Email Triage with a Judgment-Only Model
Updated 2026-09-20
The scenario
A shared inbox — support@, sales@, founders' personal mail — receives far more than anyone can read. Most messages need a quick route, not a thoughtful reply: newsletters, notifications, cold pitches, auto-generated reports. A minority need a human or a carefully drafted response today.
Why Jev fits
Triage is a routing decision repeated thousands of times a day. The outputs are closed: reply or don't, which lane, how urgent. That is precisely the yes/no + choice + score triad a judgment-only model emits — fast enough to run on every incoming message at arrival time, cheap enough that the inbox stays sorted in real time instead of in batches.
Judgment design example
Structured input per message:
| Field | Example |
|---|---|
subject | "Invoice discrepancy — March" |
sender_domain | acme-corp.com |
is_existing_customer | true |
body_excerpt | first ~500 characters |
thread_length | 4 |
Judgments to define:
- Binary: "Does this need a reply from a human or a drafted response?" → no → archive with label.
- Choice: "Which lane?" →
billing/support/sales/legal/personal/noise. - Score: "Urgency 0–100" → sorts today's queue; anything above threshold pings immediately.
Where the LLM sits
Downstream of the verdict. For messages Jev routes as needs-reply, an LLM drafts the response using full thread context — the expensive read happens only where a reply will actually be sent. Noise never touches the generation budget.
This mirrors the voice-controlled-browser demo from early coverage in spirit: Jev decides what an incoming signal is and where it goes; other components do the acting. See the front-filter recipe.
Watch-outs
- VIP override. A hard-coded allowlist (boss, key accounts, legal) should bypass the noise verdict entirely.
- Excerpt limits. Judging on a truncated body can miss "by the way, we're cancelling" at the bottom of a long newsletter forward — include a tail excerpt or a full-body flag for long threads.
- Drafts, not sends. LLM-drafted replies should land as drafts for human approval until your error rate earns more autonomy.
Sources
- "Jev: The New AI Model That's Breaking The Internet (Full Tutorial)" — Moritz (~43K views, 2 days) — includes the voice-controlled-browser demo: judgment as the routing layer.
- "Jev is HERE. How to use it" — Greg Isenberg (~293K views, 1 day).
- AI with Surya — early-access coverage of the judgment (not generation) model concept.
Unofficial fan-made handbook. Not affiliated with TypeSafe AI or jev.com.
Related Guides
Document Triage: Type and Urgency on Two Axes
Sort inbound documents — contracts, invoices, resumes, tickets — by type with Choice and by urgency with Score, so each document class gets its own downstream pipeline and deadline.
Intent Classification: The Front Door of a Chatbot
Classify user messages into intents (and urgency) before any LLM touches them — so simple requests get cheap handlers and only genuinely open-ended messages reach the expensive model.
RAG Reranking: Filtering Retrieval Results by Relevance
Use Jev-style Noul judgments to batch-score retrieved chunks for relevance and keep only the top-k — so the LLM answers from signal, not from whatever the vector search coughed up.
Validating Structured Extraction: Code First, Judgment Second
Pair deterministic code validation (schema, types, ranges) with a Jev-style semantic review pass — catch the extraction errors that parse fine but mean the wrong thing.