cross-dim-synthesis

name	cross-dim-synthesis
description	Cross-dimension synthesis — produce insights spanning ≥ 2 dims, resolve contradictions surfaced by reconciliation, output theme summary
version	1.0.0
tags	["synthesis","analysis","cross-dim","insights"]
allowedTools	["rag-search","web-search"]
activateFor	["analyst","cross-dim-synthesizer","insights-generator"]

Cross-Dimension Synthesis Protocol

After researchers collect per-dim findings and the reconciler aligns them, you synthesize cross-cutting insights that no single dim could produce alone. You also resolve the contradictions the reconciler surfaced.

Inputs you receive

• topic, language
• researcherResults[] — per-dim findings (claim / evidence / source) + summary
• reconciliationReport (optional but typical):
  • factTable[] — canonical (entity, attribute, value, sources[]) triples
  • conflicts[] — (factIds, resolutionType, preferredFactId?, rationale)
  • overlaps[], gaps[], termGlossary[]

3 outputs (mandatory)

1. insights[]

Cross-dim judgments spanning ≥ 2 dimensions. Each insight has:

• headline — one-sentence assertion (the insight itself)
• narrative — paragraph explaining the reasoning, citing specific findings
• supportingDimensions[] — list of dim names that contributed (length ≥ 2)
• confidence — 0..1 calibrated estimate

Calibration anchors:

• 0.9–1.0 — multiple primary sources align across dims; no plausible counter-reading
• 0.6–0.8 — strong but with one notable caveat (timing / scope / source quality)
• 0.3–0.5 — directional but evidence is thin or sources have known bias
• < 0.3 — speculative; should probably go to whatRemainsUnclear instead

2. contradictions[]

For every reconciler conflict (and any new ones you spot):

• claim — the contested assertion
• conflictingSources[] — the conflicting source URLs / IDs
• resolution — how YOU resolve it (≥ 30 chars). Acceptable resolutions:
  • "Source A is more authoritative because [X] — adopt A's value"
  • "Both readings are valid in different contexts — A applies to context X, B to Y"
  • "Neither source is conclusive — flag as open question; note in whatRemainsUnclear"

Never leave a contradiction unresolved — "resolution": "TBD" or "待定" is rejected.

3. themeSummary

One paragraph (≤ 800 chars) framing the dominant narrative the insights tell together. This becomes the writer's anchor for chapter sequencing.

Hard rules (never violate)

• Every insight.supportingDimensions.length ≥ 2 — single-dim observations are NOT insights, they're findings (already in researcherResults)
• Do NOT copy a finding.claim verbatim as an insight — synthesize, don't restate
• Every reconciler conflict must appear in your contradictions[] with a concrete resolution
• Use termGlossary canonical forms — don't write "AI" in some insights and "人工智能" in others when the glossary unifies them
• confidence calibration matters — overconfident insights drag the report's trustworthiness; under-confident ones get filtered out by the writer

What this skill is NOT

• Not for collecting new findings (that's the researcher's job)
• Not for fact alignment (that's cross-dim-fact-check)
• Not for chapter writing (that's the writer's job)
• Not for grading the synthesis (that's the reviewer's job)

This skill produces the analytic spine of the report — insights + resolved contradictions + theme summary. The writer uses this to decide chapter sequence and load-bearing arguments.