Verifiable AI agent receipts

Prove what your AI agent did.
Verify what changed since.

AgentProof turns a Codex session into a signed, verifiable evidence bundle: an ordered, tamper-evident record of commands and file changes with SHA-256 commitments. AgentProof Sign keeps source code local and is deliberately unanchored.

Sign for free, indefinitely. No login for the sample. No blockchain transaction.

agent-session.json
RECEIPT VERIFIED

ROOT HASH

f3c300bbe2db…aed3ff1f
ORDERHash chain valid
SIGNATUREEd25519 signed
TIMESTAMPNot in Sign
PRIVACYHashes, not source
Tamper-evident, not exhaustive.
SHA-256Ed25519Hash chainEB1Apache-2.0

From release evidence to agent evidence

approver packet → agent receipt

Your release pipeline already speaks evidence.

Teams already assemble release evidence by hand: approver packets, SHA-256 manifests and ZIP archives prepared for review. AgentProof applies that familiar evidence-bundle model to a new problem: proving what an AI coding agent did, automatically, in CI.

Built for technical review. Reproducible SHA-256 commitments, an ordered hash chain and an Ed25519 signature can support SOC 2 or ISO 27001 evidence work. AgentProof does not itself certify compliance.

The accountability gap

A canonical evidence bundle, not another activity log.

Think of AgentProof Sign as a flight recorder for AI coding sessions. It produces audit evidence that can be reproduced: SHA-256 commitments for observed commands and files, an ordered hash chain and an Ed25519 signature over the EB1 root. If something changes afterward, the verifier names the file rather than asking you to trust another account of events.

01

Capture

Run a Codex session through AgentProof. Every observed command and file change becomes a SHA-256 commitment in an ordered hash chain.

02

Seal

The canonical receipt and its SHA-256 manifest are placed in an EB1 evidence bundle and signed with Ed25519. Source, prompts and command output remain local.

03

Verify

Anyone can re-hash the repository in the browser. Green means it matches. Red names the exact file that changed.

A reviewable chain

From agent activity to independently checkable evidence.

1Codex sessionLocal activity
2Canonical receiptOrdered hash chain
3EB1 bundleSigned with Ed25519
4Verify againRepository match

AgentProof Sign

Sign a real sample. Keep the proof.

We provide a canonical sample receipt. The gateway validates its hash chain, generates an EB1 bundle through the existing HREVN core and signs its root with Ed25519.

  • Free indefinitely, with no account required for this sample
  • Repository size is irrelevant; only hash-only receipt and frequency safety controls apply
  • Duplicate sample submissions reuse the existing result
  • Raw IP addresses are not stored
  • No blockchain transaction or independent public timestamp
FREE TECHNICAL TRIALLIVE

AUTHORITATIVE DOCUMENT

{ }
agent-session.jsonCanonical sample · hash chain included

AgentProof Sign signs the bundle without anchoring it on a blockchain.

Developer workflow

Put verification next to the pull request.

The AgentProof Sign GitHub Action verifies the canonical receipt against the repository before asking HREVN to sign it. A failed comparison never creates a bundle.

.github/workflows/agentproof.yml
- uses: miguel-herrero-systems/HREVN-AgentProof-Sign-Action@v1
  with:
    receipt-path: .agentproof/agent-session.json
    repository-path: .
    github-token: ${{ github.token }}
✓ Linux, macOS and Windows runners ✓ PR comment and downloadable artifact ✓ Only the hash-only receipt leaves the runner

The standalone AgentProof Sign GitHub Action is prepared for a separate public repository. Marketplace publication remains deferred until Build Week judging is complete.

Honest boundary

Tamper-evident, not exhaustive.

AgentProof proves the integrity of what the instrumented collector observed. It does not watch actions that bypass the collector.

AgentProof can establish

  • Which canonical receipt was sealed
  • Whether its event order still verifies
  • Whether signed bytes were modified
  • Whether a repository now diverges

AgentProof does not establish

  • That the agent's decision was correct
  • That capture was exhaustive
  • That code is safe or legally compliant
  • That human review is unnecessary

AgentProof Anchor

Request managed production anchoring on Base.

Anchor has been validated end to end in a private Base Sepolia environment. It adds an independently observable public timestamp, a detached anchor receipt and signed technical JSON and PDF reports without changing the original signed bundle. It is not yet a public self-service product.

Time boundary: the anchor proves that the exact root existed no later than its public block time. It does not retroactively prove the receipt's declared issue date.

contact@hrevn.com

Questions before adoption

What a reviewer will ask.

Does AgentProof Sign prove that an agent's work is correct?

No. It proves the integrity and ordering of captured evidence and detects later divergence. Semantic correctness still requires tests, review and judgment.

Can we verify a receipt without HREVN?

Yes. EB1 includes the material and instructions needed to reproduce its root and validate its Ed25519 signature. AgentProof Sign is deliberately unanchored, so it provides no independent public timestamp.

Can existing free receipts be anchored later?

Yes, without changing the original ZIP. A later, separate anchor receipt can timestamp the existing root from that later date; it does not retroactively prove the original issue date.

Is AgentProof Anchor publicly available?

No. The workflow has been validated privately end to end on Base Sepolia, including its detached anchor receipt and signed technical reports. It is not deployed as a public self-service tier. Organizations can request a technical conversation about managed production access.

Does the receipt publish our source code?

No. The receipt contains hashes and selected metadata. Repository content remains local unless your own workflow intentionally publishes it.