Slop Scope

Slop Scope asks whether an image is AI-generated. Your browser measures the image and a JEV model judges the measured facts; this edition runs on your own computer with your own OpenRouter key, needs no browser check or external database, and enforces no usage limits.

The browser derives a fixed fact dossier from 16 measured windows and a lens-fringe field, plus a whole-image CLIP ViT-B/16 detector. The local server rebuilds two typed JEV requests from templates, an evidence panel and a verdict, and sends them to OpenRouter with your key. JEV is the only classifier.

Only the fixed enum-only evidence wire goes from the page to the local server, and the two typed requests go from the local server to OpenRouter and JEV. The image, its pixels, measurements, filename, hash and metadata text are never sent; Hugging Face sees only your address if you choose to load a model.

Starting costs nothing; each analysis is one panel request and one verdict request paid with your own key, about $0.00013 at September 2026 prices. The page shows each request's provider-reported cost, an unknown cost stays unknown, and you cap spending with a credit limit on the OpenRouter key.

The verdict is likely_ai, likely_real or uncertain, and uncertain is a valid outcome. The README notes JEV's weights are not calibrated probabilities, and an empty or inconclusive result is treated as its own category rather than forced into a yes or no.

The README lists the untested parts plainly: the <=2% wrong-class reference-frequency gate fails for f1's AI-side detector facts, no paired JEV keep run was made, the judge was never run on f1's dossiers, and no pipeline here meets a bar of at most 1% misclassified with uncertain counted as a miss.

The GPTchatly organisation publishes the code under the MIT license; the repository was created 2026-10-03 and pushed the same day. The runtime bundles onnxruntime-web (MIT), the f1 model is non-commercial due to OpenFake, and this edition needs Node.js 22.6 or later.