keineuron — a nervous system for a fleet of agents.

Every kind of event — a command about to run, an agent action, an emitted answer, untrusted input — goes through the same shape: observe the event, select a handler from a dispatch matrix, get back a verdict. Nine different concerns collapse into one mechanism instead of nine separate subsystems.

• Reflex — deterministic checks (never a model) for the things a grep or an AST walk already catches reliably.
• Sense — a tiny, always-on, training-free network for fast pattern recognition, of the same class already measured sub-millisecond on real embedded silicon elsewhere in the lab.
• Judge — a language-model judge for the fuzzy cases the first two layers correctly decline to call.

One sense channel is a formal stability check, not a heuristic: a Lyapunov-style certifier measures whether an agent's sequence of turns is converging toward an answer or diverging into an unproductive trajectory, the same mathematical notion of convergence the lab uses to certify stability elsewhere in its research. Today this channel observes and logs — it is deliberately wired as an honest, non-blocking signal until its thresholds are validated against real, labelled episodes, not a silently-enforced gate.

The architecture is reconciled against the lab's own prior work — roughly 70% of it composes pieces already proven elsewhere (a dispatch matrix, a confidence detector, a routing gate, the Lyapunov certifier) rather than inventing new machinery. Implementation is staged and in progress; reported here as an architecture, not yet a fully running system.