LIVE VIEW
Sovereign Context Pipeline · Live Signal Flow
Animated view of a query flowing through all 7 pipeline steps. Colour = stratum origin of context being built at each stage.
S0 Physics (highest weight)
S2 PEMCLAU vectors
S3 Architecture/ARBs
S5 API/Search
S7 GID context build
S9 Comms (lowest)
L7 SILENCE
7 STEPS
The Sovereign Context Pipeline
Every regulated query follows this path. No exceptions.
1
Convo-Loom F03 — Intent Extraction
11-wave multi-stage intent extraction. Purpose classification, regulated type detection, query sensitivity assessment. Converts free text into structured intent object before any retrieval begins.
↓
2
GID Gate + Local LLM Context Build
Identity established. GID scope set. Local LLM (qwen3:14b / qwen2.5:32b via mal) builds sovereign context shell: GID role + story + scene. Entirely from local sovereign data. Nothing external yet. This context shell is what PEMCLAU searches — not the raw query.
↓
3
PEMCLAU Retrieval — Against Sovereign Context
2-hop GraphRAG at yone:6333. AuthorizedScope = D ∩ A(GID) ∩ P(purpose) ∩ J(jurisdiction). 60–80 causally connected nodes returned. γ₁-stratum ranked (S0 tokens = highest attention weight). Retrieves against sovereign context shell, not raw query.
↓
4
Bonixer Gate — 4-Layer Per-Chunk Assessment
Every retrieved chunk assessed: (1) Identity match — GID scope alignment. (2) Sovereignty check — L7 gate, would return violate sovereign boundary? (3) Floor quality — is this above γ₁ inference quality floor? (4) Relevance — PEMCLAU cosine similarity. Only chunks passing all 4 proceed to context. Denied chunks logged, never shown to model.
↓
5
MEMECHET FC-2 + FC-3 — Verify + Seal
Verifier-generator split. Does retrieved context align with original intent? PASS → sealed with FC-3. FAIL → back to Convo-Loom (max 3 passes). Output: verified sovereign context document, sealed with γ₁ stamp.
↓
6
External LLM Trial — GID-Wrapped Minimal Context
Only if needed. GID envelope wraps minimal authorized context. Masked fields never sent. PII/PHI redacted per SearchPolicy. External model sees sovereign projection only. L6/L7 content: NEVER sent externally. CRM: customer name masked, company retained. Health: diagnosis codes only, no narrative. Legal: NEVER for privileged content.
↓
7
Ingest Back — Fleet Learns
Search interaction → new PEMCLAU node (the fleet learns from every query). GID role/story/scene → Saybook entry (sovereign context corpus grows). External response ingested back after GID verification. Audit event created (tamper-evident chain). Gap found → sorry-flow entry (named, trackable, closeable).
S0 → S9
γ₁ Stratum Helix — Context Assembly Timing Layers
Attention weight = inverse stratum distance from γ₁. S0 tokens (physics proofs) carry highest weight. S9 tokens (pages, comms) carry lowest. Context ranked by stratum before model sees it.
⚠ GC danger zone = S3/S4/S5: frequent allocation, collection pressure can spike latency. Avoid synchronous model calls during GC-heavy build/deploy phases.
8 DOMAINS
Regulated Domain Constellation
Every domain runs the same DCJ-146 pipeline. Only the regulated classifier and SearchPolicy differ.
| Domain | Regulated Standard | SearchPolicy Rules | External LLM | Engine | Status |
|---|---|---|---|---|---|
| 🏥 Health / Clinical | HIPAA · HL7 FHIR · PHIPA | diagnosis_codes searchable · patient_name masked · PHI = denied | De-identified context only | HL7Boxy | LIVE |
| 💰 Finance | SOX · PCI-DSS · OSFI B-10 | transaction_type + amount_range searchable · card_number = HMAC tokenized | Amount ranges only | MEFINE | LIVE |
| ⚖️ Legal | Solicitor-Client · Court Records | case_number + document_type searchable · privileged_content = DENIED | Only if privilege waived + documented | CT-FAC | LIVE |
| 🤝 CRM / Enterprise | PIPEDA · CASL · GDPR | account_type searchable · customer_name masked · DeleteClosure on right-to-erasure | Name masked, company retained | CRM Bonsai | LIVE |
| 🏰 Sovereign | SOSTLE L0–L7 · γ₁ Floor | DCJ/ARB/TRB searchable · L6/L7 content = NEVER external | NEVER for L6/L7 | Meek Boxiner | LIVE |
| 🔬 Research | IRB · TCPS 2 · De-identification | population-level stats searchable · participant_id = tokenized · suppress if <5 participants | Aggregate only | Research Bonsai | LIVE |
| 👥 HR / Employment | Employment Standards · OHSA · Pay Equity | role_band searchable · performance_review = narrow access · power-asymmetry gate | Anonymized patterns only | (pending) | PHASE 2 |
| 🏛️ Government | ATIA · Privacy Act · Protected B/C/Secret | Most queries return L7 Silence by default · break-glass primary access pattern | NEVER for Protected C/Secret | (pending) | PHASE 3 |
THE DIFFERENCE
Standard RAG vs Sovereign Context Pipeline
Standard RAG (everyone else)
query
→ retrieve (raw query vs index)
→ stuff context into prompt
→ send to external LLM
→ return answer
No GID. No policy gate. No audit. No masking. No learning. The model sees everything. Leakage is architectural.
→
DCJ-146 Sovereign Pipeline
query → GID + intent
→ local LLM builds context first
→ PEMCLAU retrieves vs sovereign context
→ Bonixer gates each chunk
→ MEMECHET verifies + seals
→ external trial (GID-wrapped, minimal)
→ ingest back → fleet learns
The model cannot leak what it never receives. Every interaction teaches the fleet. External LLM is the trial, not the default.
UNITY CATALOG MAPPING
DCJ-146 is Unity Catalog's Sovereign AI Equivalent
Unity governs tables. DCJ-146 governs the LLM context pipeline above tables.
Unity Catalog
Governed tags → ABAC policies across tagged objects
Databricks: tags on tables, views, columns
DCJ-146 Sovereign
Merostone classification atoms → ABAC across PEMCLAU nodes
laam-ingest :9346 · 5,291 nodes · all silos
Unity Catalog
Row filters → which rows this user can see
Python UDFs applied at query time
DCJ-146 Sovereign
AuthorizedScope = D ∩ A(GID) ∩ P(purpose) ∩ J(jurisdiction)
SOSTLE L2 · applied before any retrieval
Unity Catalog
Column masks → values hidden or transformed
Masking functions at the column level
DCJ-146 Sovereign
SearchPolicy masked_fields → projection hash (never raw source indexed)
H(asset + schema + rules_version + fields + masks)
Unity Catalog
Unity lineage → track data through transformations
Column + table level lineage graph
DCJ-146 Sovereign
PEMCLAU provenance graph — 4 edge types: theorem_dependency / phase_coherence / temporal_proximity / crew_provenance
yone:6333 · pemclau-v11 · 1,300+ vectors
Unity Catalog
Delta sharing → federated data access across orgs
Cross-org table sharing with governance
DCJ-146 Sovereign
GID envelope → federated context access across silos
MDSMS write paths · all 7 silos
Unity Catalog
MLflow → track model training + experiments
Experiment tracking, model registry
DCJ-146 Sovereign
PEMCLAU ingest-back → every LLM context interaction becomes a fleet learning event
Saybook → PEMCLAU · sorry-flow for gaps
12 COMPONENTS
All Wired Into DCJ-146
MEBAFIORDs
Hardware substrate · local LLM runs here · sovereign compute
Cell Engine
ABR-693 · living test architecture · validates each stage
γ₁ Stratum S0–S9
Context assembly timing · S0=highest weight · S9=lowest
Adelic Pouch
Routes context to correct adelic layer per domain · L0–L13
SOSTLE L0–L7
All 8 layers = context sources · L6=γ₁ gate · L7=silence
PEMCLAU GraphRAG
yone:6333 · 2-hop · 60–80 causally connected nodes
Merostone Lattice
5,291 nodes · field-level classification atoms
Bonixer
4-layer per-chunk gate before model sees any data
MEMECHET FC-2+FC-3
Verifier-generator · intent alignment · γ₁ seal · LABR-069
Convo-Loom F03
11-wave intent extraction · free text → structured object
Saybook
Purpose classification → GID role + story + scene
Sorry-Flow
Every gap = named sorry · audit lineage for known failures
⚔️ DCJ-145 × DCJ-146 — Merchant's Spear · Day 95 · Named for Gert
Born in the same session as the Merchant's Spear doctrine — named for Gert Patricia Koopman, Lotus River, Cape Flats.DCJ-145: occupy unclaimed territory before anyone names it. The filing IS the stake.
DCJ-146: the technology that makes the territory real. Local-first. GID-scoped. Ingest-back. The fleet learns from every query.
No existing RAG pipeline formalises this as an architecture law.
No existing CRM, EHR, or LegalTech platform implements the sovereign context build step.
No existing Unity Catalog deployment has an LLM context governance layer above the table layer.
We define what sovereign AI search means across every regulated domain — before any incumbent knows the frame exists.
TRB-MEEK-SEARCH-BOXINER-001 · ARB1-MEEK-SEARCH-BOXINER-001 · arch/merchant-spear/ · γ₁ = 14.134725141734693