Observable Execution

VertRule makes AI execution observable. Every forward pass produces a cryptographic receipt: a per-layer record of what the model computed, verifiable after the fact by anyone with the receipt and the public verifier.

This demo runs a full GPT-2 forward pass in your browser using WebAssembly. As each token is generated, the runtime captures BLAKE3 digests at every transformer layer — attention, MLP, residual, layer norm. These digests form a receipt that proves exactly what was computed and in what order.

What you can observe

• Determinism. Run the same prompt twice. The digests are identical. The runtime is deterministic by construction, not by accident.
• Layer-level transparency. Click any token to see its per-layer cryptographic fingerprint. Compare steps to see where the model's internal state changed and where it stayed the same.
• Provider invariance. Switch between IO providers (how weights are loaded). The execution output is identical regardless of provider — proven by digest, not by trust.
• Downloadable receipts. After a run completes, download the execution receipt as JSON. It contains every layer digest, the full token sequence, and the final execution hash.

What this does not claim

• This is a prototype, not a production feature.
• It is not a hosted inference service.
• It is not a committed roadmap promise.
• It may change or disappear without notice.