Content Moderation with a Judgment-Only Model

#The scenario

You run a platform with user-generated content: comments, reviews, listings, messages. Volume is high, most of it is fine, and a small fraction is spam, abuse, or policy-violating. Sending every item through a generative LLM for review is slow and ruinously expensive; sending nothing means shipping a sewer.

#Why Jev fits

Moderation is the canonical closed-answer problem. The outcome set is known in advance — allow, remove, escalate — and the workload is a firehose where 95%+ of items need only a verdict, not an essay. A judgment-only model is built for exactly this: one structured input, one decision, at a per-call cost low enough to apply to every item instead of a sampled fraction.

Full coverage is the hidden win: sampling-based moderation misses rare-but-severe violations; judge-everything pipelines do not.

#Judgment design example

Structured input per item (keep it minimal):

Judgments to define:

1. Binary: "Does this clearly violate policy?" → yes → remove, no LLM involved.
2. Binary: "Is this clearly fine?" → yes → publish immediately.
3. Choice (for the remainder): "Which risk bucket?" → hate / harassment / spam / sexual / other — routes to the right review lane or policy-specific LLM prompt.
4. Score: "Severity 0–100" → orders the human review queue so the worst is seen first.

#Where the LLM sits

Behind the filter, on the middle band only: borderline items get an LLM pass for nuanced reading and a written rationale moderators can audit. The LLM sees maybe 2–5% of volume — the items where context and nuance genuinely change the answer. See the front-filter recipe for the full pipeline shape and the Jev vs LLM split.

#Watch-outs

• Appeals need explanation. Jev returns verdicts, not rationales; if your process requires user-facing reasons, generate them downstream (LLM) or from templates keyed to the verdict.
• Adversarial drift. Spammers adapt; re-check filter precision on a sampled audit weekly.
• Never auto-action severe categories on score alone. High-severity verdicts should route to humans regardless of confidence.

#Sources

• "Jev is HERE. How to use it" — Greg Isenberg (~293K views, 1 day) — classification-first framing.
• "Jev + GPT-6 Astra = 400X Cheaper" — Jack Roberts (~21K views, 11 hours) — cheap filtering ahead of frontier models.
• "Jev: The New AI Model That's Breaking The Internet (Full Tutorial)" — Moritz (~43K views, 2 days).

Unofficial fan-made handbook. Not affiliated with TypeSafe AI or jev.com.