Jev as an NSFW prompt filter: 93% on CSAM evals, sub-cent cost, and where thresholds bite

Murky_Ad8671 · reddit · 2026-09-20

A detailed one-day experiment using Jev to pre-screen image-gen prompts for nudity, sexual content, minors and nonconsensual imagery. Setup: one request with five typed questions plus a shared tail instruction, thresholds turned into allow/block via a pure function; 300-1100ms from a Cloudflare Worker at $0.042/M input tokens — a fraction of a cent per screen. Surprises: burying explicit requests in 11,000 chars of scenery changes nothing; no language gradient (Chinese/Arabic/Persian/Hindi score like English); no jitter across runs. The fragile part is question wording as policy — a rewrite of the nonconsensual question dropped false scores from 0.85 to 0.06 while real undress requests stay at 0.97-0.99. Public dataset runs ($0.20): 0.7% false positives on OpenAI clean texts, 93% blocked on S3 (minors), 83% on adult sexual chat, 26% on I2P sexual prompts (mostly policy-level-1 suggestive content, not model misses). Post is truncated at the MMA adversarial results.

Original post →

More from coding & agent

coding & agent channel →