cross-dim-fact-check

name	cross-dim-fact-check
description	Cross-dimension reconciliation protocol — extract fact table, detect conflicts/overlaps/gaps, build figure candidate pool from N parallel research streams
version	1.0.0
tags	["reconciliation","fact-checking","quality","research"]
allowedTools	["web-search","fetch","rag-search"]
activateFor	["reconciler","fact-checker","cross-dim-reviewer"]

Cross-Dimension Fact Reconciliation Protocol

When N parallel research streams have completed and downstream synthesis (analyst / writer) is about to start, run this reconciliation pass. Do NOT produce new research — your job is to align, deduplicate, and surface conflicts.

Inputs you will receive

• topic and language
• plan.dimensions[] — what each stream was supposed to cover
• researcherResults[] — per-dimension findings + figureCandidates

5-step protocol (execute in order)

Step 1 — Extract fact table

Scan all findings, distill (entity, attribute, value, sources[]) triples.

• sources is a list of URL or finding-source strings — multiple sources increase confidence
• Each fact gets a stable id like fact-1, fact-2, ...
• Aim for ≥ 3 facts total (mission must surface core evidence)

Step 2 — Detect conflicts

Same (entity, attribute) with different values → flag as conflict.

• resolutionType:
  • preferred-one — one source clearly more credible (gov / academic > blog) → set preferredFactId
  • kept-both — sources equally credible → keep both with annotation
  • flagged-unresolved — only when no info available to decide (rare; ≤ 30% of conflicts)
• rationale ≥ 20 characters explaining the decision

Step 3 — Detect overlaps

Cross-dim claims with similar meaning (semantic, not exact-match).

• similarityScore 0–1 (judge by reading; you don't run embeddings)
• resolutionAction:
  • merge-into-cross-dim — core finding shared across dims → write once in cross-dim section
  • keep-both — different angle on same topic → keep in respective dims
  • drop-from-second — verbatim duplicate → keep in primary dim only

Step 4 — Detect gaps

plan.dimensions[i].rationale promised aspects that findings don't cover.

• severity: "critical" if gap breaks the dim's purpose
• severity: "minor" if peripheral

Step 5 — Aggregate figure candidate pool

Aggregate researcherResults[*].figureCandidates into a single deduplicated array.

• Deduplicate by sourceUrl — keep the entry with highest relevanceHint
• Cap at 20 figures total (rank by relevanceHint=high first, then caption informativeness)
• NEVER fabricate figures — only aggregate what researchers actually extracted
• Empty array is acceptable — prefer correctness over coverage

Reconciliation report (≤ 1500 chars markdown)

### 对账总览
- 事实表概要 (count + key entities)
- 冲突 (each conflict + resolution + rationale snippet)
- 重叠 (each overlap + action)
- 空白 (each gap + severity)
- 下游消费指引 (one-line guidance for Analyst/Writer)

Downstream Analyst & Writer MUST consume this — be precise and quotable.

Heading discipline: this report becomes the cross-dim chapter body when the analyst skips crossDimAnalysis. The chapter title ## 跨维度分析 is added by the report assembler, so never use # or ## headings inside this report — they would split into top-level chapters in the rendered report. Use ### or list bullets for sub-sections.

Hard rules (never violate)

• factTable.length ≥ 3
• conflicts[i].rationale.length ≥ 20
• flagged-unresolved ≤ 30% of conflicts
• figureCandidates.length ≤ 20
• Every factTable[i].id is unique
• Every conflict.factIds[j] references a real factTable[].id
• Every figure.evidenceCitationIndex is set and points to a citation
• Every figure.sourceUrl starts with https://
• If two facts share (entity, attribute) they MUST appear in some conflict.factIds

What this skill is NOT

• Not for generating new research findings (that's the researcher's job)
• Not for writing the final report (that's the writer's job)
• Not for evaluating overall quality (that's the reviewer's job)

This is an alignment pass — it fails fast if the inputs are inconsistent or thin, and it stays silent (gaps: []) when the inputs are well-covered.