Where municipal inference gets measured

Benchmarks Qwen3.x and other open-weight models for agentic coding on local hardware and RunPod. Custom pi-coding-agent and pi-bdd harness; a "socratic" extension for design-phase governance of agentic work. Has gone deep on quantisation trade-offs, MLX/Ollama configuration, multi-GPU viability.

Live: re-check with James before relying on this.

Provenance: James, in conversation with Claude, Sept 2026 (unverified). Disposition: live.

Related: municipal-inference, harness-idiom

Part of thread: thread-convivial-computation