AI transparency

What AI does — and does not do

Deterministic validation is the source of truth. The rule engine evaluates structured requirements against the normalised model on the server. The Evidence Assistant, when configured with a provider, only explains already-computed findings.

• The Evidence Assistant never authors validation results that are stored as evidence, and cannot change a recorded pass, fail or unknown.
• AI never receives raw IFC, DOCX, XLSX, or PDF bytes.
• Without provider keys the product uses a deterministic fallback explanation path.

Providers

Optional providers (Gemini via Google, OpenAI) are selected by environment configuration. Provider priority and verification steps are documented in the README. Outputs are checked against the trusted validation context before display.

Human review

Review decisions (open, acknowledged, resolved, waived) are always human-authored, with optional waiver reason and expiry. AI suggestions do not auto-waive findings.