How the numbers are made.

GDACS disaster alerts, a verified armed-conflict catalogue, IPC food-security assessments, WHO disease-outbreak notifications, and per-country macroeconomic baselines.

Detection is wired end-to-end for four shock types today — natural disasters, geopolitical conflict, food-security crises and disease outbreaks. The authoritative detected-versus-specified split is published live by the platform itself, so it can never go stale in this paragraph. Stage-2 projections are being calibrated against documented historical shocks — the empirical benchmark currently covers earthquake, hurricane and flood classes only; until that calibration is signed, Stage-2 outputs are presented as projections, not measurements.

• Several factor priors are provisional pending analyst sign-off. The SSI's constants are under active review; expect revisions.
• The remaining specified shock types are not yet wired to live detectors; the platform's live coverage read is the authority on that split at any moment.
• Stage 2 currently runs on deterministic elasticity means; formal uncertainty bands are a staged increment, not yet shipped.
• The framing is economic and supply-chain by design — the SSI is not a humanitarian impact estimate.

USGS ShakeMap intensity contours and PAGER; GDACS event footprints; geoBoundaries ADM2 districts; GHSL population.

The finite-fault correction was fitted against the published USGS ShakeMap for the M7.1 Venezuela earthquake: mean contour error fell from 21.6 km to 8.8 km, and the corrected model recovers the innermost MMI 8 ring that a point-source form misses. District exposure bucketing was checked against USGS PAGER for the M7.5 Venezuela event: the engine's MMI 6+ exposed population matched PAGER's published figure (8.0M vs 8.0M).

Fatality brackets use the Jaiswal-Wald (2010) lognormal lethality form with a global vulnerability bracket — the span of the range is the uncertainty. They are labeled modeled, order-of-magnitude: a planning signal, never a prediction. When USGS publishes PAGER for an event, PAGER is authoritative — the internal bracket is shown alongside it, never blended with it, never as a competing headline.

• The casualty model is not calibrated per country — it uses a global vulnerability bracket, which is why it reports a wide range rather than a point estimate.
• Scenario footprints are idealised concentric contours from the intensity-prediction equation; a real published ShakeMap supersedes them the moment it exists.
• Exposure counts derive from modeled gridded population, not census rolls; they are exposure to shaking, not affected-person counts.
• The engine emits only what it actually computed — surfaces without a defensible value stay empty rather than showing a synthesised one.

Google Open Buildings footprints; USGS ShakeMap ground-motion grids; Copernicus Emergency Management Service rapid-mapping gradings; radar-measured ground deformation (InSAR line-of-sight displacement) as a regional observed layer.

© European Union, 2026, Copernicus Emergency Management Service (© 2026 European Union), [EMSR884] — Venezuela Earthquake. Copernicus EMS rapid-mapping products are published under CC BY 4.0, which permits commercial use with attribution; the terms grant reproduction, distribution and adaptation with no non-commercial clause.

The fragility curve was re-fitted by maximum likelihood on three satellite-observed damage tiles — 46,564 buildings spanning peak ground accelerations of 0.25–0.84 g. An earlier coastal-amplification heuristic over-predicted observed damage roughly threefold and was removed: where the observed data contradicted the model, the observed data won.

Corrected 2 August 2026 — a claim withdrawn. This paragraph previously reported the ranking model at “AUC 0.65–0.80 within an imaged zone, ≈0.67 transferred to an unimaged area, above the bare fragility curve’s ≈0.58”. Retested against 12 Copernicus EMS activations — 401,250 buildings across 11 earthquakes — none of the three figures survives. The 0.67 was leave-one-tile-out inside a single earthquake, which is not transfer to a new event. The ≈0.58 corresponded to no validation run in the codebase. And once the undamaged class is drawn from the standing building stock rather than from an analyst’s selection, per-building skill is AUC 0.490 — indistinguishable from chance. Seven covariates were swept (footprint area, height, built density, neighbour count, spectral shape, construction epoch at 1 km and 100 m); each works inside one area with a coefficient specific to that area, and those coefficients disagree across areas.

What is measured, and what it is for. Shaking predicts the damage rate of a neighbourhood, not the fate of a building. Refitted on 2 August 2026 over 12 earthquakes and 579,357 graded buildings: at ~150 m cells, Spearman ρ +0.362 leave-one-EVENT-out (p <0.0001) — pooled across the held-out earthquakes. That is a ranking claim, and it is the only accuracy figure Watchfloor publishes for this model. Qualified 5 August 2026, and read this before using it. Holding a whole earthquake out is the right split. But the correlation is then computed over every held-out cell from every earthquake at once, so differences in LEVEL between events — a M7.8 damages more everywhere than a M6.0 — enter the number and count as skill at ordering neighbourhoods. Watchfloor never asks anyone to compare a block in Türkiye with a block in Peru; it answers within this earthquake, which blocks are worst. That is the within-event figure, and it has now been measured. Measured 6 August 2026. The twelve Copernicus EMS activations were rebuilt from scratch and the rebuild authenticates itself twice: it yields 401,250 graded buildings — the same count this paragraph has always cited — and reproduces the published pooled figure at +0.355 against +0.362. On that same rebuild, the within-event ordering is:

• +0.104 — mean across the eleven earthquakes that can be scored.
• +0.211 — mean across the nine with at least fifty 150 m cells, which is the fairer read.
• +0.687 to −0.051 — the observed range on those nine. Four of eleven are negative: in those earthquakes the model ordered neighbourhoods backwards.

Use the range, not the average. A buyer who reads “+0.21” pictures modest but dependable performance; what the model actually delivers is one strong ordering or none, and nothing measurable tells you in advance which earthquake you have. We tested three candidate explanations for the spread and none survives: number of cells (rank correlation +0.445, p = 0.17, and the second-largest event scores −0.051), share of cells with zero damage (84.7% on average, p = 0.87), and the spread of ground motion within the event (p = 0.74). And one more thing that must travel with the number. Nine of the twelve activations grade only damaged buildings, so the undamaged class is reconstructed as “every other building inside the same mapped area”. Where Copernicus mapped a small fraction of what the footprint layer sees — coverage runs from 0.3% to 282% across areas — that reconstruction is doing more work than the data. Both figures above inherit that assumption. That assumption has now been tested against real observed undamaged buildings. Measured 24 August 2026. Microsoft’s AI for Good Lab published on HDX (CC BY 4.0) a city-wide assessment of Pereira for the 10 August Colombia M7.4: 143,472 buildings with usable imagery, 752 damaged — 0.52%. Every damaged building was reviewed by a person (166 ground truth, 302 confirmed, 295 unsure) and 754 further model candidates were rejected at review. So for the first time the undamaged class is observed rather than reconstructed, and the reconstruction caveat above does not apply. Matched building-to-building against what Watchfloor predicts — 24,767 buildings, 99.7% matched at a median 2.4 m, across the fourteen 0.008° boxes holding most of the damage — the result is:

• Level: right. Watchfloor predicts a ~0.65% damage rate for Pereira against 0.52% observed — agreement within about 1.2×.
• Ordering, within this city: absent. Per building the ranking measures AUC 0.445; stratified within each box, 0.482. At the model’s own 150 m cells, restricted to the 359 cells that hold the ≥25 buildings the rate was fitted on, ρ −0.087. The quartile of cells the model called worst came in at 1.07% observed damage; the quartile it called safest, 1.43%.

This is a thirteenth activation, from an independent source, and it lands just past the negative end of the range above (−0.051). Read plainly: for this earthquake the model told you correctly how much of Pereira was damaged, and told you nothing about which parts. It is one city, and a city where shaking barely varies — only seven distinct intensity values across the whole sample — so it does not refute ordering skill between regions of a larger event. It does say that inside a city, the map is a level, not a ranking. Where an observed layer exists, Watchfloor now serves it in place of the model for exactly this reason.

Which twelve — recorded 5 August 2026. The activations behind this figure were, until that date, written down nowhere in the codebase: the list survived only in working notes, so the figure could be rebuilt by its author and by nobody else. The list is now versioned with the engine and was verified against Copernicus rather than merely copied — all twelve activation pages describe an earthquake, and the two ends of the original scan range do not, which is the control that makes it a selection rather than a list of whatever answered. The file also carries the two traps that make a rebuild wrong rather than merely hard: nine of the twelve grade only damaged buildings, so their base rate is 100% by construction, and each activation must be paired to its own USGS event by date and area — pairing by name has previously matched Java to Sumatra 700 km away.

Which individual building fell remains chance — per-building AUC 0.490 once the undamaged class is drawn from the standing building stock rather than an analyst's selection — so Watchfloor shows individual buildings only where a satellite has actually looked. A 150 m cell holds a few dozen buildings, so what it observes is a noisy draw from the rate rather than a measurement of it: on held-out events, 88% of cells with a non-zero predicted rate saw no damage at all, which is what small samples do and is not evidence the rate is wrong.

• Outside an observed assessment area, every per-building value is a probability, labeled indicative — never a statement about an individual building's condition.
• Calibration rests on a small number of tiles from one event region, so transfer to different building stocks is the open question. The ≈0.67 new-area figure this line used to cite belonged to the ranking model withdrawn above and is not a property of what ships; the figure that does ship is the out-of-sample ordering, Spearman +0.362 — but that is pooled across held-out earthquakes, so level differences between events count as skill and it is an upper bound. Within a single earthquake — which is the only comparison Watchfloor asks anyone to make — the ordering is ρ +0.104 mean over the eleven scorable events (+0.211 over the nine with ≥50 cells), ranging +0.687 to −0.051, with four of eleven negative. Use the range, not the average.
• Satellite assessments are themselves model outputs of their providers, with their own error; Copernicus gradings cover only affected buildings, so their percentages read "confirmed among assessed", not a town-wide rate.

HungerMap LIVE survey history (the only feature source — no media or macro features), plus crop-calendar lean-season timing.

On its held-out test set the model reports MAE 1.6 percentage points, direction accuracy 97.7%, and R² 0.98. Read those honestly: the model is autoregressive — it extrapolates a country's own survey history — so these scores measure trajectory extrapolation, not shock anticipation.

• Inference is skipped for countries with under 90 days of stored survey history — the model never runs on synthetic inputs.
• Coverage is bounded by where phone-survey collection operates; no surveys, no nowcast.
• Being history-only, the nowcast is blind to exogenous shocks that have not yet moved the survey series.
• Retired from scoring on 2 August 2026. The nowcast used to raise a country's food-security signal when it forecast deterioration. It no longer moves any published score — on the day it was cut it was adding +8.6 to Yemen and +12.1 to Nigeria, saturating both at 100, on inputs up to 1007 days old. The model still answers on explicit request, with its staleness guard intact; what was removed is its ability to change a number you see.

Strictly out-of-sample only: forward-chaining AUC 0.714 (train on the past, predict the next) and leave-one-country-out AUC 0.721 (train on 127 countries, predict the held-out one). No random cross-validation — it leaks.

Qualified 5 August 2026 — what those figures were measured on. Both describe the eleven-feature model. The shipped forecaster feeds it eight: gdppc, sii (State Integrity Index at election time) and traj (its 180-day change) are always imputed to the training median, and alt_hist and polc join them for countries outside the curated set. gdppc is imputed for a concrete reason: the model was trained on GDP-per-capita growth while the live series returns the level, and feeding the level drives the standardised value to minus infinity. Weighed on the shipped coefficients, the three always-imputed features carry 2.8% of the linear predictor's variance (7.2% including the two curated-only ones), against 90% for incumbent tenure alone — so this is a small correction, not a hidden one. But it is a correction: an accuracy measured on a predictor that is not the one shipped is an upper bound. Measured 5 August 2026, and no longer an estimate. The training set was rebuilt (1,427 competitive elections, 128 countries, 54.5% incumbent-loss base rate) and both validations re-run twice on the same rows: once with all eleven features real, once imputing gdppc, sii and traj at prediction time exactly as the shipped forecaster does. With eleven features the published figures reproduce exactly — 0.714 and 0.721. The shipped configuration measures forward-chaining AUC 0.693 and leave-one-country-out AUC 0.713, a cost of 0.021 and 0.008. Those are the figures that describe what Watchfloor actually predicts, and they are what this page now quotes.

And a provenance repair worth naming. Until that run, the two published AUCs were literals: the export script wrote them into the model artefact by hand and printed them as fixed text, and the “countries” count was an expression that always returned 128. Nothing recomputed them, so nothing would have noticed if they had drifted away from the data. Re-running showed the values were right — but a correct number without provenance stays correct only until something changes underneath it. The export no longer prints an accuracy it did not compute.

Every call runs the same loop: predict → lock → resolve → score. A forecast is locked as an immutable snapshot the moment it is made — the number Watchfloor is later graded on can never be retroactively edited. When the outcome is established, the entry is resolved and Brier-scored, misses included. Forecasts built from structural priors alone are flagged as such, and every entry carries its framing — "calibrated relative risk · watch flag" — rendered verbatim, never as "prediction". The track record is public, no sign-in required — read it on the front page, or take the raw feed: /api/ledger/track-record (JSON).

• AUC ≈ 0.71 is honest skill for structural signals — materially better than chance, far from certainty. Treat outputs as ranked watch flags.
• Structural features move slowly; the model does not see fast idiosyncratic events. Narrative signals are layered separately and flagged as such.
• The ledger is young. The resolved-forecast count and the scores live on the public feed itself — this page states none, so it can never overstate them.

In a backtest over 4,841 country-years (1996–2024), countries with three or more simultaneously deteriorating vectors were 7.3× more likely to experience an irregular government change the following year than countries with one or none. Risk stacks monotonically with the count.

• The lift comes from additively stacking independent signals — the backtest found no non-linear "convergence magic", and we do not claim one.
• Irregular government change is a rare event: even at high convergence, most country-years see no such change. The flag marks a danger window, nothing more.
• Each vector inherits the limits of its underlying signal — including the media-derived nature of the conflict signal (see below).