Dot Reflex 14B: the agent supervisor just went open-weight

Most agent "supervision" is a prompt. Dot Reflex 14B is a model. It reads a compressed trajectory of what an agent just did and returns one of ten control decisions: continue, verify, retry differently, replan, rollback, branch, switch model, ask a human, stop — success or failure. The supervisor role, extracted, trained, and released open-weight.

The repo is fresh off GitHub trending — created yesterday, Apache-2.0, weights on Hugging Face. It's a rank-64 QLoRA adapter on Qwen3-14B-Base: 256.9M trainable parameters (1.71% of base), 375 optimizer steps on 6,000 synthetic trajectories, about 30 minutes on a single H200. Deliberately framework-neutral — it sits beside a GPT, Claude, Gemini, or Llama worker and supervises the execution loop, so the controller doesn't share the worker's architecture.

Bench numbers: 100% recovery accuracy on the repo's Agent Recovery Bench v0 (1,000 held-out synthetic trajectories) versus 90% for the same Qwen with a fixed recovery prompt and 80% for deterministic rules. And the repo is aggressively honest about what that means: synthetic distribution only, not SWE-bench, not a production trial. Fit to the benchmark, not proof of universal reliability. Provenance receipts, a pinned base revision, sha256 checksums, a full training receipt — built like a paper.

Why it matters: the worker/supervisor split is quietly becoming the default agent architecture, and the big labs ship it as managed, closed features. Dot Reflex is the first open-weight answer — small enough to run beside any worker, clean enough to deploy. If the controller role gets commoditized the way embedding models did, value migrates up-stack to whoever owns the harness. Oracle's read: watch that ten-decision taxonomy. Early days, but a standard is taking shape.