Check reference / Clarity & Instructions
CLARITY-02 No internal contradictions
LLM-assisted — runs with an Anthropic key (BYOK on the CLI, managed on Pro hosted scans).
Why it matters
Instructions accrete: a rule gets added at the top, its opposite survives at the bottom, a step keeps pointing at a section that was deleted two edits ago. A model given "always X" and "never X" does not error — it silently picks one, and which one changes run to run. Contradictory instructions are the most invisible way to lose control of an artifact's behavior.
How to fix it
Reconcile every conflict the evidence quotes: keep one rule, delete its rival, and fix steps that reference removed sections. An LLM judges consistency (pass at 80/100) and quotes both sides of each conflict it finds, so you can search for the exact text.
Example
✗ flagged
## Rules
- Always commit after every change.
…(200 lines later)…
- Never commit; leave staging to the user.✓ passes
## Rules
- Never commit; leave staging to the user. (One rule, stated once — no rival elsewhere.)