Observable Execution

Run a transformer locally. Inspect every layer. Keep the receipt.

This demo executes a full transformer forward pass in your browser using WebAssembly. As each token is generated, the runtime records a cryptographic digest at every layer and binds those digests into a receipt you can inspect and verify afterward.

Available models:

• GPT-2 (124M)
• Qwen2 (0.5B)
• SmolLM2 (360M)

Model selection is part of the execution context. The selected provider and model identity are recorded in the receipt payload and committed via its digest.

What you can observe

• Deterministic runs under the same inputs and execution context
• Per-token, per-layer execution fingerprints
• One execution surface across multiple models
• Downloadable JSON receipts

What this is

A working research prototype you can run today.

What this is not

• Not a production feature.
• Not a hosted inference service.
• Not a roadmap commitment.

The prototype is real now. What may change is the research surface around it.