Know what your model actually computed.
Every traced answer comes back with readings from the model's own neurons, and your rules can act on them.
From a prompt to a measured, signed answer.
A model reads untrusted input
A support ticket pasted into your copilot hides an instruction to leak customer data.
An OpenAI-compatible call, or our worker beside your own weights.
Trace works with open-weight models such as Llama, Qwen, Mistral and Gemma, through an OpenAI-compatible endpoint or our sidecar beside your own weights. Hosted Qwen3-8B is being provisioned, so the sandbox replays real, recorded activations from Qwen2.5-0.5B.
- your appPrompt → open-weight modelProvenance Trace · canary readings + trace id
- measured readingsTraced inferencecanary_tripped blocks · diffs gate releases
- acts on itYour policy · your CI
Everything you need to trace the model behind your product.
Traced inference
An OpenAI-compatible endpoint that returns the completion plus canary readings and a trace id.
Canary guardrails
Use neuron readings as policy conditions. Block on canary_tripped, and fail closed when no reading exists.
Neuron-CI diffs
Diff a fine-tune against its base: label changes, unexplained neurons, canary drift against a tolerance.
Self-host sidecar
Run our worker beside your own open-weight model. It signs readings with a key you register.
Public neuron registry
Keyless reads of every registered neuron: labels with their labeler, top examples, measured ablation effects.
Verifiable manifests
Every trace and diff is signed, so anyone can check it offline against our published keys.
Trace a call, gate on it, diff the next model.
Pick an endpoint, flip the toggles, and send.
POST /v1/trace/chat/completionsPoint an OpenAI-style client here. Alongside the completion you get a provenance block: the trace_id, canary readings measured from the model's own neurons, and hashes of the prompt and completion. Message text is kept only if you opt in with provenance.retain. Every response names its backend. Set provenance.on_trip to block or block_and_lock and a tripped canary withholds the completion (finish_reason: "canary_tripped") and can lock the agent session.
POST /v1/trace/chat/completions
{
"model": "qwen2.5-0.5b-instruct",
"messages": [{ "role": "user", "content": "Summarize ticket #4821" }]
}Click “Send request” to see the response.
An illustration with real field names and reason codes — nothing is sent from this page.