Shadow fleet analyst dashboard — ranked watchlist, vessel investigation, and patrol handoff.
Live app: arktrace.edgesentry.io
- Ranked watchlist — vessels scored by causal intent, not anomaly. Each entry shows confidence, region, AIS gap count, and STS transfer candidates.
- Vessel detail — Feature attribution (SHAP) and ownership chain panels; AIS track; sanctions graph distance.
- Patrol brief — plain-language dispatch summary generated locally in the browser. No score or evidence leaves the site.
- Review workflow — triage (Watch / Escalate / Dismiss), handoff status, and review history synced across sessions.
- Pre-designation lead time — validates detections against known OFAC designation dates; typical lead: 60–90 days.
cd app && npm install && npm run dev # http://localhost:5173Select a region → watchlist loads → click any vessel → feature attribution, ownership chain, and AIS track appear.
Demo data (no pipeline required):
npx skills add edgesentry/arktrace
/arktrace-demo-dataarktrace applies Difference-in-Differences (DiD) causal modelling to identify vessels whose behaviour changed because of a sanction event — not merely anomalous vessels.
| Anomaly detection (excluded) | arktrace causal inference | |
|---|---|---|
| Unit of evaluation | A point — "does this vessel look unusual?" | A line — "did behaviour change because of a sanctions event?" |
| False positive driver | Any vessel that deviates | Only vessels whose deviation is statistically linked to a trigger |
| Lead time | Reactive | Pre-designation (see indago dashboard / C1 validation) |
| Output | Score + threshold | ATT ± 95% CI, p-value, SHAP signal breakdown |
Scoring is fully deterministic — no LLM in the pipeline. The browser generates patrol briefs via a local LLM with strict grounding constraints; the LLM cannot modify scores. See docs/ref-llm-grounding.md.
npx skills add edgesentry/arktraceApache-2.0 OR MIT