ShotProof: evidence for image review

ShotProof helps product-content reviewers inspect edited creative assets against an agreed reference. It keeps the image pair, a model proposal and the reason for a review together, so the next person can inspect what the decision was based on. The prototype ran Qwen2-VL-7B-Instruct on an AMD Instinct MI300X VF through the official AMD One-Click Notebook environment, using PyTorch and ROCm. Both actual PNG images were sent to the GPU model. Cached weights loaded offline without a new API key or model download. Each request preserves exact PNG bytes and hashes, source SVG hashes, model snapshot, raw output, token counts and synchronized generation time. Strict validation retains invalid responses for review. Only a whole-response Markdown JSON fence may be removed for parsing; semantic fields are never invented. Reviewer annotations remain separate, and agent demonstrations never count as human approval. Three recorded comparisons exercised addition, occlusion and lighting, with generation times of 1398, 1282 and 1566 ms. These are single-request observations, not a benchmark. The synthetic fixtures contain author-written labels naming expected changes, so they are not blind visual-recognition tests and establish no accuracy or customer benefit. Hidden regions remain unresolved and confidence is uncalibrated. The architectural cover is an AI-generated brand illustration, never an AMD test input. The public static app supports manual review and inspection of saved results; it performs no online inference. The real 42-second GPU recording, original evidence and reproducible source are preserved. The private GPU session was stopped after export. Demo: https://oldblog.zzw4257.cn/shotproof-amd-demo/evidence.html Code: https://github.com/zzw4257/shotproof-amd-demo