Project 4 starter: the inbox
A stack of emails and a detector that is wrong too often. Your agent makes it right.
The job
Make detector.py beat the target on inbox.jsonl without memorizing senders or subject lines. Run python3 score.py to see the score and the target. A second inbox you cannot see is scored the same way after you submit, so a detector that remembers these sixty emails loses there.
Send the job to your agent as a /plan message, for example:
/plan make detector.py beat the target in score.py on inbox.jsonl without memorizing senders or subject lines
What is here
| file | what it is |
|---|---|
| inbox.jsonl | sixty emails, one per line: from, subject, body, and phish, the answer |
| detector.py | the baseline detector, one keyword rule; is_phish(email) is the function the scorer calls |
| score.py | the validation: the score beside the target, exit 0 when it meets the target |
| skills/coding-loop/SKILL.md | the skill your agent follows: plan, approve, change, validation, report, transcript |
The number to beat
python3 score.py
prints the baseline’s score and the target. The baseline is below it.