v0.334.1 - The system card reports the first faithfulness audit

2026-11-06. Text only, in docs/ai-system-card.md. The eighth /mycelium:xai-check on the maintainer's dogfood
project measured what 0.311.0 changed. 0.311.0 holds a framework turn until its Next: line cites a source.

• What was measured. Every Next: line in the twelve most recent sessions cites a source (284 of 284, in ten
  sessions of twelve; three of twelve before 0.311.0). No turn was held for a missing citation. The first
  blind-prediction faithfulness audit ran: ten lines, one fresh reviewer given only the situation and the cited
  source. The reviewer predicted the proposed action in 8 of 10, and partly in the other 2. Citations elsewhere did
  not move: about four per twelve sessions, as in September and October.

• §4, Evaluation methodology. The Self-eval bullet said no faithfulness audit had run and named the Juniors.dev
  pilot, stopped on 2026-08-14, as the source for one. The 0.310.19 fix corrected the bullet below it and missed this
  one. It now states the audit and its limits.

• §5, Per-decision rationale. It said the rule's effect "is measured at the next audit"; it now gives the result.

• §1 and §9 name the eighth audit.

• Alternatives considered.

  • Reporting 284 of 284 alone: the hook demands that number, so it says nothing about whether the cited source
    explains the action. The blind prediction is what does.
  • Calling the rationale surface faithful in general: the audit covers the Next: line in the maintainer's own
    sessions. The card says that, and names a larger sample and an outside builder's sessions as the next step.