Model Receipt Notes

Short working notes on one narrow problem: proving which image model actually produced a picture, and what that picture cost.

The problem is more common than it sounds. Consumer chat products route between models without telling you which one answered. Resellers list model names their channel pool cannot serve. Pricing is published per token while the question people have is per image. In each case the gap is the same — the output exists, the provenance does not.

These notes come out of building and running a third-party interface for OpenAI's GPT Image 2.5, which prints the model ID the API returned under every result. That interface is at gptimage25.top if you want to see the idea working rather than read about it. The notes here are the parts that generalise past one product.

Notes

• The 16:9 frame that gets refused Why 1024×576 is rejected and 1536×864 is not, and the three simultaneous rules that decide it.
• What "keep intact" actually sends Turning a checkbox list into constraint language the model honours, and why the constraint goes last.
• A PNG corner mark with no canvas and no WASM Compositing a watermark inside an edge runtime using nothing but CompressionStream and arithmetic.