James O’Halloran uses Jev to moderate Squigdle, an anonymous daily drawing game. Players turn shared squiggles into doodles and vote on each other’s work. A server-side check runs before drawings reach the voting pool.

Put the policy into competing descriptions

  • The server renders a drawing as a PNG and sends it to Jev with its title.
  • A multiple-choice question describes acceptable drawings alongside sexual content and other moderation categories.
  • Harmless lookalikes get explicit treatment: the author allows innuendo and some anatomical jokes while describing what should trigger restriction.
  • O’Halloran reports refining those descriptions because innocent, blobby drawings can resemble prohibited content.

The model returns probabilities over those categories. Application code decides where a drawing goes next.

Keep sharing, voting and review separate

O’Halloran’s excerpt approves a drawing when p.ok >= 0.65, then marks it NSFW when p.sexual >= 0.5. He omits additional harmful-content rules from the excerpt; the remaining fallback is review.

  • Approved drawings enter public voting.
  • NSFW drawings can still be shared behind a blurred warning, outside the voting pool.
  • Potentially harmful cases wait for the author, alerted through a Discord webhook.

Those cutoffs are the author’s operating choices. He publishes no labeled threshold evaluation showing how often they accept harmful drawings or reject harmless ones.

Read the bill separately from the accuracy claim

O’Halloran reports roughly 1,000 checks, each carrying a 512×512 PNG and title, costing six cents total, or about $0.00006 per drawing. These are his deployment figures, without independent reproduction.

The headline describes 500 daily doodles; the cost report covers approximately 1,000 checks, a different quantity. His quality evidence is anecdotal: occasional phallic-looking false alarms that take a tap in Discord to resolve.

A typed verdict makes the result easy to consume. Correct moderation still depends on the question, thresholds and review process.

Caveat: the writeup exposes selected code snippets and screenshots, without raw predictions, a labeled error count or a complete policy implementation to rerun.