User repository on GitHub
Findings from building an AI-image hallucination detector on TypeSafe Jev 1.14 Box + image input: what worked, what didn't, why
- Forks
- 1
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
Jev directory
CommunityJev Tools & CLIs
Jev Box Study
0xRinne's MIT-licensed field report on using TypeSafe Jev 1.14's Box type and image input to spot AI-image hallucinations, across seven pipeline versions v1-v7b, with a live Vercel demo and raw score JSON.
- Stars
- 0
- Language
- JavaScript
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- 0xRinne
- Last verified
- Oct 2026
- Languages
- JavaScript
- Platforms
- Vercel AI GatewayBrowser
- Capabilities
- Open sourceRequires API key
Highlights
- Seven pipeline versions (v1-v7b) probe Jev 1.14's Box type and image input as an AI-image hallucination detector; observations from 14 October 2026.
- v6 caught 4 of 6 planted errors in the office image in 3 of 3 runs and flagged 0 of 26 real Wikimedia photos once two-format confirmation was added.
- Table misalignment (#4 and #5) was never caught; the report attributes it to a Jev 1.14 geometry limit.
- v7b judges each image by its own medium after a cartoon meme was flagged malformed hand 0.88 under photo rules and then went to 0 flags.
- Costs about $0.02 per image: roughly 170-250 calls and 0.5-0.6M input tokens at $0.042 per million.
Watch out
MIT-licensed, AI-assisted independent research built with Claude Code. The report calls its work small, hand-picked evaluations on a handful of images, not a benchmark or an accuracy score; recent versions get one targeted run per image, and the 26-photo false-flag set was not re-run for v7/v7b. Jev 1.14's Box and image input are experimental, header-gated features.
Questions about this entry
Short answers drawn from the linked project or page.
An MIT-licensed field report by 0xRinne on building an AI-image hallucination detector from TypeSafe Jev 1.14's Box type and image input, with seven pipeline versions, raw score JSON, the full website code, two agent skills and a live demo.
Jev Box localizes planted document errors very well, but whether it flags a real AI image depends more on how the question is framed and at what scale the crop is taken than on the model itself; finding which region to ask about is the weakest link.
4 of 6 ground-truth hallucinations caught in 3 of 3 runs (the truncated title, the purple octopus sentence, the chart contradiction, and the layering word), while the two table-misalignment errors were missed. It also falsely flagged 0 of 26 real photos.
Every earlier check assumed photo conventions such as five fingers and real joints, so a cartoon's simplified hands and deliberately mixed fonts were read as errors: malformed hand 0.88, confirm 0.80. Asking the medium first, and passing it to the everything-is-correct option, took the meme to 0 flags.
About $0.02 per image, based on roughly 170-250 Jev calls and 0.5-0.6M input tokens at $0.042 per million, as reported for the v6 pipeline.
The same code runs at hallucination-check.vercel.app; the full app is in the app/ directory, runnable locally with node dev.mjs and deployable to Vercel, with its own README.
No. The report calls its results small, hand-picked evaluations on a handful of images rather than a benchmark or an accuracy score, and recent versions get one targeted run per image; it is a field report on what worked, what did not and why.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Happy JevDay: images, box, longer context, and an opt-in to find jaggedness
X: First among models under 10B on the Decision Index v0.2.1. Liquid AI's d1-3B is built from LFM2.5-VL-3B and…
X: The classic hotdog classifier as a real-time vision decision. Fayaz Ahmed rebuilt the not hotdog app on…
X: Liquid d1 with vision
More Jev resources like this
A local Node app that asks whether an image is AI-generated: the browser measures 16 windows and a lens-fringe field and runs a bundled CLIP ViT-B/16 detector, then sends only a fixed fact dossier to JEV through OpenRouter for a verdict of likely_ai, likely_real or uncertain.
Jev Tools & CLIs#community#github#javascript
Community
GPTchatly/slop-scopeJavaScript
GPTchatly/slop-scopeJavaScriptA Chrome extension and web page that checks whether a public JavaScript or TypeScript repo's README matches its code: rules and five deterministic verifiers decide what is true, Jev labels README lines with calibrated confidence, and Haiku explains the findings.
Jev Tools & CLIs#community#github#javascript
A read-only triage dashboard that finds likely duplicate GitHub issues: lexical retrieval shortlists candidates, Jev answers four Choice questions per pair, and selected passages are shown side by side. It never closes issues.
Jev Tools & CLIs#community#catalog-import#logicrw
A localization pre-review gate that pairs source and target strings and judges meaning, terminology, tone, and placeholders as separate questions, so one failing dimension blocks a release without blending scores.
Jev Tools & CLIs#community#python#localization
Community
A Chrome extension from og2701 that crosses out verbose LinkedIn posts and writes, in red pen, the one plain sentence they are actually making. TypeSafe's Jev works out what each post is for; unsure posts, substantive posts, and posts about grief or illness are left alone.
Jev Tools & CLIs#community#github#chrome-extension
Community
og2701/in-other-wordsJavaScript
og2701/in-other-wordsJavaScriptA self-hosted answer book for everyday choices: an OpenAI-compatible LLM lists candidate options, Jev picks among them with a probability distribution, and answers can be favorited, restyled, or followed up.
Jev Tools & CLIs#community#catalog-import#logicrw
Community
csskrtao/jev-to-answerJavaScript
csskrtao/jev-to-answerJavaScriptFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.