Track record

What we will measure

Likelihood is an activity index, not a forecast probability (see the methodology). It still makes a checkable claim: a theater in a higher band should, more often than one in a lower band, be followed by the kinds of disruption the register exists to warn about. The calibration check we intend to publish:

• Unit: one (theater, day, band) observation, taken from the daily posture snapshot the feed cron already writes — the same number a customer saw that morning, not a recomputation.
• Outcome (revised 2026-09-07): whether, within the following 14 and 30 days, the theater recorded an event of a disruptive type — armed clash · attack on shipping · missile activity · blockade-class maritime incursion · cable or pipeline incident — drawn from the feed itself, so the check is reproducible from published data. It is counted at two tiers, reported separately and never pooled:
  • Tier A — curated or primary. The same types at medium confidence or better. For these types medium is, almost always, the ceiling: the one source that grades a disruptive event high is the NGA maritime warnings, which reach the feed only where a theater covers the water they warn about. The first arrived on 2026-10-01, when the Middle East shape grew to cover the Strait of Hormuz (an in-force mine warning). So Tier A means, in practice, a UCDP conflict record, and now and then an official maritime warning.
  • Tier B — machine-coded tripwire. The same types at any confidence, which in practice means GDELT’s low-confidence records. Reported beside Tier A and never merged into it, because in most theaters those same records are also driving the score being checked — Tier B is an internal-consistency check, not a validation.
  The first published version of this definition could not run. It required a high-confidence, primary-source disruptive event, and until 2026-10-01 the feed had never held one — zero rows, across every one of those five types. A check whose outcome variable cannot occur produces nothing, forever, and would have sat here looking like a plan. It is in the corrections record below. The first such row (a Hormuz mine warning) does not revive that definition: one source, in one strait, cannot carry a calibration check, so the two-tier definition stands.
• Report: per band and per tier, the observed outcome rate with a confidence interval, per theater and pooled; the same for the intensity band once recalibration ships. Bands that do not separate are a finding, not a footnote.
• Known weaknesses of the check, stated up front:
  • It is not independent of the thing it checks. The outcome is drawn from the same feed that drives the score, so a source going quiet lowers both.
  • Outcomes arrive late. The bulk conflict source lags 42–120 days, so outcome rates for recent windows stay provisional until its backfill lands.
  • A permanently violent theater passes without predicting anything. It will score “high band, high outcome rate” on base rates alone — which is exactly why the recalibration measures tempo against the theater’s own baseline.
  • Tier A is unobservable in five of the twelve geographic theaters. In the window this site ships today, the Taiwan Strait, South China Sea, East China Sea, Korean Peninsula and Baltic & Nordic record no disruptive-type event from a curated or primary source: every disruptive row in them is machine-coded, and the only ones above low confidence are breadth promotions, which Tier A excludes (next item). Those are the coercion flashpoints the product is bought for, so the first report will be able to check the active-conflict theaters and not those. Closing the gap needs a primary-source disruptive feed. The maritime-warning adapter (NGA MSI, listed on the sources page) shipped on 2026-09-16 and would grade such a row high — but 51 of the 54 security warnings in force fall outside every theater polygon, so it has produced no disruptive-type row in these theaters yet. The theaters are drawn too narrowly for the warnings that matter. Methodology 2026.09.2 widened the Middle East over the Gulf and Hormuz; the rest of the geometry proposal is still to be decided. Southeast Asia and South Asia gained shapes on 2026-09-30 (2026.09.3); their curated conflict records make both observable in Tier A from the first publish, so their record starts on 2026-10-01.
  • Confidence is not purely a source property, so the two tiers leak. The pipeline promotes a machine-coded candidate from low to medium when a high-confidence event from another source shares its theater, type and date, and — since methodology 2026.09.1 — when enough distinct outlets carried it. Such a row would satisfy Tier A while being a Tier B row wearing a promotion, so the check must exclude events tagged corroborated-by: from Tier A — otherwise it counts its own tripwires as outcomes. The breadth path is not rare (it lifts roughly a third of a day’s machine-coded rows), but every promotion is tagged, which is what makes excluding it possible; the exclusion is now what keeps the five theaters above honestly listed as blind.
  • Even Tier A means “a dataset recorded it”, not “a government confirmed it”. Its rows are curated conflict records coded from reports weeks after the fact. The stronger evidential standard the first definition reached for is the right one; the feed cannot currently supply it, and saying so is more useful than publishing a check that silently never runs.

Outcomes log

Empty. Daily posture snapshots have accrued since 2026-08-13; the first calibration report is due once a full 120-day window of snapshots exists with its bulk-source backfill complete — December 2026 at the earliest — and will be published here with the query that produced it.

Corrections record

Events withdrawn from the feed

Every event withdrawn from the public feed, with the reason. Withdrawals are soft — the row keeps its id and a retraction_reason, so a subscriber who already pulled it can reconcile against the correction rather than a gap. Reads exclude withdrawn rows everywhere.

This record covers the shared feed only. An organization’s own analyst may overrule an event’s grade for that organization (labelled analyst-adjusted on every surface it touches, with the reason); such overrides never alter the shared row, never appear here, and reach us only as field reports. When one leads us to change the shared feed, that change is recorded above like any other.

Claims we published that were wrong

The same standard, applied to our own prose. Nothing was retracted from the feed for these — what was wrong was something this site said about the instrument, which is the kind of error a reader has no way to catch. Each entry names the claim as it stood and what replaced it.

Methodology versions

Current: argus-methodology-2026.10. A comparison across versions compares two rulers; the product says so wherever it happens.