Maritime and subsea-infrastructure protection

A maritime operations centre responsible for subsea cables, pipelines and offshore energy installations has more contacts than it has attention. Most vessel behaviour is lawful and unremarkable: fishing, transit, survey, standby. The behaviour that matters, a vessel loitering above a cable route with its transponder off, a repeated survey pattern by a research hull, a small uncrewed system launched from a merchant deck, is rare, ambiguous and easy to lose in the volume.

The difficulty is not detection. Sensors already see the contacts. The difficulty is correlation across sources with different refresh rates, coverage gaps and confidence characteristics, under time pressure, with an evidential standard high enough that a decision to shadow, hail, board or attribute survives later legal and diplomatic scrutiny.

SynapseCommand consumes the feeds an operations centre already holds. It does not require new sensors, and it does not replace the systems of record. Typical inputs in this profile:

The ISR Fusion agent normalises every input onto one track picture, resolving duplicate tracks from different sensors and carrying each contributing observation forward rather than collapsing it. Where sources disagree, the disagreement is preserved and surfaced rather than averaged away.

A behaviour-assessment pass compares track history against normal patterns for that vessel class, that area and that season. Anomaly is scored, not asserted: loitering over a cable corridor is weighted differently from loitering in a recognised anchorage. The Scenario Synthesis and COA Analysis agents then develop response options, each simulated against the assets actually available, weather and sea state, legal jurisdiction and expected escalation.

Throughout, the provenance layer records which agent produced which claim, from which inputs, with which model version. Nothing enters the dossier without a traceable line back to an observation.

The watchkeeper works in a queue, not a chat window. Anomalies arrive ranked, each with a confidence band, the sensors that contributed, and the specific behaviour that triggered the score. Opening one shows the track history, the imagery, the ownership and charter picture, and the reasoning chain in plain language.

Questions can be asked in natural language through CommanderNL: which other vessels were within five nautical miles in the last six hours, what this hull did on its previous three voyages, what changes if the anomaly threshold is tightened. Every answer cites its sources, and an operator can reject a fused track, correct an identity or mark a pattern as benign; the correction is recorded and feeds assessment.

The system recommends. It does not act. No tasking order, hail, intercept, boarding or attribution statement leaves SynapseCommand without a named human authorising it, and in an assured deployment configuration, that authorisation is ML-DSA-signed and written to the hash-chained, append-only record, bound to the exact evidence state at the moment of the decision.

Recommendations that cross policy thresholds are held behind an authorisation gate that states, before the operator approves, what will happen, what the system is uncertain about, and which rule permits it. Operators can override any ranking; overrides are first-class recorded outcomes, not errors.

A completed assessment produces artefacts intended for use outside the operations room:

SynapseCommand augments the recognised maritime picture rather than competing with it. Ingestion is over standard interfaces where they exist, including STANAG-aligned formats and Link-family tracks, with adapters for national systems. Outputs are returned as structured products the existing C2 system can display, and as human-readable dossiers.

Identity, clearance and releasability are taken from the host environment's directory and PKI. There is no separate user population to administer.

This profile is normally deployed on-premises inside the operations centre's accreditation boundary, or in a national sovereign cloud where one is accredited for the classification involved. Air-gapped operation is supported for the classified enclave, with controlled one-way ingest of unclassified feeds such as AIS and OSINT.

Models, weights, logs and evidence remain inside the customer's boundary. Nothing is sent to a vendor service for inference.

The benefit is in time and defensibility rather than in replacing watchkeepers. Correlation that a small team performs manually across several consoles is performed continuously, so the interval between an anomalous behaviour appearing and a briefed, evidenced course of action reaching the duty commander shortens substantially. Because provenance is captured as the assessment is formed rather than reconstructed afterwards, the evidential quality of an attribution improves and the effort of preparing it falls.

Figures quoted anywhere on this site are illustrative planning targets from internal test scenarios, not certified operational results.

The system cannot see what the sensors do not cover. Gaps in radar coverage, dark vessels and denied areas remain gaps; the platform marks them explicitly rather than interpolating over them.

Anomaly scoring reflects the patterns present in the data available to it, so unusual but lawful behaviour will sometimes be surfaced, and a sufficiently patient adversary can behave within the envelope of normal. Attribution of intent is not something the platform performs: it establishes what happened and how confidently, and leaves the judgement of why, and what to do about it, to the commander.