# Meeting notes to action items — rubric v1

Dataset: `meeting-01.txt` … `meeting-05.txt` (plain-text notes), answer key in `meeting-notes-v1.json`. Written before any tool was run.

## Task given to each tool
From each meeting's notes, list the action items for **our side and the other party**, each with: owner, task, due date (ISO or empty) and priority (urgent / high / medium / low). Tools get only the `.txt` files.

## Scoring, per expected action
| Item | Points |
|---|---:|
| The action is found (same task, any wording) | 2 |
| Owner correct | 1 |
| Due date correct (or correctly empty) | 1 |
| Priority within one level | 0.5 |
| **Max per action** | **4.5** — 20 expected actions, dataset max 90 |

Penalties: −2 for every invented action (something nobody agreed to: a parked newsletter, a cancelled photographer, a vendor trial that was declined). −1 for assigning a task to the wrong side of the table.

## Traps planted on purpose
Each meeting has one to three traps listed in the answer key: explicitly parked items, conditional starts, superseded deadlines, and the other party's tasks. Report how each tool handled every trap.

## Also record
Setup time, run time, cost, whether the tool could take all five files at once, and whether it produced a list you could paste into a task manager without editing.

## Limitations
Five short English meetings, one business type, clean notes rather than raw transcripts. Says nothing about long meetings, audio quality or non-English speech.