Disclosure

Disclosure

This page sets out what the project claims, the limit of each claim, and the evidence for both. It follows the format of the AIQT disclosure: for each thing the project does, the claim and its limitation stand side by side, each pointing to the evidence.

Example

Entry 1

What the product is made of

Claim: The 1.0.5 chat artefacts, the Claude Skill and the portable instruction files, are prose instructions and reference material only, with no executable code and no network calls. The repository that ships them also carries code, each file listed in the 1.0.5 release manifest: the Python and shell tooling under tools/ and opf/tools/; the opf enforcement hooks (the pre-commit hook and the Claude Code deny hook), the pre-commit installer, and the CI template and script under opf/enforcement/; the site HTML pages (inline scripts included) and JavaScript under site/ and opf/site/; docs/_shell.html, an HTML shell page with an inline script; the preview hook scripts under .preview/; and the generated Claude Code hooks plugin under plugin/ and its source under .aiqt/core/hooks/.

Limitation: Instructions cannot do anything your assistant platform cannot already do. The 1.1.0 design adds a guided install and a setup doctor around that tooling, each documented before it ships.

Evidence: What the files are (opens in a new tab) The 1.1.0 directory (opens in a new tab)

Entry 2

What it can enforce

Claim: AIQT is a behavioural standard your assistant is required to follow: it announces what its guardrails catch, surfaces issues instead of hiding them, and backs its claims with evidence.

Limitation: It does not technically prevent a model from erring. Results still depend on the model, the platform, competing instructions, and the tools in the conversation. AIQT does not make an AI infallible, and the chat skill does not sandbox, monitor, or block anything at runtime; the repository's CI gates and the opt-in Claude Code hooks plugin block or warn only on the specific actions they are built to catch, within each control's stated residuals.

Evidence: What AIQT does not do (opens in a new tab)

Entry 3

Data handling

Claim: The 1.0.5 chat artefacts send nothing anywhere: they are instructions your assistant reads, with no code and no network calls. The 1.0.5 release is wider than those artefacts: its repository ships code, each file listed in the release manifest and named under What the product is made of, and on your machine none of it runs on its own. What you type still goes to your assistant's vendor under that vendor's terms, exactly as it would without AIQT. In the 1.1.0 design, routine checks run locally and send nothing to a model.

Limitation: The shipped repository code is not network-free when run: some of the checks and tools make network requests when run. Among them: the link-liveness check that verifies published framework URLs, the CWE importer (tools/import_cwe.py), which fetches the catalogue it vendors, and tools/ci-status.sh, which reads CI results through the GitHub CLI. The 1.1.0 substantive and delicate review tiers will send content to model providers. The providers, credentials, retention, and cost stay marked pending until they are documented before 1.1.0 ships. AIQT cannot intercept, encrypt, or block data that your assistant platform itself sends.

Evidence: What the 1.0.5 files are (opens in a new tab) The data flow, by tier (opens in a new tab)

Security

To report a security vulnerability, see the vulnerability disclosure policy (opens in a new tab). The privacy policy that covers this site and the other 5 is at posluns.ca/privacy (opens in a new tab).