SECTION 1 — THE FULL SPEC
{
"intent_id": "γ₁-anchored UUID", // deterministic from γ₁ + epoch
"actor": "kay | agent_id | system", // who initiated
"purpose": "what is being attempted", // human-readable goal
"domain": "health | legal | security | math | ops | family",
"sovereignty_class": "local | allied | external | sensitive",
"timestamp": "1746270000000", // epoch ms
"bfe_layer": "L1 | L2 | L3 | L4 | L5", // debt layer classification
"cloak_tier": "acquaintance | friend | family", // relationship tier
"debt_signal": "nominal | watch | break", // silo WPA state
"model_routed_to": "resolved at routing time" // filled by CLO CLOAK
}
SECTION 2 — FLOW THROUGH THE STACK
LAAM
ingest tag
→
qdrant
vector metadata
→
Redis
session tag
→
graph edge
provenance
→
CLO CLOAK
routing decision
→
audit trail
CLO review
The intent object is created at LAAM ingest and threads all the way to the audit trail. Every hop stamps model_routed_to and debt_signal at runtime.
SECTION 3 — THE 6 DOMAIN CLASSES
DOMAIN: health
Health & Medical
Patient data, clinical patterns, DCJ numbers, health fleet operations. FORCE LOCAL when patient data involved — no exceptions. PIPEDA + provincial health law applies.
DOMAIN: legal
Legal & Compliance
Amani's work, CLO decisions, ARB filings, litigation strategy, sovereign IP claims. High sensitivity. Local strongly preferred. Never to DeepSeek.
DOMAIN: security
Security & Architecture
Fleet architecture, moat strategy, sovereign IP, infrastructure details. Treat as L4/L5 minimum. Anthropic scan risk applies — switch to local Qwen for deep discussions.
DOMAIN: math
Mathematics & Proofs
γ₁ geometry, Lean4 proofs, zeta anchors, PEMCLAU graph math. Medium sensitivity. External OK for public math; local preferred for proprietary derivations.
DOMAIN: ops
Operations & Infrastructure
Builds, deploys, file ops, AKS management, routine fleet tasks. Generally low sensitivity. Good candidate for local Qwen migration in Phase 1.
DOMAIN: family
Personal & Family
Kay, Amani, personal context, family matters. Never to DeepSeek. Local strongly preferred. Private data stays private — no external routing without explicit intent.
SECTION 4 — THE 4 SOVEREIGNTY CLASSES
🟢 LOCAL
mal/, ollama-local/, lounge/ — never leaves msi01 — $0 — zero exposure — preferred class for all new work
🟡 ALLIED
cohere (Canada/PIPEDA), mistral (France/EU) — data stays in allied jurisdiction — acceptable for non-sensitive external routing
🟠 EXTERNAL
anthropic, openai, gemini, xai, perplexity — US servers — data leaves Canada — use with awareness — avoid for health/legal/security domains
🔴 SENSITIVE
deepseek — Chinese servers — RESTRICTED — never use for: DCJ numbers, fleet IP, moat strategy, patient data, Amani's work, architecture details — public knowledge only
SECTION 5 — BFE LAYERS → CLOAK RULES
L1 — ATOMIC
single-node operations, basic queries — external OK — lowest exposure — no routing restriction
L2 — SEQUENTIAL
multi-step sequences, pipelines — external OK — watch for data accumulation across steps
L3 — CONCURRENT
parallel ops across silos — prefer local — concurrency data is sensitive — fleet topology visible at this layer
L4 — COMPOSITIONAL
composition patterns = IP — local strongly preferred — graph edges at this layer reveal moat architecture
L5 — SOVEREIGN
LOCAL ONLY — no exceptions — liability data — clinical patterns — sovereign IP — never leaves msi01
SECTION 6 — WHY IT MATTERS
"Without the intent object, the graph knows WHAT happened but not WHY — and WHY is what legal needs."
Before V12: every graph edge recorded the operation. LAAM logged the call. qdrant stored the vector.
But: no provenance. No actor. No domain. No routing record.
Legal question: "Was this patient data sent to an external model?" — pre-V12 answer: unknown.
V12 answer: query the graph. Filter by domain=health AND sovereignty_class=external.
The intent object is the provenance layer. It makes the graph legally queryable.
Every operation is now: WHAT + WHO + WHY + WHERE IT WENT.