# Proofmark WebWild v3 model card

## Summary

`Proofmark/proofmark-webwild-v3` is a browser-sized binary real/AI image classifier. It retains the frozen 384-dimensional ViT representation from `OwensLab/commfor-model-384` and replaces the original classifier with a 384→32→1 GELU head. The exported ONNX model is dynamically quantized to Q8 for local WebAssembly inference.

- Input: RGB image, resize shorter side to 440, center crop 384, ImageNet normalization
- Output: one fake-image logit followed by sigmoid and Proofmark's calibrated 65% display mapping
- Size: approximately 23 MB
- Weight SHA-256: `ed17ceb332bef84d0adcc2fa537eef85ed3ac6fb32c30393c326321fbbe54683`
- Upstream revision: `6076002bf0d9dd37537f965ee2f06f826c333b61`

## Training and selection

The backbone stays frozen. The local head was trained on 4,614 image records producing 9,228 feature rows through exact and augmented views. Augmentations include random crop, JPEG recompression, blur, color jitter, small rotation, horizontal flip, and random erasing. ADM examples receive additional loss weight because earlier held-out tests identified ADM as the weakest major family.

Separate data roles were used for weights, calibration, development, and final evaluation. The final 1,749-image v3 shard and its derived web variants were untouched until the model, raw cutoff, displayed threshold, and manifest were locked. Exact and perceptual audit overlap against 8,115 prior files was zero.

## Data and attribution

- [OwensLab Community Forensics model](https://huggingface.co/OwensLab/commfor-model-384), MIT. Upstream training-data details are in the Community Forensics paper and dataset card.
- Real vs AI Corpus by Zitacron, CC BY 4.0. Proofmark uses its per-row source/license metadata.
- [TheKernel01/Tiny-GenImage](https://huggingface.co/datasets/TheKernel01/Tiny-GenImage), CC BY-NC-SA 4.0, revision `6fe1304`.

Dataset images are neither committed nor shipped with the extension. The source datasets retain their original terms. Proofmark source and model modifications are released under the repository's MIT license.

## Evaluation

At the production 65% threshold, the final untouched test scored 94.6% balanced accuracy on 1,749 originals and 91.4% on deterministic JPEG, WebP, crop, downscale, and blur variants. AI recall was 96.0% clean / 93.6% web-stress; real recall was 93.1% / 89.3%. Full per-row results and hashes are in `benchmark-results/`.

## Intended use and limitations

The model is intended for private, browser-local triage of ordinary webpage images. It should help a user decide what deserves scrutiny. It is not suitable as the sole evidence for moderation, employment, legal, academic-integrity, or other consequential decisions.

Performance can change with future generators, screenshots, collages, heavy edits, adversarial processing, unusual illustration styles, or a website distribution unlike the benchmark. The percentage is a calibrated detector score, not a literal probability with universal coverage.