claim-extraction

Purpose

Replace vague "candidate units" with atomic, traceable claims — the unit that importance ranking, evidence records, and principles build on. Output analysis/claims.jsonl, validated by validate_claims.py.

Input

• interrogation-records.yaml (Q1–Q18 + evidence_gaps) — what the source is about.
• sources/markdown/<id>.md — the source text (source_text.load_source_texts).
• sources/anchors/<id>.anchors.jsonl — real anchor IDs for source_anchors.

Procedure (two-stage; research-grounded)

1. Stage 1 — detection. Identify check-worthy sentences/passages (claims, not background). (The research recommends a fine-tuned XLM-R encoder for scale; v1 uses an LLM check-worthiness pass — no model dependency. Revisit if recall is poor on long sources.)
2. Stage 2 — decomposition. For each, emit one claims-v1 object using delayed-structure (reason first, then JSON). Classify:
   • component_class: major_claim / claim / premise / non_argumentative (Stab & Gurevych).
   • claim_type: fact / value / policy (+ causal only after in-domain validation).
   • premise_type (premises only): the 6 AAE-FG sub-types.
   • evidence_type (on linked evidence): the 5 AQE types.
   • stance: support / contest / no_relation (3-way, not binary).
   • az_zone: Teufel zones — scientific documents only, else null.
   • certainty: asserted / hedged / speculative (BioScope cue model).
   • condition / exception: nullable. Populate exception on an undercutting pattern (surface cues: unless, except when, only if, provided that, assuming, subject to, absent, if and only if, in the absence of).
   • source_anchors: real anchor IDs; support_granularity: section / page / heading.
3. Separate passes, not joint 4-tuple (full 4-tuple F1=21.39): one pass for claim+type+premise; a later pass (Step 3) for evidence linking.
4. Coverage gate (skill responsibility, not the validator): estimate the ratio of extracted claims to claimable sentences; if < 0.50, re-extract (recall is a model-capability ceiling).
5. Type sanity (advisory): a causal connective (therefore/because) in a fact claim, or if/unless phrasing with an empty condition, signals a mis-type — fix before writing.

Output

• analysis/claims.jsonl (one claims-v1 object per line), scored with score_extracted_units (same 9 dims) → analysis/claim-importance-scores.yaml. Must pass validate_claims.
• evidence/evidence-records.yaml (evidence-records-v1): ≥1 record per high-value claim binding it to source (type / strength / support_level; quote_allowed false for restricted sources). Must pass validate_evidence_records. This is the Step-3 evidence layer, authored together with claims since both come from the same source analysis.

Caveats

• condition/exception is surface-cue heuristic — no NLP dataset exists (GAP-4); treat as best-effort, not validated.
• az_zone taxonomy is from a secondary survey — schema-grade, not classifier-grade.
• LLM extraction can hallucinate; the deterministic validator + Step-1/3 faithfulness catch over-reach, not extraction alone.