lyndonkl/claude/classify-claim

Classify Claim

Five buckets

• simplified-correct: strips detail; underlying claim still holds. Keep.
• simplified-boundary: holds in common case, breaks in edge case. Fold break into post.
• wrong: flat factual error. Fix.
• contested: field actively debating. Hedge.
• overclaim: true narrowly, asserted broadly. Scope.

Examples

• "Softmax turns logits into a probability distribution" → simplified-correct (omits calibration caveats but the core claim holds).
• "Attention is O(n²) in memory" → simplified-boundary (true naive; FlashAttention is O(N)).
• "Softmax produces calibrated probabilities" → wrong (distribution ≠ calibrated probability).
• "LLMs exhibit emergent reasoning with scale" → contested (Schaeffer et al. 2023 actively disputes).
• "RAG beats fine-tuning for domain knowledge" → overclaim (true for tail factoid recall; false for reasoning).

Workflow

Per claim:
- [ ] Step 1: Match claim to nearest bucket by definition
- [ ] Step 2: Assign confidence: low (pattern-matching, need source), medium (field knowledge suggests), high (sure)
- [ ] Step 3: Write one-sentence rationale
- [ ] Step 4: If low confidence, defer final classification until cross-reference-claim confirms

Guardrails

1. Never classify wrong without a post-classification primary-source check (that's cross-reference-claim).
2. Respect [contrarian] regions: wrong downgrades to contested or overclaim inside contrarian annotations.
3. Classification is structural — goes on the claim object, not in prose.
4. If the writer's glossary defines a term differently than the field, glossary-alignment-check ran first and flagged it — apply the writer's definition for classification.
5. Default to the more conservative classification on ties. simplified-boundary beats wrong when the claim is right "most of the time."