sanguo-relationship-semantic-review

name	sanguo-relationship-semantic-review
description	Use when reviewing Sanguo relationship semantic review packets exported by run_codex_relationship_semantic_review_bridge.py. Produces source-grounded JSONL cache rows for relationship trust-zone evidence when local models are too weak or unavailable.

Sanguo Relationship Semantic Review

Use this skill only for exported relationship semantic review packets. It is a temporary Codex-side reviewer that can be replaced later by a stronger local or remote sentence-level model.

Contract

• Input is a JSON packet exported by pipelines/sanguo-rag/run_codex_relationship_semantic_review_bridge.py --mode export.
• Output must be JSONL, one reviewed cache row per input unit.
• Keep canonicalWrites=false.
• Do not invent facts from memory. Judge only whether the preserved sourceSentence explicitly supports one of the candidate pair-key relationships.
• Deterministic extractor output is only a narrowing signal. The source sentence is the truth surface.
• If the sentence is ambiguous, uses pronouns that cannot be resolved from allowedEntities, or only mentions two names without a relationship cue, return not_enough_context.

Review Steps

1. Read sourceSentence first.
2. Identify explicit people and aliases using only allowedEntities.
3. Extract relationship facts from the sentence, if any.
4. Map extracted facts back to candidate trustKey values.
5. For each candidate, emit one relationship decision.

Output Row Shape

Each JSONL row must preserve the input unit fields and add review fields:

{
  "semanticReviewUnitId": "string",
  "promptVersion": "string",
  "sentenceHash": "string",
  "sourceSentence": "string",
  "sourceRefs": [],
  "allowedEntities": [],
  "candidates": [],
  "reviewedAt": "ISO-8601 UTC",
  "reviewer": {
    "provider": "codex-skill",
    "preset": "sanguo-relationship-semantic-review",
    "model": "codex",
    "apiUrl": null
  },
  "reviewMode": "sentence-relation-extraction",
  "semanticReviewPerformed": true,
  "reviewedCandidateKeys": ["trustKey"],
  "relationships": [
    {
      "trustKey": "string",
      "relationshipType": "ruler_subject|spouse|parent_child|sibling|sworn_sibling|enemy_rival|...",
      "fromId": "string",
      "toId": "string",
      "verdict": "supported|contradicted|uncertain|not_enough_context",
      "semanticTrustScore": 0,
      "confidence": 0,
      "typeMatched": false,
      "directionMatched": false,
      "matchedFromNameZhTw": "",
      "matchedToNameZhTw": "",
      "fromEvidenceSpanZhTw": "",
      "toEvidenceSpanZhTw": "",
      "relationshipCueSpanZhTw": "",
      "cueCategory": "",
      "polarity": "",
      "mismatchReasonZhTw": "",
      "evidenceSentence": "source sentence excerpt",
      "rationaleZhTw": "繁體中文簡短理由",
      "normalizedClaimZhTw": "",
      "stableRelation": false,
      "reviewMode": "sentence-relation-extraction",
      "canonicalWrites": false
    }
  ],
  "extractedRelationships": [],
  "rawReviewerSummary": {
    "reviewer": "codex-skill",
    "notesZhTw": "繁體中文摘要"
  },
  "canonicalWrites": false
}

Scoring

• 90-100: sentence directly states the exact pair and relationship type; direction is clear.
• 80-89: sentence supports the relationship, but wording is slightly indirect.
• 50-79: some related signal exists, but not enough for stable trust.
• 0-49: unsupported, contradicted, or not enough context.

Only supported with typeMatched=true, directionMatched=true, high confidence, and explicit sentence evidence may later enter skill review. Human review still controls final 100-point lock.