124,170 Unsubtyped Dark Events. Subtype Classifiers Have Been Silent Since May 20. 2,066 of Them Are Critical.

The Setup

124,170 dark events have been recorded in the 45 days through June 23 with no subtype assigned. The most recent subtype assignment in the dataset — across all four active subtypes — is dated May 20. That is 34 days of detection without classification. The detection pipeline is running. The classification pipeline stopped.

The Chain

Dark events carry four subtypes: impossible_kinematics (959 total events, average gap 121.0 hours), nav_status (124 events, average 114.5 hours), teleport (32 events, average 104.2 hours), gnss_spoof (2 events, average 38.2 hours). Every one of those 1,117 subtyped events has a gap_start before May 20. Not a single dark event created after May 20 has received a subtype.

In the same 34-day window, the detection pipeline generated 124,170 unsubtyped events. The risk-tier breakdown:

| Risk Tier | Count | Avg Gap (hrs) | |---|---|---| | Critical | 2,066 | 32.2 | | High | 30,222 | 21.7 | | Medium | 52,535 | 19.5 | | Low (unadjudicated) | 27,802 | 16.4 | | Low (suppressed) | 11,549 | 6.9 |

The suppression mechanism is still active — 11,549 low-risk events have been automatically suppressed. Some automated processing runs. The behavioral subtype classifier specifically has produced no output since May 20.

The duration gap between subtyped and unsubtyped events is large and consistent. Subtyped events average 6 to 7 times longer gaps than the unsubtyped bulk: impossible_kinematics at 121.0 hours versus the unsubtyped average of 18.4 hours. This ratio suggests the classifier was built to target outlier-duration events, not all events equally. The 2,066 critical-tier events (avg 32.2 hours) are the closest analog — 75% longer gap than the unsubtyped average, the tier most likely to qualify for classification — but none have received a subtype in 34 days.

The Implication

Without subtype, the 2,066 critical-tier events and the 30,222 high-tier events are operationally indistinguishable beyond duration and risk score. An analyst cannot filter the critical queue for impossible_kinematics events — vessels that reappeared at physically implausible locations — because none of the current critical events carry that label. The same applies to gnss_spoof (only 2 confirmed events in the full dataset, both pre-May 20) and teleport (32 confirmed events, all pre-May 20).

The suppression logic appears dependent on subtype as a gate for higher tiers. Every suppressed event is low-risk and unsubtyped. No medium, high, or critical event is suppressed. Whether suppression above low-risk requires a subtype classification as a prerequisite, or simply never triggers on these tiers by design, is unclear — but the pattern is consistent with subtype being a dependency for automated processing above the low tier.

34 days of classifier silence means any analyst touching the current critical queue encounters 2,066 events with: risk tier, gap duration, geographic coordinates, and risk score. Behavioral category: none.

What to Watch

Whether the subtype classifier resumes and whether the 124,170 unsubtyped events will be retroactively classified. A backfill would reveal the behavioral distribution inside the current high and critical tiers — if impossible_kinematics is concentrated in the critical tier at rates consistent with pre-May 20 data, that reshapes the triage priority.

The maximum gap in the full dataset (including pre-May 20 subtypes) reaches 976.1 hours for impossible_kinematics — a vessel dark for 40 days. The critical tier's current maximum gap is 188.6 hours (7.9 days). Whether the critical-tier upper tail maps to specific behavioral subtypes is unknown without classification.

Limitations

The May 20 cutoff may be intentional — a classifier pause during a model update, schema migration, or quality control hold — rather than a failure mode. The data shows the freeze; the cause is not observable from the data alone. The 34-day gap is a fact about the classification pipeline, not necessarily a problem with the detection pipeline. Additionally, the 6–7× duration ratio between subtyped and unsubtyped events partly reflects selection bias in how the pre-May 20 classifier was deployed: it likely targeted outlier events, so the historical subtype population is not representative of the full event distribution.

Data as of June 23, 2026. Source: dark_events table, Axiom Overwatch. 45-day observation window.