Detect AI-image hallucinations with typed boxes

01 · Role in the system

What Jev does here

The v6 pipeline asks Jev 1.14 (image input, Box answers) to find single sheets, tables, objects and people, then judges masked crops at two close-up scales with one flat Choice per item - 28 hallucination types plus named real-photo look-alikes and can't tell - flagging when the hallucination options sum past a threshold, and gating windows with a mostly-text noul asked in the same call. On the author's AI office image it catches 4 of 6 planted hallucinations in 3 of 3 runs (layering word boxed exactly) and misses two table misalignments; 26 Wikimedia real photos yield 0 false flags; planted-error documents are all caught. About $0.02 per image (170-250 calls, 0.5-0.6M input tokens). The report is unusually honest: an earlier whole-image AI-likelihood noul was removed for inaccuracy, run-to-run flips are documented, and discovery - which region to ask - is named the weakest link.

02 · Control boundary

Where Jev sits

Jev Box searches for neutral discovery, then typed per-crop judgments: a flat Choice list with a summed-probability flag threshold, a same-call noul text-density gate, and deterministic code owning masking, scales and thresholds; the app reports findings only, never an AI-likelihood score.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• Small hand-picked evaluations on a handful of images, explicitly not a benchmark or accuracy score.
• Table-misalignment hallucinations were not caught in v6; discovery of which region to ask is the weakest link.
• An earlier whole-image AI-likelihood noul score was removed for inaccuracy; run-to-run nondeterminism is documented.
• The live demo (hallucination-check.vercel.app) is author-run; raw score JSONs are published in data/scores/.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• RINNECODER ↗Community · github · public · checked 2026-10-15