Every signal carries its own audit
Select a row to fly to it on the Picture tab. Select its DQ score to open the evidence card. A live single-source detection can score lower than a corroborated alert; that is the standard working.
Somewhere out there, a ship just went dark.
Live data · four minutes · every claim cited
Harbor
Maritime risk intelligence · Gimble Technologies, Inc.
Plenty of platforms detect dark ships. HARBOR shows whether you can trust the alert: every signal carries its provenance, confidence, staleness and sensor coverage, computed on live public data.
The unsolved problem is not detection. It is trust at machine speed.
Live feeds are fetched as the page loads and stamped with source and time. Switch the mode to Training to add the notional scenario, which is labeled synthetic everywhere and never drawn into the live layers.
Select a row to fly to it on the Picture tab. Select its DQ score to open the evidence card. A live single-source detection can score lower than a corroborated alert; that is the standard working.
The world's sea areas are divided among national coordinators, each publishing its own warnings in its own format. HARBOR attempts every one and shows the result per area. Warnings that state a position join the queue above and the picture. An area marked Not read contributes nothing, and the reason is given.
Twenty-four indicators of deceptive shipping, each with its common false positives and the independent corroboration it needs before anyone acts on it. HARBOR marks which ones it runs today and which it only declares. An indicator's corroboration requirement caps the independence score of every alert it raises. An alert with that requirement unmet is flagged single source; the flag does not change its band, because one source can be authoritative.
These are observable indicators for review, not findings. A vessel without AIS is not automatically part of the dark fleet, and not every shadow-fleet vessel is formally sanctioned.
Sea-state stress (Open-Meteo) charted against chokepoint news mentions (GDELT), rendered in your browser with Apache ECharts. Once the hourly collector has history, this chart switches to the persisted series for all six chokepoints.
Real Baltic vessel count (Fintraffic/Digitraffic), re-sampled every 60 seconds while this page is open. The rolling mean and ±2σ band build as samples accumulate.
A notional 14-day chokepoint series showing the full grammar: 7-day moving-average baseline, ±2σ tolerance band, and a flagged dark day where the count falls below baseline − 2σ.
Collected hourly by a scheduled function: Hormuz sea-state stress with a 24-sample moving-average baseline and ±2σ band, against GDELT news mentions. The same math as the demonstration, on persisted data.
Three kinds of statement about a hull, never merged: what an authority has officially listed, what a watchlist reports, and what HARBOR has observed. Everything is keyed on the IMO number, the one identifier a hull keeps when it changes name and flag.
Each card is assembled when you open it from everything HARBOR holds for that IMO number. The images are illustrations by vessel class: no verified, licensed photograph is on file for any hull, so none is shown. The registry seed is 38 identities from dated US Treasury releases and a UK sanctions summary; it is not the whole dark fleet.
Dark ships are found through AIS gaps, spoofed tracks, ship-to-ship transfers and satellite radar matching. Vessel photographs are copyrighted by their photographers, so each dossier links out to photo and track pages. Live algorithmic detection (Global Fishing Watch) appears on the Picture and Alerts tabs when its key is configured.
Traverse from a vessel to its stated owner, the parties an authority links it to, the network a watchlist places it in, and any company of the same name in the ICIJ Offshore Leaks Database. Every line says who asserts it. A dotted line is a name match only: a lead to check, not a finding.
Blindness is a property of a sensor portfolio, not of the ocean. SEAMS grids the Hormuz box and computes, per cell, the longest window in which no sensor in the selected portfolio can see a non-emitting vessel. By default the open-source result is measured from real Sentinel-1 acquisitions over the last 30 days, and it feeds every alert's coverage score.
Claude composes the brief from this page's live state only. Each sentence must cite evidence items; the server flags anything unsourced; and the narrative receives its own Decision-Quality™ score computed from what it cited.
Every layer declares what it is, where it came from and how it may be used. Planned integrations are declared before they exist. Select a row for the full record. The wider catalogue of sources HARBOR has assessed, connected or not, is in the source registry.