# Email triage test — rubric v1

Dataset: `email-triage-v1.json` (20 synthetic emails). Written before any tool was run.

## Task given to each tool
For each email, output: **category** (one of the 20 labels in the dataset's `expected_category` values), **priority** (urgent / high / medium / low / none) and **action** (the `expected_action` label). Tools receive only `from`, `subject`, `body`; never the `expected_*` fields.

## Scoring
| Item | Points |
|---|---:|
| Category exactly right | 2 |
| Priority exactly right | 1 |
| Priority off by one level | 0.5 |
| Action exactly right | 2 |
| **Max per email** | **5** — dataset max 100 |

Hard failures (reported separately, each one is a red flag regardless of score):
- E18 (phishing) classified as anything other than suspicious.
- E05 or E10 (urgent) given priority below high.
- Any email marked "no reply needed" that expects a reply.

## Also record
Time to set up, time to run, cost, whether the tool needed the emails pasted one by one or accepted a batch, and what the tool did with the two emails that mention money.

## Limitations
Twenty emails from one kind of business. Says nothing about volume, languages other than English, threads, or attachments.