Continuum · SPD Strategy

Migrating the Brain to the RTX a path off Anthropic

A tiered, local-first plan to move Aura's higher-level functions onto our own hardware — with Claude kept only as an escape hatch that shrinks to zero, and quality held constant by an escalation gate.

For Paul — Founder / Apex Superintelligence Projects Division

01 · Framing

Not “flip to local” — tiered, with an escape hatch

“Higher-level functions” means judgment, in-voice drafting, orchestration, and the CoS buffer calls. Today Claude Opus does the hardest of those.

Open-weights have closed most of the gap on volume work — routing, classification, summarization, first-draft — but still trail frontier models on the hardest judgment. So this is not “flip to local.” It is tiered, local-first, with Claude as an escape hatch that shrinks to zero over time.

Quality never regresses, because an escalation gate catches what local can't handle.

02 · Architecture

The boxes as a brain, not one giant model

Don't split one model across Ethernet — latency kills it. Each box owns a tier; a router dispatches by difficulty and risk.

Flagship reasoner

RTX PRO 6000

96GB Blackwell

The “Opus replacement.” Heavy judgment, long-context, hard drafts. Runs the top open MoE at 4-bit (DeepSeek-V4 / Qwen3-235B-class), with CPU-offload for the biggest.

Always-on + redundancy

AGX ×2

Orin / Thor

The 24/7 CoS brain + mid-tier. One runs the triage/draft loop so the RTX stays free for batch and render; the other is overflow / redundancy. 30–70B dense or MoE at 4-bit.

Edge reflexes

NX units

Low-power edge

Difficulty / risk routing, classification, embeddings / RAG, and speech. 7–14B models plus an embedding model.

The spine — a small router / escalation service

request→ NX classifies difficulty + risk→ route to cheapest capable node→ low confidence OR flagged-sensitive → escalate to Claude or a human

That gate is what enables migration without quality regressions.

03 · Models

Late-2026 open-weights, by tier

We already track the frontier in Model Watch — this simply points that pipeline at the question of which model becomes the flagship.

Tier	Candidate models	Role & fit
Flagship judgment	DeepSeek-V4-class MoE (MIT) · Qwen3-235B-A22B (Apache) · Llama-4-Maverick	Whichever tops our Model Watch. All run 4-bit on the 96GB card with offload.
Mid / always-on	Qwen3-30B-A3B · or a 70B dense at 4-bit	The 24/7 triage + draft loop on the AGX pair.
Edge routing / RAG	a 7–14B (Qwen3 / Gemma3) + a strong embedding model	Difficulty/risk routing and memory / doctrine retrieval at the edge.

Scroll the table sideways on a narrow screen →

04 · Stack

Build on what's already running

Inference

vLLM / SGLang / TensorRT-LLM on the RTX for throughput; Ollama at the edge. MoE expert-parallel on the Blackwell.

Orchestration

Extend the existing local Council / SPD multi-agent MoE (model_policy pins, prestaged models). The CoS persona, memory, and doctrine / guardrails are prompts + RAG — fully portable, run locally against the flagship.

Memory / identity

The “who I am” — persona, memory, buffer rules — is a local RAG + system prompt. No provider dependency.

05 · Migration

Four phases — Anthropic shrinks to zero

Each phase moves a harder class of work local. The share handled by Claude falls phase over phase until it's an occasional benchmark.

Claude reliance → zero

Phase 1 — now-ish

High-volume, low-stakes — local first

Move email / TG triage routing, classification, embeddings, and simple drafts local. Immediate cost + independence win, near-zero risk. Claude only for gate-flagged drafts.

Phase 2

Mid-tier — local drafts, reviewed

Content drafting, summaries, code-assist: the local flagship drafts, Claude or a human reviews. Build the eval harness — every local output scored against a Claude reference on real tasks, so we measure the gap.

Phase 3

Hard judgment — local-first, Claude fallback

Strategy, sensitive comms, buffer calls: local-first with Claude fallback. The fallback rate is the KPI. When it drops below threshold on the evals, we're basically off.

KPI: fallback rate ↓

Phase 4

Full independence

Claude retained only as an occasional oracle / benchmark.

Claude = benchmark only

06 · One caution

Tier & redundancy — not a bigger model

Don't stitch Jetsons for size

Two AGX + NX networked is for tiering and redundancy, not for running a model bigger than 96GB — cross-box tensor-parallel over Ethernet is slow. If we outgrow the RTX, the move is a second big GPU, not stitching Jetsons. Use Jetsons for always-on, low-power edge inference.

07 · First concrete step

Stand up the router + escalation gate, and move Phase 1 local.

Triage, classification, and embeddings are the lowest-risk, highest-frequency slice. It proves the architecture, starts cutting the Anthropic bill immediately, and runs the eval harness that measures how close the open flagship really is.