Findings

We built TQ2Q — a 4-level ("quaternary") 2-bit weight format — for CPU inference. The same files run on an NVIDIA GPU with no kernel changes: perplexity is bit-identical to the CPU, and a 35B mixture-of-experts model decodes at 59 tokens/sec on a single A10G — 4.1× its CPU speed on the same box.

Pre-computed codebase graph context hit 100% scope accuracy at 465 tokens. Model-driven MCP querying hit 46–54% at 29K tokens. The standard "give it tools and let it explore" pattern is 60× the cost for worse accuracy — because the model can read a complete picture but can't assemble one through iterative queries.

21 runs at max reasoning caught zero conceptual flaws. We changed the question — from "find a defect" to "enumerate assumptions, stress each" — and the same models at default effort started catching them. The role is the lever, not the model or the compute.

Every prompt guide says "assign a role." We A/B tested it. Personified prompts — "you are a reviewer" — underperform plain task descriptions on both recall (53% vs 73%) and cost (548 vs 362 tokens). The leaner the persona, the better the model.

The A8 planning mechanism fans out the consolidated plan to multiple coder agents. The worry was cost: does each coder re-read the full plan? No — with a byte-stable prefix on the same model, every coder lane hit cache at 97%. Cross-model cache is 0% (confirmed), but within-model the fork is cache-cheap.

A 6-model, 144-run probe of the safety-net assumption (a stronger model auditing a weaker pass). Recall saturates across the cohort at ~80% — but models discriminate sharply on precision. GLM-5.2 and MiniMax-M3 (the cheapest) had zero false alarms in every role, while grok (5–13× the cost) hallucinated defects on clean code.