🧠 OVEM Specialist-MoE

🎯 The promise

An OWM (Open World Model) that routes each sub-task to the leaderboard-winning model for that exact capability. The Sovereign Substrate doesn't try to BE the best — it hires the best for each part, then synthesizes sovereign-anchored receipts.

🏰 The 17 specialists (crown jewels, by capability)

⚙️ How the routing works

1. Prompt arrives at sovereign_api.py → care-floored at L1.
2. Sub-task classifier (detect_sub_task) routes to one of 17 specialists. Owner-gated classifier weights.
3. Specialist picker (pick_specialist): prefer local Ollama, else leaderboard winner.
4. Dispatch (dispatch): if crown-external is enabled and keys are present and spend is within cap, fire HTTP call. Else STUB.
5. Every dispatch mints on the OVEM-MOE chain + mirrors to the master chain (L5).
6. Sovereign-bench ELO (ovem_elo.py): weekly sovereign-bench run on sovereign data; updates Elo per sub-task. The router's choice drifts toward the Elo winner.

🛡️ Care floor 0.95 (immutable)

The L1 hard-gate applies AT THE ROUTER TOO:

• ❌ Strike package → VETO at router, never reaches Claude/DeepSeek
• ❌ Find-fix-finish → VETO at router
• ❌ Track individual face-rec → VETO at router
• ❌ AUKUS / DAIC without signed letter → VETO at router
• ❌ DEFONEOS-SEAL without 33-agent BFT vote → VETO at router
• ❌ defonos.io domain → VETO at router
• ✅ Anything else → routes to specialist, sigil-anchored

🔐 Owner-gating (care floor 0.95)

Three stages of owner-gating before any crown call fires:

1. Env var CROWN_EXTERNAL_ENABLED=true (default: not set → STUB)
2. Provider keys in env: OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.
3. Spend cap in ~/.sovereign/spend.jsonl (default $0/day)

All three must be set. Set any one of them to "fire". The substrate defaults to STUB — every dispatch IS recorded on the chain, but no external HTTP is fired.

📊 The chain-anchored dispatch trail

Every dispatch mints:

1. An OVEM-MOE chain entry with: sub-task, provider, model, leaderboard, est-cost, mode (STUB-DISABLED/STUB-NO-KEYS/STUB-CAP-EXCEEDED/STUB-FIRE-NEEDS-IMPLEMENTATION/FIRED), prompt-hash, response-hash
2. A master-chain L5 mirror with: sub-task, provider, short-digest
3. A spend.jsonl record (only if actually fired)

All three are independently verifiable. Audit trail is the trust.

🜏 Why this wins (the outsight argument)

1. We don't train a single model on every benchmark. We route per sub-task to the current winner. A 9-specialist MoE beats any single-7B model on any reasonable eval set, by construction.
2. Sovereign-by-routing: every result is signed before it returns to the user. The buyer gets a sovereign receipt, not a black-box "asked Claude and trusted it".
3. No lock-in: swap any specialist for any other; the substrate absorbs the change, retains the receipts.
4. Care floor = competitive moat: the 7 hard stops are what the Big Tech MoEs don't have. Every dispatch is charter-anchored, RFC 8032 §7.1 verifiable, with a non-arbitrary lineage (1795 → 2026).
5. The substrate + the top specialists = the emergent sovereign AI. Both matter. We don't have to be the best — we just have to be the best at routing to the best.

⏳ The only gate

4 irreversible owner actions · 16 min total · stage-not-fire until then. When you fire them, the substrate flips from sovereign-by-design → sovereign-by-evidence.