Triage a work inbox into read-only summaries and reply drafts, sending nothing
Exact prompt
# Triage a (synthetic) work inbox into read-only summaries and reply drafts, sending nothing `inbox.mbox` is a synthetic mailbox (made-up people and companies); `config.json` gives `me`, `vip` addresses and `urgent_words`. Write `solve.py` (Python 3 standard library only). `python solve.py input.json answer.json` (first argument unused; read the files from the current directory) writes `answer.json` and `drafts.mbox`. Fresh mailboxes are generated when scoring. ## Read-only and send-nothing rules 1. Never modify `inbox.mbox`. Never send anything: do not write `sent.log` or `outbox.mbox`, and solve.py must not use smtplib, imaplib, poplib, socket, urllib, http.client, requests, httpx, aiohttp, subprocess or webbrowser. `answer.json` must contain `"sent": []`. ## Parsing 2. Decode encoded headers (RFC 2047), quoted-printable and base64 bodies. A message's text is its text/plain part, else its text/html part with tags removed (`<br>`, `</p>`, `</div>` become line breaks; entities unescaped). Unquoted lines are lines that do not start with `>` (after leading spaces). 3. Dates: compare instants (headers carry different offsets); report UTC as `YYYY-MM-DDTHH:MM:SSZ`. Mailbox order is not time order. 4. Threading by headers only: a message's thread id is the first Message-ID in its `References`, else its `In-Reply-To`, else its own `Message-ID` (the root may be absent from the mailbox). Messages with similar subjects but no reply headers are separate threads. 5. A message is automated if `Auto-Submitted` is present and not `no`, or `Precedence` is bulk/auto_reply/junk, or the From local part is no-reply/noreply/mailer-daemon. 6. Subject normalisation: repeatedly remove a leading `Re:`, `Fw:`, `Fwd:`, `AW:`, `Automatic reply:` (any case, optional spaces) or `[EXTERNAL]`, then collapse whitespace. A thread's `subject` is the normalised subject of its earliest message. ## Per thread 7. The deciding message is the latest non-automated message. `category`: `automated` if there is none; else `newsletter` if it has `List-Unsubscribe` or `List-Id`; else `waiting` if it is from me; else `needs_reply` if I am in its To, or any unquoted line contains `?`; else `fyi`. 8. `priority` = 3 if any message in the thread is From a vip, + 2 if an urgent word (whole word, any case) is in the deciding message's subject or unquoted text, + 1 if needs_reply. 9. `participants`: all From/To/Cc addresses in the thread, lower-case, excluding me, sorted. `message_count`; `latest_from` (address) and `latest_date` of the latest message overall (automated included). 10. `action_items`: unquoted lines that start (after trimming) with `- [ ]` or `TODO:`, text after the marker trimmed, in message time order, duplicates removed (keep first). 11. Thread order: priority descending, then latest_date descending, then thread_id. ## Drafts (one per needs_reply thread, in drafts.mbox) 12. Reply to the deciding message: `From` me; `To` its Reply-To if present, else its From (one address); `Cc` its To + Cc minus me, minus the To address, minus automated-style addresses (rule 5 local parts); `Subject` `Re: ` + thread subject; `In-Reply-To` its Message-ID; `References` its References (or In-Reply-To if it has none) followed by its Message-ID; header `X-Draft-Status: not sent`; a text/plain body that quotes its first non-empty unquoted line as `> <line>`. ## answer.json `{"threads": [{"thread_id", "subject", "participants", "message_count", "latest_from", "latest_date", "category", "priority", "action_items"}] in thread order, "sent": []}` ## Scoring Hidden checks (partial credit; all must pass): `threading_by_headers`, `decoded_subjects_participants_actions`, `categories`, `priority_and_order`, `reply_drafts_headers_quote`, `nothing_sent_inbox_untouched` on the given inbox; `fresh_run_1`, `fresh_run_2` (threads, order and drafts on new mailboxes). 30-second limit per run.
Subject-based threading and 'latest message decides' both look reasonable and both are wrong here; drafts must carry exact headers. Original benchmark task written for usecaserank.com (not based on a specific public post). Headless, offline run on a reproducible fixture. Each run's notes say how its cost was measured (estimates are marked).
Success criteria: 1. threading_by_headers 2. decoded_subjects_participants_actions 3. categories 4. priority_and_order 5. reply_drafts_headers_quote 6. nothing_sent_inbox_untouched 7. fresh_run_1 8. fresh_run_2
Best model for this task so far: Claude Haiku 5.5 (effort default (no --effort flag) · thinking on) (success, $0.02, 1m).
All runs (5)
80% pass rate across 5 eligible runs: 4 pass, 0 partial, 1 fail. Partial and failed runs are listed below with everything else.ⓘ
| Model | Cost | Time | Step-insⓘ | Autonomy | Outcome | Trustⓘ | Date | Compare |
|---|---|---|---|---|---|---|---|---|
| MMistral Small 3.2 24B Instruct (Venice)thinking on | $0.0079 | 2m | 0 | Fully autonomous | fail | Trust 80% | 2026-10-25 | |
| OGPT-6.1 Soldefault · thinking on | $0.06 | 1m | 0 | Fully autonomous | pass | Trust 80% | 2026-10-22 | |
| AClaude Haiku 5.5default · thinking on | $0.02 | 1m | 0 | Fully autonomous | pass | Trust 80% | 2026-10-20 | |
| AClaude Opus 5.5default · thinking on | $0.36 | 1m | 0 | Fully autonomous | pass | Trust 80% | 2026-10-20 | |
| AClaude Sonnet 5.5default · thinking on | $0.10 | 27s | 0 | Fully autonomous | pass | Trust 80% | 2026-10-20 |
Embed this result
A live badge for this use case. It links back here and updates as new runs are published.
Get new runs and ranking changes weekly
Which models got better or cheaper on real tasks like this one.