WarrantOS
claude-provenance on GitHub and as the legacy Claude Code plugin; warrantos on PyPI and the CLI.
No claim ships without a warrant.
WarrantOS does not detect truth, and it does not try to. It enforces that every claim in an AI-assisted document carries a warrant: a source, an explicit [CITE NEEDED], or a BLOCK on the record. A four-state verdict (PASS / HOLD / BLOCK / NOT_ASSESSABLE) gates the output before it ships in enforce mode (the default report mode logs every miss without blocking), and every miss is written to an append-only ledger, tamper-evident against a previously distributed checkpoint, that you can hand an auditor.
It also catches the other way an AI document betrays itself: internal scaffold and conversational residue that bleeds from the chat into the final artefact. "Certainly! Here's the revised version", "As an AI language model, I cannot verify", "based on the information provided", "I hope this helps, let me know if you would like me to expand", a stray [TODO: ...] placeholder. A clean artefact carries its evidence and none of the machinery that produced it. WarrantOS blocks the machinery from shipping.
It governs the artefact, not the model. It runs at the writer's desk, on one document, before it ships, with zero infrastructure: stdlib-only, MIT, no API, no account. Governance platforms watch the system after the fact; WarrantOS gates the output before the fact.
Built in a personal capacity by an independent policy researcher for the people who publish AI-assisted writing under their own name and carry the reputational liability for a fabricated citation: research-integrity, policy, and academic-governance practitioners. It is a personal open-source project, not associated with, funded by, or endorsed by any employer or government. It is informed by the working paper From Citation to Epistemic Governance (Prometheus Policy Lab, in preparation): it operationalises that paper's problem framing, the gap between citation as attribution and citation as evidence, rather than its formal model.
The honest demo, reproducible. The demo ships inside the package, so you do not have to take an anecdote on trust. After pip install warrantos, one command with no setup runs WarrantOS over a bundled AI-style first draft and returns BLOCK: 6 claims detected, 0 supported, 7 boundary violations (scaffold and conversational residue).
warrantos demo
The same fixtures live in the repository at examples/honest-demo/ if you prefer to run the explicit command and inspect the inputs; tools/run_gallery.py asserts the identical verdict on every CI push.
warrantos check examples/honest-demo/draft.md \
--context examples/honest-demo/context.json \
--actor-identity examples/honest-demo/actor.json --profile final-prose
That is the gate working as designed on an unremediated AI draft: it names the epistemic debt so it can be paid down before the artefact ships, instead of going out silently. A governance tool worth trusting is one anyone can hold to its own standard.
Under the hood, claude-provenance wraps AI-assisted writing in an eight-layer pipeline so the final artefact ships clean prose, while a separate audit ledger carries the sources, the feedback, the review history, the transformations, and the structured overrides that produced it. The per-layer status dashboard tells you exactly what is built and what is not.
v0.9.5. Build state: 20 BUILT / 0 PARTIAL (unchanged at v0.9.5; was 13 BUILT / 3 PARTIAL / 2 STARTER / 2 NOT_BUILT at v0.9.1, reaching 20 BUILT at v0.9.2). All five output-integrity gates (G1-G5) and all eight foundation rows are BUILT. The final three rows closed in v0.9.2: F-policy (the normative spec docs/SPEC.md and a machine-readable six-role registry are now committed), F-compliance (a self-assessment control mapping to ISO/IEC 42001 and the NIST AI RMF in docs/COMPLIANCE.md — a documented mapping, explicitly not certified conformance), and F-metrics (shadow-log aggregation via the warrantos metrics command). Adopter-specific configuration (sensitivity tiers, retention windows) and an automated SPEC-ID conformance check remain future work, stated plainly in those docs. See docs/STATUS.md before evaluating scope.
Quickstart
Install from PyPI:
pip install warrantos # MCP server extra: pip install "warrantos[mcp]"
