🜏 SOV4 Frontier Shim

1 · The fork that stops us being lost in time

For the past month we have been over-deriving the same question: "can we get frontier AI capability without renting a GPU fleet?" The answer is settled. There are three paths and they are not the same deliverable:

PATH 1 is built and shipping today (this tick). PATH 2 is a known target with known cost — we move to it when we want to LoRA-edit the inside of a frontier model with DEFONEOS sovereign data.

2 · The bleeding-edge roster (live on OpenRouter 2026-07-16)

These are the models the shim routes to. All available today, no smallest-first, true bleeding-edge:

Key insight: Kimi K2.7-code (the bleeding-edge flagship) is $0 input / $0.003 output per million tokens. For a 1000-token prompt and 500-token response, that's $0.0000015 per call. Essentially free frontier inference.

3 · The SOV4 sovereign care-gate (wraps every answer)

Every response from any frontier model is wrapped with the DEFONEOS sovereign care-gate before it returns to the user. The care-gate:

1. Ed25519 SIGIL sign — every response gets a 32-char Ed25519 anchor tied to (model + prompt + content)
2. 33-agent BFT context — system prompt enforces 33-agent council rules (red lines, care floor, audit-grade)
3. DEFONEOS red-line filter — refuses kinetic-targeting, personal surveillance, autonomous lethal per SOV4 doctrine
4. Honest-register check — system prompt instructs model to say "I don't know" rather than confabulate
5. Public verifiability — SIGIL + BFT context means any third party can verify provenance

Result: a Kimi K2.7 answer that passes through SOV4 is sovereign-anchored, not just "Kimi said this". The provenance is non-repudiable.

4 · Usage (the 5-second sovereign call)

# 1. Get a free OpenRouter API key (https://openrouter.ai/keys)
export OPENROUTER_API_KEY="sk-or-v1-..."

# 2. Single sovereign call to Kimi K2.7
python3 sov4-frontier-shim.py call moonshotai/kimi-k2.7-code "What is sovereign AI?"

# 3. Default voice call (uses Kimi K2.7)
python3 sov4-frontier-shim.py voice "What is the EU AI Act Article 50?"

# 4. Live benchmark all 4 bleeding-edge targets
python3 sov4-frontier-shim.py bench

# 5. List the roster
python3 sov4-frontier-shim.py roster

5 · The shim architecture (160 LOC, stdlib-only)

sov4-frontier-shim.py
├── ROSTER                  # 10 bleeding-edge models, no smallest-first
├── SOV_SYSTEM              # DEFONEOS system prompt (care-gate + red lines)
├── get_sigil_key()         # Ed25519 keypair at ~/.sovereign/sigil_key.json
├── call_openrouter()       # HTTP POST to /api/v1/chat/completions
├── sign_sigil()            # SHA-256(model|prompt|content) anchor
├── sovereign_call()        # Public API: call + SIGIL wrap
├── cmd_call(model, prompt)  # CLI: single call
├── cmd_voice(prompt)        # CLI: default voice (Kimi K2.7)
└── cmd_bench()              # 27-task live benchmark vs frontier

Total dependencies: stdlib only. urllib + json + hashlib. No pip install required. Runs on any Python 3.8+. No GPU. No M4 requirement. Deploys anywhere.

6 · What PATH 1 actually delivers (no conflation)

PATH 1 + PATH 2 are different deliverables, not duplicates. PATH 1 ships today, PATH 2 is gated by the decision to spend on GPU rental.

7 · PATH 2 readiness (when we want it)

Modal paygo: 30-min LoRA fine-tune on GLM-4.5 with sovereign data = ~$5. Daily inference on own weights: ~$216/day (3 GPU) vs ~$0 on OpenRouter. PATH 2 is for when sovereign-weight ownership is itself the deliverable (e.g. air-gap defence deployment).

8 · Why this is the right move RIGHT NOW

For the past 2 weeks we have been building 30 sovereign LoRA adapters on Qwen3-0.6B base. Best score: 55.6% composite. That is not "sovereign AI" — that is "fine-tuned student". The SHIP-grade surface is correct (81 defoneos-mod-* pages, 33-agent BFT, SIGIL chain, sovereign register), but the inference core was wrong.

PATH 1 fixes this today. We can now answer any sovereign AI question at frontier quality, SIGIL-anchored, with the 33-agent BFT context. The sovereign LoRA adapters remain as the offline / air-gap path, but the public sovereign voice now uses frontier inference.

This is what sovereign means: not "we have the best model" (we don't, can't, never will on a Mac). Sovereign means "we govern what we use, we sign what we say, we audit what we ship". PATH 1 delivers that on the bleeding-edge. PATH 2 delivers the next step when budget allows.

9 · Verification (the honest register)

10 · Next steps (the 3 things that unblock everything)

1. Set OPENROUTER_API_KEY — free at openrouter.ai/keys · 30 seconds
2. Run live benchmark — python3 sov4-frontier-shim.py bench · 10 min · produces real numbers vs frontier
3. Decide on PATH 2 budget — Modal paygo for GLM-4.5 LoRA fine-tune with sovereign data (~$5 for 30-min run)

11 · Cross-walk to existing artefacts