Nautilus

Nautilus is the supervisor that sits beside a federated engine, not inside it. It is deterministic and agent-free: no model, no agent loop. Triggers are a pure function of a snapshot, unit-testable without I/O. Judgment stays in the engine as Overwatch (ACP plus an independent model). Nautilus reaches Overwatch only through an escalation; it scores the answer like any other observer.

The grammar is zndx.supervision.v1 in signals-protocol. The protocol names no project. Each engine ships an instance — config/supervision/<project>.textproto — that declares the supervision tree, restart strategies, cadences, phase gates, horizons, resource intents, and objectives. Gaius is the reference instance; Hermes, Signals, Ægir, and Atelier ship their own.

What it does

• Supervision tree. Every process names a parent and a restart strategy on its children (ONE_FOR_ONE, ONE_FOR_ALL, REST_FOR_ONE, or NONE = observe and score). Warmup is grace; silence past a cadence is MISSING; progress beats wall-clock. Recycle budgets trip a breaker that is surfaced, never silently retried.
• Phase gates. A machine decomposes a flow into phases closed by explicit gates. Forecasts resolve against that phase’s gate within that phase’s horizon, so a download succeeding is not Brier-penalized for a later serving failure.
• Positions. Processes that declare reports_positions call ReportPosition at each gate. Adapters fill Metaflow steps, Airflow task instances, and pg_cron runs as observed positions. A process that promised to report and never does is UNSUPERVISED.
• Resource intents. Occupancy, floor, and priority are declared on a phase. Nautilus emits the intent on entry (and a zero floor on exit) so Signals can merge guarantee floors before the GPU token is already needed.
• Ledger. Forecasts, resolutions, transitions, and objective verifications go first to a local write-ahead buffer, then to the project’s PostgreSQL. The supervised database being down is when the supervisor’s memory matters most.
• Federation. One binary; one instance and one ledger per project. Peers watch peers over Engine/Status. No engine writes another engine’s ledger.

Nautilus does not schedule work (YuniKorn and zndx.scheduler.v1 do) and does not host models. It observes through each engine’s zndx.engine.v1 face, so a distributed deployment can expose one protocol port per host and nothing more.

Source: weathership/nautilus. Spec: nautilus_supervision.md.