contradict

CONTRADICT — Coherence Auditor

You generate sequentially — one section, then the next. Each section is optimized for local coherence. Section A reads well. Section C reads well.

But locally coherent sections can be globally contradictory. Section A promises simplicity. Section C introduces complexity that destroys the simplicity. Both sound right alone. Together, they are a lie.

You have no built-in global consistency checker. This skill is that checker.

---

The Failure Mode You Must Recognize

You are about to deliver an answer where:

• The introduction says "simple and straightforward"
• Step 7 requires complex configuration across three services
• The summary says "no external dependencies"
• Step 4 uses two external libraries

This is the coherence trap: local fluency creates a false sense of global correctness. Each part is individually convincing, but the parts cannot coexist. You don't notice because you generate forward, not cross-sectionally.

---

The Protocol

Step 1 — EXTRACT: Write Every Claim as a Standalone Statement

Read the entire output. Extract every claim, promise, estimate, and assertion. Write each as a standalone sentence that could be true or false:

CLAIM EXTRACTION
────────────────────────────────────────
C1: [section/paragraph ref] — "[exact claim]"
C2: [section/paragraph ref] — "[exact claim]"
C3: [section/paragraph ref] — "[exact claim]"
...
────────────────────────────────────────

Types to catch:

• Promises: "This approach is simple," "No configuration needed," "Handles all cases"
• Estimates: "Takes about a week," "Under 100ms latency," "Minimal effort"
• Scope claims: "No external dependencies," "Works offline," "Supports all browsers"
• Characteristic claims: "Highly scalable," "Easy to maintain," "Secure by default"

Minimum count: Extract at least 10 claims for a substantive answer. If you found fewer, you're extracting too loosely — tighten your filter.

Artifact: Numbered claim list with section references. Step 2 cross-compares these.