Methodology & AI Limitations

Delta Signals are distilled from open-source global news media and official hazard feeds (significant earthquakes from the U.S. Geological Survey and volcanic-ash advisories) — there is no covert or private-source collection. The pipeline is automated and runs continuously:

1. Ingest and screen — articles and hazard reports are pulled on a rolling basis and screened for relevance before any enrichment.
2. Enrichment — a model reads each article for what happened, where, when (the time it happened, kept apart from the time it was reported), who is said to be responsible and what was hit — each with the basis it rests on (stated in the article, reported without a figure, not stated).
3. Placing it — the places the reporting names are resolved only as far as they allow, on a three-step ladder: reported (country or province), localized (town or city) or pinpointed (street block or facility). Each event says where it stopped and why (stage_reason), a coordinate that contradicts the reported country is dropped rather than drawn, and the published coordinate is rounded to the level actually resolved.
4. Folding — reports of the same event are folded into one record, so the feed counts events rather than articles.
5. Assessment — every event gets a severity and a GEOINT-relevance score. Only events that could reach a reader's shortlist (GEOINT relevance 6 or more, or reporting from 3 or more outlets) also get a satellite-collection recommendation — sensor, resolution and what to look for. A missing recommendation means the question was not asked, not that the answer was no: the API returns collection.observable as null in that case, never false.
6. Source-linking — every event keeps the reporting it was built from; the facts it states keep the article each came from.

The Daily World Brief and the Delta Agent analyst are produced on top of this same corpus.

Each signal carries several distinct scores — they measure different things and should not be collapsed into one number:

• Event Severity (1–10) — the scale of the physical harm and footprint of a new event, not how newsworthy or how morally serious it is. This distinction is load-bearing: an individual crime, an arrest, a court ruling or a political announcement is scored 1–3 by design however grave it is, because it has little physical footprint. A high-severity event is also not necessarily observable from space.
• GEOINT Relevance (1–10) — how useful the event is for geospatial/satellite analysis. It is a different question from severity, but it is capped by it: severity 1–3 limits GEOINT relevance to 4, severity 4–5 limits it to 6, and only severity 6+ can reach 10. So a low-footprint event cannot come out highly GEOINT-relevant, by construction.
• What each rating rests on. Both ratings are made by an AI model reading the reports against the scales above. They are our own scales: no public body issues or endorses them, and we found no public method that scores single attacks, explosions or protests by severity. Where a report gives a figure, hovering over the rating in the event detail shows which public threshold that figure reaches, and says so when the rating is higher than the reported figures alone support:
  • 1 or more killed — the UCDP Georeferenced Event Dataset (Uppsala University) records an event of organized violence when it results in at least 1 direct death.
  • 10 or more killed, or 100 or more people affected — EM-DAT's entry criteria (CRED / UCLouvain) for recording a disaster.
  • More than 100 or more than 1,000 killed; more than 80,000 or 800,000 displaced — the Orange and Red flood alert levels of GDACS (United Nations and European Commission). For earthquakes, USGS PAGER uses 100–999 and 1,000+ estimated fatalities for its Orange and Red alerts.
  • A structure reported destroyed or damaged — the damage classes of the Copernicus Emergency Management Service (Destroyed, Damaged, Possibly damaged). Here the class comes from what the reports say, not from an image.
  • What was hit — the 16 critical infrastructure sectors listed by the U.S. CISA.
  When no casualty or damage figure was read from the reports, the hover says so: that rating rests on the model's reading alone. For GEOINT relevance the hover shows the ceiling set by severity, how precisely the event is located, whether it leaves a mark a satellite can read, and the image interpretability the recommendation calls for on the NIIRS scale (the public scale prepared by the IRARS Committee). We use “GEOINT” as defined in U.S. law (10 U.S.C. § 467) and NGA's GEOINT Basic Doctrine; that is a definition of the term, not a scoring method.Beside these AI readings, the event detail shows two labels worked out by fixed rules instead of a model's judgement: the Delta Impact Scale (the highest public threshold the reported figures cross) and Delta GEOINT Readiness (whether satellite imagery can check the event). How each is worked out.
• Satellite observability — two separate fields. collection.observability is a resolution class (0–3), not a probability: 3 means a ~10 m image would show the mark (flooding, burn scars), 2 needs ~3 m, 1 needs ~0.3 m, 0 means nothing would show from orbit. collection.observable says whether the place is known precisely enough to frame an image of that size; null means it was not assessed, which is not the same as false. sort=geoint ranks observable events first, then by GEOINT relevance, then by how precisely they are placed.
• Location precision — stage (reported / localized / pinpointed) and stage_reason; location_level is the level actually resolved and location_level_reported the level the reporting named; plottable is false when the place is too coarse to draw as a point.
• Reporting — how many articles were folded into the event and how many distinct outlets carried them. Distinct outlets is a breadth count, not an independence count: syndicated reprints of one statement are not collapsed, so independence is not asserted. Where an article carries a NATO (Admiralty) grading, the best source-reliability and information-credibility grades are shown beside it.

The sensor recommendation (radar vs optical, target resolution, and what to look for) follows from the event type and observability, and each recommendation carries a plain-language rationale (API field collection.reason) so you can see why a given sensor was suggested, not just which one. Ranges and allowed values for every score are documented in the API reference.

Everything below is produced by a large language model:

• Signal enrichment — what happened and where the reporting says it happened (the place names — resolving them to coordinates is deterministic), category, severity, GEOINT relevance, escalation trend, market-impact reasoning, and satellite-collection recommendations where no deterministic rule decides them.
• The Daily World Brief — a synthesized narrative over the day's signals, produced under quality checks (see below).
• Delta Agent — an agentic analyst that reasons over the corpus and returns a sourced brief.
• Signal assessment — the per-signal remote-sensing deep dive (assess_signal), which reasons about what imagery could confirm or refute a specific event.

Deterministic code — not a model — decides everything that gates or prices a result: plan entitlements, token charges, the count of distinct outlets carrying an event, how far a place is resolved, whether an event is observable (a comparison of the place's precision with the size of the mark), and the quality checks below. Those checks also constrain model output rather than trusting it: contradictory enrichment fields are suppressed before they reach you, and GEOINT relevance is capped by severity. What a model still judges is the reading of the article — what happened, how severe, what would show from orbit.

Enrichment is grounded in the source reporting rather than the model's own memory, which reduces (but does not eliminate) fabrication. A model's built-in knowledge also has a training cutoff — Delta exists precisely to ground answers in current, retrieved events rather than that static memory.

By default, no. Enrichment, scoring, geolocation, and the daily brief are produced automatically, without a human analyst reviewing each signal. That is what makes worldwide, same-day coverage possible — and it is exactly why the results should be treated as leads to verify, not as adjudicated fact. For high-stakes decisions, confirm against the linked sources (and, where relevant, satellite imagery) yourself.

• False positives / noise — an event may be mis-categorized, over- or under-rated on severity, or duplicated across reports.
• Mislocation — geolocation can resolve to the wrong place, especially for same-name locations; a signal may stop at a coarser step of the ladder (reported, at country or province level) when the reporting does not name a finer place.
• Source-dependent — coverage and framing inherit the biases and gaps of the underlying news reporting; absence of a signal is not evidence that nothing happened.
• AI error — scores and summaries are model estimates and can be wrong; the source links are the ground truth, not the generated text.
• Observability is a recommendation — the satellite-collection suggestion is an assessment of what could be imaged, not a guarantee of a usable image (weather, revisit, and licensing still apply).
• Market tags are informational — any market-impact tag is context, not investment advice.

The current event record starts on 2026-09-23. It is checked against automated consistency rules — for example, that an event is drawn as a point only when its place is resolved to a town or finer, that it is called satellite-observable only when its place is precise enough to frame, that every stated figure keeps the article it was taken from, and that one article is not counted in two events.

Scores and enrichment are validated against allowed ranges and value sets, and the Daily World Brief passes a deterministic quality gate before publishing: a truncated or structurally-incomplete brief is held back and the previous good brief keeps serving, rather than publishing a half-finished one. You can check current data freshness and pipeline status at any time via the API (GET /api/v1/status) or the MCP status://current resource.

Every signal links the reporting it was built from (the latest article from each outlet; every article on plans that include source detail); when you export or republish, keep those attributions. Satellite imagery shown or searched in Off-Nadir Delta comes from open programs (e.g. Copernicus Sentinel-1/2), which are free to use with attribution to the providing agency. Follow each provider's attribution and licensing terms when you redistribute derived imagery or products.

What we collect from news publishers, how long article text is kept, what is shown and what is not, and the legal position we take are set out in the Source & Content Policy.

Use Delta to find, prioritize, and orient — what is happening, where it is concentrating, and what to look at next — then verify the specifics against the linked sources before acting. Do not treat a single signal, score, or generated summary as confirmed fact, and do not use the service for unlawful surveillance or targeting of individuals.