EMI — Environmental Screening Score: Methods Note

Version 1.29 — 2026-09-11

Point of Nexus — EMI (emi-data.com). This note describes the method behind the map: what is scored, from which sources, through which formulas and constants, and how well the results agree with independent data. Where the live product and this note disagree, the product is authoritative and the note will be corrected.

1. What EMI is, and what it is not

EMI (formally, the Environmental Metrics Index) answers one question for any coordinate on Earth: an Environmental Screening Score from 0 to 100, where higher is better, blended from six environmental pillars. Above 60 renders green, 31–60 yellow, 30 and under red. The scale never inverts anywhere in the product. Scored responses carry two markers, scale and blend, naming the rules in force (Section 7).

EMI is a screening instrument: it tells you where to look harder, not what to conclude. Three things it is not:

• A calibrated risk model. Every saturation constant is an anchor — measured at a named real place from the same data the score reads, never fitted against health, loss or regulatory outcomes. The pillar weights are declared judgements, not fitted parameters.
• A due-diligence product. A low score is a reason to commission a site assessment, not a substitute for one. A high score is not a clearance.
• A forecast. Every input is a climatology, a multi-year retrieval or a retrospective observation; nothing in the score moves with today's weather.

Its validated strength is ranking places; its measured weakness is level in specific regimes (Section 5).

2. Six pillars, one composite

Each pillar is scored 0–100 (higher = better), in two classes that combine asymmetrically:

A quality pillar measures a level whose whole range is informative. A hazard pillar measures an event burden whose top of scale means "the model found no burden": a measured absence of hazard may confirm the score, never raise it. The weights are final, declared judgement constants — the order is reasoned, the decimals are not; Section 5 measures their influence. A hazard weight prices how hard a bad reading pulls the score down, not how much a good one could lift it.

Q̄     = Σ (W_q / ΣW_q') · P_q     the QUALITY BASE: weighted mean over surviving quality pillars
P*ᵢ   = min(Pᵢ, Q̄)               for hazard pillars; quality pillars enter unchanged
mean  = Σ (Wᵢ / ΣWⱼ) · P*ᵢ       weights renormalised over surviving pillars
pullᵢ = p · (Wᵢ / the largest pillar weight) · max(0, mean − P*ᵢ)   one pillar's claim, in points
UERS  = mean − the largest pullᵢ,  rounded to an integer

Each pillar makes a claim on the score: how far below the mean it reads, priced by its declared weight. The score takes the strongest of those claims, not the lowest reading. A heavy pillar reading a little low outweighs a light pillar reading very low — so air at 12/100 decides the lean where green at 3/100 does not, because the product declares air worth 8/26 and green 3/26. Taking the strongest claim is the same as taking the lowest of the six scores each pillar would produce on its own, which is what makes the number move smoothly: a pillar getting worse always lowers the score or leaves it, never raises it.

The one-sided clamp min(Pᵢ, Q̄) makes the combination asymmetric. A hazard reading at or above the quality base enters at the base and contributes zero net points; a reading below the base drags the mean at its full declared weight and can supply the lean. A hazard held at the base can never be the pillar the score leans on: the mean of the clamped set never rises above the base, so that hazard's claim is zero. The clamp can only lower a score or leave it — never raise one. It introduces no tuned constant: it rests only on the quality/hazard partition and the weights already declared.

Why the clamp is one-sided, and what it displaces. It serves one principle: a measured absence of hazard must not add points, and the red band must be reachable on real ground. Without it, a "modelled, no flood found" reading alone is worth about 13 composite points at a typical flood-covered city centre, and none of the 300 most populous city centres reads red. Each other design examined fails a leg of that principle. Treating a class-0 flood reading as unmeasured fabricates an absence from a measurement, and red stays unreachable. Clamping only flood, or flood and fire, applies the principle selectively: a continuous urban core still reads green on a heat score of 100. Weighting each hazard by how informative its reading is still adds points mid-range and gives every point weights of its own that no reader can check. Multiplying a hazard factor into the quality score — the shape the INFORM Risk Index uses, where risk is a geometric mean of its dimensions and a zero in any one zeroes the whole — is a published precedent that runs the other way; under it one hazard reading of 0 serves a composite of 0 whatever the air, water and green read, the compensation-free extreme this note rejects as firmly as the fully compensable plain mean. Publishing each burden on its own and never combining is the purest design, but the map carries one number per hexagon, and a genuine floodplain would then wear a green quality-only hexagon with its flood reduced to a badge. The clamp adds no tuned constant, keeps every reading published and counted in confidence, and removes only a hazard's power to add points. Its price — quality-pillar error propagating about twice as hard at clamped points — is measured in Section 5.

The pull-back p is 0.45 when exactly 2 pillars survive, 0.30 for 3–6; clamped hazards still count as survivors. Without the pull a plain mean would be fully compensable — parkland beside a refinery would read "moderate". EMI publishes no composite with fewer than 2 surviving pillars, nor with fewer than 2 surviving quality pillars: hazard readings alone cannot testify that a place is good. The individual pillar readings still appear, with a note that they are not enough.

Heat sits hazard-side contingently: its anomaly reading is untrustworthy exactly where it reads best — a continuous urban core has no cooler neighbourhood to be measured against (Section 5's YCEO comparison). If a re-based heat baseline passes its pre-registered external check, heat's class must be re-argued, not inherited.

The lean is proportional to the declared weight of the pillar that supplies it. Air leans at the full p; heat at 5/8 of it; flood at 4/8; green, water and fire at 3/8. Across a sample of 21 world points green reads lowest at two thirds of them but supplies the lean at 12; scaling bounds a light pillar's share by its declared weight, so it can lower the score but not decide it.

Absence is never zero — and never 100. A measurement nobody took is never scored as a number: the term drops and the surviving weights renormalise; a pillar with nothing measured drops out and is reported by name as dropped. A published number always means somebody looked and measured. A separate per-pillar confidence signal (high/medium/low) reports how much is known and never touches any score.

Withheld is not dropped. A pillar enters the composite only when at least half of its declared sub-metric weight produced a reading. Below that floor the pillar is withheld: its measured values are shown, with the reason stated, and nothing from it enters the composite.

3. Data sources

BAKE = pre-processed offline into planet-wide data artifacts; RUNTIME = consulted while answering a request. Every scoring input is baked and read from EMI's own storage, and no country code takes part in scoring: the emissions term searches every registered source within 25 km of the point, whatever country it stands in. No third-party service is called when a score is computed, and the clicked coordinate never leaves EMI's systems. A point outside the baked coverage reports each affected pillar as not measured, by name.

The basemap is built on OpenStreetMap; no OpenStreetMap data enters any score — no fallback, no partial-coverage substitution, no live-lookup path. Address search is not built on OpenStreetMap either: it reads the baked place index in the row above. Where a source has no value for a point, the term drops; nothing stands in for it. Each scored quantity rests on exactly one source, so nothing in EMI is triangulated across sources (Section 6). The one place two sensors meet is fire's burn confirmation, and they do not meet as agreement: the CLMS and MODIS records cover different years and the pillar takes the strongest single reading, never a consensus of them.

The surface-water record is JRC Global Surface Water v1.5, covering 1984–2024. JRC reprocesses earlier years where it finds errors, mostly along shorelines and in gradient areas, so where a reading differs from the previous edition (v1.4, 1984–2021), part of the difference is JRC's correction, not the ground changing. Measured at the edition change, over the 236.4 million cells the record then held: 34.76 million carry a different value than under v1.4, and at the 99th percentile of the changed cells the water share differs by 0.0857 — 8.6 points on the linear surface-water term, roughly 0.4 points of the published score (estimate, from the Section 2 weights and the 2:2:1 sub-weights in Section 4). 20,795 cells with a v1.4 reading have no v1.5 reading at all: they report the term as not measured and the water pillar renormalises over its other scored terms — never written down as zero.

Shown but not scored: recent temperature and rainfall. The point report publishes the recent monthly mean air temperature and total rainfall for the cell (Copernicus ERA5-Land), each set against the 1991–2020 normal for the same calendar month — a July only against thirty Julys. These readings enter no score and refresh monthly. The newest months arrive preliminary and are replaced by the final reanalysis roughly two months later; a month still preliminary says so on screen. A twelve-month change verdict is withheld until twelve months of readings exist; until then the report states how many are in hand.

Shown but not scored: recent air. The point report also publishes a modelled monthly mean PM2.5 for the cell from the Copernicus Atmosphere Monitoring Service (CAMS) — a model estimate, not a ground-station reading — set against the same calendar month in up to three earlier years. CAMS publishes no 30-year normal for this product, so the comparison stays CAMS-against-CAMS; only the years that produced a value are averaged, and the report names how many. The model was upgraded across those years (generating process 144 in 2015 to 161 in 2026), so part of any year-to-year difference is the model, and the comparison is never extended into a trend. Air is scored from the WashU ACAG annual mean in the table above, a different product. These readings refresh monthly.

4. Normalization and anchors

Every constant is one of three kinds: anchored (measured at named coordinates, recorded alongside the constant), a source fact (published by the upstream provider), or a judgement (labelled as such). A constant that could not be measured is never guessed: the term drops.

Air — PM2.5 (0.70): linear from the WHO 2021 annual guideline (5.0 µg/m³) to 59.40 µg/m³, the measured p99 over all 632,877 baked cells.

Air — emissions (0.30): an inverse-square distance-decay burden over Climate TRACE assets within 25 km, mapped through a two-anchor log window — the shape EPA EJScreen uses for proximity scores. Floor: the median burden among cells with any registered source nearby. Ceiling: the largest burden any cell carries. Both anchors are distribution-anchored, not health-anchored — no authority publishes a guideline for CO₂e proximity, and the term measures the presence and size of large registered emitters, never exposure or dose. The radius and distance floor are judgements.

Heat — LST anomaly (0.60): the cell's summer-mean MODIS surface temperature minus a distance-weighted reference of its surroundings (23.74 km effective radius), scored 100·(1 − clamp(A/7.84, 0, 1)). The 7.84 K saturation is anchored at Sakarya, Türkiye — the largest of nine measured coordinates. Cooler than the surroundings clamps to 100.

Heat — thermal extremity (0.40): distance of the 30-year ERA5-Land climate normal from a temperate band of 282.15–288.15 K (9–15 °C), in either direction, saturating at 21.17 K — Verkhoyansk, the worst measured extreme. Cold dominates: Verkhoyansk saturates at 0 while Kuwait City lands near 47.

Flood: the score falls with the chance of at least one modelled river flood over a 30-year residence, computed from WRI Aqueduct's published return periods (a source fact): the worst mapped class — the 5-year floodplain — reads 0, and ground with no modelled inundation at any published return period reads 100. That 100 is a statement about one model's riverine search, not a certificate of dry ground (the caveats below and the class-0 row say what it covers), and under Section 2's rule it confirms the composite without raising it. The 30-year horizon is the one FEMA quotes for its high-risk zones; the 100-year floodplain carries a 26% chance over such a stay and reads 74. An unrecognised class drops the term rather than being read as lower hazard.

Each cell's class is read from the worst modelled pixel within one pixel of the cell's centre — a 3×3 block of the model's ~1 km pixels, the smallest window that holds every pixel the cell's own footprint can touch plus one pixel of margin, the width of a modelled river channel at this resolution. A cell on a channel the model leaves dry takes its bank's class; this is a statement about the ground within one pixel of the cell, not a floodplain search. Whether a cell has a reading at all is decided by its centre pixel alone, so the neighbourhood never creates a record over open sea.

The ladder has nine rungs: class 0, ground with no modelled inundation at any published return period, and eight return-period classes. The ninth published return period — 2 years — is excluded for the reason given under the table. Each score is 100 × (1 − P₃₀/P₃₀(5 yr)) with P₃₀ = 1 − (1 − 1/RP)³⁰.

The source also publishes a 2-year return period, absent above because its raster carries no inundation anywhere on modelled land: the scale is normalised against the worst class the data can actually reach.

Two caveats travel with every flood reading, both properties of the source:

• The hazard is undefended. WRI simulates inundation without flood protection: dikes, levees, barriers and pumped drainage are not modelled, so protected ground still reads as hazard. Dutch polders (the Alblasserwaard and Flevoland cells) read as inundated at a 10-year return period.
• The hydrology is a 1960–1999 climatology — a static product published in 2020, not current climate, not a projection, carrying no trend. A place whose flood regime has changed since 1999 is described by its earlier regime.

Water (2:2:1 sub-weights): dryness : surfaceWater : waterDistance at 2 : 2 : 1. Dryness scales soil water linearly against 0.3883 m³/m³, the reading measured at the Amazon wet-control point — wetter ground, higher score. SurfaceWater is linear in the JRC occurrence-band fraction, falling as the share of water rises; ground the survey never observed drops the term. WaterDistance is linear in distance: 100 where the surroundings hold permanent water, 0 at 100 km.

RainDeparture — the bounded ratio (r−e)/(r+e), floored where monthly totals fall below 5.0 mm because the source's own quantisation dominates there — is shown beside the score and not scored. It measures this season against the place's own normal, so it reads near-perfect wherever the climate is currently behaving: across a 340-point world sample its mean was 91.9, median 94, minimum 46. It can go low for a dry spell, not a dry place — information about now, not the ground, so it lives with the other shown-not-scored readings in Section 3.

One disclosed limit inside the scored terms: surfaceWater scores water underfoot as exposure, so ground with no surrounding water reads a perfect sub-score — a Sahel dryland with a measured water fraction of 0 takes 100 on that term. That mints base points from a measured absence of burden at sub-metric grain, against the principle the aggregation enforces at pillar grain; it stands, disclosed, until the term is re-founded on the survey's permanence band.

What the distance term measures, precisely: the distance to the nearest ground holding at least 10 hectares of permanent water — water present in all twelve months of 2024 in the JRC seasonality band — or to the sea, searched to 100 km. The 10-hectare floor is a measured choice: seeding on the mere presence of permanent water made 11 of 13 settled test cities read 0 km, and the floor is the smallest of five tested that separates dry cities from waterfronts (Şanlıurfa reads 29.7 km, Ankara 11.8, Konya 6.9; Istanbul's Bosphorus front, Cairo and Yakutsk read 0). Three limits travel with it: it is measured cell to cell on a ~7.45 km grid (honest band about ±3.7 km, and a cell holding water reads 0 wherever inside it the water sits); the floor is blind to shape (a permanent canal of the same area counts as a lake); and the coastline is inherited from a coarser land mask, so sea distance is quantised at ~19.7 km. Where the search finds nothing within 100 km — the deep Sahara, the Rub al Khali, Alice Springs — the term is reported as not measured and the pillar renormalizes over its other two scored terms: never 0, never 100, because "searched, and nothing found" is a statement about the search.

What the surface-water term measures, precisely: occurrence is the share of the 1984–2024 satellite record in which a 30 m pixel was seen to be water, averaged over the pixels the survey actually observed inside the 1 km disk. A pixel under water three months a year reads about 0.25 — seasonal and intermittent water count, and water that has since dried (the Aral Sea) keeps counting while it sits inside the record. Permanence is the survey's separate seasonality band, read by the distance term above and by nothing else: two bands, two questions, neither substitutes for the other.

Fire: P = 100 · (1 − w · f · m) — a product, not an average, because fire requires dangerous weather and available fuel together. w saturates at FWI 82.74 (Attica, Greece); f, a rainfall fuel proxy, at 23.7 mm/month (Central Yakutia, the driest real fire regime measured). The confirmation multiplier m = 0.5 + 0.5·c asks whether this ground has been seen to burn like a fire regime. c is the strongest of three readings, each divided by its own saturation constant: the worst whole calendar year and the worst single month of the CLMS record, at 0.1083 apiece, anchored at the cell of the 2021 Evia North megafire — which burned within a single month, so its worst month is its worst year — and the worst whole calendar year of the MODIS record over 2001–2018, at 0.5915, anchored at the cell of the 2017 Knysna fire, the only named catastrophe that window contains. A reading that is missing is left out of the comparison, never entered as a zero, and where no reading exists no confirmation is applied at all. The 0.5 floor is a judgement.

Green: greenAccess = 100·clamp(g/0.68953, 0, 1), linear in the tree/shrub/grass fraction of the 1000 m disk, computed over the area actually measured — unmeasured ground never dilutes it, and a disk with nothing measured returns no value. The 0.68953 saturation is the p90 of the green fraction across dense-urban disks (n = 10,611,411).

5. Validation and sensitivity

Sensitivity of the combination step, measured on the current aggregation: 208 real points sampled from the live map, weights drawn from a Dirichlet distribution centred on the live vector, pull-back from a uniform range, 2,500 draws per scenario, fixed seed 20260824. In the widest joint scenario the mean colour-band flip rate is 21.08% and the mean Spearman rank correlation against EMI's ordering is 0.9277: the ranking is the stable part, the six weights moving together are the sensitive axis, and the flips concentrate at points near the band thresholds (60/30). The record is scripts/sensitivity/results-2026-09-05.md. Two structural facts are derived from the declared weights rather than sampled: at a point where every hazard sits at the base, the three quality pillars carry the composite alone, so their noise and bias propagate roughly twice as hard — about 0.5 published points per point of air and 0.3 per point of green, against roughly 0.31 and 0.21 under a symmetric blend — which raises the priority of the air follow-up below and of water's still-missing external validation; and the hazard weights price only how hard a below-base reading bites.

The anchors and the band thresholds are perturbed too, on the same 208 points and the same seed. Four axes move: the six weights, the pull-back, every saturation anchor and ramp endpoint — each multiplied by U[0.8, 1.2], ±20%, drawn independently, with every point re-scored from its raw readings through the pillar curves — and the two band thresholds, 30 and 60, each moved by ±5 score points independently. The anchors alone flip the average point's colour band in 5.08% of draws and barely touch the ordering (Spearman 0.9917); the thresholds alone flip 10.55%, almost all of it at the 60 line (10.02%, against 0.77% at 30) and without moving a single score. All four axes together flip 23.83% of draws at Spearman 0.9243, and the flips stay concentrated at the edges: points within 2 points of 30 or 60 flip 45.35% of the time, points 11 or more away 5.82%. The record is scripts/sensitivity/results-2026-09-10.md, which also names what still stands unperturbed — the sub-metric weights inside each pillar, the withholding gates, and the two emission-window anchors, whose raw burden is not carried on the wire.

AIR PM2.5 vs 161 ground monitors (79 US EPA AQS, 82 EEA stations across 20 European countries including Türkiye): Pearson r = 0.933, MAE = 3.61 µg/m³, mean bias −3.06 µg/m³ (n = 161); 95.7% agreement on the 5 µg/m³ WHO cut. Three caveats. At high pollution the satellite-derived input reads below ground monitors — mean bias −8.45 µg/m³ in the 25–35 µg/m³ range and −16.35 above 35, concentrated in Balkan and Anatolian winter-inversion basins, where a reassuring air score deserves extra care. The top half of the curve is untested: no monitor in this sample sits near the 59.4 saturation anchor, and none sits in South Asia or Africa, where the highest baked values sit — that ground is covered by the separate check below, not by this one. And the comparison is not fully independent: ACAG is itself calibrated against ground monitors, so this tests EMI's full path from source to score, not the source's own method.

AIR PM2.5 vs 49 reference-grade monitors in South Asia and Africa (OpenAQ; 35 in India, and 38 of them read above 35 µg/m³ — the ground the US/European check could not reach): as collected, Pearson r = 0.830, Spearman ρ = 0.802, mean bias −1.78 µg/m³, MAE 12.14 (n = 49); dropping every monitor-year whose hourly feed pinned at the instrument ceiling leaves 27 monitors and reads r = 0.846, ρ = 0.885, bias +0.71 µg/m³, MAE 10.09. Any statement about level leans on the cleaned reading, because a pinned hour inflates the monitor and never EMI. On the 35 µg/m³ screening call the two agree at 45 of the 49. Levels are the weak part, and here the scatter runs both ways: individual sites differ by up to ~63 µg/m³ as collected and ~48 cleaned — EMI reads about 25 high at four Punjab stations and 35 to 63 low in Delhi, Lahore and Lucknow, a few hundred kilometres apart, so no single scale factor corrects it. Two limits travel with every figure here: the Indian ground truth spans 2016–2022, not the decade EMI serves, and 70 of the sample's 199 monitor-years contain pinned hourly readings. This check adds no uncertainty interval above 35 µg/m³ — the table still has no row there (Section 6).

HEAT vs Yale YCEO SUHI v4 at 44 city centres on six continents: r = 0.688 (n = 44). The gap is a construct difference, not noise: YCEO's reference is non-urban by construction, EMI's includes whatever surrounds the point, so a megacity core absorbs part of its own heat island.

GREEN vs MODIS NDVI at 113 globally stratified points: Spearman ρ = 0.676, Pearson r = 0.590 (n = 113). The ceiling is structural: EMI's green measure excludes cropland, NDVI counts it, so intensively farmed points differ by construction.

FLOOD vs two independent references, neither of which shares model ancestry with the ECMWF family EMI depends on. First, a global terrain map — height above the nearest drainage, derived from the Copernicus 30 m elevation model, CC0 — at 900 cells, 100 per flood class: low ground (within 5 m of its drainage) makes up 82.2% of the cells in class 4 or worse against 49.0% of class-0 cells, a 33.2-point gap (95% CI 22.4–43.6), so the model's "modelled wet" and "modelled dry" separate in a way terrain agrees with. But the rungs above class 0 do not order with terrain: Spearman ρ = −0.066 (n = 900), and classes 1 to 8 all sit at a median 0.4–1.5 m above drainage with no trend. Second, England's statutory river floodplain map (Environment Agency Flood Zones 2 and 3, Open Government Licence) at 240 inland, river-only points, 80 per zone: the model flags a 1%-a-year floodplain (class 4 or worse) at 7 of 80 zone-3 sites — hit rate 0.087, critical success index 0.080 — and reads class 0 at 73 of the 80; it never flags a point outside the mapped floodplain (0 of 80), and 57 of the 297 inland points asked (19%) were outside the model's domain and dropped as not measured. The reading: the model runs on roughly 1 km cells and resolves "wet or dry" but its return-period rung is not parcel position, and it omits rivers below its size threshold — exactly the ground England's dense small rivers occupy. A class-0 reading therefore means "no modelled hazard", not "no floodplain", and the error runs one way, towards under-detection: a non-zero class is trustworthy, a zero is not a certificate. Neither check met its pre-registered bar, so the flood pillar still reads "not independently checked" and this disagreement is published rather than relabelled (report: docs/validation/T91-FLOOD-VALIDATION.md; reproduce: node scripts/validation/t91-flood-hand.mjs --verify and node scripts/validation/t91-flood-ea.mjs --verify).

Rainfall normal vs two gauge records — the 1991–2020 mean monthly precipitation that WATER shows beside the recent months and FIRE scores as its fuel term — against CHIRPS v2.0 (CC0, within 50° of the equator) and GPCC v2022 (CC BY 4.0, poleward of 50°) over the identical window and statistic, at 345 stratified cells: Pearson r = 0.883, Spearman ρ = 0.932, MAE 17.3 mm/month, mean bias +9.7 mm/month; the two sides agree on the 23.7 mm/month fuel cut at 94.8% of cells and on the 5 mm/month rain-departure floor at 96.8% — consistent with its pre-registered bar. The reanalysis reads wetter than gauges: mildly within 50° (+4.5 mm/month, n = 225) and by about 20 mm/month poleward of 50° (+19.6, n = 120), a level bias that fire's fuel term absorbs by saturating at 23.7 mm/month — where it disagrees at the cut it reads slightly more fuel, never less. This checks a shown quantity and a fire input; two of water's three scored terms — the distance to permanent water and standing water — are checked next, and the third, soil dryness (2/5 of the pillar), remains unchecked (report: docs/validation/T91-WATER-VALIDATION.md; reproduce: node scripts/validation/t91-rain-chirps-gpcc.mjs --verify).

WATER distance-to-permanent-water vs an independent lake inventory — the scored waterDistance term (1/5 of the pillar) against HydroLAKES v1.0 (CC BY 4.0; every lake and reservoir of at least 10 ha, the same floor) and the Natural Earth coastline (public domain), at 400 cells stratified by served distance, 40 per 10 km decile: Spearman ρ = 0.722 (n = 400) on the full sample and 0.595 on the 276 cells the sampling did not steer, below the ρ ≥ 0.70 half of the pre-registered consistent bar. The test is one-sided: the inventory holds no rivers and no reservoir built after 2016, so EMI reading nearer than the inventory is the expected shape (188 cells), and only EMI reading farther than the inventory by more than one grid cell (7.45 km) counts as a miss. The miss rate is 0.135 over the full sample and 0.196 over the 276 cells the sampling did not steer — the fair headline, because the 124 steered cells were chosen for having no mapped lake within about 94 km and cannot miss by construction — and 0.196 crosses the pre-registered finding bar (> 0.15), so the unsteered subset fails on both counts and this is a published disagreement. What the misses look like: all 54 nearest polygons are natural lakes, most of them small (median 0.38 km²; 41 of 54 under 1 km²), almost all in dry interiors (Australia holds 13 of the 54, from 42 sampled cells); of the four at or above 10 km², three are closed-basin salt lakes and one, Mar Chiquita, is a saline lake with a strongly fluctuating shoreline. Whether such a lake holds water in all twelve months — which is what the served term requires — decides whether the served number is right, and this check cannot tell "not permanent" from "not seen". What it does not show: the sea side agrees throughout (0 of 65 coastal cells miss), the near end agrees (40 of 40 cells served under 10 km), and where EMI reports no permanent water within 100 km the inventory also finds none within 90 km at 200 of 224 such cells — withholding lands on genuine absence. The finding re-labels nothing: the tier moves only on the two terms that carry 4/5 of the pillar (report: docs/validation/T91-WATER-VALIDATION.md §2; reproduce: node scripts/validation/t91-water-hydrolakes.mjs --verify).

WATER standing water vs an independent land-water mask — the scored surfaceWater term (2/5 of the pillar), the 1984–2024 surface-water occurrence share of the ground within 1000 m, against the MODIS and SRTM land-water mask for 2024 (MOD44W v061, openly shared; one yes/no per 250 m pixel for a single year held against a 41-year occurrence mean — a construct difference stated before the run) at 300 points, 60 per served bin: Spearman ρ = 0.699, Pearson r = 0.575, MAE 0.162, mean bias +0.083 (n = 300), and the two sides agree on whether any water is present at 78.3% of points (128 both, 107 neither, 63 EMI-only, 2 mask-only). That sits between the pre-registered bars — under the consistent bar on both halves (ρ ≥ 0.70 and presence agreement ≥ 85%), above the finding bar (ρ < 0.50) — so it is neither corroboration nor a finding: where EMI serves no water the mask agrees at 58 of 60 points, where EMI serves more than half the ground as water the mask sees water at 54 of 60, and the disagreement sits between, where EMI serves 5–50% water and the mask reads none at 49 of 120 points — ground a single year's mask and a 41-year occurrence share must read differently, which only a same-construct multi-year comparator would settle. The tier does not move; the checks of soil dryness (against SMAP L3 soil moisture) and rain received (against GPM IMERG) are pending, so water still reads "not independently checked" (report: docs/validation/T91-WATER-VALIDATION.md §3; reproduce: node scripts/validation/t91-water-mod44w.mjs --verify).

FIRE fire-weather climatology vs a second, independently forced Fire Weather Index — the served w input, the 1991–2020 mean of each year's maximum daily FWI, against the same statistic from GFWED v2.0 (the same index computed from MERRA-2 reanalysis forcing in place of ECMWF's; terms of use as recorded in the validation report), at 939 cells stratified across all ten served deciles: Spearman ρ = 0.903, Pearson r = 0.919, relative bias −0.135 (n = 939) — consistent with its pre-registered bars (ρ ≥ 0.80, |bias| ≤ 25%). Two things travel with the headline. The bias is a slope, not an offset: a line through the pairs reads served ≈ 10.6 + 0.66 × reference, so the served value sits 7–17% above the reference in the low-to-middle deciles and 12–25% below it from decile 7 up — about three quarters of the reference on the strongest fire-weather ground. The two sides agree on the 82.74 saturation anchor at 85.5% of cells, an agreement carried by the 691 cells both put below it: the reference would call 244 cells saturated where the served scale calls 116, so the anchor is met in fewer places than a MERRA-2-forced index would meet it. And the frame is vegetated land only: the reference computes nothing on desert, ice and cold-desert ground, which removed 42 of the 100 cells drawn from the served top decile, so desert fire weather is checked by nothing here. Both sides are reanalyses: this corroborates a ranking and bounds the dependence on forcing, and it measures no error against observed fire weather.

FIRE burn confirmation vs mapped fire perimeters — the two CLMS burn records (worst whole calendar year over 2019–2025 and worst single month from July 2018) against agency-mapped fire perimeters in the United States (MTBS; terms of use as recorded in the validation report) and Canada (the Canadian National Fire Database, Open Government Licence – Canada), at 485 cells drawn one third each from served-confirmed, served-partial and served-zero ground, both sides read at the 0.1083 confirmation cut. The annual record, the term the bars were pre-registered on: hit rate 0.950, false-alarm rate 0.252, critical success index 0.720, Spearman ρ = 0.766 (n = 485) — consistent with its bars (H ≥ 0.70, ρ ≥ 0.60). The monthly record, reported beside it as characterisation: hit rate 0.839, false-alarm rate 0.208, critical success index 0.688, Spearman ρ = 0.765 (n = 483; two Canadian cells drop for an undated perimeter), which meets the same bars. The disagreement runs one way: a mapped perimeter encloses unburnt ground that a 300 m burnt-pixel product does not flag, so on the 113 cells where both agree a burn happened the served fraction reads a median 0.54 of the perimeter fraction, and all six misses are large boreal wildfires mapped at 11.8–21.3% of the cell and served at 4.4–10.8% — the low tail of that same ratio. The burn factor under-reads and does not invent: no cell the served record reads as exactly 0 holds a mapped perimeter above the cut, and the false alarms sit on cropland and boreal ground where burning below the mapping floor (500–1,000 acres) or in 2025, a year the archives do not hold, leaves no perimeter. The headline hit rate weights the three served classes equally; re-weighted to their measured share of the ground, the hit rate is an estimate of 0.60 for the United States and 0.81 for Canada (inputs: the class shares and per-class rates in the report), which clears the finding bar and not the consistent bar in the United States. Two countries only: nothing here speaks to Mediterranean, tropical or savanna fire regimes. The third burn record — MODIS, the worst whole calendar year over 2001–2018 — is checked against the same perimeters over the same 2001–2018 window at 439 cells, both sides read at its own 0.5915 cut: hit rate 0.968, false-alarm rate 0.124, Spearman ρ = 0.868 (n = 439), consistent with the same bars, with the served-confirmed class short of its design (105 of 166 cells, because ground burnt beyond 59.15% in one year is under 2% of the candidates and the query budget ran out — a shortfall that can lower the hit rate and not raise it) and the served fraction reading a median 0.94 of the perimeter fraction where both agree, so the under-read of about half belongs to the CLMS record and not to this one.

What this means for the fire pillar's tier. Under the rule in force for a product-form pillar, fire reads corroborated on the fire-weather leg: an independent dataset corroborates the ranking, and the fuel factor's rainfall check above and the burn factor's perimeter checks stand beside it. It reaches measured only when all three factors carry a measured error against observation; the burn factor's three records and the fuel factor's rainfall are now each checked, and what remains is fire weather — both sides of that check are reanalyses — so no error size is claimed, and the pillar's confidence sentence names the datasets it was compared against and where it reads lower (report: docs/validation/T91-FIRE-VALIDATION.md; reproduce: node scripts/validation/t91-fire-gfwed.mjs --verify, node scripts/validation/t91-fire-perimeters.mjs --window clms --verify and node scripts/validation/t91-fire-perimeters.mjs --window modis --verify).

6. Known limitations

• Hazards not covered. EMI does not model earthquake, coastal or storm-surge flooding, wind, or landslide hazard; check those separately.
• Flood covers river flooding only. Coastal flooding, storm surge and rainfall that never reaches a river are outside this layer.
• Flood answers almost everywhere, and a "0" is a finding, not a blank. The WRI Aqueduct layer reaches 99.61% of artifact cells — 1,217,497,813 of 1,222,281,444 — because the artifact's cell set follows the model's own domain: every res-9 cell whose centre pixel the river-flood model covers has a record, plus every cell that already carried another measurement. The raster itself covers about 87.48% of land, a different denominator. On the remaining 0.39% — cells carried for another measurement where the source publishes no data at all — the pillar drops and is reported as not measured; open ocean and ice outside the model carry no record and no score. Most modelled ground reads class 0 — 1,083,393,391 cells — and that is a measurement: the model ran here and found no inundation reaching 0.1 m at any return period within one pixel of the cell. 11.01% of modelled cells are wet at some return period.
• Class 0 means "no water at least 10 cm deep", not "no water". Ground whose deepest modelled inundation is 4 cm reads 0; 46,197,709 cells sit in that band — wet, but under the source's 0.1 m floor. That population is counted, not estimated: the flood layer's own run receipt tallies it (class0WithShallowDepth, v2-flood-aqueduct-nbr3-260910).
• Class 0 means "no modelled hazard", not "no floodplain". The river model runs on roughly 1 km cells and leaves out rivers below its size threshold, so ground on a smaller river's floodplain reads 0 because that river is not in the model, not because the model looked and found it dry. Measured on England's statutory river-floodplain map: 73 of 80 sites in the 1%-a-year zone read class 0, and no point outside the mapped floodplain was flagged — the error runs one way, towards under-detection (Section 5). A non-zero class is trustworthy; a zero is not a certificate.
• A cell on a large river's channel can read 0 — and such a reading does not publish at high confidence. The permanent channel appears excluded from the modelled inundation surface. Reading each cell from the worst pixel within one pixel of its centre (Section 4) moves most channel cells onto their bank's class; where a class-0 cell is nevertheless itself mostly permanent water (surface-water share above 0.5), the flood pillar publishes with its confidence capped below "high" and a reason naming the situation; the score is unchanged, and a cell missing either reading carries no cap. Measured: 1,866,264 such cells — 0.94% of the 199,191,942 class-0 cells that also carry a surface-water reading — sit on majority-water ground; most are lakes, reservoirs and coastal water, where "no river flood depth" is the correct answer. A channel wider than the one-pixel reach still reads 0 beyond its bank — a known limit of this layer.
• Fire's burn confirmation reaches back to 2001 and no further; a fire before that leaves no trace in the score. Three records answer the question and the strongest of them confirms: the CLMS worst whole calendar year (2019–2025, because 2018 and 2026 are part years and cannot hold a worst year), the CLMS worst single month (July 2018 to February 2026), and the MODIS worst whole calendar year over 2001–2018. Each is divided by its own bar — 10.83% for the two CLMS readings, 59.15% for the MODIS one — and a reading that is missing drops out of the comparison instead of entering it as a zero; where none of the three exists, no confirmation is applied at all. Pedrógão Grande, Portugal and Knysna, South Africa both burned catastrophically in June 2017 and are inside the score through the MODIS record; the Camp Fire of November 2018 is inside it through the monthly record. The factor only ever discounts, so unconfirmed ground publishes less fire hazard than confirmed ground with the same weather and fuel. Inside the windows there is no fading: the worst year, or the worst month, counts the same in 2003 or 2025. The MODIS record is closed and does not refresh — its readable archive ends, and its window is cut at 2018 where the CLMS record opens — so it can only add confirmation, never withhold it; recent fires are watched by CLMS, which reads whole years to 2025 and months to February 2026. The year-by-year burn chart on the point report is the same MODIS series over its full 2001–2022 span, shown and not scored.
• A barren desert does not score worst-possible on GREEN. Where the measured disk is barren-dominated, with neither green cover nor built presence, the pillar is withheld with the stated reason structural_zero instead of scoring 0.
• The heat reference is neither water-masked nor elevation-controlled. Inland-water pixels cool the reference, and relief within the 23.74 km radius reads as thermal anomaly. Both errors overstate heat only — the score biased downward — because the cool side clamps to 100.
• Heat is unavailable at many coastal and monsoon-belt points, big cities included — the widest coverage hole in the product. The anomaly needs a clear-sky summer composite at the point (MODIS warm seasons 2019–2023) and a neighbourhood reference, computed only where at least ~253 km² of land inside the 23.74 km radius was itself measured. Persistent warm-season cloud removes the first; coast, island and narrow-peninsula geometry removes the second. Either way the anomaly is unmeasured — never 0 — and since it carries 0.60 of the pillar, heat is withheld: its temperature reading stays on screen, its score out of the composite. The project's own 51-city heat cross-check lost five city centres exactly this way (Mumbai, Dhaka, Jakarta, Ho Chi Minh City, New York Midtown), and Miami Beach withholds heat the same way — hot, humid, crowded places where a thermal reading matters most.
• Per-score uncertainty is a validation tier, not an error bar. Each pillar's confidence object names the external check its input survived (measured / corroborated / unquantified). A pillar can also cap its own confidence where a specific combination of its readings undermines its claim — a ceiling one step below "high", never arithmetic on any value, with its reason named, and not raised when either reading it depends on is unmeasured; the one such cap is flood's channel-zero case above. Only AIR carries a number: a disagreement interval conditioned on the reported value (<25 µg/m³: MAE 3.35, n = 155; 25–35: MAE 10.31, bias −10.09, n = 6). No interval is reported above 35 µg/m³. Monitor pairs on that ground now exist — the South Asia and Africa check above holds 38 of them — but their sites miss by up to ~48 µg/m³ in both directions and the clean row that would anchor a number is not yet stable, so the absence is stated, not filled with a zero row.
• Coverage edges. pm25Annual is unmeasured poleward of 68°N/−58°S, so high-Arctic points drop the term rather than reading clean; the fire burn record misses 4.17% of cells around Greenland; watercourses narrower than the 30 m surface-water survey can read as if no water were present on that term. The heat anomaly's edge is wider than all three, falls on cities, and has its own entry above.
• Single source per quantity, throughout. Nothing in the system can be triangulated, each anchor stands on few measured points, and Section 5's sensitivity analysis measures how much the published score leans on where each anchor sits, not whether it sits in the right place — moving an anchor is not the same as checking it against a second source, and there is no second source. Fire's burn confirmation reads two sensors, but over disjoint years and as a maximum, so it widens the record rather than corroborating any single reading in it.
• Several scored terms descend from one reanalysis family. Water's soil-dryness term, heat's thermal-extremity term and the two factors inside fire's one scored term — the fire-weather climatology and the rainfall normal that stands in for fuel — share one lineage: Copernicus ERA5-Land, and the Copernicus CEMS Fire Weather Index, which is itself computed from ERA5. An error in that family moves those three pillars together rather than cancelling out, and the recent temperature and rainfall readings shown beside the score come from it too. Agreement between those pillars is never read as corroboration: every pillar publishes its agreement field as unavailable, and nothing in the product turns an unchecked quantity into a number. Measured outside that family: air's satellite PM2.5 annual mean and its emissions inventory, green's WorldCover and GHS-BUILT-S land cover, water's standing-water share and distance to permanent water (EC JRC Global Surface Water, whose sea mask is the one ERA5-Land input in that chain), fire's CLMS and MODIS burned-area confirmation, and the MODIS surface temperature behind heat's anomaly. Flood belongs to neither list — open global flood models share modelling ancestry with that family, which is why its external checks (Section 5) use a terrain map and an engineering floodplain map that share ancestry with neither.
• Grid and point run one model; the overlay is short one water reading. The map overlay and a clicked point run identical formulas from the same scoring configuration, so a hexagon and the point inside it can never disagree by version. They can differ by input: the overlay scores water without the distance-to-permanent-water term — that table is read per point, not per viewport — and water renormalizes over its other two readings, so a hexagon and a click on the same ground can land a few points apart, and at a band threshold in different bands. The gap is bounded at ≈3.00 composite points, ≈6.05 where the composite leans on water. The clicked point carries every measured term and is the exact one, and the map's legend says so where the shading is explained.

7. Versioning

• Model described: the scoring model live on the v14 data artifact.
• Data artifacts: core v14-planet-260910 (21 B / 7 fields per cell; 1,222,281,444 res-9 cells in 13,514 chunks — every cell whose centre pixel the river-flood model covers, plus every cell that carried a measurement before); heat baseline heat-baseline-260820; ERA5-Land sidecars v8-{expected-rain, rain-received, dryness, temp-normal}-260821; fire weather v8-fire-weather-260821; PM2.5 v10-air-pm25-260823; burnt area v10-burned-area-annual-260823 (worst whole year) and v10-burned-area-monthly-260823 (worst single month, 551,268 cells), plus v1-burned-area-modis-pre2019-260906 (worst whole year 2001–2018, 523,366 cells, a closed window); emission assets v11-emission-assets-260902; permanent-water distance v3-gsw-waterdistance-r6-260903/f10ha (13,398,221 cells, 10 B each); display-only monthly/annual series v9-*-by-month-260822 / v9-burned-area-by-year-260822.
• Scale and blend: scored responses (/api/score, compact and detail, and /api/grid) carry scale: "benefit-100-260903", naming the ruler — higher is better — and a blend field naming the aggregation revision that combined the pillars. A stored response whose markers differ from the current ones, or that carries neither, was produced under different rules and is not comparable number-for-number or band-for-band.
• Change policy: a substantive method change produces a new version of this note; the note always describes the method in force, and scores never change silently.
• This version (1.29, 2026-09-11): the data artifact moves to v14-planet-260910. The cell set widens from 236,486,188 to 1,222,281,444 cells — every res-9 cell whose centre pixel the river-flood model covers, plus every cell that already carried a measurement — so unbuilt land the model covers now has a record and a composite where it read "insufficient quality pillars" (the Sahara core among it); open ocean stays without a record. The flood class is read from the worst modelled pixel within one pixel of the cell's centre (3×3 neighbourhood maximum) instead of the centre pixel alone — 11.01% of modelled cells wet at some return period, against 6.81% before; Section 4 states the operator, Section 6 the coverage and the narrowed channel-zero count. The record shrinks from 25 fields to 7: the eighteen OpenStreetMap-derived tallies, scored by nothing since 2026-08-29 and shown nowhere since 2026-08-27, are no longer baked; the four land-cover and water fields are re-measured with 0 mismatches on every cell that had them, and the heat reading is carried byte-identically (231,546,321 measured cells) and reads "not measured" — never 0 — on every new cell. No formula, constant or anchor changes. Existing points whose 1 km disk now holds newly measured cells move: on a sample of the two artifacts, about one point in eight changes by a point or more, in both directions, because the green, water and built shares are means over the whole disk rather than its previously measured part. Three of the four control cities are byte-identical; Karabük moves from 51 to 54.
• Version 1.28 (2026-09-10): disclosure only — Section 5 adds water's standing-water check: the served surface-water occurrence share against the MODIS and SRTM land-water mask for 2024 at 300 points, between its pre-registered bars (ρ 0.699 against ≥ 0.70, presence agreement 78.3% against ≥ 85%, finding bar ρ < 0.50 not crossed), neither corroboration nor a finding; the rainfall paragraph now names soil dryness as the one unchecked water term; no formula, constant, score or validation tier changes.
• Version 1.27 (2026-09-10): disclosure only — Section 5's fire paragraph adds the third burn record's check: the MODIS worst-year record against the same mapped fire perimeters in the United States and Canada over 2001–2018 at 439 cells, consistent with its pre-registered bars, with the served-confirmed shortfall (105 of 166 cells) and the near-parity with the perimeter fraction (0.94) named; all three burn records the fire pillar reads are now checked; no formula, constant, score or validation tier changes.
• Version 1.26 (2026-09-10): Section 5 adds fire's two independent checks — the fire-weather climatology against a second, independently forced Fire Weather Index at 939 cells, and burn confirmation against mapped fire perimeters in the United States and Canada at 485 cells; both consistent with their pre-registered bars, with the top-end compression and the under-read of about half named — and water's distance-to-permanent-water check against an independent lake inventory at 400 cells, a published disagreement (miss rate 0.196 on the unsteered cells). The fire pillar's validation tier moves from "not independently checked" to corroborated on the fire-weather leg, and its confidence sentence names both comparisons; water's and flood's tiers do not move; no formula, constant or score changes.
• Version 1.25 (2026-09-10): disclosure only — Section 5 replaces flood's "no external comparison" with two independent checks (a global terrain map at 900 cells, England's statutory floodplain map at 240 sites; both below their pre-registered bars, so the pillar still reads "not independently checked") and adds the rainfall-normal check against two gauge records (345 cells, consistent, with the poleward wet bias named); Section 6 states that a class-0 flood reading means "no modelled hazard", not "no floodplain", and the class-0 sentences on the panel and the point report say the same; no formula, constant, score or validation tier changes.
• Version 1.24 (2026-09-10): disclosure only — Section 5 now reports the sensitivity of the saturation anchors (±20% each) and of the band thresholds (±5 points each) alongside the weights and the pull-back, Section 6 states what moving an anchor can and cannot establish, and Section 5 adds the air check against 49 reference-grade monitors in South Asia and Africa, the ground the US/European check could not reach; no formula, constant or score changes, and no uncertainty interval is added above 35 µg/m³.

Özet (Türkçe)

EMI — Çevresel Tarama Skoru: Yöntem Notu, sürüm 1.29, 11.09.2026.

EMI (resmî açılımıyla Environmental Metrics Index), Dünya üzerindeki herhangi bir koordinat için 0–100 arası tek bir Çevresel Tarama Skoru üretir; yüksek değer daha iyidir ve ölçek üründe hiçbir yerde tersine dönmez. Skorlu API cevapları yürürlükteki kuralları adlandıran iki işaret taşır: cetveli adlandıran scale ("benefit-100-260903") ve birleştirme kuralını adlandıran blend. İşaretleri bugünkülerden farklı olan ya da hiç taşımayan kayıtlı bir cevap, bugünkü sayılarla ne sayı sayıya ne bant bantta karşılaştırılabilir. Skor bir tarama aracıdır: nereye daha dikkatli bakılmalı sorusuna cevap verir, ne sonuç çıkarılmalı sorusuna değil. Kalibre edilmiş bir risk modeli, durum tespiti ürünü veya tahmin değildir; her doygunluk sabiti, adı kayıtlı gerçek yerlerde ölçülmüş bir çapadır. Tek bir saha için kesin cevap gerekiyorsa yerinde inceleme yaptırın.

Skor iki sınıfa ayrılmış altı sütundan hesaplanır: hava (8/26), sıcaklık (5/26), taşkın (4/26), yeşil (3/26), su (3/26), yangın (3/26); ağırlıklar nihai birer yargı sabitidir — sıralama savunulur, ondalıklar değil. Kalite sütunlarının (hava, su, yeşil) ağırlıklı ortalaması "kalite tabanı"dır. Tehlike sütunları (sıcaklık, taşkın, yangın) tabanda ya da üzerinde okuduklarında tabana kırpılır ve ortalamaya net sıfır puan katar: "aradık ve tehlike bulamadık" skoru teyit eder, yükseltmez. Tabanın altında okuyan tehlike ise tam ağırlığıyla ortalamayı aşağı çeker. Ortalama sonra en kötü sütuna doğru çekilir; çekim o sütunun beyan edilmiş ağırlığıyla ölçeklenir — düşük ağırlıklı bir sütun skoru tek başına belirleyemez, tabana kırpılmış bir tehlike hiçbir zaman "en kötü sütun" olamaz. Ürünün ilk kuralı: ölçülmemiş bir değer asla bir sayı olarak puanlanmaz — eksik veri düşer, adıyla bildirilir, kalan ağırlıklar yeniden normalize edilir. İkiden az sütun ya da ikiden az kalite sütunu kalırsa bileşik skor üretilmez: tehlike okumaları tek başına bir yerin iyi olduğuna tanıklık edemez. Alt-metrik ağırlığının en az yarısı ölçülmemiş bir sütun alıkonur: değerleri gerekçesiyle gösterilir, skora girmez.

Örnek çapa: PM2.5 eğrisinin sıfırı DSÖ 2021 kılavuzu 5,0 µg/m³, tavanı 632.877 hücreden ölçülen p99 = 59,40 µg/m³'tür; ölçülemeyen sabit tahmin edilmez, terim düşer. Altlık harita OpenStreetMap üzerine kuruludur; hiçbir OpenStreetMap verisi hiçbir skora girmez ve skor hesaplanırken hiçbir üçüncü taraf hizmeti çağrılmaz: her girdi EMI'nin kendi depolamasından okunur, tıklanan koordinat EMI'nin sistemlerinden dışarı çıkmaz. Adres araması da OpenStreetMap'i değil, EMI'nin kendi pişirilmiş yer dizinini okur. Bir kaynağın değeri olmayan terim düşer; yerine hiçbir şey konmaz.

Nokta raporu, 30 yıllık normallerin yanında son ayların okumasını da yayımlar: aylık ortalama sıcaklık ve toplam yağış (Copernicus ERA5-Land), aynı takvim ayının 1991–2020 normaliyle karşılaştırılır — temmuz yalnız temmuzlara karşı. Bu değerler hiçbir skora girmez ve aylık tazelenir; en yeni aylar ön veri olarak gelir, nihai analiz yaklaşık iki ay sonra yerini alır ve ön veri olan aylar bunu ekranda yazar. On iki aylık değişim yargısı, on iki okuma birikene kadar verilmez. Hava için ikinci bir okuma daha vardır: CAMS'in ürettiği aylık ortalama PM2.5 tahmini — model çıktısıdır, istasyon ölçümü değildir — aynı takvim ayının önceki en çok üç yılıyla karşılaştırılır; CAMS bu ürün için 30 yıllık normal yayımlamaz. Model bu yıllar arasında güncellendi (üretim süreci 2015'te 144, 2026'da 161), bu yüzden karşılaştırma asla bir eğilime dönüştürülmez. Skorlanan hava, tablodaki WashU ACAG yıllık ortalamasıdır; CAMS de kaynak tablosunda kendi satırını taşır (Copernicus CAMS küresel atmosfer bileşimi tahminleri, CC BY 4.0).

Yüzey suyu kaydı JRC Global Surface Water v1.5'tir (1984–2024). JRC hata bulduğu eski yılları da yeniden işler; v1.4'ten (1984–2021) farklı okunan yerde farkın bir kısmı zeminin değişmesi değil, JRC'nin düzeltmesidir. Sürüm değişiminde, kaydın o sırada tuttuğu 236,4 milyon hücre üzerinde ölçülen fark: 34,76 milyonu farklı değer taşır; değişen hücrelerin 99. yüzdeliğinde su oranı 0,0857 ayrışır — terimde 8,6 puan, yayımlanan skorda yaklaşık 0,4 puan (ağırlıklardan türetilmiş tahmin). v1.4'te okuması olan 20.795 hücrenin v1.5'te okuması yok: terim "ölçülmedi" olarak bildirilir, sıfır yazılmaz. Yüzey suyu terimi, 1984–2024 kaydında suyun görülme payıdır: mevsimlik sular da sayılır ve kurumuş sular (Aral) kayıtta kaldıkça saymayı sürdürür. Kalıcılık ise aynı taramanın ayrı "seasonality" bandıdır ve onu yalnız uzaklık terimi okur: iki bant, iki ayrı soru. Uzaklık terimi, en az 10 hektar kalıcı su tutan en yakın zemine ya da denize uzaklıktır; arama 100 km'ye kadardır, skor doğrusaldır (çevrede kalıcı su varsa 100, 100 km'de 0) ve su sütununda kuruluk : yüzey suyu : uzaklık 2:2:1 ağırlık taşır. 10 hektarlık taban bir ölçümdür: yalnız "kalıcı su var mı" sorulduğunda 13 yerleşimden 11'i 0 km okuyordu; denenen beş tabandan kuru şehirleri su kıyılarından ayıran en küçüğü budur (Şanlıurfa 29,7 km, Ankara 11,8, Konya 6,9; Üsküdar, Kahire ve Yakutsk 0). Sınırları: ölçüm ~7,45 km'lik ızgarada hücreden hücreyedir (±3,7 km), taban şekle kördür (aynı alandaki kalıcı bir kanal göl sayılır) ve denize uzaklık ~19,7 km'lik adımlarla ölçülür. 100 km içinde hiçbir şey bulunmazsa terim "ölçülmedi" olur ve sütun diğer iki terim üzerinden yeniden normalize edilir — 0 da 100 de yazılmaz. Yağış sapması gösterilir, puanlanmaz: bu mevsimi yerin kendi normaliyle karşılaştırır ve iklim olağan seyrindeyken hemen her yerde yükseğe yakın okur (340 noktalık örneklemde ortalama 91,9, ortanca 94, en düşük 46) — zemin değil, şimdiki zaman bilgisidir. Yüzey suyu teriminin beyan edilmiş sınırı: çevresinde hiç su olmayan zemin o terimde tam puan okur; terim, taramanın kalıcılık bandına oturtulana dek bu hâliyle servis edilir.

Birleştirme adımının duyarlılığı bugünkü birleştirme üzerinde ölçülüdür: 208 gerçek nokta, senaryo başına 2.500 çekiliş, sabit tohum 20260824. En geniş ortak senaryoda ortalama bant değişimi %21,08, sıralama korelasyonunun ortalaması 0,9277 — sıralama kararlı taraftır, duyarlı eksen altı ağırlığın birlikte oynamasıdır ve oynaklık bant eşiklerine (60/30) yakın noktalarda toplanır (kayıt: scripts/sensitivity/results-2026-09-05.md). Doygunluk çapaları ve bant eşikleri de oynatılır: her çapa ve her rampa ucu bağımsız olarak U[0,8, 1,2] ile çarpılır (±%20) ve her nokta ham okumalarından yeniden puanlanır; 30 ve 60 eşikleri ise bağımsız olarak ±5 puan kaydırılır. Tek başına çapalar ortalama noktanın bandını çekilişlerin %5,08'inde değiştirir ve sıralamayı neredeyse hiç bozmaz (Spearman 0,9917); tek başına eşikler %10,55 değiştirir — bunun neredeyse tamamı 60 çizgisindedir (%10,02; 30 çizgisinde %0,77) ve hiçbir skoru oynatmadan. Dört eksen birlikte %23,83 değiştirir (Spearman 0,9243); kayıt: scripts/sensitivity/results-2026-09-10.md. Ağırlıklardan türetilen yapısal sonuç: tüm tehlikeler tabana kırpıldığında skoru üç kalite sütunu tek başına taşır ve gürültüleri yaklaşık iki kat sert yansır (hava puanı başına ~0,5, yeşil başına ~0,3 yayımlanan puan). Dış doğrulama: EMI'nin raporladığı PM2.5, 161 bağımsız yer istasyonuna karşı r = 0,933, MAE = 3,61 µg/m³ vermiştir. Üç not: yüksek kirlilikte uydu kaynaklı girdi istasyonların altında okur (25–35 µg/m³ aralığında ortalama −8,45, 35 üzerinde −16,35; Balkan ve Anadolu kış inversiyon havzalarında yoğun — oradaki güven verici hava skorlarını ihtiyatla değerlendirin); eğrinin üst yarısı test edilmemiştir (59,4 çapası yakınında istasyon yok); karşılaştırma tam bağımsız değildir (kaynak istasyon verisiyle kalibredir). Güney Asya ve Afrika artık ayrı bir kontrolle ölçülüdür: OpenAQ'nun referans nitelikli 49 istasyonu (35'i Hindistan'da; 38'i 35 µg/m³ üzerinde okuyor — ABD/Avrupa kontrolünün erişemediği zemin). Toplandığı hâliyle r = 0,830, ρ = 0,802, ortalama sapma −1,78 µg/m³, MAE 12,14; saatlik beslemesi cihaz tavanına dayanmış istasyon-yılları düşülünce geriye 27 istasyon kalır ve r = 0,846, ρ = 0,885, sapma +0,71 µg/m³, MAE 10,09 okunur — seviyeye dair her ifade bu temizlenmiş okumaya dayanır, çünkü tavana dayanmış bir saat istasyonu şişirir, EMI'yi asla. 35 µg/m³ eşiğinde 49 istasyonun 45'inde aynı tarafta kalınır. Zayıf taraf seviyelerdir ve sapma burada iki yönlüdür: tek tek noktalarda fark toplandığı hâliyle ~63 µg/m³'e, temizlenmiş hâlde ~48 µg/m³'e ulaşır — dört Punjab istasyonunda EMI yaklaşık 25 yüksek, Delhi, Lahor ve Leknev'de 35 ile 63 arası düşük okur; birkaç yüz kilometre arayla iki yön, yani tek bir ölçek katsayısı bunu düzeltmez. İki sınır her rakamla birlikte gider: Hindistan yer gerçeği 2016–2022'yi kapsar, EMI'nin sunduğu on yılı değil; ve örneklemin 199 istasyon-yılının 70'i tavana dayanmış saatlik okumalar içerir. Bu kontrol 35 µg/m³ üzerinde sayısal aralık eklemez. Sıcaklık 44 şehir merkezinde YCEO ile (r = 0,688), yeşil 113 noktada MODIS NDVI ile (ρ = 0,676, r = 0,590) çapraz kontrol edilmiştir; farklar kavram ayrımlarıdır — EMI'nin yeşil ölçüsü tarımı dışlar, NDVI sayar. Sıcaklığın tehlike sınıfı da bu ölçüm sınırına bağlıdır: kesintisiz bir kent çekirdeğinin karşılaştırılacak daha serin bir komşuluğu yoktur; yeniden temellendirilmiş bir referans dış kontrolü geçerse sınıf miras alınmaz, yeniden tartışılır. Taşkın, model soyunu paylaşmayan iki bağımsız referansla karşılaştırılmıştır. Birincisi küresel bir arazi haritası — en yakın drenaja göre yükseklik, Copernicus 30 m yükseklik modelinden türetilmiş, CC0 — her taşkın sınıfından 100 olmak üzere 900 hücrede: drenajına 5 m içinde kalan alçak zemin, 4. sınıf ve üstü hücrelerin %82,2'sini, 0. sınıf hücrelerin %49,0'ını oluşturur; 33,2 puanlık fark (%95 GA 22,4–43,6), yani modelin "modelde ıslak" ile "modelde kuru" ayrımı araziyle uyumludur. Ama 0'ın üstündeki basamaklar araziyle sıralanmaz: Spearman ρ = −0,066 (n = 900); 1–8. sınıfların tümü drenajın 0,4–1,5 m üstünde, eğilimsiz. İkincisi İngiltere'nin resmî nehir taşkın ovası haritası (Environment Agency 2. ve 3. Taşkın Bölgeleri, Open Government Licence), kıyıdan uzak, yalnız nehir kaynaklı 240 noktada, her bölgeden 80: model yılda %1 olasılıklı taşkın ovasını (4. sınıf ve üstü) 3. bölgedeki 80 yerin 7'sinde işaretler — isabet oranı 0,087, kritik başarı indeksi 0,080 — ve 80'in 73'ünde 0. sınıf okur; haritalanmış ova dışında hiçbir noktayı işaretlemez (80'de 0), sorulan 297 iç noktanın 57'si (%19) modelin alanı dışındadır ve "ölçülmedi" olarak düşmüştür. Okuma: model yaklaşık 1 km'lik hücrelerde çalışır, "ıslak mı kuru mu" sorusunu çözer ama tekrarlanma basamağı parselin konumu değildir ve boyut eşiğinin altındaki nehirleri dışarıda bırakır — İngiltere'nin sık küçük nehir ağı tam o zemindir. 0. sınıf okuması bu yüzden "modelde tehlike yok" demektir, "taşkın ovası yok" değil; hata tek yönlüdür, eksik saptama yönünde: sıfır olmayan bir sınıf güvenilirdir, sıfır bir belge değildir. İki kontrol de önceden kayıtlı eşiğini geçemediği için taşkın sütunu hâlâ "bağımsız kontrol edilmedi" okur ve bu uyuşmazlık etiket değiştirilmeden yayımlanır (rapor: docs/validation/T91-FLOOD-VALIDATION.md). Yağış normali — suyun son ayların yanında gösterdiği, yangının yakıt terimi olarak puanladığı 1991–2020 aylık ortalama yağış — aynı pencere ve aynı istatistikle iki ölçüm kaydına karşı 345 katmanlı hücrede ölçülmüştür: CHIRPS v2.0 (CC0, ekvatorun 50° içinde) ve GPCC v2022 (CC BY 4.0, 50° ötesinde): r = 0,883, ρ = 0,932, MAE 17,3 mm/ay, sapma +9,7 mm/ay; 23,7 mm/ay yakıt eşiğinde %94,8, 5 mm/ay yağış sapması tabanında %96,8 uyum — önceden kayıtlı eşiğiyle tutarlı. Yeniden analiz ölçümlerden daha yağışlı okur: 50° içinde hafifçe (+4,5 mm/ay, n = 225), 50° ötesinde yaklaşık 20 mm/ay (+19,6, n = 120); yangının yakıt terimi 23,7 mm/ay'da doyduğu için bu seviye sapmasını soğurur ve eşikte ayrıştığı yerde biraz daha fazla yakıt okur, asla daha az. Bu kontrol gösterilen bir niceliği ve bir yangın girdisini sınar; suyun puanlanan üç teriminden ikisinin — kalıcı suya uzaklık ve yüzey suyu — kontrolü aşağıdadır, üçüncüsü olan toprak kuruluğu (sütunun 2/5'i) henüz kontrol edilmemiştir (rapor: docs/validation/T91-WATER-VALIDATION.md). Suyun kalıcı suya uzaklık terimi (sütunun 1/5'i), aynı 10 hektar tabanını kullanan bağımsız bir göl envanteri (HydroLAKES v1.0, CC BY 4.0) ve Natural Earth kıyı çizgisiyle (kamu malı), servis edilen uzaklığa göre katmanlanmış 400 hücrede (her 10 km'lik dilimden 40) karşılaştırılmıştır: tam örneklemde ρ = 0,722 (n = 400), örneklemenin yönlendirmediği 276 hücrede 0,595 — önceden kayıtlı tutarlılık eşiğinin ρ ≥ 0,70 yarısının altında. Sınama tek yönlüdür — envanterde nehir de 2016 sonrası baraj da yoktur, bu yüzden EMI'nin envanterden daha yakın okuması beklenen biçimdir (188 hücre); yalnız EMI'nin envanterden bir ızgara hücresinden (7,45 km) fazla uzak okuması ıska sayılır. Iska oranı tam örneklemde 0,135, örneklemenin yönlendirmediği 276 hücrede 0,196'dır — adil başlık ikincisidir, çünkü yönlendirilen 124 hücre yaklaşık 94 km içinde haritalanmış göl bulunmadığı için seçilmiştir ve yapı gereği ıskalayamaz — ve 0,196 önceden kayıtlı bulgu eşiğini (> 0,15) aşar: yönlendirilmemiş alt küme iki ölçütün ikisinde de kalır ve bu yayımlanmış bir uyuşmazlıktır. Iskaların en yakın 54 poligonunun tamamı doğal göldür, çoğu küçüktür (ortanca 0,38 km²; 54'ün 41'i 1 km² altında), neredeyse hepsi kurak iç bölgelerdedir (örneklenen 42 Avustralya hücresinden 13'ü); 10 km² ve üstündeki dördünden üçü kapalı havza tuz gölü, biri (Mar Chiquita) kıyı çizgisi güçlü dalgalanan tuzlu bir göldür. Böyle bir gölün on iki ayın tamamında su tutup tutmadığı — servis edilen terimin şartı budur — servis edilen sayının doğru olup olmadığını belirler ve bu kontrol "kalıcı değil" ile "görülmedi"yi ayıramaz. Göstermediği şey: deniz tarafı baştan sona uyumludur (65 kıyı hücresinin 0'ı ıskalar), yakın uç uyumludur (10 km altında servis edilen 40 hücrenin 40'ı) ve EMI'nin 100 km içinde kalıcı su bildirmediği 224 hücrenin 200'ünde envanter de 90 km içinde hiçbir şey bulmaz — alıkoyma gerçek yokluğa denk düşer. Bulgu etiket değiştirmez: kademe yalnız sütunun 4/5'ini taşıyan iki terimle hareket eder (rapor: docs/validation/T91-WATER-VALIDATION.md §2). Suyun durgun su terimi, bağımsız bir kara-su maskesine karşı — puanlanan surfaceWater terimi (sütunun 2/5'i), 1000 m içindeki zeminin 1984–2024 yüzey suyu görülme payı, 2024 yılına ait MODIS ve SRTM kara-su maskesiyle (MOD44W v061, açık paylaşımlı; 250 m'lik piksel başına tek bir yılın evet/hayır'ı 41 yıllık bir görülme ortalamasına karşı — çalıştırmadan önce belirtilmiş bir yapı farkı) her servis diliminden 60'ar olmak üzere 300 noktada karşılaştırılmıştır: Spearman ρ = 0,699, Pearson r = 0,575, MAE 0,162, ortalama sapma +0,083 (n = 300); iki taraf herhangi bir su bulunup bulunmadığında noktaların %78,3'ünde uyuşur (128 her ikisi, 107 hiçbiri, 63 yalnız EMI, 2 yalnız maske). Bu, önceden kayıtlı eşiklerin arasındadır — tutarlılık eşiğinin iki yarısının da altında (ρ ≥ 0,70 ve varlık uyuşması ≥ %85), bulgu eşiğinin (ρ < 0,50) üstünde — dolayısıyla ne doğrulama ne bulgudur: EMI'nin su bildirmediği yerde maske 60 noktanın 58'inde uyuşur, EMI'nin zeminin yarıdan fazlasını su olarak bildirdiği yerde maske 60'ın 54'ünde su görür ve uyuşmazlık aradadır — EMI'nin %5–50 su bildirdiği 120 noktanın 49'unda maske hiç su okumaz; bu, tek bir yılın maskesi ile 41 yıllık görülme payının yapı gereği farklı okumak zorunda olduğu zemindir ve bunu ancak aynı yapıda çok yıllı bir karşılaştırıcı çözer. Kademe hareket etmez; toprak kuruluğu (SMAP L3 toprak nemine karşı) ve alınan yağış (GPM IMERG'e karşı) kontrolleri beklemededir ve su hâlâ "bağımsız kontrol edilmedi" okur (rapor: docs/validation/T91-WATER-VALIDATION.md §3). Yangın havası klimatolojisi — servis edilen w girdisi, her yılın en yüksek günlük FWI değerinin 1991–2020 ortalaması — aynı indeksin ECMWF yerine MERRA-2 yeniden analiziyle zorlanmış ikinci bir hesabıyla (GFWED v2.0; kullanım koşulları doğrulama raporunda kayıtlıdır) on servis diliminin tamamına yayılmış 939 hücrede karşılaştırılmıştır: ρ = 0,903, r = 0,919, göreli sapma −0,135 (n = 939) — önceden kayıtlı eşikleriyle tutarlı (ρ ≥ 0,80, |sapma| ≤ %25). İki not başlıkla birlikte gider. Sapma bir kayma değil bir eğimdir: çiftlerden geçen doğru servis ≈ 10,6 + 0,66 × referans okur; servis edilen değer alt-orta dilimlerde referansın %7–17 üstünde, 7. dilimden yukarıda %12–25 altındadır — en güçlü yangın havası zemininde referansın yaklaşık dörtte üçü. İki taraf 82,74 doygunluk çapasında hücrelerin %85,5'inde uyuşur; bu uyumu iki tarafın da çapanın altına koyduğu 691 hücre taşır: referans 244 hücreyi doymuş sayarken servis edilen ölçek 116'sını sayar, yani çapa MERRA-2 zorlamalı bir indeksin karşılayacağından daha az yerde karşılanır. Çerçeve yalnız bitki örtülü karadır: referans çöl, buz ve soğuk çöl zemininde hiçbir şey hesaplamaz — bu, servis edilen en üst dilimden çekilen 100 hücrenin 42'sini düşürmüştür — ve çöllerdeki yangın havasını burada hiçbir şey kontrol etmez. İki taraf da yeniden analizdir: bu kontrol bir sıralamayı teyit eder ve zorlamaya bağımlılığı sınırlar; gözlenmiş yangın havasına karşı hata ölçmez. Yanma teyidi — iki CLMS kaydı, 2019–2025'in en kötü tam takvim yılı ve Temmuz 2018'den bu yana en kötü tek ay — Amerika Birleşik Devletleri (MTBS; kullanım koşulları doğrulama raporunda kayıtlıdır) ve Kanada'da (Kanada Ulusal Yangın Veritabanı, Open Government Licence – Canada) kurumlarca haritalanmış yangın sınırlarıyla, üçte biri teyitli, üçte biri kısmi, üçte biri sıfır okuyan zeminden çekilmiş 485 hücrede ve iki taraf da 0,1083 teyit eşiğinde okunarak karşılaştırılmıştır. Eşiklerin önceden kayıtlandığı terim olan yıllık kayıt: isabet oranı 0,950, yanlış alarm oranı 0,252, kritik başarı indeksi 0,720, ρ = 0,766 (n = 485) — eşikleriyle tutarlı (H ≥ 0,70, ρ ≥ 0,60). Yanında nitelendirme olarak bildirilen aylık kayıt: isabet oranı 0,839, yanlış alarm oranı 0,208, kritik başarı indeksi 0,688, ρ = 0,765 (n = 483; tarihsiz bir sınır yüzünden iki Kanada hücresi düşer) — aynı eşikleri karşılar. Uyuşmazlık tek yönlüdür: haritalanmış bir sınır, 300 m'lik yanmış piksel ürününün işaretlemediği yanmamış zemini de kapsar; bu yüzden iki tarafın da yangın gördüğü 113 hücrede servis edilen oran sınır oranının ortanca 0,54'ünü okur ve altı ıskanın tamamı hücrenin %11,8–21,3'ünde haritalanıp %4,4–10,8'inde servis edilen büyük boreal yangınlardır — aynı oranın alt kuyruğu. Yanma çarpanı eksik okur, uydurmaz: servis edilen kaydın tam 0 okuduğu hiçbir hücre eşiğin üstünde haritalanmış sınır taşımaz; yanlış alarmlar ise haritalama tabanının (500–1.000 akre) altında ya da arşivlerin tutmadığı 2025 yılında yanan tarım ve boreal zemindedir. Başlıktaki isabet oranı üç servis sınıfını eşit tartar; sınıflar zemindeki ölçülmüş paylarına göre yeniden tartıldığında isabet oranı ABD için 0,60, Kanada için 0,81 tahminidir (girdi: rapordaki sınıf payları ve sınıf başına oranlar) — ABD'de bulgu eşiğini aşar, tutarlılık eşiğini aşmaz. Yalnız iki ülke: Akdeniz, tropik ve savan yangın rejimleri hakkında burada hiçbir şey söylenmez. Üçüncü yanma kaydı — MODIS, 2001–2018 aralığının en kötü tam takvim yılı — aynı sınırlarla aynı 2001–2018 penceresinde 439 hücrede, iki taraf da kendi 0,5915 eşiğinde okunarak karşılaştırılmıştır: isabet oranı 0,968, yanlış alarm oranı 0,124, ρ = 0,868 (n = 439) — aynı eşiklerle tutarlı; servis-teyitli sınıf tasarımının gerisinde kalmıştır (166 hücre yerine 105, çünkü bir yılda %59,15'in üstünde yanan zemin adayların %2'sinden azdır ve sorgu bütçesi tükenmiştir — isabet oranını yalnız düşürebilecek, yükseltemeyecek bir eksik) ve iki tarafın da uyuştuğu yerde servis edilen oran sınır oranının ortanca 0,94'ünü okur; yani yaklaşık yarı eksik okuma bu kayda değil CLMS kaydına aittir. Yangın sütununun kademesi: çarpım biçimli bir sütun için yürürlükteki kural gereği yangın, yangın havası kontrolüyle teyitli (corroborated) okur — bağımsız bir veri kümesi sıralamayı teyit eder; yakıt çarpanının yağış kontrolü ve yanma çarpanının sınır kontrolleri onun yanında durur. measured kademesine yalnız üç çarpanın üçü de gözleme karşı ölçülmüş bir hata taşıdığında ulaşılır; yanma çarpanının üç kaydı ve yakıt çarpanının yağışı artık tek tek kontrol edilmiştir, geriye yangın havası kalır — o kontrolün iki tarafı da yeniden analizdir — bu yüzden hata büyüklüğü iddia edilmez; sütunun güven cümlesi hangi veri kümeleriyle karşılaştırıldığını ve nerede daha düşük okuduğunu söyler (rapor: docs/validation/T91-FIRE-VALIDATION.md).

Bilinen sınırlar: EMI deprem, kıyı veya fırtına kabarması taşkını, rüzgâr ya da heyelan tehlikesini modellemez; bunlar ayrıca kontrol edilmelidir. Taşkın sütunu yalnız nehir taşkınıdır ve WRI Aqueduct Floods (2020) haritalarını okur: katman artefakt hücrelerinin %99,61'ine (1.222.281.444 hücrenin 1.217.497.813'üne) cevap verir, çünkü artefaktın hücre kümesi modelin kendi alanını izler: merkez pikseli nehir taşkını modelince kapsanan her res-9 hücrenin bir kaydı vardır, ayrıca daha önce başka bir ölçüm taşıyan her hücrenin. Kaynağın hiç veri yayımlamadığı %0,39'da — başka bir ölçüm için taşınan hücreler — sütun "ölçülmedi" okur; modelin dışındaki açık okyanus ve buzun kaydı da skoru da yoktur. Yokluk güvenlik demek değildir. Her hücrenin sınıfı, hücre merkezinin bir piksel çevresindeki en kötü modellenmiş pikselden okunur (3×3 komşuluk maksimumu): modelin kuru bıraktığı bir nehir yatağı üzerindeki hücre kıyısının sınıfını alır; bir hücrenin kaydı olup olmadığına ise yalnız merkez piksel karar verir, komşuluk açık denizde kayıt üretmez. Merdiven 0–8 arası dokuz basamaktır: bir basamak modellenmiş taşkını olmayan zemine, sekizi dönüş periyotlarına aittir; kaynağın yayımladığı dokuzuncu periyot — 2 yıllık — rasteri modellenmiş karada hiç baskın taşımadığı için dışarıda kalır. Skor, 30 yıllık oturma süresinde en az bir modellenmiş taşkın olasılığı arttıkça düşer: 8. sınıf (beş yılda bir su basan ova) 0, modelin hiçbir dönüş periyodunda 10 cm'ye ulaşan baskın bulmadığı zemin 100, 100 yıllık taşkın ovası %26 olasılıkla 74 okur. O 100 tek bir modelin nehir aramasına dair bir ifadedir ve birleştirme kuralı gereği skoru teyit eder, yükseltmez; "0" bir boşluk değil bir bulgudur. 0. sınıf "modelde tehlike yok" demektir, "taşkın ovası yok" değil: model yaklaşık 1 km'lik hücrelerde çalışır ve boyut eşiğinin altındaki nehirleri dışarıda bırakır; İngiltere'nin resmî nehir taşkın ovası haritasında yılda %1 olasılıklı bölgedeki 80 yerin 73'ü 0. sınıf okumuş, haritalanmış ova dışında hiçbir nokta işaretlenmemiştir — hata tek yönlüdür, eksik saptama yönünde. İki kaynak uyarısı her okumayla gelir: tehlike korumasızdır (setler, bentler ve pompalı drenaj modellenmez) ve hidroloji 1960–1999 klimatolojisidir — 1999'dan beri rejimi değişen bir yer önceki rejimiyle tarif edilir. Takasın iki yönü de haritada görünür: modellenmiş hücrelerin %11,01'i bir dönüş periyodunda ıslaktır (merkez piksel okumasında %6,81'di) ve 1.083.393.391 hücre 0. sınıf okur; nehir yatağı üzerinde olup yine de çoğunlukla kalıcı su olan 0. sınıf hücreler — yüzey suyu okuması da olan 199.191.942 sınıf-0 hücrenin 1.866.264'ü, %0,94 — yüksek güvenle yayımlanmaz. Çıplak çöl yeşil sütununda en kötü skoru almaz; sütun structural_zero gerekçesiyle alıkonur. Sıcaklık referansının su maskesi ve yükselti kontrolü yoktur; iki hata da yalnız sıcaklığı abartır. Sıcaklık birçok kıyı ve muson noktasında ölçülemez — üründeki en geniş kapsama boşluğu: anomali, noktada bulutsuz bir yaz kompoziti (MODIS 2019–2023) ve 23,74 km yarıçapta en az ~253 km² ölçülmüş komşuluk ister; yoksa anomali ölçülmemiştir — asla 0 değil — ve sütunun 0,60'ını taşıdığı için sıcaklık alıkonur: okuması ekranda durur, skora girmez. 51 şehirlik çapraz kontrolde beş merkez böyle düştü (Mumbai, Dakka, Cakarta, Ho Chi Minh Şehri, New York Midtown); Miami Beach'te de aynı. Yangının yanma teyidi 2001'e kadar geriye uzanır, daha öteye değil. Soruyu üç kayıt yanıtlar ve en güçlüsü teyidi verir: CLMS'in en kötü tam takvim yılı (2019–2025; 2018 ile 2026 eksik yıllardır, en kötü yılı taşıyamazlar), CLMS'in en kötü tek ayı (Temmuz 2018 – Şubat 2026) ve MODIS'in 2001–2018 aralığındaki en kötü tam takvim yılı. Her okuma kendi eşiğine bölünür — iki CLMS okuması için %10,83, MODIS okuması için %59,15 — ve eksik bir okuma karşılaştırmaya sıfır olarak girmez, karşılaştırmadan düşer; üçü de yoksa hiç teyit uygulanmaz. Haziran 2017'de büyük yangın geçiren Pedrógão Grande (Portekiz) ile Knysna (Güney Afrika) artık MODIS kaydı sayesinde skorun içindedir; Kasım 2018'deki Camp yangını ise aylık kayıtla içeridedir. Teyit çarpanı yalnız indirim yapabildiği için teyitsiz zemin, aynı hava ve yakıta sahip teyitli zeminden daha az tehlike yayımlar ve pencereler içinde sönümleme yoktur. MODIS kaydı kapalıdır ve yenilenmez — okunabilir arşivi biter, penceresi de CLMS kaydının açıldığı 2018'de kesilir — bu yüzden yalnızca teyit ekleyebilir, hiçbir teyidi geri alamaz; yakın tarihli yangınları CLMS izler ve tam yılları 2025'e, ayları Şubat 2026'ya kadar okur. Rapordaki yıl yıl yanma grafiği aynı MODIS serisinin tam 2001–2022 aralığıdır: gösterilir, puanlanmaz. PM2.5, 68°K/−58°G ötesinde ölçülmez ve terim temiz okumak yerine düşer. Skor başına belirsizlik üç kademeli bir doğrulama bandıdır; yalnız hava sayısal aralık taşır (25 µg/m³ altı MAE 3,35 / n=155; 25–35 arası MAE 10,31 / n=6) ve 35 üzerinde aralık verilmez: o zeminde artık istasyon çifti vardır (Güney Asya/Afrika kontrolü, 38 istasyon), ama tek tek noktalar iki yönde ~48 µg/m³'e varan farklar taşıdığı ve sayıyı taşıyacak temizlenmiş satır henüz oturmadığı için aralık yayımlanmaz; yokluk sayıyla doldurulmaz. Birkaç puanlanan terim tek bir yeniden analiz ailesinden gelir. Suyun kuruluk terimi, sıcaklığın aşırılık terimi ve yangının tek puanlanan teriminin iki çarpanı — yangın havası klimatolojisi ile yakıtın yerine geçen yağış normali — aynı soydan gelir: Copernicus ERA5-Land ve kendisi de ERA5'ten hesaplanan Copernicus CEMS Yangın Havası İndeksi. Bu ailedeki bir hata, bu üç sütunu birbirini götürmek yerine birlikte kaydırır; skorun yanında gösterilen son ayların sıcaklık ve yağış okumaları da aynı aileden gelir. Bu sütunların birbiriyle uyumu asla teyit sayılmaz: her sütun agreement alanını unavailable olarak yayımlar ve üründe hiçbir yerde denetlenmemiş bir nicelik sayıya çevrilmez. Bu ailenin dışında ölçülenler: havanın uydu kaynaklı yıllık PM2.5 ortalaması ve emisyon envanteri, yeşilin WorldCover ve GHS-BUILT-S örtüsü, suyun yüzey suyu payı ile kalıcı suya uzaklığı (EC JRC Global Surface Water; bu zincirdeki tek ERA5-Land girdisi deniz maskesidir), yangının CLMS ve MODIS yanma teyidi ve sıcaklık anomalisinin altındaki MODIS yüzey sıcaklığı. Taşkın iki listede de yoktur: açık küresel taşkın modelleri bu aileyle model soyunu paylaşır; bu yüzden taşkının dış kontrolleri, ikisiyle de soy paylaşmayan bir arazi haritası ve bir mühendislik taşkın ovası haritasıyla yapılmıştır. Bu not, dağıtımdaki v14 veri artefaktı (v14-planet-260910: hücre başına 21 bayt, 7 alan; 1.222.281.444 hücre) üzerindeki modeli tarif eder; esaslı bir yöntem değişikliği notun yeni bir sürümünü üretir ve not her zaman yürürlükteki yöntemi anlatır.