NOTE

Synthesis Intelligence Layer

authorclaude-sonnet-4-6 aliases titleSynthesis Intelligence Layer statusactive date2026-04-30 typepermanent

Synthesis Intelligence Layer

The Synthesis Intelligence Layer is the "Mind" of the spec-agentic-source-orchestrator. It governs what ingested material is allowed to become permanent knowledge by enforcing epistemic gates at every stage of the synthesis pipeline.

The complementary "Body" (infrastructure, state machine, SQL schemas) is built by Codex; see codex-orchestrator-build-handoff-2026-04-29.

Scope Statement

This note covers the four intelligence components that the Chronicler (Claude) owns within the orchestrator: the epistemic risk classifier, conflict resolution templates, provenance block generator, and synthesis quality rubric. It does not cover the ingestion infrastructure, policy enforcement, or MCP server wiring.

Components

1. Epistemic Risk Classifier (epistemic_classifier.py)

Classifies every claim in a synthesis draft against the T0–T5 risk tiers defined in spec-agentic-source-orchestrator §3. The classifier evaluates each claim's supporting evidence chunks and returns the most restrictive applicable tier.

Tier Gate Logic
T0 — Fabrication Zero evidence chunks supplied
T1 — Weak Evidence No chunk meets the min_similarity_threshold
T2 — Unmarked Inference Inference language present but no [Derived] annotation
T3 — Stale Evidence All strong chunks older than freshness_threshold_days
T4 — Conflict Negation mismatch detected against existing wiki claims
T5 — Verified At least one fresh, high-similarity chunk supports the claim

MCP tool: classify_synthesis_draft — accepts a claims list (each with text and chunks), returns per-claim tiers plus an overall_tier for the draft.

2. Conflict Resolution Templates (conflict_templates.py)

Arbitration prompt templates for the three conflict types identified in §5 of the master spec:

Conflict Type Description
direct_contradiction A says X, incoming source says ¬X
version_skew Both true but for different versions or time periods
scope_overlap Incoming claim overlaps but does not contradict existing note

Each template produces a structured Conflict Report (JSON) with four resolution options: Update, Narrow, Parallelize, Escalate. Choosing Escalate forces auth_required: true, triggering an AUTH_REQUIRED pattern-human-in-the-loop gate.

MCP tool: get_conflict_resolution_template — accepts conflict_type and context variables, returns the filled arbitration prompt.

3. Provenance Block Generator (provenance.py)

Generates the YANP provenance frontmatter block required on every permanent note promoted through the orchestrator (§7). Links the Wiki note back to the specific Supabase chunk IDs and source record IDs that supported it.

provenance:
  source_record_ids: ["sr_abc"]
  chunk_ids: ["chk_1", "chk_2"]
  retrieved_at: "2026-04-30T12:00:00Z"
  acting_agent: "claude-chronicler"

MCP tool: build_provenance_block — accepts chunk and source record IDs; optional render_yaml: true flag returns the YAML text ready for frontmatter insertion.

4. Synthesis Quality Rubric (synthesis_rubric.py)

Checks a synthesis draft for YANP atomicity compliance before the promotion step:

  • Word count within bounds (50–800 words)
  • Section count ≤ 5
  • Scope Statement section present
  • No concept-boundary crossing phrases detected

Returns passed: bool, atomicity_score (0.0–1.0), and per-issue suggestions.

MCP tool: run_synthesis_rubric — accepts draft_text, returns the full rubric result.

Integration Point

All four tools are registered in 02_System/vulture-ingest/server.py alongside the existing ingestion tools (propose_source_intake, orchestrate_ingestion, execute_source_crawl). The synthesis tools consume policy thresholds from 02_System/pipeline-policy.yaml at runtime — changing min_similarity_threshold or freshness_threshold_days in the policy file propagates automatically to the classifier.

Next Seam

Once Codex completes the infrastructure seam (schema migrations, index_crawled_source, verify_source_index, promote_synthesis_candidate), the first End-to-End ingestion test can run: ingest a source → classify claims → check rubric → generate provenance → promote to 01_Wiki/.

References