Σ₀ Local-Model Adapter

1. The decision: what "best local model" means here

The box is the constraint: 8 GB VRAM, one model process at a time. That hard-caps the candidate set. Grounded comparison of what actually fits (web, June 2026):

Resolution (Ouro stays default): Ouro-1.4B remains the Σ₀-native default — its recurrent depth + Q-exit is what Σ₀ is. Qwen2.5-Coder-7B is registered as the opt-in capability lever: when you want raw coding accuracy and it fits the VRAM budget, flip one flag and it leads, wrapped by the convergence loop so it stays Σ₀-compliant.

Sources: Microcenter — local LLMs by memory tier · Tembo — best local LLM for coding 2026 · daily.dev — best local LLM models 2026 · qwen.ai — Qwen3-Coder-Next · Ouro paper arXiv 2510.25741.

2. The contract

apps/lantern-garage/lib/local-model-registry.js declares each local backend:

Source of truth: built-in defaults in the module, overlaid by data/models/local-registry.json (operator-editable, ~60 s TTL — no consumer restart needed beyond that).

Why selfConverges is the key field

The convergence loop must sit above the backend, not inside it:

Convergence Core (Reason / Act)

   │

   ├─ selectChain(taskType)            ← registry: best LOCAL model that fits VRAM

   │

   ├─ if  selfConverges(lead)  → call model directly      (Ouro: Q-exit inside)

   │  else                     → loopedReason(model)       (Qwen: cdfExit wraps it)

   │

   └─ Verify gate (existing grounding / requireVerified)

Before this adapter, the LOOP_REASONER path was blind to the model: it would double-loop a self-converging Ouro and skip the grounding loop for a non-self-converging model. The registry makes that decision a capability lookup. Unknown models default to selfConverges=false → wrapped (grounding by default).

3. Selection rules

selectChain(taskType, opts) returns the VRAM-fitting models eligible to lead, best-first:

• Default order = rank ascending → Ouro leads every task it's eligible for.
• Capability-first (LOCAL_CAPABILITY_FIRST=1 or opts.capabilityFirst) =

capabilityScore descending → the strongest model that still fits the box leads. (A model that doesn't fit VRAM_BUDGET_GB can never lead — the gate is applied first, so capability-first on aGB box still yields Ouro.)

• Kernel is strict: "default"-tagged general-fallback models are not

widened into the kernel task. The Σ₀ Convergence Core path only uses models that explicitly opt into kernel.

Config flags

The authoritative lead: resolveLocalLead(intent, opts)

selectChain answers "what's the capability order?" — but keystone chat needs ONE authoritative lead that folds in the operator's OLLAMA_MODEL pin and the caller's served-fallback chain. resolveLocalLead(intent, { pin, fallback }) is that single resolver:

1. An OLLAMA_MODEL pin the registry does not manage (a custom model) → it leads.
2. Otherwise the registry's capability-gated #1 leads; the pin stays a candidate.

This fixed a real bug: a stale OLLAMA_MODEL=ouro:latest (the legacy default) used to front-jump the chain and silently defeat the capability swap — keystone chat stayed pinned to Ouro while the adapter had already picked the better model. The lead is now re-asserted as the chain head after the leaderboard reorder, so neither a stale pin nor measured win-rate can displace the deterministic, VRAM-gated swap pick. LOCAL_CAPABILITY_FIRST=0 still restores Ouro-first (rank-order) cleanly. The chosen local model + reason is surfaced in the chat reply signature (a ⇄ <model> chip).

4. Serving the backends (one process at a time)

TheGB box serves one local model on :11434. The two backends are mutually exclusive at runtime; the registry remains the source of truth for capabilities regardless of which is currently up.

Ouro (default, Σ₀-native):

OURO_NATIVE=1 OURO_ADAPTER=<dir> python scripts/ouro_serve.py   # speaks Ollama on :11434

Qwen (capability lever):

ollama pull qwen2.5-coder:7b          # real Ollama on :11434

LOCAL_CAPABILITY_FIRST=1 LOOP_REASONER=1 npm start --prefix apps/lantern-garage

LOOP_REASONER=1 is what actually wraps Qwen in the convergence loop (Ouro ignores it — it Q-exits internally). When VRAM grows, registering the next model (e.g. Qwen3-Coder-Next) is a one-line data/models/local-registry.json add.

5. Where it plugs in

All wiring is additive and falls back to prior behavior on any error.

6. Σ₀ alignment

• Reason: better local-model selection (fit-aware, capability-aware) — one

selection layer, not a parallel router.

• Verify: the selfConverges contract keeps the convergence loop honest —

every non-self-converging model is grounded by loopedReason(); the self-converging one isn't redundantly looped.

• No sprawl: no new memory system, no new agent ecosystem, no second serving

path. One contract, feeding the existing loop + the existing transport.