Run each phase in order. Phases 1–2 take seconds. Phase 4 extracts layered semantic knowledge. Phase 4b refines confidence. Phase 5 clusters tag vocabulary. Phase 6 reviews clusters. Phase 7 applies canonical labels.
Decrypt synapse.dat, extract sheet 2, and store all matrix items in the database.
Read dll.config and map all matrix IDs to their source groups (Cardio, Allersode, Spinal, etc.).
Extract layered semantic knowledge from each item (body/mind/spiritual/environmental/remedy) and store as typed graph edges.
Model:
Sample:
Resubmit low-confidence tags to stronger models for CONFIRM / RAISE / LOWER / REASSIGN / REMOVE decisions. Full audit trail stored in tag_refinement_history.
Pass 1:
Pass 2:
Pass 3:
Confidence threshold:
Max passes:
Sample:
Community filter:
Embed unique LLM tag values with text-embedding-3-small, then cluster with HDBSCAN per layer to discover canonical vocabulary.
Min cluster size:
Send each HDBSCAN cluster to the LLM to generate a precise canonical label, 1-sentence description, coherence score, and keep/rename/split/reject suggestion. Run this before Phase 6 so the review table shows meaningful names.
Model:
Embedding pre-filter skips clearly distinct pairs before touching the LLM. Only pairs with cosine similarity ≥ embed threshold AND LLM confidence ≥ merge threshold are merged. Resume-safe — already-compared pairs are automatically skipped.
Model:
Merge threshold:
Embedding pre-filter:
0.78
cosine-similarity — pairs below this skip the LLM entirely
Embed Phase 5b cluster descriptions with text-embedding-3-small, then HDBSCAN per layer to build a second-level vocabulary hierarchy — grouping related clusters (e.g. leaky_gut + gut_permeability) into shared vocab themes.
Min group size:
Approve, rename, or reject each discovered cluster. Approved canonical labels will be applied in Phase 7.
No pending clusters — run Phase 5 first, or all clusters already reviewed.
Push all approved canonical labels back to matrix_item_tag and mark is_normalized = TRUE. A DB backup is taken first; rollback resets only this run's rows.
No approved clusters yet.
Run Leiden graph algorithm on the item co-occurrence network. Items sharing similar tag profiles cluster into thematic communities (liver · hepatic, grief · trauma, digestive · gut, etc.).
Resolution:
Min shared tags:
Stopword cutoff:
max tag freq
Send each Leiden community's items + their enriched tags to the LLM. Generates a clinical name, tagline, main themes, and practitioner insights for every community. Makes the knowledge graph queryable by Corvus.
Model:
Mine directed semantic edges between items within the same community: CAUSES · TREATS · SUPPORTS · DISRUPTS · INDICATES · LEADS_TO · etc. Transforms the knowledge graph from named clusters into a traversable network Corvus can walk to answer 'what causes X?' or 'what treats Y?'
Model:
Min confidence:
Two-step: (1) association rule mining on the Phase 9b graph — finds statistically significant multi-item co-occurrence patterns ({heavy_metal} → DISRUPTS → {vitamin, mineral}) at zero API cost. (2) LLM names each pattern and explains the clinical mechanism (~$0.10 on Flash).
Model:
Mine directed semantic edges between communities (not just within them). Gives Corvus a chapter-level navigation map: Heavy Metal cluster → DISRUPTS → Methylation cluster, Gut Dysbiosis cluster → CAUSES → Immune Hypersensitivity cluster, etc. ~78 community pairs, ~$0.05 on Flash.
Model:
Ask LLM to describe each relationship in free-form mechanism language, then cluster those descriptions with HDBSCAN. The data teaches you its own relationship vocabulary — archetypes like 'biochemical_conversion', 'drainage_cascade', or 'energetic_resonance_bridge' emerge from the graph itself, not from your predefined list.
Model:
Phase 9e's emergent archetypes are frequently rediscovered under slightly different names — 'homeopathic_cardiac_frequency', '..._synergy', '..._resonance' are the same mechanism. This bucket-compares archetypes by layer_affinity, pre-filters with embedding cosine similarity, then asks the LLM to confirm true mechanism equivalents vs. genuinely distinct mechanisms that share domain vocabulary. Merges remap item_relation.pattern_id and are fully reversible per run_id.
Model:
Merge threshold:
Embedding pre-filter:
0.8
cosine-similarity — pairs below this skip the LLM entirely
Generate deep, multi-channel knowledge records for every enriched matrix item using DeepSeek v4 Pro. Pass 1 produces mechanism-level knowledge per category (enzymes, cofactors, depletion chains, crystal/homeopathic/flower-essence fields). Pass 2 adds cross-domain connections, convergence patterns, and quality scores for Corvus pattern-recognition and evidence-convergence reasoning.
Model:
Core knowledge: mechanisms, interactions, syndromes, category-specific fields
Deep connections: cross-domain bridges, convergence patterns, quality scores
Targeted remediation: re-enrich items below quality threshold using their own gap list from Pass 2
Threshold:
— items below this completeness or coherence score are re-enriched
Two-step: (1) deterministic vertebra pass — regex-extracts spine level (C1-C7/T1-T12/L1-L5/sacrum) from matrix items already literally named that way, zero API cost. (2) LLM classifies every remaining item's organ/body_region/chakra against a closed vocabulary (reuses the existing ORGANS/BODY_REGIONS taxonomy, adds new CHAKRAS/VERTEBRAE). Prepares matrix_item_tag for the future interactive body-map visualization — see docs/BODY_MAP_DATA_PREP.md.
Model:
Cross-references Phase 12's organ tags against Phase 4's existing body_system classification. Free deterministic pre-filter finds mismatches, then two INDEPENDENT LLM stages (Stage 1: deepseek-v4-pro with max reasoning effort; Stage 2: deepseek-v4-pro standard, not shown Stage 1's verdict) each judge confirm/reassign/remove. Only applied when both stages agree — disagreement is flagged for manual review, never guessed. See docs/BODY_MAP_DATA_PREP.md.
Model:
Run a phase to preview data
No data yet.
Each run is isolated by run_id. Rollback deletes only that run's LLM tags and resets those items so they can be re-enriched. Source-confirmed tags (dll_config) and manual overrides are never touched.
No runs yet. Click Refresh after enrichment.