feed-the-machine 1.5.0 → 1.6.0

package/ftm-researcher/references/synthesis-pipeline.md

@@ -1,165 +1,165 @@
-# Synthesis Pipeline
-
-5-phase pipeline that takes raw findings from finder agents and produces a structured disagreement map.
-
----
-
-## Phase 1: Normalize & Deduplicate
-
-Input: Raw findings from all finder agents (7 agents x 3-8 findings each = 21-56 findings)
-
-Steps:
-1. Flatten all findings into a single list
-2. Group by semantic similarity (same claim from different agents)
-3. For each group:
-   - Merge into a single canonical claim
-   - Track which agents found it (agent_count)
-   - Track source type diversity (source_diversity_score = unique source types / total sources)
-   - Flag circular sourcing: if all sources in a group cite the same original source, mark as circular=true
-4. Output: unique_claims[] sorted by agent_count DESC, source_diversity_score DESC
-
-### Semantic Similarity Heuristics
-
-Two claims are considered semantically similar when:
-- They make the same factual assertion about the same subject, even with different wording
-- One is a subset of the other (e.g., "X uses Y" vs "X uses Y for Z")
-- They cite the same source for the same conclusion
-
-Two claims are NOT similar when:
-- They address different aspects of the same topic
-- They reach different conclusions about the same subject
-- One is general and the other is specific with additional qualifying conditions
-
-When merging, keep the most specific version as the canonical claim.
-
----
-
-## Phase 2: Adversarial Review (ftm-council)
-
-Input: Top claims from Phase 1 (all claims with agent_count >= 2, plus any high-confidence unique claims with confidence > 0.8)
-
-Council invocation:
-- Send claims as a structured prompt to ftm-council
-- Ask: "Evaluate each claim. For each: Is the evidence sufficient? What would make this wrong? Are there alternative explanations? Rate confidence 0-1."
-- Council runs Claude + Codex + Gemini independently, then reconciles
-
-Output: claims[] with council_verdict (agreed | contested | insufficient_evidence), provider_disagreements[]
-
-### FALLBACK (if Codex/Gemini unavailable):
-
-Spawn 2 standalone agents on the review model:
-
-**Devil's Advocate:** "Your job is to find reasons each claim is WRONG. Search for counter-evidence, flag single-source claims, identify logical gaps."
-
-**Edge Case Hunter:** "Your job is to find where each claim BREAKS. Scaling limits, security concerns, accessibility gaps, failure modes under load."
-
-Both receive all claims and return challenge_findings[]
-
----
-
-## Phase 3: Pairwise Rank (for contested claims)
-
-Input: Claims marked as "contested" by council
-
-For each pair of conflicting claims:
-- LLM-as-judge prompt: "Given research question Q, Claim A says [X] with evidence [E1]. Claim B says [Y] with evidence [E2]. Which claim is better supported? Why? Consider: source authority, evidence specificity, logical coherence, relevance to the question."
-- Tournament bracket: winners advance, losers are demoted to "minority view"
-
-Output: ranked_claims[] with rank_position, judge_rationale
-
-### Ranking Criteria (in priority order)
-
-1. **Source authority**: Primary sources and peer-reviewed research outweigh blog posts and forum answers
-2. **Evidence specificity**: Concrete data points (benchmarks, case studies with numbers) outweigh general assertions
-3. **Logical coherence**: Claims with clear causal reasoning outweigh correlational arguments
-4. **Relevance to question**: Claims that directly address the research question outweigh tangentially related findings
-5. **Recency**: For fast-moving topics, newer evidence outweighs older evidence (all else equal)
-
----
-
-## Phase 4: Reconcile — Disagreement Map
-
-Input: All processed claims (normalized, council-reviewed, ranked)
-
-The Reconciler agent produces structured output in 4 tiers:
-
-### Tier 1: Consensus Claims
-3+ agents agree, council agreed, multiple source types.
-- Highest confidence. Present as established findings.
-- Include: canonical claim, supporting agents, source count, source diversity, council verdict, confidence score
-
-### Tier 2: Contested Claims
-Council disagreed, or pairwise ranking was close.
-- Present BOTH sides with the specific disagreement.
-- Include: claim_a, claim_b, agents_for_a, agents_for_b, council positions, rank winner, judge rationale
-
-### Tier 3: Unique Insights
-Found by 1 agent only, not contradicted.
-- High value OR hallucination — flag for user judgment.
-- Include: claim, agent_role, confidence, source, note flagging single-source status
-
-### Tier 4: Refuted Claims
-Council rejected, or pairwise loser with low evidence.
-- Still present briefly — knowing what's wrong is valuable.
-- Include: claim, rejection_reason, original_agent
-
----
-
-## Phase 5: Render
-
-Produce both:
-- **Structured JSON artifact** (see output-format.md for schema)
-- **Rendered markdown** for user display (see output-format.md for template)
-
-The JSON artifact is the primary output for skill-to-skill consumption. The markdown is for human reading.
-
----
-
-## Reconciler Agent Prompt
-
-```
-You are the Reconciler — the final judge in a multi-agent research pipeline.
-You receive findings from 7 research agents that have been normalized,
-deduplicated, and adversarially reviewed.
-
-Your job is NOT to average or blend. Your job is to JUDGE:
-- Which claims are strong? (multiple independent sources, council agreement)
-- Which claims are contested? (present both sides, don't pick a winner)
-- Which claims are unique insights? (valuable if true, flag for verification)
-- Which claims should be rejected? (weak evidence, circular sourcing, council rejection)
-
-Produce a structured disagreement map, not a smooth summary.
-The user should see WHERE agents agreed, WHERE they disagreed, and WHY.
-
-INPUT:
-- normalized_claims: [list of deduplicated claims with agent_count and source_diversity]
-- council_verdicts: [list of claims with agreed/contested/insufficient verdicts]
-- pairwise_rankings: [list of contested claim pairs with winners and rationale]
-- credibility_scores: [list of claims with scored credibility from score_credibility.py]
-
-OUTPUT FORMAT:
-Return a JSON object with these exact keys:
-{
-  "consensus": [{ claim, supporting_agents, source_count, source_diversity, council_verdict, confidence }],
-  "contested": [{ claim_a, claim_b, agents_for_a, agents_for_b, council_verdict, provider_positions, rank_winner, judge_rationale }],
-  "unique_insights": [{ claim, agent_role, confidence, note }],
-  "refuted": [{ claim, rejection_reason, original_agent }]
-}
-
-RULES:
-- A claim needs 3+ agents AND council agreement to be consensus
-- A claim with 2 agents but council agreement goes to consensus with a "moderate confidence" flag
-- A claim with council disagreement ALWAYS goes to contested, even if 5 agents agree
-- A single-agent claim with confidence > 0.8 goes to unique_insights
-- A single-agent claim with confidence <= 0.5 goes to refuted
-- Everything else goes to unique_insights with appropriate flagging
-- NEVER merge contested claims into a smooth middle ground — preserve the disagreement
-```
-
----
-
-## Pipeline Skip Rules
-
-- **Quick mode**: Skip Phases 2, 3, 4. Orchestrator does a single-pass synthesis directly from normalized findings.
-- **Standard mode**: Skip Phase 2 (council). Run Phases 1, 3, 4, 5.
-- **Deep mode**: Run all 5 phases.
+# Synthesis Pipeline
+
+5-phase pipeline that takes raw findings from finder agents and produces a structured disagreement map.
+
+---
+
+## Phase 1: Normalize & Deduplicate
+
+Input: Raw findings from all finder agents (7 agents x 3-8 findings each = 21-56 findings)
+
+Steps:
+1. Flatten all findings into a single list
+2. Group by semantic similarity (same claim from different agents)
+3. For each group:
+   - Merge into a single canonical claim
+   - Track which agents found it (agent_count)
+   - Track source type diversity (source_diversity_score = unique source types / total sources)
+   - Flag circular sourcing: if all sources in a group cite the same original source, mark as circular=true
+4. Output: unique_claims[] sorted by agent_count DESC, source_diversity_score DESC
+
+### Semantic Similarity Heuristics
+
+Two claims are considered semantically similar when:
+- They make the same factual assertion about the same subject, even with different wording
+- One is a subset of the other (e.g., "X uses Y" vs "X uses Y for Z")
+- They cite the same source for the same conclusion
+
+Two claims are NOT similar when:
+- They address different aspects of the same topic
+- They reach different conclusions about the same subject
+- One is general and the other is specific with additional qualifying conditions
+
+When merging, keep the most specific version as the canonical claim.
+
+---
+
+## Phase 2: Adversarial Review (ftm-council)
+
+Input: Top claims from Phase 1 (all claims with agent_count >= 2, plus any high-confidence unique claims with confidence > 0.8)
+
+Council invocation:
+- Send claims as a structured prompt to ftm-council
+- Ask: "Evaluate each claim. For each: Is the evidence sufficient? What would make this wrong? Are there alternative explanations? Rate confidence 0-1."
+- Council runs Claude + Codex + Gemini independently, then reconciles
+
+Output: claims[] with council_verdict (agreed | contested | insufficient_evidence), provider_disagreements[]
+
+### FALLBACK (if Codex/Gemini unavailable):
+
+Spawn 2 standalone agents on the review model:
+
+**Devil's Advocate:** "Your job is to find reasons each claim is WRONG. Search for counter-evidence, flag single-source claims, identify logical gaps."
+
+**Edge Case Hunter:** "Your job is to find where each claim BREAKS. Scaling limits, security concerns, accessibility gaps, failure modes under load."
+
+Both receive all claims and return challenge_findings[]
+
+---
+
+## Phase 3: Pairwise Rank (for contested claims)
+
+Input: Claims marked as "contested" by council
+
+For each pair of conflicting claims:
+- LLM-as-judge prompt: "Given research question Q, Claim A says [X] with evidence [E1]. Claim B says [Y] with evidence [E2]. Which claim is better supported? Why? Consider: source authority, evidence specificity, logical coherence, relevance to the question."
+- Tournament bracket: winners advance, losers are demoted to "minority view"
+
+Output: ranked_claims[] with rank_position, judge_rationale
+
+### Ranking Criteria (in priority order)
+
+1. **Source authority**: Primary sources and peer-reviewed research outweigh blog posts and forum answers
+2. **Evidence specificity**: Concrete data points (benchmarks, case studies with numbers) outweigh general assertions
+3. **Logical coherence**: Claims with clear causal reasoning outweigh correlational arguments
+4. **Relevance to question**: Claims that directly address the research question outweigh tangentially related findings
+5. **Recency**: For fast-moving topics, newer evidence outweighs older evidence (all else equal)
+
+---
+
+## Phase 4: Reconcile — Disagreement Map
+
+Input: All processed claims (normalized, council-reviewed, ranked)
+
+The Reconciler agent produces structured output in 4 tiers:
+
+### Tier 1: Consensus Claims
+3+ agents agree, council agreed, multiple source types.
+- Highest confidence. Present as established findings.
+- Include: canonical claim, supporting agents, source count, source diversity, council verdict, confidence score
+
+### Tier 2: Contested Claims
+Council disagreed, or pairwise ranking was close.
+- Present BOTH sides with the specific disagreement.
+- Include: claim_a, claim_b, agents_for_a, agents_for_b, council positions, rank winner, judge rationale
+
+### Tier 3: Unique Insights
+Found by 1 agent only, not contradicted.
+- High value OR hallucination — flag for user judgment.
+- Include: claim, agent_role, confidence, source, note flagging single-source status
+
+### Tier 4: Refuted Claims
+Council rejected, or pairwise loser with low evidence.
+- Still present briefly — knowing what's wrong is valuable.
+- Include: claim, rejection_reason, original_agent
+
+---
+
+## Phase 5: Render
+
+Produce both:
+- **Structured JSON artifact** (see output-format.md for schema)
+- **Rendered markdown** for user display (see output-format.md for template)
+
+The JSON artifact is the primary output for skill-to-skill consumption. The markdown is for human reading.
+
+---
+
+## Reconciler Agent Prompt
+
+```
+You are the Reconciler — the final judge in a multi-agent research pipeline.
+You receive findings from 7 research agents that have been normalized,
+deduplicated, and adversarially reviewed.
+
+Your job is NOT to average or blend. Your job is to JUDGE:
+- Which claims are strong? (multiple independent sources, council agreement)
+- Which claims are contested? (present both sides, don't pick a winner)
+- Which claims are unique insights? (valuable if true, flag for verification)
+- Which claims should be rejected? (weak evidence, circular sourcing, council rejection)
+
+Produce a structured disagreement map, not a smooth summary.
+The user should see WHERE agents agreed, WHERE they disagreed, and WHY.
+
+INPUT:
+- normalized_claims: [list of deduplicated claims with agent_count and source_diversity]
+- council_verdicts: [list of claims with agreed/contested/insufficient verdicts]
+- pairwise_rankings: [list of contested claim pairs with winners and rationale]
+- credibility_scores: [list of claims with scored credibility from score_credibility.py]
+
+OUTPUT FORMAT:
+Return a JSON object with these exact keys:
+{
+  "consensus": [{ claim, supporting_agents, source_count, source_diversity, council_verdict, confidence }],
+  "contested": [{ claim_a, claim_b, agents_for_a, agents_for_b, council_verdict, provider_positions, rank_winner, judge_rationale }],
+  "unique_insights": [{ claim, agent_role, confidence, note }],
+  "refuted": [{ claim, rejection_reason, original_agent }]
+}
+
+RULES:
+- A claim needs 3+ agents AND council agreement to be consensus
+- A claim with 2 agents but council agreement goes to consensus with a "moderate confidence" flag
+- A claim with council disagreement ALWAYS goes to contested, even if 5 agents agree
+- A single-agent claim with confidence > 0.8 goes to unique_insights
+- A single-agent claim with confidence <= 0.5 goes to refuted
+- Everything else goes to unique_insights with appropriate flagging
+- NEVER merge contested claims into a smooth middle ground — preserve the disagreement
+```
+
+---
+
+## Pipeline Skip Rules
+
+- **Quick mode**: Skip Phases 2, 3, 4. Orchestrator does a single-pass synthesis directly from normalized findings.
+- **Standard mode**: Skip Phase 2 (council). Run Phases 1, 3, 4, 5.
+- **Deep mode**: Run all 5 phases.