We introduce the Neural Control Plane (NCP), a closed‑loop system that continuously observes language‑model execution, synthesizes candidate weight‑space intervention rules, evaluates them, and keeps a versioned table of admitted policies. NCP decouples the inference data plane of a pretrained model from the control policy that selects temporary parameter interventions.\ \ W‑ISA supplies transactional execution and parameter restoration, while NCP provides telemetry collection, candidate synthesis, admission criteria, priority scheduling, and rule‑lifetime management. The optimization state consists of the set of active rules and an archive of evaluated candidates; base‑model parameters are restored after each request.\ \ In a local SmolLM‑135M‑Instruct experiment, NCP generated three candidates from development telemetry and evaluated them on separate task examples and a 4,095‑target WikiText‑2 slice.\ \
- One projection reduced validation token repetition from $0.4133$ to $0.2191$, with a perplexity change of $-0.0574\%$.\
- All candidates produced zero newly correct exact answers, so the admission gate archived all three and retained the original policy as the sole active rule.\ \ Repeating the process with a fresh run and resuming the saved table reproduced the measured outcomes. NCP offers an executable framework for studying recursive self‑improvement at the control‑policy level, emphasizing explicit rule synthesis, verification, and state transitions.\ \ Review: NCP demonstrates a practical pathway for fine‑tuning model behavior via weight‑space interventions, laying experimental groundwork for future adaptive model control.