Moderating a site with Jev

It is below the 83.3% single-human average the post cites, and only the 97% majority-vote line clears it. It does beat every other decision model tested, including Jeeves 9B at 68% and Laya Multilingual at 53%, so the honest read is best-in-class among classifiers, not good enough to skip human review.

Not repeatable. Jev 1.13.0 and Laya have one run each, the GPT and Claude entries are means of four, Gemini and DeepSeek means of three. So the 79% versus 96.25% gap is real in direction but the narrow spreads between adjacent models should not be read as an ordering.

The site auto-moderates, and the interesting mechanism is rule changes: you open a post arguing for new subreddit rules and Jev decides whether enough people back them. The write-up does not document an appeal path, an escalation tier, or what a user sees when their content is rejected.

The author's own figures: about $15 per 30-day month for 10,000 posts and comments a day. He then extrapolates that rate to the 4.4 billion posts and comments Reddit reported for its year, landing near $18,000 a month, which is his projection rather than a measured or contracted number.

Tom Nick wrote it on 27 September 2026. Jevdit is a prototype social network where Jev moderates content, built around the argument that anonymous posting without moderation quickly fills a site with content no family-friendly blog would print.

The piece argues computers can now understand language well enough to tell whether someone genuinely hates on someone or is only joking, which is what moderation needs. It cites a paper on sarcasm detection in German soccer tweets, and the author reran the test against newer SOTA models and some decision models including Jev.

According to the write-up, SOTA models are now on average better at understanding a joke than a human neighbour. Jev was not as good yet, but clearly outperformed the other decision models on the same irony test.