Audits

This directory holds habitable's audit-as-artifact record. The same way the evidence engine treats a photo's hash and timestamp as committed, checkable facts rather than assurances, we treat the project's own security and accessibility posture as committed artifacts rather than claims in a README.

A tool that asks tenants under threat of retaliation to trust it owes them evidence of its own. So the reviews live here, in version control, where anyone can read them, diff them across releases, and see when something changed.

What gets committed here

• Security reviews — findings from internal and (eventually) external review, with their resolution or accepted-risk rationale.
• Dependency audits — pip-audit output, capturing known-vulnerable dependencies and the bump or mitigation taken.
• CodeQL results of note — static-analysis findings worth recording, with triage (true positive and fixed, or false positive and why).
• Accessibility conformance — the Accessibility Conformance Report lives at ../accessibility/ACR.md (VPAT 2.5, Rev 508), covering the WCAG 2.2 A/AA success criteria, the Revised Section 508 software and support-documentation criteria, and the Functional Performance Criteria.
• Threat-model baselines — content-pinned freezes of ../threat-model.md handed to external reviewers, each with a maintainer re-review and the residual risks to confirm. See threat-model-baseline.md (current: B1).

How to start a review of habitable — scope, run instructions, and synthetic data — is in onboarding.md.

These artifacts are regenerated and re-committed on each release, so a given tag carries the audit state that was true for it. pip-audit and CodeQL also run in CI on every change, but the committed snapshots here are the durable, citable record.

Initial self-review (2026-06-17, v0.1.0)

This is the first entry, and it is deliberately modest.

• Stage. habitable is alpha and concept-stage. The architecture and the evidence method are documented and partly implemented; the project is not yet fit for relied-upon use in a real case.
• Cryptography. The cryptographic work uses well-reviewed primitives via the cryptography library (SHA-256, RSA/PKCS#1 v1.5, Ed25519, X.509) and standard RFC 3161 trusted timestamping (CMS SignedData over TSTInfo). We have written no novel cryptography. The standard timestamp path is exercised fully offline by a local RFC 3161 issuer; the bundled Ed25519 dev TSA is explicitly non-production and always reports an untrusted chain.
• No external audit yet. No third-party security audit has been performed. This self-review is not a substitute for one, and nothing here should be read as an independent attestation.
• Known limitations are tracked honestly in the threat model at ../threat-model.md — including relay metadata exposure, the absence of any duress-safe state, coercion/forensic exposure, and what a timestamp does and does not prove.
• Quality gates. make verify runs the full gate: ruff, mypy --strict, and pytest with property-based convergence tests and tamper-detection tests against clean, altered, and chain-broken fixtures, at roughly 85% coverage. pip-audit and CodeQL run in CI.

Findings addressed pre-audit

Issues the project found and fixed in its own review, recorded here so the trail is honest before external eyes arrive. These predate any third-party audit.

Date	Severity	Finding	Resolution
2026-06-17	Medium (privacy)	A custody-actor identity (an importing peer's fingerprint, details.from) and the tenant's original source filename (details.source) were carried in the exported, signed bundle.json, weakening the "exports name no one" guarantee in threat model §4.	Moved both to a vault-only private_details field that is never hashed and never exported; the union keeps them for its own audit. Previously-produced packets still verify. Regression-guarded by tests/test_guards.py; released in v0.2.0.
2026-06-17	Low (robustness)	The verifier could crash on hostile input: invalid-UTF-8 bundle bytes, and a malformed custody chain. Found by the fuzz/property harness.	Both are now clean rejections (VerificationError / a failed verdict), never a crash.
2026-06-17	Low (interop)	The RFC 3161 client assumed SHA-256 and RSA and mis-read a name-form PKIStatus, so it failed against some real public authorities.	Follows the token's own digest algorithm, dispatches RSA + ECDSA, accepts int/name PKIStatus; verified against DigiCert and FreeTSA in a network-gated CI job.

The privacy finding post-dates the freeze of threat-model baseline B1: the baseline text did not over-claim (§4 promised only that actor/salt/signature are dropped), but the fix strengthens the spirit of that section. The baseline document remains valid as frozen; this fix does not require a new baseline.