Tamper‑evident decision records for AI, anchored to RFC 3161

Description and use

tg‑attest provides high‑risk AI systems with tamper‑evident decision logs, anchoring each evidence set to an RFC 3161 TSA and using Merkle chains to prove integrity.

Problem

High‑risk AI systems must log decisions with evidence that is tamper‑evident and admissible as evidence, but existing LLM observability tools only record that a document was retrieved, not its content at the time of decision.

Outcome

tg‑attest enables high‑risk AI systems to produce tamper‑evident, verifiable decision records that satisfy EU AI Act logging requirements, reducing audit data transfer to ~8 KB per record and allowing auditors to independently verify evidence integrity with standard tools.

Workflow

• 1. Capture decision data and all evidence items at decision time.
• 2. Compute a hash for each evidence item and store the content.
• 3. Build a Merkle tree over the evidence hashes and compute the epoch root.
• 4. Batch epoch roots into a Merkle epoch and generate a Merkle inclusion proof for each evidence item.
• 5. Submit the epoch root to an RFC 3161 TSA to obtain a signed timestamp token.
• 6. Store the decision record JSON, Merkle proofs, and TSA token in a repository.
• 7. To verify, fetch the TSA token, verify the signature with the TSA’s CA certificate, check the Merkle inclusion proof, and confirm the epoch root matches the stored hash.

Evidence

tg‑attest records the point‑in‑time state of every piece of evidence a decision consumed, hashes each record into a chain, batches them into Merkle epochs, and anchors the epoch root to an RFC 3161 timestamp authority.

Source

The Merkle layer isn’t gold plating. A regulator asks about one loan. Without inclusion proofs you hand over the whole ledger to answer. With them you hand over 8 KB.

Source

An auditor verifies with a CA certificate they fetch themselves, not one I supply. The timestamp checks out with stock openssl and nothing installed.

Source