Consolidate Research

Consolidate Research

Synthesize research outputs from multiple sources into a unified, confidence-tiered, pattern-aware report with provenance tracking and quality metrics. Works standalone or manifest-driven from create-research-brief.

Shared references: consolidation-manifest-schema.md (schema), pattern-registry.md (patterns + defaults)

Workflow Overview

1. RECEIVE INPUTS ─── Manifest detection, source identification
2. NORMALIZE INPUTS ── Model-specific extraction → common claim format
3. TRIAGE CLAIMS ───── Atomic decomposition, dependency graphs, provenance
4. RECONCILE ───────── Provenance-weighted disagreement, citation dedup
5. SYNTHESIZE ──────── Cross-domain synthesis, structural artifact merging
6. GENERATE OUTPUT ─── Pattern-specific template, quality metrics, freshness
7. SELF-REVIEW ─────── 7 mandatory checks before finalizing

Step 1: Receive & Validate Inputs

Manifest Detection

Input contains YAML block with `consolidation_manifest:` label?
│
├─ YES (manifest-driven)
│   ├─ Parse manifest fields: research_id, pattern, models, coverage_matrix
│   ├─ Extract consolidation config: mode, verification_priorities
│   ├─ Load pattern-specific output template from Pattern Registry
│   ├─ Note research chain context (upstream_id, inherited_constraints)
│   └─ Use model identifiers from manifest — do NOT guess sources
│
└─ NO (standalone mode)
    ├─ Ask for or infer: topic, sources, consolidation mode
    ├─ Derive pattern from content using Pattern Registry decision tree
    ├─ Use pattern defaults from Quick Reference table
    └─ Confirm inferred pattern + mode with user before proceeding

Source Identification Heuristics

When manifest is absent and source is unlabeled, apply capability-aware signals:

If uncertain after heuristic check, ask user to confirm. When manifest is present, always use manifest.models[].model_id.

Validation

Research outputs provided?
├─ NO → Ask: "Please share the research outputs to consolidate."
└─ YES → Sources identifiable?
         ├─ NO → Ask: "Which source produced each output?"
         └─ YES → ≥2 sources?
                  ├─ NO → Warn: single-source consolidation has limited value.
                  │        Offer: proceed as structured review, or add sources.
                  └─ YES → Proceed to Step 2.

Step 2: Normalize Inputs

Convert heterogeneous model outputs into a common claim format before triaging.

Extraction Patterns by Source

Normalized Claim Format

claim: {id: "C-{seq}", text: "...", type: factual|causal|quantitative|recommendation|unique|reasoning_chain|structural_artifact,
  source_model: claude-opus-4-6|gemini-3.1-pro|gpt-5.2-deep|gpt-5.2-chat|human,
  provenance: {channel: web_search|file_search|site_restricted|quick_validation|internal_document,
    site_restrictions: [], context_documents: [], citation_quality: 0-5, temporal_marker: "YYYY-MM-DD"|null},
  depends_on: [claim IDs]}

Citation Quality Scale

Step 3: Triage Claims

Decompose normalized claims into an atomic claims matrix with dependency tracking.

Claim Types

Claims Matrix

Rows = claims, columns = sources. (qN) = citation quality. Flag same-primary-source convergence separately.

             │ Claude │ Gemini │ GPT-5.2D│ GPT-5.2C│ Status
─────────────┼────────┼────────┼─────────┼─────────┤
Claim A      │ ✓ (q4) │ ✓ (q5) │  ✓ (q3) │         │ Convergent (diff sources)
Claim B      │ ✓ (q4) │ ✓ (q4) │         │         │ Convergent (same primary?)
Claim C      │ ✓ (q3) │        │         │         │ Unique
Claim D      │ ✓ (q4) │ differ │  differ │         │ Contested

Claim Dependency Graphs

Claim A (factual) ──supports──► Claim B (causal) ──supports──► Claim C (recommendation)

Rules: Recommendation confidence CANNOT exceed supporting facts. If a fact drops tiers, propagate to ALL dependents. Document chains in output.

Step 4: Reconcile

Provenance-Weighted Disagreement Protocol

Execute in order — stop at first resolution:

Confidence Tier Assignment

Key rule: Same-primary-source agreement caps at Tier 2 — prevents "3 models citing same blog" inflation.

Dependency Propagation

After reconciling individual claims: (1) Walk dependency graph root-to-leaf. (2) Supporting claim dropped → dependent drops too (floor: Tier 3). (3) Recommendations on Tier 3 facts → flag: "requires independent verification." (4) Document all propagation in Methodology Notes.

False Confidence Audit

Apply to all Tier 1 candidates (mandatory in Adversarial mode):

If concerns → downgrade to Tier 2 with explicit note.

Step 5: Synthesize

Transforms reconciled claims into integrated insights. Mandatory — not optional.

Cross-Domain Synthesis Pass (4 Required Prompts)

Tag all outputs claim_type: cross_domain_synthesis. Document pass execution even if yield is zero.

Structural Artifact Merging

Step 6: Generate Output

Template Selection

Manifest present with pattern → use directly. No manifest → infer from content via Pattern Registry decision tree; confirm with user.

Universal Sections (All Patterns)

Pattern-Specific Sections

In addition to universal sections, each pattern adds:

"For Downstream" Actionability

Quality Metrics

Compute and include in every output:

Freshness Model

freshness_model:
  research_id: "{from manifest or generated}"
  research_pattern: "{pattern_id}"
  topic_volatility: high|medium|low   # from Pattern Registry
  estimated_half_life: 3_months|6_months|12_months
  recommended_refresh: "YYYY-MM-DD"   # execution date + half-life
  staleness_indicators: ["signal 1", "signal 2", "signal 3"]  # 3-5 observable triggers
  source_dates: {oldest: "YYYY-MM-DD", newest: "YYYY-MM-DD"}

Step 7: Self-Review (Mandatory)

Execute ALL checks before finalizing. Report results in the Self-Review Results section.

Consolidation Modes (7)

Mode Quick Reference

Mode Selection Logic

Priority: (1) User explicit > (2) Manifest recommended_mode > (3) Infer from context > (4) Pattern default.

Context inference signals: "high stakes"/"executive" → confidence_weighted | "due diligence"/"comprehensive" → gap_driven | suspiciously aligned outputs → adversarial | "quick synthesis" → standard | strategic/nuanced → depth_first | new domain/landscape → breadth_first | "agentic"/"hands-off" → agentic.

When inferred, confirm: "I'll use [Mode] based on [reasoning]. This will [brief description]. Proceed?"

Pattern x Mode Defaults

Mode Procedures

Standard: (1) Identify convergent claims. (2) Flag divergent for reconciliation. (3) Apply disagreement protocol. (4) Tier by confidence (same-primary-source caps at Tier 2).

Adversarial: (1) Flag suspicious alignment on contested topics. (2) False Confidence Audit on ALL Tier 1. (3) Identify cross-source critique points. (4) Stress test: what falsifies aligned claims? (5) Downgrade where audit reveals concerns.

Gap-Driven: (1) Derive expected coverage from manifest or pattern template. (2) Audit actual vs expected. (3) Document total gaps (no coverage) + partial gaps (single-source). (4) Prioritize by downstream impact.

Confidence-Weighted: (1) Citation Quality Scale on every claim. (2) Per-claim confidence = (sum scores) / (max possible). (3) Weight by quality, not source count. (4) Single quality-5 > three quality-1.

Depth-First: (1) Find deepest source (usually Claude/human). (2) Lead with reasoning chains. (3) Validate key chain claims via other sources. (4) Fill breadth gaps from cataloger (Gemini). (5) Feature strategic insights prominently.

Breadth-First: (1) Find most comprehensive source (usually Gemini). (2) Lead with coverage completeness. (3) Add depth on high-importance areas. (4) Update with recency from GPT-5.2. (5) Completeness is primary metric.

Agentic: Autonomous end-to-end with minimal user intervention. (1) Ingest all + manifest. (2) Execute Steps 2-5 autonomously. (3) Web search to resolve conflicts. (4) Targeted searches for coverage gaps. (5) Draft with cross-domain synthesis. (6) Full self-review with web verification. (7) Compute metrics; revise if needed. (8) Final output. Requires: claude-opus-4-6, thinking: adaptive, effort: max. Use when: all models done, topic bounded, manifest present, user wants hands-off.

Research Chain Protocol

When manifest.research_chain.upstream_id is set: (1) Request upstream consolidated output if not provided. (2) Validate downstream vs upstream consistency. (3) Flag contradictions. (4) Propagate inherited_constraints. (5) Apply chain validations (below). (6) Note in Methodology: "Part of chain: [upstream] → [this]". If no upstream → standard consolidation.

Chain-Specific Validations

Constraint Propagation Rules

Upstream constraints are HARD by default — cannot contradict without explicit justification. If downstream evidence challenges an upstream constraint, flag as "constraint tension" (do not silently override). Document all propagated constraints in Methodology Notes.

Integration Notes

Manifest-driven (after create-research-brief): Manifest provides ALL config — skip derivation. Pattern template pre-selected. Coverage matrix enables precise gap detection. Chain context available. Model capabilities known — provenance tagging is precise. Verification priorities pre-identified.

Standalone: Derive config from content + user input. User may paste a manifest manually. Pattern detection via Registry decision tree. Source ID via heuristic table. Coverage from pattern defaults. Works with any source types.