ratchet-reduce

name	ratchet-reduce
description	Extract structured claims from source documents into graph concepts. Adapted from JP's 6R REDUCE phase.
skill_type	graph-extraction
related_skills	["ratchet-reflect","ratchet-reweave","ratchet-verify"]

Ratchet REDUCE — Exhaustive Claim Extraction

Extract every distinct claim from a source document and add to the A2 graph.

Invocation

/ratchet-reduce [file_path]          # extract from specific file
/ratchet-reduce inbox                # process all unprocessed source docs
/ratchet-reduce --entropy-first      # process lowest-entropy docs first

EXECUTE NOW Steps

1. Read source fully. If >2500 lines, chunk into 350-1200 line segments.
2. Source entropy check: Score 1-5 (1=formal spec, 5=raw braindump). If ≥4, quarantine — do NOT extract yet.
3. Hunt for extractable claims by category:
   • DEFINITION: "X is Y" / "X means Y"
   • CONSTRAINT: "X must/shall/cannot Y"
   • RELATIONSHIP: "X depends on/enables/contradicts Y"
   • ASSERTION: "X causes/implies/requires Y"
   • PATTERN: "When X, do Y" / "The pattern is..."
4. Duplicate check: For each candidate, check if already in A2 graph:
   • MATCHED → skip
   • PARTIAL → flag for enrichment
   • MISSING → extract as new concept
   • CONTRADICTS → add CONTRADICTS edge
5. Classify each claim: OPEN (needs investigation) or CLOSED (standalone, ready)
6. Present findings — report by category with counts
7. Extract — create concept nodes with:
   • name: IS the claim (test: "this concept argues that [name]")
   • description: 1-2 sentences, adds info beyond name
   • tags: [source_type, domain_area]
   • trust_zone: A2_3_INTAKE initially
   • source_doc: path to origin
8. Create SOURCE_MAP edges from concept to source document

Quality Gates

INVALID skip reasons (these are BUGS):

• "validates existing approach"
• "already captured"
• "obvious"
• "near-duplicate" (without checking for enrichment)

Expected yields:

Source Size	Expected Outputs
~100 lines	5-10 claims
~350 lines	15-30 claims
~1000+ lines	40-70 claims

Zero extraction from a domain-relevant source is a BUG.

Entropy Scoring Reference

Score	Source Type	Action
1	Formal spec, schema, contract	Extract first (highest priority)
2	Structured doc with definitions	Extract second
3	Mixed structure + prose	Extract with moderate selectivity
4	Chat log, informal notes	QUARANTINE — process after 1-3 complete
5	Raw braindump, stream of consciousness	QUARANTINE — last priority

Pipeline Chaining

After REDUCE → /ratchet-reflect [created concepts] to weave into graph.