# Meeting extraction fixture

Origin: original fictional meeting text emitted by the deterministic Python generator in scripts/evals/meeting-transcript-to-action-items/generate_fixture.py. Dataset and labels: CC0 1.0 Universal, https://creativecommons.org/publicdomain/zero/1.0/. Code: MIT. No third-party meeting corpus, AMI annotations or MeetingBank text is included. The AMI annotations license was not verified, so this set uses the card's synthetic-only fallback.

Generator: Pareto authored the generator and fixed intended labels before rendering text. No separate text-generation API was used. All names, organizations, meetings and dates are fictional. There are 10 test and 2 separate dev meetings; each has 5 action items and 2 decisions. Two test transcripts are substantially longer, using repeated background discussion. This is a controlled retrieval and commitment test, not a representative conversational benchmark.

Labeling: labels set by the generator, spot-checked by us (the authoring agent checked the emitted dev labels against source lines, including delegated owners, null owners and relative dates). No human labeler or second annotator has checked this set yet. Human spot-check remains required; there is no human agreement statistic. Wording on the page does not imply human adjudication.

Written rubric: retain explicit accepted or assigned open tasks, including an absent person's written commitment. Exclude suggestions, completed history, cancelled proposals and repeated mentions. Use the final agreed owner, not automatically the speaker. Owner and deadline are null when unspecified. Resolve tomorrow and Friday against the provided Monday meeting date. Decisions require explicit agreement. Cite the commitment or decision line.

Headline: action-item F1 with one-to-one maximum-cardinality matching on cited line within plus/minus 2 and exact owner. Unit: meeting. Sum tp/fp/fn across repeats; bootstrap meetings, not action items. Task wording is not semantically judged: a correct line and owner can earn credit despite a poor paraphrase. Decision matching uses cited line only. Duplicate predictions count as false positives. Invalid/missing tool calls receive empty arguments and all gold actions become false negatives.

Secondary: action precision/recall, decision precision/recall, recall for owners who are not the cited speaker, invented-date count and rate (incorrect non-null due dates divided by predicted actions), found actions and fixed gold count. Count metrics use the median per meeting across repeats then sum; rate metrics pool counts. An omitted date is not an invented date; due-date recall is not measured. Empty predictions have invented-date rate zero, so read this beside action recall.

Pre-freeze: dev-only Pareto checks, no baseline calls and no test tuning. Prompt is frozen before the scored test. Small n and formulaic text can create a ceiling and do not support workplace-quality claims. Varied paraphrases satisfy the runner's output-variation check but do not establish statistical independence.

Safety: obtain recording consent and permission to upload a real transcript. Never use this recipe for employee evaluation. The model produces draft notes, not approved assignments; a person reviews owners, decisions and dates before any task creation.