Skip to content

SPEC-EVO-007 — AI Assistant

Purpose

AI may accelerate research operations, but human judgment and repository evidence remain authoritative.

Permitted assistance

AI MAY propose new objects, relationships, tags, summaries, questions, contradiction candidates, and migration patches. It MAY explain validation failures and generate review bundles.

Requirements

  1. AI MUST NOT silently modify canonical repository content.
  2. Every proposed change MUST be reviewable as a patch or staged object set.
  3. Proposals MUST distinguish quoted evidence, extracted facts, inference, and speculation.
  4. Source references MUST be preserved when available.
  5. Acceptance MUST require an explicit human or authorized policy action.
  6. The repository MUST record whether content was AI-assisted when the project’s provenance policy requires it.
  7. AI confidence scores MUST NOT be treated as truth scores.
  8. Rejected proposals SHOULD be retainable outside the canonical object graph for audit or learning without polluting published knowledge.
  9. Sensitive or private repositories MUST support disabling external model transmission.

Safe write workflow

inspect → propose → validate → review → accept → commit

No step may collapse propose and accept without an explicitly configured, auditable policy.

Acceptance criteria

  • An AI-generated proposal cannot appear in the canonical graph before approval.
  • Reviewers can see the exact proposed diff and evidence basis.
  • Offline or AI-disabled operation retains all core QLens functions.
  • Model substitution does not require changes to canonical object contracts.