
- Do not add the already-completed review in M07 or a new survey task in M03 as pending actions.
- Report per-excerpt correctness for task/status, owner, and deadline. Count a fully correct excerpt only when all applicable fields are correct. Report any invented additional active actions separately.
- Report counts out of ten and show mistakes with evidence. This small constructed demonstration does not establish performance on real meetings or audio transcription.
