2,241 Vessels Made an Unexplained 94-Degree Course Change at 7 Knots Last Month. They Were Avoiding Something That Wasn't Broadcasting AIS.

A solo maneuver inference is not a Rule 17 deviation — those require a known counterpart. This is what the data looks like when the counterpart is not on AIS.

The Setup

Over the last 30 days the infer-ghost-vessel-hourly cron emitted 3,160 inferences. Average course change: 93.5 degrees. Maximum: 179.8 degrees (a near-complete reversal). Average speed at the moment of the maneuver: 7.1 knots. Average geometric confidence: 0.81. Unique vessels involved: 2,241.

Every one of those 3,160 inferences was classified as head_on — the model's only currently emitted phantom-encounter type.

The Chain

The pipeline runs on AIS positions and reaches into a counterfactual: when a vessel makes a sharp course change at slow speed in open water, was there an AIS-broadcasting counterpart nearby? If yes, the encounter belongs in pairwise_encounter and gets COLREGS-scored. If no — no AIS counterpart within geometric proximity — the maneuver becomes a solo_maneuver_inference with a phantom-target hypothesis.

Mechanically:

1. track_point ingestion identifies a course change exceeding the threshold (here, ~30°+ from a stable bearing).
2. The slow-speed gate (≤ ~10 kt) selects for maneuvering states — fast, gradual course changes are routine ocean-passage adjustments and are excluded.
3. The detector queries pairwise_encounter and the local AIS lattice for any plausible counterpart at the maneuver moment.
4. If no counterpart, the model fits a phantom: where would a non-broadcasting vessel have to be, on what bearing, at what speed, for the observed course change to be a rational COLREGS-shaped avoidance?
5. The inferred phantom position, course, and a confidence score are written to solo_maneuver_inferences.

The output for the last 30 days is what 2,241 unique vessels did when they apparently saw something the AIS network did not.

The Implication

A 94-degree turn is large. Routine open-ocean course adjustments rarely exceed 15–20 degrees. A 94-degree turn at 7 knots is a vessel that has slowed, looked, and committed to a new bearing. That posture is consistent with a near-encounter avoidance.

If the model's hypothesis is right for even half of these 3,160 inferences, the implication is that the AIS network has substantial blind spots — and crucially, those blind spots are populated by something. Fishing fleet boats below the AIS-mandate threshold. Naval and coast guard vessels with deliberate emissions discipline. Sanctioned tankers with their transponders off. The detector cannot tell you which.

For maritime risk teams, the directional claim is: 2,241 unique vessels in 30 days observed at least one non-broadcasting counterpart at close enough range to require an evasive turn. That is a lower bound on dark-fleet density, mediated by the geometry of the broadcasting fleet. It is not a count of dark vessels — multiple inferences can resolve to the same phantom — but it is a count of broadcasting vessels that responded to one.

For a sanctions analyst, the highest-quality signal is the subset of the 2,241 with multiple inferences in 30 days. A vessel with three or four solo maneuvers all classified as head-on phantom encounters in known sanctions-evasion corridors is materially different from one with a single inference in the open Atlantic.

What to Watch

• Geographic clustering of inferences. Overlay the observer_lat/lon onto the chokepoint polygons (Bab-el-Mandeb, Hormuz, Suez S, Cape Agulhas). Phantom-encounter density inside known sanctions-evasion corridors versus the open ocean is the key disaggregation.
• Repeat-observer vessels. Vessels with 3+ phantom inferences in 30 days. Cross-reference with vessel_sanctions_risk and flag-hop history.
• The first non-head-on encounter type. The detector emits only head_on today; when crossing or overtaking phantoms come online, the spatial signal becomes much sharper. Until then, the head-on bias is methodological.
• SAR overlap when Copernicus credentials return. A solo maneuver inference within 10 nm of a Sentinel-1 pixel detection within 6 hours is corroboration. That is the unambiguous tip-and-cue signal.

Limitations

A 94-degree turn at 7 knots can also be: a vessel under tow being repositioned, a fishing vessel turning to set or haul gear, a piloted vessel responding to current or wind in tight water, a tug-and-barge unit slewing across a channel. Confidence (0.81 average) is geometric — it measures how well a phantom can fit the maneuver, not whether the phantom was physically there. The classifier's single-class output (head_on) is a model artifact, not a finding. A proper accuracy estimate requires either SAR corroboration (currently offline) or a manual analyst-labeled validation set, which the platform does not yet have at scale. Treat the 2,241 figure as an upper bound on broadcasting vessels that responded to a phantom, not as a count of dark targets.

Data current as of 2026-05-07. Sources: Axiom Overwatch solo_maneuver_inferences table, infer-ghost-vessel-hourly cron, derived from public AIS positions via AISStream.io. Read more at https://www.axiomoverwatch.io/blog.