Guardian intelligence platform
Structured decision-support for wreck pollution risk.
More than 8,500 potentially polluting WWII wrecks lie on the world's seabed. No government can survey them all. The Guardian platform exists to answer the question that makes action possible: which wrecks, in which order, and why.
Built for coastal states, port authorities, development banks and environmental agencies — every score is evidence-grounded, every change is audited, and the platform is explicit about what it has observed versus what it has inferred.
Wrecks assessed to date
Monitoring regions
Authoritative geodata sources
Active mission — Rabaul
01 · ASSESS
Every wreck gets a defensible number.
The WERP v6 methodology (Wreck Environmental Risk Prioritisation, with drift-aware environmental sensitivity) scores each wreck on hull condition, pollutant payload and environmental exposure, amplified by the live stresses acting on it:
Risk = ( WCS + PHS + ESI ) × RPM
Phase 1 builds the assessment from historical and open-source intelligence — loss circumstances, fuel and cargo records, survey literature — hydrated against authoritative geospatial datasets. Phase 2 replaces inference with observation: in-water survey evidence upgrades each score component individually, with the survey provenance recorded against it.
Unsurveyed wrecks carry an explicit deduction — the platform never presents inference with the confidence of observation. How the scoring works, in full →
What an assessment contains
- Ranked risk score and band — Critical, High, Medium, Low — on fixed, published thresholds.
- Component evidence — how the vessel sank, what it carried, what surrounds it, what is stressing it now.
- A fact-checked narrative dossier — machine-drafted, credibility-checked, with the scores composed programmatically: the language model never writes a number.
- A visible methodology stamp — version, revision, scoring date and verification state, wherever the score appears.
- Exportable briefs — stamped PDF, HTML and structured formats for institutional use.
02 · MONITOR — GUARDIAN SENTRY
The assessment doesn't stop when the report is written.
Wrecks fail when the environment stresses them. Guardian Sentry watches the forces that matter across 40 monitoring regions, continuously, and feeds what it observes back into each wreck's risk score.
Seismic
USGS earthquake feeds checked hourly. Each event is run through ground-motion physics — peak ground acceleration at the wreck's location and depth — not a magnitude lookup table. Only shaking that could actually stress a hull moves a score.
Storm
NOAA National Hurricane Center and GDACS cyclone feeds every six hours — including the Western and South Pacific basins where our wrecks are concentrated. Wave models estimate the shear stress a storm actually delivers to the seabed at each wreck's depth.
Satellite
Sentinel-1 radar oil-slick detections via SkyTruth Cerulean (CC-BY-4.0), screened for confidence and size, every six hours. Radar sees slicks day and night, through cloud.
Observed stress events decay over years rather than vanishing — the earthquake happened; its effect on the hull fades, it doesn't un-happen. Every alert lands on an analyst desk with escalate, watch and dismiss actions, and significant developments can be issued to host-state authorities as sealed, hash-verified Sentry notifications.
03 · ATTRIBUTE
Attribute, don't inflate.
A satellite slick detection is a question, not a verdict: something released oil here — was it one of our wrecks? Guardian treats it that way. Each detection becomes its own tracked object with a full lifecycle — detected, analyst-confirmed, under survey, attributed — and a ranked list of candidate source wrecks scored on proximity, fuel inventory and depth.
Detections refine attribution; they do not inflate risk scores. A slick near a wreck raises questions for the analyst — it changes a government-facing risk number only in one tightly-defined case: a recurring release where that wreck is the only plausible source. Ambiguous detections never score, no matter how often they recur.
That restraint is deliberate. A risk register that quietly inflates on every satellite pass stops being believable. Guardian's numbers move on evidence, and every movement is logged.
The analyst desk
- Slick triage — every active detection in one ranked table.
- Candidate review — confirm, dismiss with reasons, group detections across satellite passes.
- Human supersedes machine — once an analyst has classified a detection, automated labels defer to that judgement, and the supersession is audited.
- Escalation — confirmed developments become formal notifications to the responsible authority, not just dashboard pixels.
04 · MODEL
Where would the oil go?
Guardian is coupled to an independent spill-drift engine built on operational ocean forecasts. When a slick goes under survey, the engine runs the release backwards — winds, currents, timing — and produces its own ranking of which candidate wreck the oil most plausibly came from, in seconds.
The two methods are kept independent by design. The engine never reads Guardian's candidate rankings; Guardian never blends the engine's answer into its own. When two independent methods agree on a source, that is corroboration worth acting on. When they diverge, that disagreement is honest information about uncertainty — and the platform shows both.
The same engine generates seasonal drift envelopes for every assessed wreck — where a release could actually reach, season by season — which ground the environmental sensitivity component of WERP v6 scoring.
Honest about limits
The drift capability page states plainly what is operational, what is in build, and where the model is still being calibrated against commercial benchmarks. Some environmental receptor layers remain structured placeholders pending authoritative sources — and the platform says so wherever they are used. We would rather show you the caveat than have you find it.
05 · SOVEREIGN BY DESIGN
Your wrecks, your waters, your data.
Locally processed AI
Assessment and narrative generation run on language models hosted on Guardian's own infrastructure. Wreck intelligence is not sent to third-party AI services for scoring.
Scoped access
Institutional evaluators receive read-only accounts scoped to their regions, enforced in the data layer itself — not just hidden in the interface. Regional partners see their waters, completely, and only their waters.
Audited, always
Every score recomputation writes a before-and-after audit record with its trigger. Issued Sentry notifications are sealed with a cryptographic hash. Nothing moves silently.
The platform is structured decision-support, not an oracle. It tells you what it observed, what it inferred, and how confident it is in each — and it leaves the decision where it belongs, with the responsible authority.
Go deeper
Science & methodology
The scoring model in full: components, phases, data provenance, and how the methodology is governed.
Resources & publications
The WERP white paper, protocol definitions, case studies and area reports.
The Rabaul Mission
The platform applied: a proof-of-concept deployment to one of the most concentrated wreck sites in the world.
See it with your own portfolio in mind
Institutional evaluators can be provisioned with a scoped, read-only account to explore the live platform — the assessed inventory, the monitoring picture, and exportable briefs.