QYLTRACK INTELLIGENCE / EVIDENCE SYSTEMShello@qylai.com

SOFTWARE FOR INCOMPLETE, CONFLICTING MOTION DATA

See what happened between the pings.

QYL is building systems that reconstruct motion through sensor gaps, score behavior against what was expected, and keep the evidence needed to check the result. Maritime is the first public evaluation.

Discuss a track problem See the maritime case ↓
TRACK FIELD / SYNTHETICWEST ← LONGITUDE → EAST
Synthetic AIS vessel track fieldDozens of faint vessel paths cross a dark plot. One orange path stops, leaves an empty interval, and resumes under a different identity.LAST PINGAIS GAPNEW IDENTITYNORTH ↑TIME →
The background paths are generated at build time from a committed fixed seed. They are a visual model, not Global Fishing Watch observations.

PRODUCT DIRECTION / IMPLEMENTATION IN PROGRESS

Track intelligence that can be checked.

The planned QYL product connects raw observations to an analyst's decision. Each case will keep the sensor record, reconstructed track, anomaly score, and confidence interval together.

01 / INGEST

Keep each report.

The source record, clock, identity, and sensor label will stay attached.

02 / ALIGN

Use one clock.

Reports will be ordered in time without discarding their source fields.

03 / ESTIMATE

Fill the gap.

A motion model will estimate the missing path and the vessel's return.

04 / SCORE

Measure departure.

The track will be compared with a named baseline and a held-out vessel set.

05 / REVIEW

Open one case.

The source track, estimate, score, and interval will share one review record.

FIRST PUBLIC PROGRAM / MARITIME

Scoring a vessel that goes dark

A gap says only that the receiver heard nothing. A useful flag has to describe what the vessel did after the silence and state how often that assessment is wrong.

A
Normal vessel track with regular AIS pingsEARLIERLATER
REGULAR REPORTINGOne identity follows its usual route.
B
AIS pings stop, leaving an empty time intervalNO PINGSEARLIERLATER
AIS SILENCEThe dashed segment is unknown water.
C
A vessel track reappears off course with a conflicting identityIDENTITY CONFLICTEARLIERLATER
CONFLICTING RETURNThe new identity does not fit the route.
D
Proposed predicted path across the gap with widening uncertaintyEARLIERLATER
PROPOSED OUTPUTScore and interval await evaluation.
anomaly scorenot measuredconfidence bandnot measured
A normal track disappears. A conflicting identity returns off course. The detector is intended to compare that return with a predicted path, but no AIS score is shown before the evaluation exists.

FIRST PUBLIC EVALUATION / AUGUST 2026

No AIS result is public yet.

The available QYL code has no vessel anomaly dataset or calibration bins, and no baseline has been named. No score is presented here. The source and one-command run will be linked when they exist so reproducibility can be checked outside QYL.

evaluation command pending publication
detector against baselinenot measuredcalibration errornot measured
RELIABILITY DIAGRAMMEASURED CURVE PENDING
Reliability diagram with no measured curvePredicted confidence is on the horizontal axis and observed frequency is on the vertical axis. Only the ideal diagonal reference is shown.PREDICTED CONFIDENCE / PERCENTOBSERVED FREQUENCY / PERCENTNO MEASURED BINSPUBLIC AIS RUN REQUIRED
The dashed diagonal is the target: stated confidence should match the observed frequency. A solid measured curve will appear only after a held-out AIS evaluation is published.
SCORE DISTRIBUTIONTHRESHOLD PENDING
Anomaly score distribution placeholderAn empty score axis shows where ordinary tracks, a decision threshold, and the flagged tail will appear after evaluation.ORDINARYMORE ABNORMALTHRESHOLD NOT FITNO SCORES PLOTTED
The flagged tail stays empty until the detector, baseline, and decision threshold are run on the same held-out vessel set.

DATA BOUNDARY

What the evaluation will run on

The intended source is vessel activity and identity data made available through Global Fishing Watch. Use requires a Global Fishing Watch account and acceptance of its terms; published work must credit the organization. Global Fishing Watch states that its underlying commercial raw AIS cannot be downloaded freely.

No region or month has been frozen. Until that happens, QYL claims no data access and no geographic coverage.

METHOD NOTEIMPLEMENTATION PENDING

From track fragment to reviewable score

The planned detector compares each observed segment with its predicted track. A trajectory-prediction residual measures where the vessel returns; reconstruction error measures how poorly the full sequence fits ordinary movement. Conformal prediction will wrap those scores with confidence bands whose miss rate can be checked on held-out vessels. This AIS method has not been implemented in the current codebase.

VESSEL TRACK→RESIDUAL→SCORE + INTERVAL

COMPANY DIRECTION

One evidence system. More than one sensor picture.

QYL is being built for operational data that arrives late, drops out, or disagrees with itself.

The common unit is a time-ordered track with its source, identity, model output, and uncertainty intact.

That structure can support maritime awareness first, then other missions where motion is observed through imperfect sensors.

Broader claims will follow public evaluations. The AIS detector and its error-rate calibration are still in development.