Evidence that tells on itself

Tellproof turns an AI system's runtime telemetry into signed, hour-by-hour evidence that a bank, an insurer or an auditor can check without trusting you. No prompt or completion ever leaves your machine. And every check can be shown to fail: an edited hour, a guardrail that stopped catching or a CI gate that tests nothing gives itself away.

pip install tellproof tellproof-agent     # or: npm install tellproof
tellproof-agent run --local               # dashboard on 127.0.0.1:4319

Then add tellproof.init() to your app. Python and Node.js; OpenAI, Anthropic and Amazon Bedrock are instrumented automatically. No account needed.

Signed, content-free evidence

Every hour of model calls becomes a signed batch chained to the one before: calls and tokens per model and region, guardrail decisions, personal-data redaction, agent tool use, changes and anomalies. Counts and short names only.

Checks that can fail

Guardrail canaries send known-bad probes through your own guardrail and record whether it caught them. check --self-test shows every evidence objective fails on bad evidence and that tampering with the log is detected.

Governance as code

Declare models, regions and registers in tellproof.policy.json, next to the code. tellproof-agent check evaluates eleven evidence objectives and fails CI when one fails; aibom writes a CycloneDX AI bill of materials.

Evidence for coding agents

tellproof-agent ci records which gates ran on each commit, how many tests ran, and whether a must-fail canary showed the gate can fail, in a signed in-toto statement that counts the commits a coding agent wrote.

Verify, without trusting anyone

Packs are DSSE-signed in-toto statements over canonical JSON. Three independent verifiers (Go, the browser and Python) agree on every published test vector. Verify a pack in your browser, or offline with tellproof-verify.