Portable evidence for AI work, verifiable offline. Integrity, not truth
One file. No verification server. No network required.
Quick start · What it proves · Current release · Adoption review · Documentation
proofbundle runs as a plugin in Claude Code and Codex. Ask your agent to check a receipt, and it explains which signature and integrity checks passed, what was not checked, and what the receipt cannot tell you. For decision and outcome receipts, it verifies against an issuer public key you trust. An evidence bundle uses a separate check that does not by itself establish whether you should trust its signer.
Claude Code
claude plugin marketplace add b7n0de/proofbundle
claude plugin install proofbundle@proofbundleCodex
codex plugin marketplace add b7n0de/proofbundle
codex plugin add proofbundle@proofbundleThen ask for /proofbundle:verify receipt.json in Claude Code, or for the skill proofbundle:verify in Codex. For a decision or outcome receipt, give the issuer's public key when your agent asks for it. The plugin needs uv, and network access to PyPI on its first start. Plugin 0.4.0, experimental in both hosts. How the plugins work
v6.6.0 · Beta · Audit status and limits
Known limitations for 6.6.0 · Earlier risk record · Release notes · Release scope
What was checked, and what remains open
A feature release following 6.5.0. The installed package adds proofbundle scitt match (experimental) and lets decision verify and outcome verify check the action bytes, the case, an approval record, a consumption state, the effect bytes, the bound decision and the expected executor that the relying party supplies. The command line no longer accepts abbreviated options, and several checks are stricter. Before upgrading, review the CHANGELOG entries for the command line, related targets, executor identifiers, eval-result v0.1 claims and evidence packs.
- Inputs of the relying party.
decision verifytakes--action,--case,--approval,--consumedand--received;outcome verifytakes--executor-id,--effect,--decisionwith--decision-pub, and--case executed. A check whose input is missing is not performed, and neither command is a complete implementation of the draft. - SCITT statement identification.
proofbundle scitt matchreports the axes of draft-gruszka-scitt-statement-identification-03 for a reference and a candidate COSE_Sign1. It is experimental, and no verify path imports it. - Producers and closed sets. Producers sign only what their own verifier reads, and the checks of evidence packs, vector sets, installed builds and pre-tag receipts refuse members they do not name.
In the tagged tree, check audit_artifacts/660/pre_tag_receipt_v6.6.0.json for the recorded audit command and result.
The package being published and its closing audit passing are separate facts. An audit that was not run makes no statement about the absence of defects.
Install the verifier, download an example, then verify the local file.
python -m pip install proofbundle==6.6.0
curl -fsSLo receipt.json \
https://raw.githubusercontent.com/b7n0de/proofbundle/v6.6.0/examples/example_bundle.json
proofbundle verify receipt.jsonPython 3.10 or newer. Installation and download use the network. Verification reads the local file only.
Exit codes and the tamper demo
These exit codes apply to proofbundle verify, not to every command in the package.
| Exit code | Meaning |
|---|---|
0 |
Verified |
1 |
Verification failed |
2 |
Malformed input or usage error |
3 |
Relying party policy not met |
decision verify, outcome verify and relation-statement verify have separate contracts in their own --help.
To try deliberate tampering, install the evaluation extra and run the demo.
python -m pip install 'proofbundle[eval]==6.6.0'
proofbundle demoThe demo checks an honest receipt, tampered variants and a sample swap. It exits with a nonzero code if a tamper is accepted.
| What verification can establish | What it does not establish |
|---|---|
| Which key signed the content | Whether you should trust that key |
| Whether signed content has changed | Whether the reported result is true |
| Whether supplied proofs and requested policy checks pass | Whether the work was correct or complete |
A valid signature does not make a reported result true. Checks depend on the receipt format and the policy you request.
A PASS means: the signature is valid, the explicitly checked artefacts match the signed digests, and a trust policy passed if you gave one with --policy (without it the output says POLICY: NOT_EVALUATED). Truth, code quality and provenance from a particular AI model are not confirmed by it.
| I want to | Start here |
|---|---|
| Verify a receipt | Quick start |
| Create evaluation evidence | Evaluation walkthrough |
| Add receipts to Inspect AI | Inspect plugin |
| Gate a push or pull request from an AI coding agent | Plugin catalog |
| Assess proofbundle for adoption | Adversarial review guide |
Other workflows and optional features
| Workflow | Package or reference |
|---|---|
| Evaluation receipts and preregistration | proofbundle[eval] · Claim format |
| Inspect AI | proofbundle[inspect] · Plugin setup |
| Agent review disclosures | Profile inventory · Conformance examples |
| RFC 3161 and OpenTimestamps | proofbundle[anchors] · Anchor guide |
| ML-DSA-44 witness cosignatures | proofbundle[pq] · Anchor guide |
| TEE attestation bridge | proofbundle[experimental] · Experimental bridge |
Shipped features do not all have the same maturity. Agent review disclosures are self declarations. Anchor and enclave paths have experimental boundaries. Check the predicate inventory for the exact profile before relying on it.
| Where | Install | What you get |
|---|---|---|
| Claude Code | claude plugin install proofbundle@proofbundle |
Verify and review receipts from your agent session |
| Codex | codex plugin add proofbundle@proofbundle |
The same skills in Codex |
| pytest | proofbundle[pytest,eval] |
A signed receipt for a test run |
| Inspect AI | proofbundle[inspect,eval] |
A signed receipt for an eval |
Receipt emission in pytest and Inspect AI is opt-in and requires an explicit pass threshold.
Commands, threshold and output
pytest
python -m pip install 'proofbundle[pytest,eval]==6.6.0'
PROOFBUNDLE_EMIT=1 PROOFBUNDLE_THRESHOLD=0.9 pytestInspect AI
python -m pip install 'proofbundle[inspect,eval]==6.6.0'
PROOFBUNDLE_EMIT=1 PROOFBUNDLE_THRESHOLD=0.9 inspect eval task.py --model <model>These commands use POSIX shell syntax. Run pytest in your test project; replace task.py and <model> with your Inspect task and configured model.
0.9 is an example, not a default. pytest records pass_rate; Inspect selects the first available metric unless you set PROOFBUNDLE_METRIC. The comparison defaults to >=; set PROOFBUNDLE_COMPARATOR to change it. The threshold verdict is recorded in the receipt and does not change the test or eval exit status. Without PROOFBUNDLE_THRESHOLD, receipt emission is skipped.
Receipts go to the current directory by default, and the plugin prints the output path. Set PROOFBUNDLE_OUT to choose a file or directory. Set PROOFBUNDLE_KEY to the path of a file containing a 32-byte raw Ed25519 seed for a reusable signing key. Otherwise, an emitted receipt uses a temporary key and is not tied to a lasting identity.
| Your question | Reference |
|---|---|
| How do I implement the format? | Specification · Conformance |
| How do I integrate my workflow? | pytest and Inspect plugins · Other integrations · Glossary · Plugins |
| Which keys and claims should I accept? | Policies · Trust anchors |
| How is security assessed? | Threat model · Security policy |
| How was this release prepared? | Release process · Pre tag audit |
How it works
Your evaluation, review, decision or action
↓
Canonical statement, signature and supplied proofs
↓
One portable receipt file
↓
Offline verification and explicit policy checks
The verifier checks the evidence it receives. Expected subjects, trusted keys, freshness requirements and acceptance policies come from the relying party. Missing evidence is not evidence that omitted work never happened.
Capabilities and maturity
The core supports signed receipts and Merkle inclusion proofs. Evaluation receipts can bind metrics, thresholds, provenance and commitments. Selective disclosure can hide selected values while preserving the checks supported by its profile.
Decision receipts record a verdict over named evidence. That does not establish that the decision was correct. Outcome, relation, run ledger, trust pack and verification summary profiles have their own maturity limits.
The Rust cross verifier is experimental and advisory. Agreement on recorded cases does not prove either implementation correct, and the Rust tool is not part of the Python package.
Security and trust
The core uses cryptography and rfc8785, rather than implementing its own cryptographic primitives. The test approach includes external vectors, mutation checks and parser fuzzing. Those are test signals, not a proof of correctness.
Receipt signatures are Ed25519, not post quantum. ML-DSA-44 witness cosignatures and the experimental renewal path do not turn the payload signature into a post quantum signature. See the anchor documentation.
Build provenance and package attestations answer questions about the build and its bytes. They do not establish the truth of an evaluation or replace a security audit. See the release process.
Report a vulnerability · Conformance · Adoption review
The OpenSSF Scorecard is a heuristic, not a product verdict. Read the per check explanations and self assessment.
Standards and interoperability
proofbundle complements other evidence systems. A format mapping or an open proposal is not the same as adoption by the upstream project.
Tool comparison · Receipt envelope profile · in-toto mapping · SCITT mapping · Related work
The interop discussion with inspect-receipts records a specific envelope comparison. It must not be read as evidence of a second independent implementation of every predicate.
Scope, citation and contributing
proofbundle is not a hosted transparency service, a complete in-toto client, a trusted execution environment, a consensus system or a compliance product by itself.
Release scope records what belongs to this release. Deferred work is not a delivered capability.
Use CITATION.cff for citation metadata. The software archive has concept DOI 10.5281/zenodo.21110642. The Technical Note has concept DOI 10.5281/zenodo.21230466. Software and Technical Note versions are separate records.
Contributing guide · Code of Conduct · Good first issues · Security reports
MIT license · Part of b7n0de, Verified AI Work