Cryptographic provenance for every AI decision your systems make. Drop-in SDK. Tamper-evident manifests. Verifiable by any party, anywhere, at any time. Designed for the regulatory window opening in August 2026.
One line of code. Three stages. Every AI decision attested at the substrate.
One line replaces your existing LLM client. Works alongside your Anthropic, OpenAI, or Gemini integration. No rewrites.
Your call routes through the ByteTrace substrate. Every byte of the AI's reasoning lands in a signed trace — prompt, response, confidence, latency, model identifier, request ID.
You receive the AI's response plus a tamper-evident JSON manifest. Attach it to your output. Deliver it to auditors. Anyone with our open-source verifier can confirm it themselves.
Existing tools watermark the output. ByteTrace attests every decision step.
| C2PA | OBS TOOLS Langfuse · Helicone |
AGENT FW LangChain · AutoGen |
ByteTrace | |
|---|---|---|---|---|
| Signs the AI output | ✓ | — | — | ✓ |
| Records every prompt + response | — | ✓ | partial | ✓ |
| Records confidence at each step | — | — | — | ✓ |
| Bit-replayable trace | — | — | — | ✓ |
| Tamper-evident signature | ✓ | — | — | ✓ |
| Open verifier — any party can run | ✓ | — | — | ✓ |
| Counterfactual replay | — | — | — | ✓ |
| Hardware-deployable variant | — | — | — | ✓ |
| Satisfies EU AI Act Article 50 | partial | — | — | ✓ |
A signed JSON manifest. Human-readable. Machine-verifiable. Every byte of every AI decision is recoverable from this file.
{ "schema": "bytetrace.manifest.v0.3", "artifact_kind": "AI-mediated decision", "created_at": "2026-05-18T03:42:18Z", "provenance": { "trace_sha256": "a77aa42b865b3f1c8e2d…", "llm_calls": 4, "silicon_ops": 81, "llm_call_details": [ { "prompt_excerpt": "Classify sensor reading …", "value": "actionable_anomaly", "confidence": 0.95, "model_id": "claude-opus-4-7", "request_id": "msg_01GYgvErNqWX…", "latency_ms": 4803 }, // + 3 more decisions, each fully attested ] } }
Same substrate. Same manifest format. Different regulatory authority, different decision class, different deployment shape.
Deterministic rules-of-engagement with auditable LLM escalation on ambiguous classifications. Air-gappable when policy requires; network-enabled when it permits.
FDA-cleared device firmware with LLM consultation on edge cases. Local audit trail satisfies 21 CFR 820 traceability. No patient data leaves the device.
Process-control AI with sealed-device deployment. Ambiguous sensor readings escalate to AI judgment; every decision is signed locally and submittable to regulators without exposing operational data.
AI-mediated trading decisions, risk gates, fraud-detection escalation. Bit-replayable manifests support regulator-mandated reconstruction of any flagged decision after the fact.
AI-assisted research, drafting, and document review with chain-of-custody for every AI judgment. Defensible audit trail for discovery, regulatory inquiry, and client review.
Procurement-ready AI provenance for benefits adjudication, permitting decisions, citizen-facing services. Compatible with agency record-retention and FOIA-style disclosure requirements.
EU AI Act Article 50 comes into force August 2026. Providers of generative AI shipping into the EU must mark output in a machine-readable format. Penalties: up to 3% of global annual turnover or €35M. ByteTrace is a stronger compliance pathway than C2PA — attestation at every decision step, not just the final output.
C2PA signs the output — the final image, video, or text file. It tells you "this was generated by AI." ByteTrace signs every decision step the AI made to produce the output — which model, which prompt, which confidence value, at each step. C2PA is one stamp at the end; ByteTrace is a chain of receipts along the way. They are complementary; ship both for maximum attestation.
Negligibly. The substrate adds <20ms of overhead per call — orders of magnitude less than the LLM call itself (typically 500–5000ms). Signed manifest generation happens in parallel with the response stream; no blocking at the API boundary.
No. The substrate is a recording wrapper, not an executor. Heavy workloads happen between your application and the LLM provider's cloud; ByteTrace records the call and signs the trace. Containers scale horizontally; per-call overhead remains rounding error against LLM latency.
You choose. Default destination is your own S3 bucket (the trace never leaves your control). We retain only audit metadata — manifest hash plus signature — never the underlying trace content. Enterprise plans support BYO keys and on-prem trace storage.
Anthropic and OpenAI at launch. Google Gemini and major open-weights providers in v0.5. ByteTrace acts as a transparent proxy — your existing vendor API key passes through; we attest the call only.
Yes. Enterprise and On-Premise deployments support full air-gap. The hardware variant (Olimex AgonLight2-class microcontroller) is available for sealed-device deployment where no network connectivity is permitted at all.
JSON. The schema is published alongside the customer SDK and documented for integrators. We are actively engaging with the C2PA technical committee to position substrate-level attestation as the next-generation layer above output-level signing.
Tensorpunk Labs (TPL) — research and development on verifiable AI infrastructure. ByteTrace is TPL's first commercial product, building on substrate research conducted internally at Tensorpunk Labs. Cloud SaaS, on-premise, and hardware-deployable variants.
We are onboarding design partners selectively, starting with regulated industries and EU-market AI integrators. Send us your operational context and we will respond within one business day.