Traceseal — The Agent Accountability Platform

EU AI Act, Article 50 — transparency obligations for AI systems take effect 2 August 2026 . Receipts are how you prove compliance.

Platform

Trust infrastructure for the agent economy.

Agents are doing real work — writing code, moving money, touching production. Traceseal is the accountability layer underneath: prove what ran, verify it against ground truth, replay it under governance.

/ RECEIPTS<br>Verifiable receipts<br>what ran, signed

Every agent invocation emits a signed, self-contained receipt: skill identity, content-addressed manifest, input/output hashes, sandbox profile, operator signature.

ed25519 signatures<br>SHA-256 content hashes<br>canonical JSON — any tamper breaks the seal

/ VERIFICATION<br>Ground-truth verification<br>did it actually work

Attestation goes beyond "it ran": commands re-execute inside a kernel-namespace sandbox, and forbidden actions are enforced by the sandbox itself, not merely observed.

bwrap isolation — read-only root, no network<br>enforcement-as-attestation<br>independent witness cosigning

/ GOVERNANCE<br>Replayable governance<br>audit at any time

Receipts chain into a transparency log. Any third party can re-verify the whole history with one command — no access to the operator's machine, keys or audit trail.

public transparency log<br>chained run receipts<br>exit codes: 0 valid &middot; 1 invalid

How it works

Three roles. One verifiable proof.

Every receipt has three sections: what ran, who authorised the code, and who vouches for the execution. Any third party can verify all three.

01<br>Sign<br>publisher &rarr; skill

The publisher signs the skill bundle with an ed25519 key. The signature covers every source file via a content-addressed manifest, recorded in the transparency log.

Algorithm: ed25519<br>Covers: manifest over all source files<br>Published: transparency log entry

02<br>Seal<br>operator &rarr; execution

The operator runs the signed skill inside a kernel-namespace sandbox. The runtime records inputs, outputs, timing and sandbox configuration as hashes, then signs the record.

Sandbox: bwrap isolation<br>Records: input/output content hashes<br>Signed: operator ed25519 signature

03<br>Verify<br>anyone &rarr; receipt

Any third party installs traceseal-verify and runs one command. The verifier checks the operator's signature over the canonical JSON. No operator connection. No trust assumptions.

Input: receipt.json only<br>Dependency: cryptography<br>Exit codes: 0 valid &middot; 1 invalid

Regulation is arriving

The EU AI Act makes agent transparency mandatory.

Article 50 transparency obligations take effect on 2 August 2026. Organisations deploying AI systems will need to demonstrate what their systems did — not assert it. Signed execution receipts turn that from a documentation exercise into a one-command verification. Read Article 50 &rarr;

02 AUG 2026EU AI Act &middot; Article 50 applies

Built in the open

Open spec. Open verifier. Public log.

Everything a third party needs to hold an agent to account is public — the receipt spec, the verifier, the runtime attestation tooling, and the transparency log itself.

traceseal-verify PyPI v1.0

The independent verifier. One command, one receipt.json, zero trust assumptions. Ships with the open receipt spec.

pypi &rarr; github &rarr; spec &rarr;

traceseal-observe PyPI

Runtime attestation: records execution inside the sandbox, hashes inputs and outputs, and signs the receipt with the operator key.

pypi &rarr;

traceseal-langchain PyPI

Drop-in LangChain integration — signed receipts for chain and tool invocations without changing your agent code.

pypi &rarr;

alibi open source

The trust harness: reproducible behavioural probes that grade agent CLIs, emitting signed, witness-cosigned Traceseal receipts.

github &rarr;

Transparency log: log.traceseal.io

Get started

Verify a receipt in 30 seconds.

No account. No API key. Just a receipt and the open-source verifier.

$ pip install traceseal-verify<br>Talk to us &rarr;