🛡 Contradicto Revelio

"Two rules enter. One rule leaves."

Instruction layers accrete. A style plugin demands compression; a workflow demands outputs be pasted verbatim. A global file says "always ask first"; a project file says "never ask, decide". Each rule was right when written. Together they make every compliant action a violation — and a model that obeys widely pays that cost on every turn.

What to scan

All instruction sources as the model sees them together: user-global memory files, the project's own memory files (root and nested, CLAUDE.md or AGENTS.md), rule files, output styles, active plugins' skills, hook message text, and agent definitions. Contradictions live between sources more often than within one.

Anti-patterns to find

1. Direct contradiction. Rule A: "compress all output." Rule B: "paste command outputs verbatim — a summary reads as missing." Obeying either violates the other.
2. Precedence vacuum. A general rule and a specific rule collide ("never edit generated files" vs "this generated folder is hand-tuned; edits survive") with nothing saying which wins where.
3. Cross-source drift. The same rule stated in two files at different strictness ("prefer feature branches" vs "direct commits to main are blocked") — the model can't tell which is current.
4. Obligation vs prohibition. One rule requires an action ("update the doc in the same commit") that another forbids in context ("another team owns these docs — never edit them"). The obligation becomes undischargeable (→ also Finite Recursum class 4).
5. Stale half of a superseded pair. A decision inverted ("X is now adopted") while the old prohibition ("never use X") survives elsewhere. The model finds both; recency isn't visible in a file.
6. Style-vs-substance collisions. An output style (brevity, persona, formatting) that conflicts with content contracts (verbatim verdicts, exact error strings, structured reports).

Fix prescriptions

Detection technique

Contradictions rarely share vocabulary, so grep alone misses them. Build the obligation table: extract every MUST/NEVER/ALWAYS-class rule into (source, trigger, required-or-forbidden action), then sort by action. Collisions cluster on the same action with opposite polarity. For large layers, fan the extraction out per source and join the tables.

Bounds

• Scope: all instruction sources as the model sees them together — user-global memory, project memory, rule files, output styles, hook text, agent definitions. Cross-source pairs are the point; other plugins' files are still out of scope.
• Cap: 15 pairs, ranked by how often both halves are live in the same turn.
• Done when the obligation table has been built once and sorted for collisions — not when no further pair can be imagined.
• Quote once, at read time; mark anything you cannot re-confirm UNVERIFIED rather than re-reading to check yourself.
• See skills/opus-pocus/references/casting-safety.md for the full rules.

Report format

Per finding: the PAIR (fileA:line × fileB:line) · class (1–6) · both quoted rules · the prescription with which rule wins and why. End with the pair count by class. SCANNED NOTHING if fewer than two instruction sources were found, grading N/A — a single memory file cannot contradict itself across sources.