NewsPlanetAI Observatory · Case study

The problem

Everyone describes the chokepoint. Few can measure it.

Reporting on the world's straits leans on governments, operators, and interested parties. Public numbers quietly measure different things. The Observatory gives an independent physical observation, and states exactly what it saw, what it missed, and what it can't conclude.

How it works

Detect what's there. Abstain when you can't.

Clouds, glint, coastline and radar clutter all fake convincing detections, so the hard rule is knowing when to abstain. Gate-killed detections aren't deleted; they're kept for audit. Corrections are appended, never silently rewritten.

Observation pipeline

Sentinel-1 radar and Sentinel-2 optical from public catalogs, aligned to fixed observation areas, screened for land, cloud, clutter and known structures, then counted per pass.

Evidence & calibration

Preregistered blind draws are reviewed through a purpose-built light table; Wilson confidence intervals size the false-positive and false-negative error, and an instrument must pass both before graduation. Two-layer source and surface fingerprints catch stale provenance, and fail-closed publication barriers block anything that doesn't match.

Publication

Versioned JSON renders instrument pages, maps, evidence galleries and dated snapshots, each with its measured error printed beside the count.

What the instruments see

The Observatory map with its three standing instruments — Three instruments, trained on the world's chokepoints, each with its lens published.

Annotated Kerch Strait, the navigable span and the way in and out — Kerch Strait: the orientation is published, where ships enter, and the one span they must pass.

A false-colour Sentinel-2 SWIR image of the Subiya coastline — Open imagery in false colour, reading what a true-colour frame alone can't show.

The event-dated snapshots index, fires, tankers, anchorages — Event-dated snapshots: fires, tankers, anchorages, each carrying its provenance in frame.

The hard part

The hard part was never detecting bright objects. It was learning when to abstain, so a precise-looking count can't become scientifically weak evidence.

First-order observation, not a claim

A metric, retired before publication

An experimental radar-count track was retired before publication after two blind adjudication attempts failed validity review, leaving no defensible way to distinguish vessel-like returns from bright water texture. The verification system kept plausible-looking but unqualified counts out of production.

Stock, not flow

The instrument measures presence at one instant, not traffic, not vessel identity, not what happened between passes. The page says so, plainly.

Error published, not hidden

Detector loss and false-return rates are measured on cluttered passes and printed beside every count. The lens is published.

Challenged at every stage, I ratify

Adversarial tooling built the pipelines, generated stress tests, and challenged my interpretations at every stage, but none of it counted as evidence. I set the doctrine and retain publication authority.

Where it stands

Live, maintained, and open, the kind of auditable system I build.

The Observatory runs on event-driven refreshes and calibration triggers, free and public. If you need auditable research, OSINT, or decision tools where unsupported certainty is a real risk, that's the kind of system I build. Get in touch.

Open the Observatory ↗ Get in touch

Everyone describes the chokepoint. Few can measure it.

Before, during, after, every image carries its own receipts.

Detect what's there. Abstain when you can't.

Live, maintained, and open, the kind of auditable system I build.