Experiments: what was tried, and what it cost

Most of what has been tried in this project did not ship. That is worth writing down for two reasons: the negative results are the map of where Sentinel-2's 10 m resolution limit actually sits, and every one of them cost real compute that nobody else needs to spend again.

This page is the register. It covers everything that was measured, whether it ended up in the pipeline or not. Every experiment has runnable code in the repository; nothing was deleted on failure.

How to read the verdicts:

• shipped : part of the current main workflow
• partial : useful measured signal, but not strong enough to promote
• rejected : underperformed the alternative or failed its control
• superseded: previously shipped, but since replaced

The register

Experiment	Verdict	One line
In-domain Pakistani training chips	shipped	Tripled large-array recall in Punjab, 0.18 to 0.55. The single biggest lever found.
Building prior from VIDA Open Buildings	shipped	Makes "no building nearby" a usable false-positive signal.
Quadrats as training data, not just correction	shipped	Became roofclf, now half the main workflow.
roofclf, a per-building classifier	shipped	0.879 AUC (0.834 within size band) on 27 quadrats where segmentation scores near chance.
SPPI as a corroborating second opinion	shipped	Zero-training spectral index; agreement with roofclf sets an internal floor on the atlas's headline figure.
Coverage ratio stratified by size and density	shipped	Replaced precision as the multiplier, and a flat ratio with a measured per-stratum one.
Placement-split precision and recall	shipped	Unpooling rooftop from ground-mount moved both, in opposite directions.
OSM and a recall-corrected total: combine, don't add	shipped	Mapped misses were counted twice; Germany's Best fell 4.8%, Pakistan's rose 0.4% net.
End-to-end validation against a complete register	shipped	Germany ran nationally at last: 99.75% of MaStR capacity covered, and segmentation recovers about a third of what it can see.
p_unmapped from a geolocated register	shipped	Replaced Germany's p_unmapped: 0.0 floor with chance-corrected register evidence, split by placement.
Recall measured against any candidate, not the same placement	shipped	Rooftop recall was understated up to 24x because a big array overruns its footprint and the finder gets labelled ground.
The module constant, measured against a register	shipped	France put DEFAULT_KWP_PER_M2_MODULE at 0.150 kWp/m2 against the assumed 0.180. Germany could not do this at all.
Reading a register back to the imagery epoch	shipped	Dated ODRE vintages remove a 53% inflation that comes purely from PV installed after the flight.
Localizing the detector by retraining on France	shipped	Rank correlation against the register 0.29 to 0.45, closing half the gap to Germany. Slope barely moved.
The 400 m2 floor, measured from outside	shipped	Recall climbs with size (rho +0.83) while the sub-metre control stays flat (-0.29). Retraining on French data does not move it.
v5 against v4 on a third country	shipped	Nigeria is in neither checkpoint's training data: pooled recall 0.565 to 0.728, strictly dominant, McNemar p=0.000. The v5 gain is not only French.
Localizing the detector on Vietnam	shipped	v9, trained with Vietnam's own OSM labels, against v5 on the held-out Central Highlands: pixel IoU 0.724 to 0.762, recall on arrays of 500 m2 and up tied (0.973 against 0.960). The localized checkpoint ships.
The 5 km2 ground cap, outside Germany	shipped	Vietnam's five largest mapped parks, 5.6 to 8.3 km2, are 92 to 99% detected. An evidence-gated exception keeps them: Verified 6,258 to 8,180 MWp.
Small leads on steep ground left out	shipped	A third of Yemen's leads stood on ground steeper than 12°, against 0.5% of its highland OSM solar. Leads under 1,000 m2 above 15° now leave the lead files (Yemen 355 of 2,288); larger ones are kept as steep_review, unless above 30° (any size, from 60 shadows the owner flagged in Wadi Hadramawt; 57 now go). Across twelve countries 10,098 of 156,170 leads go and 3,298 stay as steep_review; 2 of 22,370 candidates that found mapped solar would. The owner judged 153 drawn leads on imagery: none is solar (148 not, 5 unsure).
Open ground at low probability, and pans and ponds, left out of the leads	shipped	Open-ground leads below 0.9 peak probability are 0-1.3% OSM-mapped in most countries and hold a registered German unit 0-5% of the time (10-38% above 0.9); none of 18 drawn was solar; they leave the lead files at a cost of at most 1.5% of mapped area. JRC surface-water history removes pans and ponds where switched on (Pakistan, Egypt, Namibia); off where solar stands on water (India, Vietnam).
Very large outlines counted only once judged; a third of the leads for Silver	shipped	Outlines above 100,000 m2 get a lead file of their own and count towards capacity only once mapped or judged solar: India's Best 99,542 to 82,574 MWp (81 judged parks of 1,789 instead of 1,650 at a sampled share). Silver needs a third of a lead file's leads and blobs with a verdict, not every lead: the top third held 69% (Pakistan) and 87% (Yemen) of what was mapped later and 46% of Germany's register hits.
Pricing low-probability open ground apart in the precision table	partial	Built, not adopted: a third placement table raises the model's priced capacity in 12 of 13 countries (the confident ground outlines then get their own, higher mapped share) but moves Best by under 0.5% (South Africa -35.5 MWp). Owner decision.
Localizing the detector on India	rejected	v10, trained with India's own OSM labels, against v5 on the held-out Surat-Bharuch-Vadodara box: recall on arrays of 500 m2 and up 0.938 to 0.911, pixel IoU 0.686 to 0.663. v5 ships. India's labels are 84% ground-mount, a population v5 already covers.
Compose scenes ranked by cell coverage	shipped	Scenes covering the whole cell now rank before partial ones, and a whole band lost to failed reads now raises. Rewa's cell went from 87.4% to 0.0% empty; composites built earlier were not rebuilt.
Solar-thermal removed from the OSM labels	shipped	OSM tags CSP plants and water heaters with the same generator:source=solar. Morocco's pull was half thermal (1,697 of 3,440 features, the Noor CSP complex). Filtered for every country after India.
Mapped park outlines in Best get the Verified gate	shipped	An OSM outline swapped in for a model candidate is exempt from the merge cap. It is now always ground, and above 5 km2 it is credited at its detected share unless the ground screen keeps it, the rule Verified already applied. India's Best went from 75,856 to 72,271 MWp: Khavda's two outlines 3,789 to 796 MWp, and 101 park outlines VIDA had placed on rooftops moved to ground (model rooftop 4,109 to 843 MWp). Verified unchanged.
Widening India's compose away from buildings	shipped	Every cell within three of the building and label selection (+3,573) imaged on the same checkpoint. Best 72,271 to 72,705 MWp: Khavda's outlines imaged whole (+330), about 220 MWp of false positives on Himalayan slopes and snow edges, and a precision floor diluted by the new unmapped candidates (-140). Found unmapped parks in Rajasthan as leads. Verified unchanged.
Unmapped parks above the merge cap, priced by review	shipped	In India the 100,000 m2 cap was excluding solar parks OSM does not hold: 30 of the 30 largest new-lead outlines and 50 of 60 drawn by area from the rest are parks. Admitted at that share outside the Himalaya, and mapped outlines above 5 km2 now credited at their detected share in Verified instead of dropped: Best 72,705 to 99,542 MWp, Verified 67,593 to 70,523 (owner-reviewed).
India against its register	shipped	MNRE's state-wise register (MoSPI eSankhyiki), January 2026: ground-mounted Best 0.88 of registered utility PV (0.79 on an AC basis), rank correlation 0.95 over 19 states, 14 within a factor of 1.5. Rooftop 7.5%: household systems are below the sensor.
Localizing the detector on Egypt	shipped	v12, trained with Egypt's own OSM labels, against v5 on the held-out Cairo Governorate: recall on arrays of 500 m2 and up 0.043 to 0.217 (1 and 5 of 23), pixel IoU 0.066 to 0.062. The localized checkpoint ships, on a thin holdout (seven positive chips, four installations of margin).
Label cells cover a park's whole outline	shipped	compose imaged one cell per mapped solar feature, by its representative point, so a park crossing a cell edge was half imaged and the ground screen dropped it. Every cell an outline touches is now imaged. Egypt gained five cells and two parks: Verified 2,375 to 3,315 MWp.
Localizing the detector on Yemen	partial	v13, trained with Yemen's own OSM labels, against v5 on the held-out 'Amran Governorate: both find 0 of 25 arrays of 500 m2 and up, pixel IoU 0 for both. The tie ships v13 under the rule, on no evidence either way.
Retraining Yemen's detector on the mapped leads (v25)	shipped	v25, trained on the pull holding the 827 features the owner mapped from v13's leads. The 'Amran holdout is now one farm area of small arrays no model sees (6 of 170 against 2). Against all 2,191 mapped arrays of 400 m2 and up, at v13's pixel precision, it finds 0.152 against 0.128 (+58/-6), rooftop 0.46 against 0.41, and 0.067 against 0.031 of the ground arrays outside both models' training images (+8/-0), while marking half as many pixels. Verified 151.3 to 277.5 MWp, Best 324.7 to 457.3, almost all from the new mapping.
Yemen's mapped PV at Sentinel-2 resolution	rejected	1,697 of 1,892 labels are ~600 m2 solar irrigation arrays in Tihama farmland, mapped in 2023. Inside-to-ring reflectance contrast is 2% per band against 8 to 25% in Egypt and Vietnam, no offset explains it, and v13 lights up 2 of 1,466 mapped features nationally. Training them as masks failed. Revised 2026-10-23: the cause is the labels' geometry, not the sensor. The Tihama polygons are four-corner rectangles drawn around plots; the same size of array, traced as panels in 'Amran, reads 8 to 21% darker than its ring, and the Tihama rectangles do hold a PV-like pixel when read as bags (next row).
Farm ground PV from pixels, weak OSM labels as bags	shipped	A footprint-free pixel model (bands, SPPI, local contrast) for arrays beside fields, which roofclf cannot see (6.6% of 'Amran's array area lies on a footprint) and segmentation does not (0 of 283). On 306 owner-traced 'Amran arrays: held-out pixel AUC 0.983, precision 0.63 at 62% array recall (a floor: unmapped and post-imagery arrays). Tihama's oversized OSM rectangles train it as multiple-instance bags (held-out bag AUC 0.934), and 88 of 150 Tihama leads were confirmed and mapped. Built with a cropland-stratified calib-candidates --farmland mode. Since 2026-10-24 a standard field on every atlas, in Best inside its calibration domain; recalibrated the same day on seven farm areas ('Amran and six Tihama areas), each cell priced by the region it resembles in cropland, buildings and elevation: 169.1 MWp in Yemen (90% 77 to 327). One box around all areas put Tihama's ratio (1.19, against 'Amran's 0.17) on highland terraces and reached 3.5 GWp, above Yemen's import ceiling.
Farm ground PV on Pakistan's farm areas	rejected	groundclf fitted on Pakistan's first eight farm areas (four landscapes, mapped 2026-10-29) plus 35 building quadrats: held-out pixel AUC 0.81 to 0.84 where dozens of arrays are mapped, but object precision 2 to 5% at every threshold, and at the deployed one Jhang f5 (102 installations) and Swabi (57) show no array found. Ground arrays median 44 to 106 m2, against 'Amran's 581. Not scored nationally; Pakistan's farm ground PV stays unmeasured.
Localizing the detector on South Africa	shipped	v14, trained with South Africa's own OSM labels, against v5 on the held-out Western Cape (452 positive chips, 1,742 mapped polygons of 400 m2 and up): recall on arrays of 500 m2 and up 0.744 to 0.794, 1,000 m2 and up 0.708 to 0.858, pixel IoU 0.364 to 0.671. The first decisive localization win of the 2026-10 queue; v5 still finds slightly more arrays below 1,000 m2, at half the pixel agreement.
Localizing the detector on Morocco	partial	v11, trained with Morocco's own OSM labels (49 train chips, accepted by the owner below the chain's floor of 50), against v5 on the held-out Drâa-Tafilalet: recall on arrays of 500 m2 and up 0.994 to 0.992, pixel IoU 0.938 to 0.935. A tie, shipped to v11 under the rule; the holdout is three parks (Noor IV, Noor Erfoud, Noor Zagora) with no small arrays, so it cannot separate the checkpoints where they differ.
Localizing the detector on Poland	shipped	v15, trained with Poland's own OSM labels, against v5 on a held-out Lublin region (706 chips, 361 with PV): recall on arrays of 500 m2 and up 0.927 to 0.960, 1,000 m2 and up 0.826 to 0.917, pixel IoU 0.571 to 0.645. A checkpoint-picker bug (a killed run's last.ckpt) first selected the wrong epoch; caught before publication and fixed.
Localizing the detector on Namibia	shipped	v17, trained with Namibia's own OSM labels (mostly traced panel rows) plus South Africa's chips, against v5 on a held-out Erongo region (36 chips): recall on arrays of 500 m2 and up 0.766 to 0.952, 1,000 m2 and up 0.179 to 0.795, pixel IoU 0.456 to 0.662. It also reads the Cuvelai basin's seasonal pans as ground-mounted panels, about 54 MWp of Best.
Localizing the detector on Turkey	partial	v19, trained with Turkey's own OSM labels (mostly traced panel rows), against v5 on a held-out Gaziantep region (164 chips): recall on arrays of 500 m2 and up 0.970 to 0.968, 1,000 m2 and up 0.953 to 0.951, pixel IoU 0.548 to 0.581. A tie, shipped to v19 under the rule; the holdout's large, well-mapped parks barely separate the checkpoints.
Quadrat candidates from a mapped-solar heatmap	partial	OpenInfraMap's solar heatmap, rebuilt from the OSM pull per 1.5 km box and per 1,000 VIDA buildings, ranks boxes where a quadrat may already be close to complete. Against Germany's register (count completeness per municipality): 71% of 38 candidates sit in municipalities at least 30% complete, against 10% of all mapped boxes, and 21% against 3% at 50%; mapped density ranks completeness far more than adoption (rho 0.94 against 0.14, partly by construction, since both share the mapped count). A candidate generator, not a certificate: four in five picks sit in municipalities under 50% complete, and every one still needs the imagery sweep. In Pakistan only 404 of 263,442 boxes hold any mapped small PV, and four of the six picks are existing quadrats.
roofclf on French residential PV	rejected	0.627 AUC within size band against Pakistan's ~0.834. A 20 m2 array is a fifth of a Sentinel-2 pixel.
13x the roofclf labels, from OpenPVMapper	rejected	16,435 positives against 1,231 moves AUC by -0.005. The model scores 0.683 even on its own label source.
Authoritative footprints instead of VIDA	partial	VIDA misses 58% of French buildings. Fixing that is worth +0.027 AUC, which is 14% of the gap to Pakistan.
roofclf transfer to Germany	partial	A foreign model transfers (0.70-0.72) and it barely matters which country; in-domain German training is worth +0.08.
Pakistan's roofclf on Yemen's roofs	partial	Ranks Yemen's four building boxes with PV as well as Yemen's own fit (median AUC 0.809 against 0.813; pooled at equal shares 0.818, now roof-classifier --extra-table), but saturates on rural stone houses, roofless shells and greenhouse tunnels: 13 of the 24 windows it scores highest showed no panel on Esri imagery, every window with panels was a town. Ranking is not Yemen's problem; the level is. It drew ten PV-rich town boxes (calib-candidates --pv-rich). The first two mapped, Yafi' and Qa'tabah town (2026-10-29), measured it: AUC 0.79 and 0.89 on household roofs, but scores summing to 3.6 per PV roof (170 over 47) against 1.115 in Pakistan's quadrats, about 3.2 times the PV there (high-leaning: the boxes were picked on their scores; Qa'tabah, picked on the fewest, reads 2.5).
The sub-400 m2 estimator against a complete register	partial	Beaten by total roof area in Germany (23% vs 58% error), beats it 3.4x in Pakistan. Regime-specific, and a perfect national total hid 47% per-municipality error.
A register-calibrated German roofclf half	partial	52.37 GWp against a registered 54.29, but still behind a roof-area baseline (48.4% vs 37.8% error), so it is reported beside the atlas rather than inside it.
Register coordinates as roofclf labels	rejected	+0.023 building AUC, 15.5 points WORSE municipal error: the geolocated band is 6.9 of 49.0 GWp, so the estimand and the labels disagree.
Register counts as a known class prior	partial	Fixes the label problem (33.4% against 70.6%) but ties the roof-area baseline on representative municipalities. The 18.2% first result was a PV-dense sample.
Size-stratified adoption intensity	shipped	First estimator to beat the German roof-area baseline: 35.4% against 37.8%, using the 200-400 m2 band and no classifier.
Size-dependent OSM rooftop conversion	shipped	Ground reconciled at 0.91; rooftop was 2.5x over because large polygons are roof outlines (0.051 kWp/m2) not arrays (0.200). Germany Verified 38,508 -> 25,921 MWp.
RID as a control on the size relationship	partial	Exhaustive labels confirm the band but refute the mechanism, and price OSM's size bias at +0.729 against a true +0.349.
OpenPVMapper as an external reference	partial	91% of the register's sub-72 kWp capacity, but 0.67 recall against hand-mapped truth and a model output throughout.
SPPI alone as Germany's sub-400 estimator	rejected	Orders German roofs better than the trained classifier does (33.7% against 36.8% municipal error) and still loses to crediting roof area with no classifier at all (31.2%).
Dividing register p_unmapped by an OSM positive control	rejected	The control is contaminated by the same sub-30 kWp coordinate suppression it was meant to absorb.
OSM geometry dissolve and closest-match dedup	shipped	Nested plant/generator ways were double-counting real installations.
Recall correction and credible intervals	shipped	Turned a structural floor into an estimate with a stated interval.
Recall correction for roofclf, not just segmentation	shipped	The roofclf half was counting only the roofs it flagged; correcting it moved Best 16.6 to 18.3 GWp.
Quadrat-bootstrap uncertainty on the atlas	shipped	The headline figure now carries a 90% range.
Solar-glint corroboration	shipped	Calibrated likelihood ratios per size bucket, reward-only, never demoting.
Tile-major glint fetching	shipped	22x faster, numerically identical output.
Glint tile copies counted once	shipped	Overlapping tiles deliver one pass twice, and a glint seen in both had validated a pose on its own. Pakistan's pose survey: 346 to 336 fits, none of the rest moved.
France glint pose survey	shipped	Every OSM installation of 1,000 m2 or more: 1,277 of 6,337 get a pose (20.2%, Pakistan 23.2% at that size), now on the France atlas with each point's OSM id, name and place. No French control yet.
India glint pose survey	shipped	A spread sample of 2,000 installations of 1,000 m2 or more: 312 get a pose (15.6%, 18.1% weighted to the real size classes), every one over 50,000 m2 read in parts. A quarter of the points face east at low tilt on April to June dates; imagery mostly shows north-south rows there, so probably single-axis trackers, by inference and not yet reviewed by the owner. No Indian control yet.
A fresh Planetary Computer token for every read	shipped	The client cached a 45-minute token and replaced it only with under 60 s left, the cause of the silent scene loss in long glint pulls.
Pre-boom epoch check	shipped	A persistent 2021 signal demotes a candidate.
Vegetation veto	shipped	596 of 5,132 leads vetoed, and specifically the right ones.
Hard-negative mining for segmentation	shipped	Bi-temporal confirmed negatives cut false-positive pixels 28.6%.
Fraction-regression head	partial	Beats segmentation on small arrays by 23x in the residential quadrat, never promoted into a published number.
Time-series step detection below 400 m²	partial	First non-zero recall under 500 m², but the capacity claim failed its own control.
Glint as a direct detector	partial	About 9% on model leads against about 1% on random buildings.
Per-pixel glint amplitude trim	partial	Moves pixel IoU slightly; threshold gating by glint does not.
roofclf hard-negative retrain from known false positives	rejected	n=6 is not enough signal; oversampling trades overall skill for it.
Quadrat-supervised fraction retrain	rejected	A large in-sample win the holdout does not support.
Two-season 20-band stacking	rejected	No recall gain anywhere; slightly worse on large arrays.
Boom-window (2021 vs current) stacking	rejected	Inconclusive: training collapsed on 332 chips.
Sentinel-1 corner reflection	rejected	Backscatter enhancement indistinguishable from speckle.
Snow-cover contrast for small rooftop PV	rejected	Sound physics, wrong calendar: 0.096 usable scenes per commune-winter outside the Alps.
Cell-aggregate glint density	rejected	Wrong statistic for the question.
Missed-installation glint recovery	rejected	Control false-validation rate exceeded the recovery rate.
Roof-axis orientation prior	rejected	Flat concrete roofs do not constrain a tilt frame's azimuth.
Standard-pose matched filter	rejected	The densest pose bin covers under 10% of installations.
Glint PSF matched filter	rejected	The PSF is real and measured, but every filter variant scores below the aperture statistic it would replace.
Verified-negative false-spike rate	shipped	2.0% against 8.7-20.3% on model-negative controls, which reopens the spike-rate density estimator.
Super-resolution, three variants	rejected	No gain, and hallucination risk on a detection task.
Two-endmember spectral unmixing	rejected	0.659 AUC, worse than both SPPI and roofclf, with a 92x scale spread.
Epoch jump / step change as roofclf features	rejected	Measured at exactly zero effect, and worse in the reflectance variant.
Keeping more than the median composite	rejected	Per-pixel p10/p90/std off the same scenes the median already downloads: +0.0003 AUC, 16 of 30 folds, sign test p=0.57.
Bilinear resampling of the native-20 m bands	shipped	The model's biggest coefficients sit on bands replicated nearest-neighbour from 20 m. De-blocking them: +0.0041 AUC within size band, 20 of 30 folds, p=0.061.
Regression sharpening of the 20 m bands	rejected	Predicting SWIR from the visible bands costs 0.0047 AUC, 7 of 30 folds, p=0.008. The SWIR signal is not synthesisable.
Footprint-constrained spatial unmixing	rejected	Cuts reflectance error 67% on synthetic sub-pixel buildings and loses 0.0323 AUC within size band on real ones, 2 of 30 folds, p=0.000.
Top-of-atmosphere instead of Sen2Cor	rejected	An exact coin flip, 15 of 30 folds, p=1.00, at triple the fold-to-fold spread.
Temporal unmixing	rejected	PV pixels vary as much as PV-free roofs (ratio 1.02), so there is no damping to measure. The gain tracks background dynamism as predicted, at 0.0015 AUC.
Gradient boosting instead of the linear model	partial	Ranks WORSE (-0.023 within size band, 4 of 30 folds) and calibrates BETTER (per-quadrat rate error 0.031 to 0.017, 20 of 29 quadrats).
Footprint shape as model features	rejected	-0.0001 AUC, 14 of 30 folds, p=1.00. The 0.8593 that made it look best was a difference of medians.
Area-weighted training loss	rejected	Improves the area ratio it optimises (1.084 to 1.035) and loses on both ranking (-0.0111 AUC) and count calibration (0.0308 to 0.0395).
A pre-boom epoch as the roof's own control	partial	+0.0234 within size band over the SHIPPED read path (25 of 29 folds, p = 0.0002, placebo null) but only +0.0067 over the noise-reduced one (23 of 29, p = 0.002). The two levers are not additive, and the old epoch is not yet deep enough.
Within-footprint pixel distribution	rejected	Extremes and spread of the same pixels the zonal mean averages: -0.0006 to -0.0015 within size band, 11 to 12 of 29 folds, p = 0.27 to 0.46.
Per-stratum deployment thresholds	rejected	Equal precision per size/density stratum COSTS 3.0 to 9.2 points of recall at matched precision (4 to 7 of 25 quadrats, p = 0.001 to 0.029).
Per-stratum score recalibration, one global cut	rejected	The theoretically-correct version. Pooled +1.3 points of recall, median per-quadrat +0.0000, 12 of 24 quadrats, p = 1.00.
Post-hoc recalibration of roofclf	rejected	Recovers 38% of gradient boosting's calibration gain at p=0.069. No monotone map reaches it, so that gain is a reordering.
Medoid instead of band-wise median compositing	rejected	A real artifact, but one date's spectrum loses more to noise than it gains in coherence: -0.0074 AUC.
Context-relative spectral features	partial	Within-cell z-scores alone TIE the shipped model's ranking while doubling its rate error: ranking is relative, calibration is absolute.
Local-contrast brightness (plus_local_contrast)	rejected	Unmeasurable since 2026-08-09 through a NaN bug; fixed and measured at +0.0000 AUC, 13 of 30 folds.
Cell-level brightness as a feature	rejected	Predicted to help calibration; makes it significantly WORSE, closer in 7 of 29 quadrats, p=0.008.
Snow-covered roofs in Germany	partial	Answers France's opportunity objection at a 10.3% scene rate; the contrast itself is 6 scenes, bimodal, p=0.39.
The composite reducer: mean, not median	shipped-pending	+0.0286 AUC within size band, 25 of 30 folds, p=0.0001. Larger than every feature block in this register combined.
Multi-frame super-resolution	rejected	Sub-pixel diversity is 1.3 m, a quarter of what fusion needs; applying the shifts is WORSE than not.
SEN2SR single-image super-resolution	rejected	-0.0367 AUC within size band, 6 of 30 folds. Invented texture dilutes real contrast.
GlobalBuildingAtlas building height	rejected	d'=+0.651 alone and +0.0006 on top of the model: height is a proxy for footprint area.
The spectral SNR budget	shipped	Measured, not argued: the noise is roof heterogeneity at 20-40x the sensor's, and the linear spectral limit is already reached.
Local background conditioning	rejected	Cuts noise 23-43% and signal faster, at every scale from 31 m to 369 m, because PV adoption is spatially clustered.
PV-free neighbours as the background	rejected	Undoes the clustering (0.94x of raw separation, against 0.75x) but not the loss: visible and SWIR contrast is neighbourhood-level. -0.0001 to -0.0016 AUC, 10 to 13 of 29 folds.
Glint geometry as a scene-level SNR lever	rejected	The apparent gain is solar elevation: partialling it out leaves +0.025 (p=0.70).
Compositing in a high-sun window	rejected	Roofs are 21% brighter pre-monsoon and the contrast is unchanged (0.98x): d-prime is illumination-invariant.
Yard features for small ground-mount	partial	The building index brackets 98.5% of the population, but detection lands at 1-2% precision.
Yard-SPPI and roofclf AND-gate for ground-mount	rejected	The rooftop floor's construction does not transfer: 2% precision, and the best operating point turns the roofclf side off.
Parcel label for roofclf	shipped	Counting PV in the yard, not just on the roof. 80% of what it recovers turns out to be rooftop PV overhanging an undersized footprint, not ground-mount.
Yard feature block in the roofclf model	rejected	Loses to roof-only features even against the parcel label it was built for: 0.8712 against 0.8734.
SPPI as a roofclf input feature	rejected	0.8736 to 0.8734 AUC. It helps by disagreeing, not by being a column.
Glint-date imagery as a roofclf feature	rejected	Only 7-24% of rooftops can ever glint into a near-nadir view; size-controlled AUC moves 0.7875 to 0.7879.
Opportunity-normalised glint sensitivity	shipped	Sensitivity varies ~2x with opportunity inside a size bin; now modelled per target instead of pooled.
Glint-mined roofclf hard negatives	rejected	126 negatives at 1.2% contamination moved held-out AUC by 0.0003. Quantity, not quality, is the binding constraint.
OSM as a complete reference in Germany	rejected	3.6% complete by unit count; implied kWp/m² unstable by more than the constant itself.
A sub-400 m² level without quadrats	rejected	The same roofs in 2021 are not negatives: roofclf's 2021 score predicts today's PV at AUC 0.83 against 0.87.
Google's AlphaEarth satellite embeddings in roofclf	rejected	Adding the 2025 embedding: -0.0003 AUC, 14 of 29 folds, recall at precision 0.5 down 0.017. The pre-boom 2021 embedding does as well, so what little it adds is propensity.
AlphaEarth change as a ground-mount precision check	partial	2017-to-2025 change separates real ground arrays from their surroundings at AUC 0.96, and a blind review puts the real share of unmapped ground detections at about 0.15, where the precision table books 0.0. Not applied; needs your review.
AlphaEarth as a second >= 400 m2 detector	partial	Not better overall: lower precision where PV is mixed with farmland (0.65-0.70 against 0.83-0.88), but equal precision and higher rooftop recall (0.51 against 0.35) on the quadrats, and it finds different arrays: the union lifts quadrat rooftop recall to 0.60.
Probability-weighted capacity pricing	shipped	Pricing roof area by score instead of by a flag cuts the per-quadrat PV-area error from 0.386 to 0.094 below 400 m2 (20 of 29 quadrats). The default since 2026-10-25: Pakistan's Best 24.4 to 34.6 GWp, floor unchanged. Untested in sparse national cells; --weighting threshold restores the old count.
Prevalence estimators for prior shift	rejected	EM prior adjustment and adjusted counts collapse to zero in most quadrats: they differ by more than their base rate.
Gradient boosting, priced end to end	rejected	Its better per-quadrat adoption rate does not survive the capacity chain: 12 of 29 quadrats closer, p = 0.46.
The pre-boom epoch already on disk	partial	composite_1 covers 4,467 of 4,473 Pakistani cells, so no download is needed; +0.0024 within size band (20 of 29, p = 0.061) and nothing under probability pricing. Not deployed.
THOR at 8 px tokens instead of TerraMind's 16 px	partial	No pooled recall gain at matched precision: rooftop 400-5,000 m2 0.585 against 0.574 (p = 0.08), ground 0.928 against 0.933. Sharper outlines (pixel IoU 0.656 against 0.633) at 2.2x the training cost per batch. Not adopted.
THOR on TerraMind's own token grid	rejected	Ties v14 on rooftop recall (0.573 against 0.574), 1 point behind on ground. With 320 m SWIR tokens instead, it loses everywhere (rooftop 0.29).
Checkpoint selection by val mIoU	partial	v14's second-best epoch finds 0.621 of rooftop arrays against the shipped epoch's 0.565, at precision 0.717 against 0.736: epoch noise is as large as the backbone effect. A recall-aware selection rule is untested.
OlmoEarth embeddings as a >= 400 m2 detector	partial	At segmentation's false-alarm budget it matches segmentation's precision in both countries, but its recall does not transfer: far higher in Pakistan (quadrat rooftop 0.65 against 0.35), half in South Africa's held-out Western Cape (0.22 against 0.43, and 0.26 with Western Cape training added), where it also fires on substations. Not adopted: it does not transfer, a restrictive licence, ~0.5-1 TB per country.
One segmentation model pooled from every country (v18)	rejected	At matched pixel precision it finds fewer installations than v14 over 14 holdouts: rooftop 400-5,000 m2 0.474 against 0.573 (+144/-609, p < 1e-60), ground 0.910 against 0.916. It wins only Egypt (23 installations, its own training data) and loses in South Africa and France, whose share of the corpus it halved. As trained; longer training is untested.
Which training labels teach rooftop PV as background	partial	v14 over 10,473 training chips it never trained on, from 15 sources. Vietnam's chips hold about 10 unmapped PV roofs per mapped rooftop array and Germany's national chips about 16 (83% of registered rooftop units of 72 kWp or more sit on background pixels), against 0.1 to 0.3 in France, South Africa and Pakistan. Poland's labels are clean of it. Supports the harmful-label part of v18's loss for those two sources; a second pooled corpus burns their unlabelled roofs as ignore.
TerraMind's own input standardisation (v21)	rejected	Every checkpoint saw DN / 10000, never TerraMind's pretraining standardisation. A frozen-feature probe said that cost a lot (held-out token AP 0.469 against 0.737), but fine-tuned with v14's recipe otherwise unchanged it ties v14 on rooftop recall (0.575 against 0.574 at matched precision, p = 0.87) and is 0.6 points behind on ground. It converges in half the epochs; its one regional effect beyond epoch noise is a loss in France.
Training on arrays down to 100 m2 (v24)	rejected	v16 retrained with Pakistan's 4,048 mapped arrays of 100 to 400 m2 burned as PV instead of ignore. At the production threshold it finds more small rooftop arrays in the calibration areas (0.052 against 0.0375 at its later epoch) only because it flags more pixels everywhere (43% more there). At v16's pixel precision it finds 0.032 and 0.036 at its two saved epochs, against 0.0375, and nothing changes below 100 m2 (about 0.1%). Small rooftop PV stays roofclf's job.
PSF weights, buffered footprints, context and model-form blocks	rejected	Eleven more feature, weighting and training-set variants; none better than +0.0005 within size band.
External corroboration: nightlights and a wealth index	partial	VIIRS nightlights and Meta's Relative Wealth Index correlate with published capacity (r=0.76, r=0.66), but partial correlation shows most of that tracks the same building-density confound (0.54, 0.35). Kept as a plausibility citation, not promoted to a stratification input.
Unrestricted national roofclf capacity	superseded	18 to 37 GWp, rejected as miscalibrated; replaced by the domain-restricted estimate.
Low/Central/High/All-PV bracket atlas	superseded	Replaced by the two-tier evidence atlas, which sorts by standard of proof instead.
Out-of-domain AND-gate extrapolation	superseded	Published 2026-08-11 to 2026-08-15, then dropped: the one Best-estimate component not measured where it was applied, and unreviewable by eye under stale imagery. -62 MWp.
The Ceiling tier	superseded	A threshold change roughly doubled it with no new validation, so it stopped bounding anything.
National dashboard bundle	superseded	Replaced by plain per-artifact result pages; the CLI still works.

What worked, and why

In-domain training chips

Training on Germany alone gave 0.18 per-installation recall at or above 1,000 m² in Punjab. Adding 274 Punjabi chips, merged with scripts/merge_chip_index.py and oversampled twice so Germany's larger chip count does not swamp them, took that to 0.55. Nothing else tried has moved the domain gap comparably, which is the empirical argument for the mapping flywheel: more verified Pakistani installations are worth more than any architectural change tested here.

A per-building classifier instead of a better segmenter

The decisive finding of this project is that below roughly 400 m² the question has to change. Segmentation asks "where is the polygon", and at 10 m ground sampling distance a 50 m² rooftop array does not have one. roofclf asks "does this building carry PV", which turns an unresolvable localisation problem into a per-footprint classification problem with a strong size prior and reflectance features. On the 27 mapped quadrats it reaches 0.879 AUC, and 0.834 with roof size controlled for, where the segmentation raster scores close to chance on the same buildings. That gap, not any architectural change to the segmenter, is why the pipeline now runs two detectors.

Requiring two detectors to agree

SPPI is a zero-training spectral index from the literature. As a standalone capacity instrument it is unusable (an 18x scale spread across quadrats, 4.7x over-prediction in arid Quetta), and as an extra column in roofclf it does nothing at all (0.8736 to 0.8734 AUC). What it is good for is disagreeing: requiring roofclf and SPPI to agree lifts precision from 0.496 to 0.540 at matched recall, and the gain concentrates in exactly the low-adoption places where roofclf alone is known to over-predict. That is why agreement between the two sets the internal floor under the evidence atlas's headline figure, rather than either model on its own.

One estimator, applied to both halves

The segmentation half already corrects for missed installations: each detection represents 1/recall real installations within its size class. The roofclf half did not. Its coverage ratio was measured only over flagged roofs, effectively assigning zero capacity to PV on missed roofs, even though roofclf contributed roughly four-fifths of the published estimate.

Using the existing labels from 16 trusted quadrats, area recall was estimated as the share of mapped PV area on buildings flagged by roofclf, by roof-size bin and building-density stratum. It is 0.808 below 400 m² and 0.978 at or above 400 m², rising from 0.34 for the smallest PV-carrying roofs to 0.99 for the largest. Area recall is used because missing a 300 m² array matters far more to capacity than missing a 20 m² one.

The correction raised the two roofclf components from 6,372 to 7,890 MWp and from 7,031 to 7,189 MWp, increasing the published Best estimate from 16,609 to 18,280 MWp. The internal floor was left unchanged because it is defined by agreement between two independent detectors and should not extrapolate to installations neither observed.

Coverage ratio and recall are refit within the same bootstrap replicates because they share the same quadrats and labels. The correction also remains conservative, chiefly because mapping completeness is certified only for the imagery epoch, so some recent installations on unflagged roofs can still be missed.

OSM and a recall-corrected total: combine, don't add (2026-10-20)

Segmentation's recall-corrected figure divides every detection by the measured recall of its size class, so it already stands in for the installations the model missed. That recall is measured as the share of OpenStreetMap-mapped installations some detection matches. Best then added the mapped installations no detection matched, which are exactly those misses, so in expectation they were counted twice. Only the two countries with a recall correction are affected, Pakistan and Germany; every other atlas adds OSM to an uncorrected, detected-only figure, which is right as it stands.

Best now takes, per placement and size bin, the larger of the satellite total and the satellite-detected part plus the mapped misses: mapped misses fill in the satellite's estimate of its misses instead of stacking on it. The reduction comes out of the satellite figure's unlocated scale-up, never below what a cell actually detected, so the mapped installations stay in their own cells. An installation also counts as detected at the recall table's own 100 m, not the atlas's 30 m.

The same audit found a gap in the other direction in Pakistan. Inside the area where roofclf replaces segmentation on rooftops, a mapped rooftop array that segmentation had also detected was counted nowhere: not from the map, because a detection matched it; not by roofclf, which skips mapped buildings; and not by segmentation, whose rooftop figure is discarded there. 687 installations, 466 MWp, of which the per-cell floor had restored 36. They are now counted from the map.

Both atlases were rebuilt after a legacy rebuild (--legacy-osm-addition) reproduced the published data blocks exactly, so the moves are the fix and nothing else. Germany: Best 88,709 to 84,473 MWp (-4.8%), 1,571 MWp absorbed from the rooftop scale-up and 2,414 from ground-mount; it is still about 9.7 GW above its ~74.8 GW register, so this explains roughly a third of that excess. Pakistan: Best 24,330 to 24,407 MWp (+0.3%): 233 MWp absorbed from the ground-mount scale-up, more than offset by the restored rooftop arrays. Verified did not move in either, since it holds no satellite estimate.

A third gap, closed the same day: the sub-400 m² roofclf populations dropped every flagged roof within 30 m of any segmentation candidate, including rooftop candidates inside that same domain (whose figure Best discards) and 124 oversize blobs excluded from capacity altogether. They now dedup only against candidates whose figure Best uses (sub400_capacity.dedup_candidates; --legacy-seg-dedup reproduces the old run, and did so exactly before the new one was built). The count-based guess was ~300 MWp; measured, it is 31.6 MWp in the central estimate and 23.4 MWp in the agreement floor, because two-thirds of the 17,670 freed roofs sit next to an OSM-mapped installation and correctly stay excluded now that it is counted from the map. Pakistan Best 24,406.8 to 24,438.3 MWp, Verified 6,970.7 to 6,994.1 MWp. A guess from counts and a national average was a tenth right; measure before quoting.

The vegetation veto, and why the obvious version fails

Manual review of countryside leads found many green fields flagged as PV. Measuring NDVI on the composite the model actually read does not separate them: 150 suspect leads had a median NDVI of 0.10 there, statistically indistinguishable from confirmed PV's 0.04. The field a validator sees as green today was dark fallow, harvested or flooded paddy soil when the dry-season median was built. This is a season mismatch, not a spectral confusion the model could have avoided.

The instrument that does discriminate is the annual vegetation cycle: every crop field greens up at some point in the year and a panel never does. scripts/veg_annual_ndvi.py samples a year of scenes per lead and reports the 95^th percentile. A positive control on ten leads the free version had already flagged found eight crossing 0.3 within the year, confirming the catches are real vegetation.

External corroboration from nightlights and a wealth index

Two proxies that had never appeared anywhere in this codebase before -- VIIRS nighttime-lights radiance and Meta's Relative Wealth Index -- were correlated against the published atlas's per-cell Best estimate to ask a narrower question than validation: does detected PV track the same "built-up, electrified, and (for RWI) relatively wealthy" signal an independent, non-imagery-derived dataset would predict. Neither proxy sees solar panels; agreement is evidence the estimate is plausible, not evidence it is correct, and it is cited that way in Validity and limitations, not used as a calibration input.

scripts/build_pv_external_comparison.py computes log(1+radiance) vs. log(1+MWp) and, separately, RWI vs. log(1+MWp), per 0.1° cell, plus the same correlation after partialling out each cell's building footprint area (roof_area_m2) -- the obvious shared confound, since a bigger built-up cell is both brighter and has more roof to put PV on. It had run once, 2026-08-12, and was never registered here. Rerun 2026-08-26 against the then-current 18,826.7 MWp atlas (14% higher than the 16,441.4 MWp the first run saw): nightlights at Pearson 0.761 raw, 0.545 once roof area is controlled for; the wealth index at 0.661 raw, 0.345 controlled. Both partial correlations had barely moved since the first run (0.548 to 0.545, 0.373 to 0.345), which read at the time as a stable property of the estimate's geography.

Rerun 2026-10-21 against the current 24,438.3 MWp atlas, and that stability did not hold. Nightlights correlate at Pearson 0.755 raw (Spearman 0.750), dropping to 0.477 once roof area is controlled for (n=4,463 cells); the wealth index at 0.748 raw (Spearman 0.656), dropping to 0.449 (4,428 of 4,463 cells have RWI coverage). The raw nightlights figure is unchanged, but what it carries beyond roof area fell by 0.07 while the wealth index's rose by 0.10. The step is the 2026-09-21 area-weighted zonal refit and national rescoring, not the 2026-10-20 fixes: the same script on the 24,330.0 MWp atlas between them gives 0.473 and 0.447, within 0.004 of today. So the partial correlations describe one estimate's geography, not a fixed property of PV placement in Pakistan: the current estimate follows relative wealth more closely, and lit built-up area less, than the August one did. The reference-hex rank correlation rose over the same step (0.82 to 0.87; capacity map).

The honest reading is that both proxies carry some information about PV placement beyond "this cell has more buildings" -- the partial correlations are positive and well clear of zero -- but a majority share of the raw correlation is exactly that confound, not new information. That is consistent with Capacity density's own three failed attempts at finding any coarse per-cell proxy (candidate density, roofclf's predicted rate, SPPI agreement rate) for which quadrat-regime a cell resembles: neither nightlights nor RWI has been tried in that specific role, and this result is why they are not expected to do much better than the three proxies that already failed there. The same script also recomputes the reference-hex comparison already published on the capacity map via the same grid match, as an internal consistency check rather than a second citation of it.

A legally-complete register as a precision instrument (2026-09-01)

Germany's own accuracy check had a hole in it: with no instrument for p_unmapped, its calibration table shipped 0.0, pricing every un-mapped candidate at zero and holding est_mwp_cal to an OLS slope of 0.038 against the register.

MaStR closes it over part of the range. Coordinates are published only at or above 30 kWp — zero of the 4.17 million units below that carry one, a privacy policy rather than missing data, while the fill rate is 99.7% above the 72 kWp / 400 m² segmentation floor. So a registered installation's address point falling inside an unmapped candidate is direct evidence that the candidate is real, for exactly the population segmentation targets. Measured per placement, the term runs from 0.061 to 0.688 for rooftop and 0.000 to 0.246 for ground, chance-corrected against the same polygons displaced 500-1,000 m.

Two things generalise beyond Germany. The register is silent below 30 kWp, which is the roofclf domain, so this improves the instrument that was already working and does nothing for the one that needed help most — a register does not remove the need for mapped quadrats. And the natural refinement, dividing by an OSM-mapped positive control the way the glint inversion does, fails here: German OSM rooftop PV is the 3.6%-complete enthusiast-mapped tail, which skews below 30 kWp and so carries no coordinates, making the control contaminated by the very suppression it was meant to absorb. Full derivation and the rejected control: Validating against a complete register.

The recall denominator, found by fixing something else (2026-09-02)

Correcting Germany's p_unmapped moved est_mwp_cal the right way and pushed est_mwp_rc_roof from 0.262 to 3.11 against the register — from understating truth to overstating it threefold. Two errors had been cancelling, and removing one exposed the other.

The cause was not either hypothesis first written down. It was that derive_placement_tables restricted both sides of the recall measurement by placement: a rooftop reference installation only counted as found if the finding candidate was itself classified rooftop. Precision and recall are asymmetric here. mapped_frac asks "is this candidate real", so its corroboration must come from references of its own placement. Recall asks "was this real installation detected at all", and how postprocess labelled the candidate that found it is irrelevant.

Measured on Germany, rooftop recall against same-placement candidates versus against any candidate:

Bin	same placement	any candidate	factor
500-1k m²	0.128	0.167	1.3x
1k-5k	0.214	0.268	1.25x
5k-50k	0.096	0.693	7.3x
>50k	0.036	0.852	23.9x

The mechanism explains why the error grew with size: a large array overruns its imagery-derived VIDA footprint, building_overlap_frac collapses, and the candidate that correctly found a rooftop installation is classified ground_adjacent or no_building — the same footprint undersizing the parcel label exists to handle. 1/recall then inflated those candidates by up to the 20x clamp.

Pakistan is affected too, which only a complete register could have revealed: rooftop recall there moves 0.423 to 0.808 in the 5k-50k bin and 0.065 to 0.952 above 50k, and ground 0.107 to 0.417 in 500-1k. Its published figures come from the checked-in calibration table and do not move until that is deliberately re-derived, which needs the glint sample and calibration boxes. Direction: Pakistan's est_mwp_rc is overstated.

Two hypotheses were written down first and both were measured and refuted: that oversize rooftop reference features deflated the top bin (they recall at 0.841, no different from the rest), and that count-recall applied to area inflated the estimator (the area/count ratio is 1.01 to 1.08, negligible). Recorded because the wrong diagnosis was the plausible one.

A second register, and what it could measure that the first could not (2026-09-04)

Germany's register settled the question above the 400 m² floor and was structurally silent below it: MaStR publishes no coordinates under 30 kWp, which is roofclf's entire domain. France's register is censored in the same direction, below 36 kW, but two features of it reach where Germany could not.

It publishes dated year-end vintages. Because a small installation cannot be dated individually inside an aggregate row, the current snapshot cannot be filtered to a past epoch, but a past snapshot can simply be downloaded. Interpolating per-commune small-PV capacity between straddling year-ends puts the reference on the same day as the imagery a label was drawn against. That correction is worth 53%: the same fourteen communes read 0.230 kWp/m² against today's register and 0.150 against the register as it stood when the aerial imagery was flown. The difference is entirely PV France installed in between, and any comparison that skips the step measures growth and reports it as physics.

Fourteen communes were swept exhaustively by hand, on sub-metre imagery, with solar-thermal collectors tagged separately from PV. That makes the register's small-unit aggregate and the mapped sub-200 m² polygons the same population, and dividing one by the other measures DEFAULT_KWP_PER_M2_MODULE externally for the first time. It comes out at 0.150 kWp/m² against the 0.180 the project assumes, ratio 0.83, stable at 0.147 under the widest reasonable definition of what counts as PV. The German attempt at the same measurement failed and was recorded as a negative result: 3.6% OSM completeness and an implied constant spanning 0.02 to 0.99 on mapper convention.

Two findings fell out of it that were not the point of the exercise.

The below-floor share is not transferable, now measured twice rather than argued once. Germany's page inferred that from dispersion across municipalities. France settles it: 17.3% of all registered French PV capacity sits below 72 kWp, 31.1% of low-voltage-connected capacity, against Germany's 65.5% of rooftop. France's fleet is far more ground-mount-heavy, so the practical cost of the floor is genuinely smaller there. Two complete registers differ by a factor of two to four on the quantity this project once transferred as a constant.

Solar thermal is a systematic confusion for imagery-based detection, and it is invisible without the tag. 381 of 3,335 mapped features are hot-water collectors: black glazed rectangles on south-facing roofs producing no electricity. 137 OpenPVMapper polygons in these communes sit on one. Any validation whose ground truth does not separate them books that as a detection error or, worse, as capacity.

France also cannot do one thing Germany can. Its register carries no coordinates at any size, so the p_unmapped precision instrument has no French counterpart, and no rooftop/ground attribute, so every share is bracketed rather than stated.

Where the sub-400 m² instrument stops working (2026-09-04)

roofclf is the half of this project that has no external check, so France was the place to get one. It did not survive it.

Fitted on 13 exhaustively hand-mapped French communes (44,314 buildings, 1,231 with PV, leave-one-quadrat-out, parcel label), it reaches median fold AUC 0.710 and 0.627 within size band, against Pakistan's 0.857 and roughly 0.834. Building footprint size alone gets 0.673, so spectral reflectance is worth about 0.037 AUC here. At a threshold tuned for 50% precision it flags 16 buildings out of 44,314. Segmentation scores exactly 0.500 in all 13 folds.

The cause is not the model. The median mapped French installation is 20 m² against a 100 m² Sentinel-2 pixel, so the array is a fifth of a mixed pixel that is mostly roof tile. The control that isolates this is OpenPVMapper, which recalls the same installations at 0.67 from sub-metre aerial imagery with no size gradient across 0 to 400 m²: the annotations are findable, the sensor cannot find them.

This bounds roofclf rather than refuting it. Pakistani quadrats contain installations that are small relative to the 400 m² segmentation floor but still large relative to a 10 m pixel. French residential PV sits below even that, and the instrument degrades to little more than a size prior there. The practical consequence for the published Pakistani atlas is a caution, not a correction: the sparse-density stratum, already the least well-supported part of the coverage-ratio fit, is the one nearest this regime.

It also answers, negatively, the "independent test of the 400 m² floor" that had been left open. The floor is real, it is a sensor limit, and below roughly a pixel of module area there is no Sentinel-2 instrument here at all.

Localizing the detector: what French chips bought (2026-09-12)

France was the first country this project inferred zero-shot -- the v4 checkpoint was trained on Germany, Pakistan, Punjab and Gujarat and had never seen a French roof -- and the first where a complete register could score that decision both before and after retraining. Same 20,501 communes, same 93.0% capacity coverage, so the two runs are directly comparable.

estimator	zero-shot slope	retrained slope	zero-shot ρ	retrained ρ	Germany slope	Germany ρ
est_mwp_det	0.127	0.160	0.262	0.446	0.340	0.661
est_mwp_exp	0.162	0.170	0.290	0.449	0.388	0.656

The gain is in placement, not magnitude. Spearman rose 55% and closed about half the gap to in-domain Germany; slope moved 5 to 25% and total recovered capacity went from 26% to 34% of registered above-floor low-voltage capacity. Spearman is scale-free, so unlike slope it cannot be explained away by conversion constants or by France's proxy denominator (the register has no rooftop/ground field, so "above floor, low voltage" still contains BT-connected ground-mount). The zero-shot model was putting capacity in the wrong communes; the retrained one largely is not.

What changed underneath is more interesting than the totals. The candidate population did not simply grow, it changed shape: 13,419 candidates became 39,462, the median candidate shrank from 8,301 to 1,400 m², rooftop share doubled from 26.9% to 50.3%, blobs at or above 10,000 m² fell from 6,036 to 2,076, and total detected area went down, 364 to 255 km². polygonize_chips merges touching thresholded pixels with no upper bound, so a sheet of weak false positives becomes one multi-hectare "installation"; a zero-shot model tuned on Pakistani and German industrial roofs produced exactly that over France. Three times as many objects, six times smaller, with a third less total area, is the signature of a detector matched to a fleet of small rooftop arrays.

A prediction recorded before the baseline ran was half right and worth keeping as calibration on this kind of forecast: the slope band (0.15 to 0.35) was about right for the uncalibrated estimators but anchored on the wrong comparator, and the rank-correlation collapse -- flagged in advance as "the more interesting failure" -- was the thing that actually happened, at 0.29 against Germany's 0.66.

The caveat, stated because it limits the claim. France is 73.6% of the retrained corpus (18,577 chips against 6,661 from every other region combined), so this result confounds "French data is present" with "the corpus is 3.8x larger". A France-capped run of about 3,200 chips, matching Germany's share, would separate them and has not been run. The checkpoint (v5_combined_france, epoch 25, early-stopped at 33) is applied to France only; Pakistan and Germany keep v4 and their published figures are unaffected.

The detection floor, measured against sub-metre truth (2026-09-12)

The 400 m² floor was always an argument from the sensor, four Sentinel-2 pixels, checked against quadrats labelled off the same class of imagery. The fourteen French communes are swept on sub-metre IGN orthophotos, so for the first time a missed installation is provably a miss rather than an annotation gap. mapped_vs_earthpv measures recall per size bin against them.

Recall climbs monotonically with installation size, Spearman +0.83 (p = 0.042), from 0.000 below 20 m² to 0.095 in the 200 to 400 m² bin. OpenPVMapper, reading the same installations inside the same boundaries from sub-metre imagery, is flat: -0.29 (p = 0.58), recalling 0.58 to 0.80 in every bin. That contrast is the whole point of the measurement. A gradient present in a 10 m detector and absent in a sub-metre one, against identical truth, is the sensor rather than the mapper.

The finding survives retraining, across three models. v4 zero-shot (no French chips), v6 (3,201) and v5 (17,059) return pooled count recall of 0.0118, 0.0103 and 0.0118, with gradients of +0.89, +0.89 and +0.83, even though v5 tripled the national candidate count and cut median candidate area from 8,301 to 1,400 m². In these communes only 28 candidates appear at all, and the median distance from a 400 m²-plus mapped installation to the nearest one is 807 m.

This corroborates roofclf's French result from the segmentation side, by a fully independent route: the classifier found no signal in the pixels, and the segmenter finds no polygons in the same places.

What it does not settle. The 400 m²-plus bin holds 44 installations, recall 0.045 with a 95% Wilson interval of 0.013 to 0.151, far too thin to claim EarthPV fails above its own floor. These communes were chosen to be exhaustively mappable, not to hold large arrays. The transferable result is the gradient against the control, not the level in any one bin, and it bounds the sensor rather than the method: Pakistani arrays are small against the 400 m² floor but still large against a 100 m² pixel.

Authoritative footprints: the French cadastre against VIDA (2026-09-12)

roofclf reads VIDA Open Buildings, which is imagery-derived. That is the right choice where no authoritative footprint layer exists, which is the case across most of the project's target geography, but it is not free. Measured over the fourteen hand-mapped French communes against the DGFiP cadastre published through Etalab:

	Buildings	PV-bearing	Median footprint	AUC	Within size band
VIDA	44,314	1,231	163 m²	0.7103	0.6274
Cadastre	103,697	1,759	53 m²	0.7369	0.6548

VIDA finds 43% of the buildings and 70% of the PV-bearing ones. In Toussieu specifically it returns 672 footprints where the cadastre returns 3,366, with a median of 146 m² against 22 m², and 1.9x less total built area. Bad footprints degrade three things at once: the roof-area feature, which is the single strongest predictor; the zonal spectral means, which get averaged over the wrong pixels; and the label itself, which is mapped PV intersected with the footprint.

Fixing all three is worth +0.027 AUC within size band, closing 14% of the gap to Pakistan's 0.821. Real, worth adopting for future French work, and not remotely enough to make the instrument deployable. The other 86% is the sensor.

Two operational findings came out of this. Overture carries essentially the same French footprints (3,826 in the Toussieu bbox against the cadastre's 3,366 inside the boundary) but costs a global S3 scan at roughly 166 seconds per commune against the cadastre's two, and Overture prunes its release directory to the last two releases, so configs/aoi.yaml's pinned 2026-06-17.0 no longer exists and fails as "No files found", which reads like an empty area rather than an expired release. The cadastre is published per commune keyed by INSEE, which is already the unit the French quadrats use.

building_table gained an optional buildings override for this; the default keeps the VIDA fetch, so no existing caller changed.

Glint tile copies and token expiry (2026-10-22)

Both defects surfaced while validating the France pose survey's fetch against the old path on real data.

One pass, counted twice. Neighbouring MGRS tiles overlap by about 10 km, so one Sentinel-2 pass arrives as two to four items. Their granule sensing times differ by about 14 s. tile_scene_series_batch dedupes on the sensing minute, so the copies survived whenever they straddled a minute: the France pilot read 308 items for 126 dates. A glint seen in two copies then counted as two dates that agree on one panel plane, which is the validation criterion itself.

Re-analysed from cached series, after first confirming each published summary reproduced exactly:

• country2000: 290 to 281 fits.
• Second regional top-up: 33 to 32 fits.
• First top-up and the 500-target sensitivity study: unaffected. The 500-target study was pulled per target, so its curve and the rank multipliers stand.
• Surviving fits: none moved.

glint.annotate_spikes now collapses rows sharing a day and a view azimuth, so every analysis path is safe, including series pulled the old way. tile_scene_series_batch( dedupe_items=True) avoids downloading the copies at all: each target is read from the one item whose footprint holds it deepest. The default fetch is unchanged, so earlier pulls still reproduce.

The 662-candidate glint sample behind Pakistan's p_unmapped was pulled through the same path, but its series were never cached, so it cannot be checked without a re-pull.

The token. Search signs every asset URL with a cached 45-minute SAS token, and planetary_computer replaces it only once it has under 60 s left. A tile group read late, or for longer than the token had left, failed at open, and the guards turned each failure into a silently missing scene. glint._asset_href now re-signs the bare URL at the moment of use, and a deliberately stale href opens.

Small leads on steep ground (2026-10-31)

The owner flagged Yemen's released leads: shadows on steep ground, where shadows from neighbouring slopes meet, and road junctions. The 2,288 leads were set against Copernicus GLO-90 slope at each lead's representative point, with Yemen's 3,229 mapped solar polygons as the real-solar reference.

• Steep ground separates. 33% of leads stood above 12° against 0.3 to 0.4% of mapped solar. Most mapped solar is in flat Tihama, so the comparison was repeated in the highlands alone (above 1,000 m): 32% of leads against 0.5% of 565 mapped polygons.
• What the steep leads are. 9 of 10 drawn at random were not solar on Esri and the composite: shaded hillsides, cliff edges, the shadow below an escarpment, gullies. Many sit beside composite pixels masked in every scene, which is terrain shadow the scene classification removed (331 leads within 30 m of one, 6 mapped polygons).
• Roads are rarer and higher ranked. 50 leads lie 30% or more within 8 m of a paved OSM road, mostly rooftop and ground-adjacent leads the building prior lifts to the top of the file. The first lead in the file is one of them and looks like a real rooftop array built after Esri's 2024 capture, so a flat road buffer would delete the best lead. No road rule was adopted.

The rule (owner): a lead under 1,000 m² on ground steeper than 15° is left out; a larger one stays, marked steep_review, because a solar farm can stand on a mountainside and a mapper has to look at it (export.steep_lead_screen, applied by scripts/release_leads.py and earthpv export).

Validation (scripts/validate_steep_lead_screen.py, results/leads_steep_screen/). In Yemen the rule removes 355 leads and keeps 319 as steep_review. Of the solar known to be real, 2 of 3,252 mapped polygons (0.04% of mapped area) and 0 of the 12 model candidates that found mapped solar fall under it. A random 24 of the removed leads and 20 of the kept steep leads were judged on the four-panel sheets (composite, 10 m detail, false colour, Esri): no solar among either; 20 removed leads were shaded slopes, cliff edges and gullies and 4 could not be settled (no or cloudy Esri, a village edge). Across all twelve rebuilt lead files (summary.csv) the rule removes 6,306 of 156,170 leads, from 0.02% in Poland to 5% in India and 15.5% in Yemen, and keeps 7,091 as steep_review. Of the solar known to be real it would remove 785 of 544,641 mapped polygons (659 of them small German rooftop outlines on hillside villages; at most 0.04% of any country's mapped area) and 2 of the 22,124 candidates that found mapped solar (one each in Germany and South Africa). Six of India's removed leads miss the rule when re-measured from the released file, all by rounding: five outlines just under 1,000 m² are stored as 1,000.0, and one point re-read at 7 decimals lands on the neighbouring 90 m pixel. A further 65 removed and 44 kept leads drawn from the other eleven countries were judged the same way: none was solar. The owner then judged all 153 on the same sheets (verdicts_owner.csv, 2026-11-01): none is solar. Of 89 removed leads, 86 are not solar and 3 could not be settled (a hillside behind a Red Sea resort, a Yemeni village edge with greenhouse tunnels, a South African mountain shadow or mine wall); of 64 kept steep_review leads, 62 are not solar and 2 could not be settled (a building in a Polish forest clearing, a German farmstead whose barn roof could carry panels). With no solar among 89, the solar share of the removed leads is below 3.3% at one-sided 95% confidence, for the sample's mix of countries (India, with most of the removed leads, gave 8). The assistant's own verdicts (verdicts_assistant.csv) agree on 141 of 153; of the other 12, 11 are leads the assistant left unsure and the owner judged not solar, and neither reviewer called any lead solar. The larger steep leads are almost all false alarms too (62 of 64), so most of the 7,091 steep_review leads are likely not solar; keeping them in front of a person is the owner's choice, so that a mountainside solar farm is not lost.

The owner's flagged shadows and a 30° cap (2026-11-01). The owner then sent 60 leads from the first release that they had judged shadow on steep ground, all in Wadi Hadramawt (data/yemen-false-postivies-shadow.geojson; per lead in owner_flagged_yemen.csv). The 15° rule had already removed 38 of them (median 35°). Another 20 were 1,000 m² or larger and so kept as steep_review, on ground of 24° to 46° (median 37°): cliffs, with 19 of the 20 above 30°. The last two stand on flat ground (6° and 11° at the lead's point) with 32° and 19° within 90 m: shadow thrown onto the valley floor at the foot of a cliff.

Leads above 30° are now left out whatever their size (owner; export.STEEP_CAP_DEG). On solar known to be real the cap costs almost nothing. None of the 13,344 model outlines of 1,000 m² or more that found mapped solar in Germany, India, Vietnam and Poland stands above 30° (these are the only countries with any such outline above 15°, and only one, in India, is above 25°). Four mapped polygons of that size in the twelve countries do: three in Germany, one in India, 9,186 m² together. The 23 kept steep_review leads above 30° that the owner had judged were all not solar. The cap removes 3,792 of the 7,091 steep_review leads (India 3,056, South Africa 245, Pakistan 146, Germany 124, Yemen 117) and leaves 3,298. A cap at 25° catches the same 19 of the owner's 20 for about 1,300 more leads, one found-mapped outline and five mapped polygons; a cap at 35° catches only 13. The validation re-run under both rules (results/leads_steep_screen/cap30/summary.csv) puts 789 of 545,468 mapped polygons under them (785 small, 4 larger above 30°), at most 0.03% of any country's mapped area, and 2 of 22,370 candidates that found mapped solar. With the cap the twelve files hold 145,826 leads. Yemen's rebuilt file also reads the OpenStreetMap pull of 2026-10-31: 246 of its leads had been mapped since 2026-10-24 and left it as well (1,570 leads; the candidates that found mapped solar there rose from 12 to 258).

The cliff-foot pair was left alone. A slope read at one point cannot see a cliff beside the outline; the owner declined a neighbourhood test for now. Leaving out small leads with ground steeper than 25° within 90 m would catch one of the two and 65 more Yemeni leads (3.4%), but also 0.2% of Yemen's mapped solar polygons: farms on valley floors under the cliffs. At 15° within 90 m, 13.6% of the leads go against 0.9% of mapped solar.

Open ground at low probability, and pans and ponds (2026-11-01)

After a survey of what other PV maps do against false positives (masks for water, snow, roads and slope; a small classifier on each detected outline; hard negatives; human review: Microsoft's India solar farms, Global Renewables Watch, Zhang et al. 2022, Kruitwagen et al. 2021, TransitionZero), two cheap checks were measured on all 13 published countries (results/fp_filter_survey/, scripts there).

Peak probability. Model outlines were re-cut from the probability rasters as postprocess.polygonize_chips draws them (reproducing every published model-geometry candidate exactly), because an outline replaced by its OSM polygon carries the highest probability of every fragment it absorbed. The share of outlines within 100 m of mapped OSM solar is flat from 0.3 to 0.9 inside every placement and size bin and rises only above 0.9 (pooled over countries, ground 5k-50k m²: 0.04-0.06 below 0.9, 0.25 above), so a plain cut at 0.4 or 0.5 shortens the lead list without cleaning it. Germany's register, which needs no OSM, separates by placement: unmatched no_building outlines below 0.9 hold a registered unit (any of 30 kWp and more, or ground) 0-5% of the time in every size bin against 10-38% above 0.9 (chance by displacement 0.1-0.2%), while ground_adjacent outlines below 0.9 hold one 15-35% of the time (rooftop arrays overrunning an undersized VIDA footprint). The rule no_building and peak below 0.9 has an OSM mapped share of 0-1.3% in most countries (4-7% in France, Germany, Poland) against 2-53% for the no_building outlines it keeps; dropping it loses at most 1.5% of the mapped area the model finds (France; Pakistan 6 of 9,920 installations, India 6 of 3,576) and none of India's 81 owner-confirmed parks above the merge cap; 18 of 18 random outlines of 1,000 m² or more (Pakistan, South Africa) were not solar on Esri (Claude's reading). It removes 35-60% of most countries' leads. Adopted for the lead files (export.low_confidence_rule), at any size.

Surface-water history. JRC Global Surface Water (1984-2020, 30 m; surface_water.py): a non-rooftop lead at least a quarter ever wet. Clean in Namibia (197 leads, 21 of the 67 MWp its leads carry, none of 57 matched real outlines), Egypt (93 leads, none of 37 matched; 9 of 9 flagged leads checked on Esri not solar: salt flats, wetland, ponds) and mostly in Pakistan (477 leads; 9 of 4,916 matched, 4.5% of mapped ground area: a Karachi-coast plant and one near Tarbela). Harmful in India (7.6% of mapped ground area: Khavda's salt marsh, Omkareshwar and Ramagundam floating) and Vietnam (17%; 3 of 9 flagged leads were real pond arrays). Rooftop flags are noise over dark city roofs. Adopted per country (lead_water_screen, on for Pakistan, Egypt, Namibia). In Namibia's Cuvelai the two screens together remove 26.5 of the 53.7 MWp of leads; 18 MWp of confident, never-wet pans remain.

Pricing low-probability open ground apart (2026-11-01)

The lead rule as a precision-table split: no_building candidates below 0.9 peak probability get their own placement table (capacity_calibration.LOWCONF_GROUP; calibrate-candidates --confidence-split 0.9, or scripts/split_calibration_tables.py, which splits the ground bins of a checked-in table keeping its recall rows and other evidence). Their own mapped share is about zero, so their credit falls (South Africa 164 to 13 MWp, Germany 706 to 162 with register p_unmapped measured per group), but the confident ground outlines are then priced at their own, higher mapped share, and they are the larger ones: the model's priced capacity RISES in 12 of 13 countries (India +875, South Africa +616 or 20%, Egypt +152, Germany +195 MWp; France -11). Through the atlas, where OSM and the per-cell Verified floor absorb most of it, Best falls slightly: South Africa 9,755.6 to 9,720.1, Egypt 3,378.6 to 3,376.5, Namibia 530.2 to 528.2 MWp (controls reproduce each published page exactly). Not adopted: no published table carries it; off by default.

Very large outlines need a verdict, and a third of the leads is enough (2026-11-01)

The owner saw many false positives among the small leads and asked for two changes: the very large outlines in a lead file of their own, never counted towards capacity until they are mapped in OpenStreetMap or a person has judged them on Sentinel-2 imagery; and a Silver rule asking for a reasonable share of the leads and blobs mapped, not every lead.

Blobs. An outline above the 100,000 m² merge cap (oversize) now goes to <country>_pv_blob_leads.geojson, largest first (release_leads.split_blobs; the pv-new-leads-2026-11-01b files were split in place with --resplit, same leads, same date): 100,556 leads and 3,127 blobs over twelve countries, France's 29 already apart. What a blob is varies by country: 81 of the 90 the owner judged in India were parks, 41% of Germany's 581 hold a registered unit, 3 of Pakistan's 121 of July have been mapped since, none of Yemen's 29. The cap already kept blobs out of capacity everywhere but India, which since 2026-10-29 admitted 1,650 of them at its review's census and sample shares. density.load_oversize_admission now admits only outlines with a pv verdict, each whole: India's 81 judged parks, 185.0 of the 569.4 km² credited before. India's segmentation estimate falls from 60,935 to 41,716 MWp, Best from 99,542 (90% 80,394 to 128,501) to 82,574 MWp (90% 70,770 to 107,295), Verified unchanged at 70,523; the 1,699 unjudged blobs are queued for a verdict each (oversize_review.py draw --all). Against MNRE's register ground-mounted Best goes from 0.88 to 0.73 of registered capacity; if the review sample is a guide, judging the remaining parks would close most of that. No other country's figures move.

Judged, 2026-11-02. The review became a default step: oversize_review.py draw queues every blob the lead screens keep (--all is gone), and data/oversize_review/index.html lists 3,050 outlines over thirteen countries. The owner judged all of India's 1,550 in one sitting: 1,346 parks, 204 not, none unsure, 87% against the sample's 0.879. India now counts 1,427 parks, 577.1 km². Best's per-cell floor becomes Verified plus the judged parks in the cell, because at Verified alone it hid 3,128 MWp of them. India's segmentation estimate goes from 41,716 to 61,324 MWp, Best from 82,574 to 103,037 MWp (90% 83,676 to 132,969), Verified unchanged at 70,523; against MNRE ground-mounted Best goes from 0.73 to 0.91 of registered capacity (0.82 AC), as the sample predicted. No other country's figures move.

Eleven more countries judged, 2026-11-03. The owner judged every outline queued for Egypt, France, Morocco, Namibia, Nigeria, Pakistan, Poland, South Africa, Vietnam, Yemen and Zambia, 913 of 919 (six left unsure stay out): 176 are parks. What a blob is depends on the country: Poland 113 of 142, Yemen 5 of 6, Vietnam 32 of 56 and France 12 of 29 were parks, against Egypt 4 of 23, South Africa 5 of 249, Pakistan 2 of 41 and at most one in Morocco, Nigeria, Zambia and Namibia, where sheets of false alarms (pans, rock, dark fields, mines) dominate. Best rises in ten atlases, Verified in none: Poland 7,781 to 9,110 MWp, Egypt 3,379 to 3,810 (an 8.1 km² park in New Valley), Vietnam 8,861 to 9,119, South Africa 9,756 to 9,910, Yemen 457 to 586 (the Aden, Mocha and Red Sea parks), France 13,369 to 13,447 (net of -8 MWp from re-running density under the 11 October rule), Pakistan 34,129 to 34,142, and Nigeria, Morocco and Zambia by 5 to 6 MWp each. On 6 November the owner judged Turkey's 774 (one left unsure): 396 are parks, 89.8 km², and Best rises from 13,103 to 17,593 MWp, Verified unchanged. The same day the owner judged Germany's 581: 149 are parks, 32.4 km², and none of the 43 lying on composite cells that lost their bands. Germany's Best rises from 84,473 to 86,010 MWp: +1,619 from the parks and -82 from re-running density (last run 15 September) under the 11 October rule, in four cells where a mapped outline above the cap was priced as rooftop. Every queued blob in every country now has a verdict but Poland's five and South Africa's and Turkey's one left unsure.

How much of a lead file to check (scripts/lead_share.py, results/lead_share/). Where a reference exists, the real installations sit at the top of the ranked file: the top third held 69% of the leads mapped since in Pakistan (385 of v3's 4,435 leads of 2026-07-16) and 87% in Yemen (246 of 2,259 from 2026-10-28, a week of the owner's mapping), with 4.1% and 2.1% real below it; in Germany, where the register needs nobody to have looked, it holds 46% of the 6,408 leads with a unit inside (35.9% below). Leads under 1,000 m² were 3.9%, 3.2% and 6.9% real. Adopted: Silver and Gold need a third (atlas.LEAD_SHARE_MIN) of a lead file's leads and of its blobs with a verdict and the model retrained on that mapping once (validation.retrained_on_leads with share and blobs_share), replacing "every lead in the current file checked" (2026-10-31) and the retrain rule's "over half". Pakistan's turn is recorded at share: 0.5, the owner's "over half", and it stays Gold; France, Germany and Yemen still show Bronze. The banner says it softly, as the owner asked: leads need to be mapped in OpenStreetMap to localize the model.

Gold also needs the rooftop classifier checked by people (owner, 2026-11-02): random validation (cells drawn at random away from the calibration areas, logged in results/roofclf_random_validation_log.csv) and point data inspection (the per-building scores from scripts/export_roofclf_points.py). Every row of that log is blank, so Pakistan moved from Gold to Silver with no figure changed, and no atlas is Gold. atlas.derive_validation_score now stops at Silver: a person's check cannot be read from files on disk, so Gold can only be declared (validation.score in the country's config).

What did not work

Two-season stacking

A 20-band two-season stack (dry-season base plus a contrast season per cell) was built to push detection below 1,000 m², on the theory that PV is spectrally stable across seasons while vegetation and roofs swing. The full path is wired and TerraMind duplicates its pretrained patch embedding into both season slots.

On the same Punjab validation installations, recall for the 1,000+, 500 to 1,000 and 250 to 500 m² buckets was 0.51, 0.17 and 0.14 seasonal, against 0.55, 0.16 and 0.14 for the production 10-band model. Small buckets unchanged within noise, large slightly worse. Likely causes: too few in-domain chips to learn a temporal signal, the tiny backbone's capacity, and post-monsoon versus dry season simply not differing enough in arid Pakistan. A later variant using a 2021 pre-boom epoch instead of a contrast season (boom-window stacking) was inconclusive rather than negative: training collapsed on only 332 usable chips, so it did not test the hypothesis either way.

Sentinel-1 corner reflection

A tilted PV row over flat ground forms a dihedral corner reflector, which should produce strong radar backscatter, persistently, since Sentinel-1's orbit geometry is fixed year-round and radar is not blocked by cloud. Tested on 17 glint-validated installations spanning the full observed azimuth range, pulling two years of Sentinel-1 RTC.

Median enhancement rate was 3.2% (VV) and 1.7% (VH) of scenes, within plain speckle noise. Critically, ascending and descending passes gave near-identical rates, 1.7 against 1.8%, with no correlation to the implied row axis. A real corner reflector should show sharp asymmetry between orbit headings, and its absence says this is not a usable channel at these sites through a simple per-footprint aggregate.

A lighter use of Sentinel-1 remains untried and is not ruled out by this: multi-temporal backscatter variance separates permanent structures from seasonally changing fields, and greenhouse metal frames give a bright return, the opposite of PV, making radar a cheap post-hoc false-positive filter.

Two glint routes to density, both negative

1. Cell-level spike counting

The idea was that a dense neighbourhood of sub-pixel residential arrays might produce more cell-level glint spikes, even when individual arrays rarely glint strongly enough to detect alone. This was tested on a fully mapped Lahore cluster with up to 120 installations within a 300 m block.

• Zero-PV controls: 1.0 mean spike
• PV-bearing cells: 1.45 mean spikes
• Median for both: 1.0
• The 120-installation hotspot produced only one spike in two years

The signal was too weak to distinguish PV-rich cells reliably. This likely reflects the statistic rather than the underlying glint physics, since a 90^th-percentile cell measure only responds when a large fraction of the cell brightens at once.

2. Recovering missed installations

A second approach tested whether glint could validate real PV installations missed entirely by the segmentation mask, allowing their area to be added back. It was evaluated on 43 missed German installations and 208 missed Lahore installations, each compared with matched non-PV controls.

• Germany: 37.2% validation on missed PV vs 20.8% on controls
• Lahore: 5.3% validation on missed PV vs 8.7% on controls

The separation is not strong enough for reliable recovery. In Germany, the control false-validation rate remains high, while in Lahore the control rate is actually higher than the missed-PV rate, indicating no useful discriminative signal.

A measured glint PSF, and the control set that mattered more (2026-08-17)

The astronomical version of the sub-pixel problem: measure the instrument's point-spread function from strongly-validated glints, then matched-filter against it, the way stellar photometry pulls faint sources out of noise. Full write-up: Glint PSF and matched filtering.

The PSF part works. Fitting a forward model (polygon rasterised at 1 m, Gaussian-blurred, block-averaged onto each scene's own grid at the target's true sub-pixel position) over 68 targets below 500 m² gives sigma 0.65 px, 90% CI 0.60 to 0.70, against an optical theory range of 0.49 to 0.62 px implied by ESA's stated MTF. The residual excess is covered by a measured per-scene source displacement of 0.72 px median, so the fitted kernel is an effective one that already contains co-registration.

The detector part does not. Every variant tested, centroid-pinned, offset-fitted, and point-source, scores below the p98-minus-annulus statistic already in the pipeline (AUC 0.485 to 0.620 against 0.648 to 0.655), significantly so over all sizes. The reason is visible in the fit itself: sigma climbs from 0.65 to 2.20 px with installation area while variance explained collapses from 0.49 to 0.06, because a specular glint comes from whichever patch satisfies the mirror condition on that date rather than from the whole array. Matched filtering is optimal for a known shape at a known position and a glint has neither, which is the one property stars have that makes the analogy work.

The valuable result was the control set built to test it. 600 verified negatives, buildings inside the 27 Rule-1 complete quadrats carrying has_pv == 0, give a false-spike rate of 2.0%, against the 8.7 to 20.3% previously measured on merely model-negative controls. An ablation puts the per-pixel SCL cloud veto at half of that improvement (4.5% with it disabled) and unmapped real PV in the old controls at the rest. Against true detection rates the instrument separates by 15x at 100 to 500 m² and 7.9x below 100 m², so the spike-rate density estimator's stated blocker, that false rate equals or exceeds true rate below 500 m², does not hold against verified negatives.

Super-resolution

Three feasibility tests, run in sequence by scripts/run_sr_experiments.sh: guided fusion of the 20 m bands to 10 m, multi-image super-resolution from repeated overpasses, and internal-learning single-image super-resolution. None improved detection, and the last carries an obvious hallucination risk on a task whose whole output is "is there a panel here". Scripts are kept for reference.

Choosing the imagery date so panels glint

A well-motivated idea that the geometry refuses. roofclf reads a dry-season median composite, and a median is built to suppress exactly the transient specular events that mark a panel, so reading the dates when panels should glint ought to raise the signal-to-noise ratio, especially in dense blocks where many would brighten at once.

Measured two ways. First the ceiling, from real granule sun and view angles over two years at all 23 quadrats: because Sentinel-2 views near-nadir, the pose that reflects sunlight into the sensor is a narrow locus, and only a median 13.2% of a plausible south-facing installed population (range 6.7 to 23.6%) can land on it at all. A textbook south-facing array at tilt 30 misses by 8.6 degrees on every scene in the archive. The single best date reaches 1.0 to 1.8% of rooftops, so the "one optimal date" framing fails specifically.

Then the feature itself, on the Lahore quadrat (13,500 buildings, 3,432 with mapped PV, the densest ground truth here and exactly the dense-urban case the idea targets). It separates PV standalone at 0.613 AUC, but 0.528 within roof-size band, i.e. nearly all of that was size. Added to roofclf's own features under a spatial holdout, size-controlled AUC moves 0.7875 to 0.7879.

Full derivation, the pose-window figure and before/after imagery of the best-case buildings: Solar glint. Two narrower versions survive untouched: per-locality pose calibration, and glint's existing role corroborating individual large arrays.

Snow-cover contrast, and the one commune that nearly sold it (2026-09-08)

The physics here is better than the glint idea's, and it targets a weakness this project can name precisely. Panels are tilted, smooth and dark, so they shed and melt snow faster than the roof around them: under lying snow a small array stops being a dark patch on a dark slate roof and becomes a dark patch on a white one. At 10 m, for a 20 m² array in a 100 m² pixel, that is the difference between a 8% reflectance deficit against a slate roof, inside BRDF noise, and a 19% deficit against snow on a brighter, higher-SNR background -- roughly 2.4x the signal, and aimed squarely at the dark-roof contrast problem that makes French rooftops harder than Pakistani ones.

It fails on opportunity, which is the same thing that killed choosing the imagery date so panels glint, and it was measured the same way: not by building a detector, but by counting how many chances a roof actually gets. scripts/snow_opportunity.py reads SCL on its native 20 m grid inside a rasterised mask of each commune's VIDA footprints -- snow over farmland says nothing about whether roofs were white -- and counts a scene as an opportunity when snow exceeds 5% of roof pixels while cloud stays under 20% of them. Both conditions matter, and the cloud one is easy to forget: in France a snowfall usually arrives with the system that hides it.

Across 1,556 scenes, 14 communes and four winters (2021/22 to 2024/25): 84 opportunities, of which 79 are in a single commune. Saint-Martin-de-la-Porte, alpine Maurienne at roughly 1,000 m, returns 19.8 per winter and peaks at 99.7% roof snow. Chambery returns 0.50 per winter, Grenoble, Langoat and Saulny 0.25 each, and nine of the fourteen communes return zero in four winters. Excluding the one alpine commune leaves 0.096 opportunities per commune-winter, about one usable scene per commune per decade.

Three things then compound, and the first is decisive:

• The signal is where the fleet is not. On a deliberately generous definition of mountain departements (Alps, high Pyrenees, Jura, Vosges, Massif Central), only 10.8% of France's 34.6 GWp of registered PV sits there.
• It would need a retrain, not a processing change. Snow currently causes false positives here: growth.py records 91 of 155 km² of spurious "vanishing" PV above latitude 34 as winter snow, and imagery.py excludes SCL class 11 from composites by design. The contrast could not be harvested by the current model, only by one trained on snow-bearing labels that do not exist.
• Winter geometry adds a false-positive mode. At 45 N in December the sun sits near 20 degrees elevation, so shadows become large relative to buildings at 10 m, precisely while the method hunts dark patches on bright roofs.

The methodological lesson is about the pilot, not the physics. The alpine commune was run first as a smoke test and on its own it argued the opposite conclusion, roughly 20 opportunities per winter. It is the outlier, not the norm, and a single-site pilot would have justified building the whole thing. Opportunity has to be counted over the population the method would be applied to.

One variant survives and is worth stating: Germany, which has more reliable lowland snow and is already the largest region in the training corpus, so the retrain objection is weaker there. If the snow hypothesis is ever tested properly it should be tested there, not in France. Raw counts: results/france_snow_opportunity.csv.

The 20 m bands, and what sharpening them costs (2026-09-19)

Six of the ten composite bands are 20 m at the sensor: B05, B06, B07, B8A, B11 and B12. The composite is written at 10 m, and odc.stac was replicating each 20 m value nearest-neighbour into a 2x2 block. Measured on cell 0122_0077, B11 and B12 are constant across 2x2 blocks in 100.0% of the raster, against 0.0% for B02 and B08.

That matters more than it sounds, because of where the classifier's weight sits:

Coefficient	Value	Native
b11_mean	+4.33	20 m
swir_vis_ratio	-3.92	uses SWIR
ndvi	+3.73	10 m
ndbi	+3.65	uses SWIR
b12_mean	-3.23	20 m
b02_mean	+2.78	10 m

The four largest are SWIR or SWIR-derived. So the model leans hardest on bands that carry no independent information at the grid it is scored on, and a 100 m² building's SWIR value comes from a 20 m cell whose centre can be 10 m away from it.

Two fixes were measured against the 30 production quadrats, leave-one-quadrat-out, paired per fold, changing nothing but how the same downloaded pixels reach the 10 m grid:

Variant	AUC	Within size band	Folds better	Sign test
Baseline (nearest)	0.8575	0.8206	--	--
Bilinear from the true 20 m grid	0.8574	0.8322	20 of 30	p = 0.061
Regression sharpening	0.8570	0.8060	7 of 30	p = 0.008

They point in opposite directions, and that is the finding. Regression sharpening predicts each 20 m band from the four native-10 m bands and adds back the residual, and it hurts significantly: -0.0047 AUC, -0.0052 within size band, better in only 7 of 30 folds. The interpretation is the useful part. If SWIR were largely a function of the visible bands, synthesising its 10 m structure from them would be nearly free; that it costs skill says the SWIR signal the classifier uses is genuinely independent of the visible bands, and replacing real-but-coarse SWIR with a visible-derived estimate throws it away. This is a property of PV worth remembering: it is spectrally distinctive in the shortwave infrared precisely where the visible bands cannot see it coming.

Simply removing the blockiness, by taking each band back to its true 20 m grid and resampling bilinearly, moves the other way: +0.0041 AUC within size band, better in 20 of 30 folds, p = 0.061, with the median within-size AUC going 0.8206 to 0.8322. That is suggestive rather than established, and it is free. It says the defect worth fixing is the misregistration, not the missing detail.

Shipped as imagery.BAND_RESAMPLING / annual_composite(resampling=...), default "20m-bilinear", with --resampling nearest to reproduce the old behaviour. Every composite records which it used in an earthpv_resampling tag; a composite with no such tag predates this and is nearest. The two must not be mixed within one AOI: the change is small but systematic, and a model calibrated on nearest composites and scored on bilinear ones is a domain shift rather than an improvement. That is enforced rather than documented -- compose inherits an AOI's existing mode and gives only a fresh AOI the new default, because a country-scale run is a restart loop and the first pass after a default changes is where silent mixing happens. Pakistan, Germany, France, Zambia and Nigeria (20,160 cells between them) are all nearest today and stay that way until someone recomposes one wholesale, which on p=0.061 evidence is not worth the bandwidth. SCL keeps nearest regardless, because interpolating between class 4 and class 8 invents class 6.

Unmixing the pixel against a known footprint (2026-09-19)

A zonal mean over a building's pixels is a mean of MIXTURES: a 100 m² roof is one 10 m pixel it shares with road, yard and neighbours. That dilution is the mechanism behind nearly every negative result in this register. Blind two-endmember unmixing was rejected at 0.659 AUC, but this is a different problem and a better-posed one, because the abundances are known: the footprints say what fraction of each pixel belongs to which building, so

y_p = sum_j A_pj r_j + (1 - sum_j A_pj) b_p

is a linear inverse problem for the per-building reflectance r, solved for a whole quadrat at once with a ridge toward the zonal mean (preprocess.unmix_buildings).

On synthetic data it works exactly as intended. With buildings deliberately offset from the pixel grid and 36 to 196 m² in size, so the median building fills 38% of its brightest pixel, unmixing cuts mean reflectance error against truth by 67% (0.0975 to 0.0319) and recovers 84% of the true dark/bright contrast where the zonal mean recovers 28%.

On real quadrats it is the worst result in this register: -0.0205 AUC and -0.0323 within size band, better in 2 of 30 folds, sign test p = 0.000. Not noise, not a tie, a substantial loss.

Two explanations, and the second is the interesting one:

• The footprints are not that good. The solve assumes A is exact. VIDA is imagery-derived and systematically undersized -- this project already measured 117,003 m² of mapped rooftop PV overhanging VIDA outlines across 27 quadrats, which is 80% of what the parcel label recovers. A wrong A does not merely fail to help, it attributes the wrong pixels to the building.
• The mixture is not purely a nuisance. The parcel label shipped because PV beside a building counts toward it, and the yard carries label-correlated signal. Unmixing does the exact opposite: it strips the surroundings out in order to isolate the roof. Running it against a parcel label is close to self-defeating, and the size of the loss suggests the context is worth more than the purity.

The honest summary is that dilution is real but removing it is not the same as recovering the signal, and an instrument that assumes its geometry is exact inherits every error in that geometry. Kept as roof-classifier-side machinery behind building_table(preprocess="unmix") for anyone who arrives with authoritative footprints -- the French cadastre is the obvious candidate, since it already bought +0.027 AUC within size band over VIDA there -- but off, and not recommended on an imagery-derived layer.

Raw numbers for both: results/roofclf_preprocess_ablation.json and results/roofclf_preprocess_*_folds.csv.

L1C against L2A: the correction is not the problem (2026-09-19)

Sen2Cor's surface-reflectance retrieval assumes a Lambertian surface and constrains aerosol from dark targets. A PV module is specular AND dark, so on the face of it L2A models the wrong physics over exactly the target of interest, in a geometry-dependent way, and it clips the saturation a real glint produces. If that costs anything, the fix is free: use L1C top-of-atmosphere reflectance instead.

Getting the data was the awkward part and is worth recording. Planetary Computer publishes L2A only, and Earth Search's sentinel-2-l1c assets point at ESA's requester-pays s3://sentinel-s2-l1c, which needs AWS credentials. Google's gcp-public-data-sentinel-2 mirrors the same SAFE archives and is anonymously readable, so imagery.l1c_composite runs discovery on STAC and reads pixels as JP2 over HTTPS from GCS, on the parent composite's own pixel grid. L1C carries no usable cloud mask (QA60 is empty from baseline 04.00), so the mask comes from the L2A scenes of the same solar days. Serially this costs about 12 minutes a quadrat, nearly all of it JP2 opens; a thread pool over the 120 (scene, band) reads brings it to 108 s.

The composite validates against L2A exactly as atmospheric physics requires: TOA/BOA is 1.42 in the blue, 1.19 green, 1.09 red, 0.97 NIR, 0.93 and 0.89 in the SWIR -- strong Rayleigh path radiance at short wavelengths giving way to absorption at long ones.

And it makes no difference. Over the same 30 quadrats, same rows, same features:

	AUC	Within size band
L2A (shipped)	0.8575	0.8206
L1C	0.8429	0.8102
Paired delta	+0.0004, 15 of 30 folds, p = 1.00	-0.0039, 12 of 30, p = 0.46

Fifteen folds each way is as null as a result can be, so the Lambertian objection, while physically real, costs the classifier nothing measurable.

The interesting part is the spread rather than the centre. The paired median is zero while the median AUC drops 0.0146, and the fold-to-fold IQR is about three times the other variants tested here (-0.0274 to +0.0128 within size band, against -0.0007 to +0.008 for bilinear resampling). A few quadrats get much worse under L1C while the typical one is unchanged. That is what an absent atmospheric correction should look like in a leave-one-quadrat-out design: path radiance varies with aerosol and geometry BETWEEN quadrats, so removing the correction adds cross-quadrat variance without touching per-building discrimination. The correction earns its place in transferability, not in contrast -- which is precisely the property a model trained on some quadrats and applied to a whole country depends on.

Kept as building_table(preprocess="l1c") plus scripts/compose_l1c_quadrats.py, because the reader is the reusable part: it is the project's only route to TOA reflectance, and glint work has a standing interest in saturation that L2A discards.

Temporal unmixing: the dry-season window removes its own signal (2026-09-19)

The third and last unmixing attempt, and the only one that needs neither the footprint nor a known endmember spectrum. Its constraint is that a PV array's AREAL FRACTION of a pixel is constant while the background's reflectance is not, so taking the temporal spread of

y_p(t) = f_p . r_PV + (1 - f_p) . b_p(t)

gives spread(y_p) = (1 - f_p) . spread(b_p), and f_p is one minus the ratio of a pixel's temporal spread to its local background's. That is self-normalising against roof colour, where absolute reflectance is not: a dark roof and a panel look alike in a median composite, and the claim is that they move differently through a season.

On synthetic data the estimator is near-unbiased, recovering 0.492 for a true fraction of 0.50 and 0.247 for 0.25, and refusing (100% NaN) when the background is held still.

On real quadrats the premise is simply false. scripts/pv_temporal_invariance.py compares per-pixel temporal MAD inside mapped PV against PV-free roof pixels in the same quadrat, which tests the assumption directly and needs no classifier:

	Median across 22 quadrats	Range
PV pixels / PV-free roof pixels	1.02	0.89-1.23
Open background / roof pixels	1.12	0.91-2.02

A PV pixel varies over time exactly as much as the roof beside it, so there is no damping to measure; and the open background moves only 12% more than a roof, so there is barely a denominator to divide by. Only 1 quadrat of 22 has a background moving more than 1.5x a roof. At 70-150 DN of MAD against 2,000-2,800 DN of reflectance, what is being measured is a common floor of per-scene radiometric residual and BRDF, not surface dynamics, and a floor is added AFTER mixing so it is not attenuated by f at all.

The ablation agrees, over 30 quadrats leave-one-quadrat-out: +0.0003 AUC (17 of 30 folds, p=0.46) and +0.0011 within size band (18 of 30, p=0.26). Damping plus size alone reaches 0.7388 / 0.5863, against a size-only baseline of 0.7441 / 0.5773.

The conditional test is what makes this more than another null. Registered before the run: if the mechanism is real but starved, the gain should rise with how much a quadrat's background actually moves. It does. Spearman between the within-size gain and open_over_roof is +0.401 (p = 0.071, n = 21 folds), and splitting at the median:

	Gain within size band	Folds better
Dynamic half (open/roof >= 1.13)	+0.0015	9 of 11
Still half (open/roof < 1.13)	-0.0001	4 of 10

So the physics is visible in the structure of the result and worth nothing in its magnitude. Treat that cautiously: n is 21, p = 0.071, and the single most dynamic quadrat (malok, open/roof 2.02) runs against the trend at -0.0013, so the relationship is not driven by its own extreme.

The most useful thing here is why the denominator is missing, because it is self-inflicted. annual_composite takes a DRY-SEASON window of the twelve least-cloudy scenes, chosen precisely to suppress phenological and atmospheric variation and produce a stable composite. That is the right choice for everything else the pipeline does, and it removes exactly the background dynamics this estimator needs. The honest next test is a full-year stack including the monsoon, where cropland actually swings; that is different data, not a different estimator, and it is still open. Until then this is rejected for the window the project actually composites.

A methodological note worth keeping: the first precondition check measured temporal amplitude over WHOLE quadrats and looked ample (median 148 DN, all 31 quadrats clearing the bar). The denominator that matters is the background near a BUILDING, which is other roofs and pavement, and measuring that population gave 1.12. Checking the precondition on the wrong population is the same error the snow-cover experiment records, in a different costume.

Artifacts: results/pv_temporal_invariance.json, results/roofclf_temporal_unmix.json, results/roofclf_temporal_unmix_folds.csv.

The model class: ranking and calibration disagree (2026-09-19)

Every feature block this project has measured -- SPPI, glint, the yard, epoch jump, temporal statistics, temporal unmixing, sharpening, spatial unmixing, L1C -- was priced against fit_logistic, an L2-regularised LINEAR model. A linear score cannot express a conjunction, and "is this a PV roof?" reads like one: dark AND spectrally flat across the visible AND the right SWIR drop AND large AND in a dense-adoption area. Six blocks landing within 0.005 of each other is consistent with the functional form being the constraint rather than the features, so it is worth testing directly.

Two declared HistGradientBoostingClassifier configurations, conservative and larger, fixed in the script rather than selected on the held-out folds, on the same 30 quadrats and the same table.

Model and features	AUC	Within size band	Median per-quadrat rate error
Logistic (shipped)	0.8574	0.8206	0.0308
Logistic + shape	0.8593	0.8196	0.0309
GBM small	0.8582	0.8006	0.0204
GBM large	0.8526	0.7979	0.0177
GBM large + shape	0.8618	0.7921	0.0165

On ranking the linear model wins, and not marginally. Paired per fold, every GBM variant loses: -0.0089 to -0.0121 AUC and -0.0163 to -0.0231 within size band, better in only 4 to 6 of 30 folds, p <= 0.002. The larger configuration is worse than the smaller one, which is the signature of a flexible model learning each quadrat's idiosyncrasy rather than what transfers between them. Note that gbm_large_shape has the highest median AUC in the table (0.8618) while losing on 25 of 30 folds: a difference of medians is not a paired gain, which is the same trap the shape block below fell into.

So the linear model is not the bottleneck, and the feature experiments were fair. That is worth having established, because it is the assumption every one of them rested on.

On calibration the answer reverses, and for this project that matters. The atlas consumes an aggregate adoption rate, not a ranking, and the register already records that ranking transfers across quadrats while absolute rates do not -- which is why the coverage-ratio-by-size-and-density machinery exists at all. Measured per quadrat, GBM predicts the adoption rate substantially better: median absolute error 0.0308 to 0.0165, closer to truth in 20 of 29 quadrats for gbm_large (Wilcoxon p = 0.006) and 21 of 29 with shape (sign test p = 0.024).

The obvious objection is that a model can tighten rate_ratio by hedging toward the global mean, which would be worthless. It is not doing that: the spread of predicted rates across quadrats is 0.0806 against a true 0.0798, and the correlation with the true rate RISES from 0.776 to 0.869. It keeps the dispersion and gets the level right more often.

The two objectives genuinely disagree, and nobody had noticed because only AUC was reported. Nothing is being swapped on the strength of one run: the lead product and the precision-thresholded population that feeds the coverage ratio both depend on per-building ranking, where the linear model is better, while the capacity half depends on the rate, where it is not. The honest next step is to price a GBM end to end through sub400-capacity against the same quadrats, not to change the classifier.

Footprint shape, settled

plus_shape had scored 0.8593 against the shipped 0.8574 in the ablation table and had sat unadopted since 2026-08-09, the best-looking block never promoted. Given the paired test every rejected idea got, it is noise: -0.0001 AUC, 14 of 30 folds, p = 1.00, and -0.0003 within size band. The apparent gain was a difference of medians across folds rather than a per-fold improvement. Closed.

Artifacts: results/roofclf_model_class.json, results/roofclf_model_class_folds.csv.

Aiming at the aggregate: weighting and recalibration (2026-09-19)

The model-class test found that gradient boosting ranks worse and predicts per-quadrat adoption rates better. Two cheap interventions follow from it, neither touching the feature set, both aimed at the aggregate the atlas consumes rather than at AUC.

Variant	AUC	Within size band	Median per-quadrat rate error	Area ratio
Baseline (shipped)	0.8574	0.8206	0.0308	1.084
Area-weighted loss	0.8506	0.8046	0.0395	1.035
Isotonic, global	0.8570	0.8199	0.0268	1.095
Isotonic, by size tercile	0.8560	0.8166	0.0253	1.128
Platt	0.8574	0.8206	0.0287	1.070
Gradient boosting, for reference	0.8526	0.7979	0.0165	--

Area weighting does what it promises, and that is not what the product needs. Capacity is area-weighted while the fit treats a 30 m² shed and a 390 m² warehouse as equally important rows, so weighting the likelihood by roof area is the obvious correction. It improves the quantity it optimises -- the predicted-over-true flagged roof AREA moves 1.084 to 1.035 -- and it is worse at everything else: -0.0111 AUC, -0.0160 within size band, and a COUNT rate error that rises 0.0308 to 0.0395. Trading the population that carries the labels against the aggregate loses more than it gains at this sample size. Available as fit_logistic(sample_weight=...), off by default, and None reproduces the unweighted fit bit-for-bit.

The calibration gain is not recoverable by recalibration, which is the useful half. A monotone map cannot change a ranking, so if gradient boosting were merely a better-calibrated version of the same score, isotonic or Platt would capture it for free. They do not. The best variant, isotonic fitted within roof-area terciles, cuts the median per-quadrat rate error 0.0308 to 0.0253 and is closer in 19 of 29 quadrats, but at Wilcoxon p = 0.069 it is not significant, and it recovers only 38% of the gradient-boosted improvement (which reached 0.0165 at p = 0.006). Platt moved AUC by exactly 0.0000 in 0 of 30 folds, which is the clean check that a strictly monotone map reorders nothing; isotonic moves it by 0.0002 only because it creates ties.

So gradient boosting's advantage is a genuine reordering of buildings, not a calibration curve, and it cannot be had cheaply. That sharpens an open question: if the aggregate matters more than the ranking for the capacity half, the GBM has to be deployed properly, plumbing and all, rather than approximated by post-processing the linear model.

Calibrators were fitted NESTED -- inside each leave-one-quadrat-out fold the 29 training quadrats were split into five blocks by quadrat, and the calibrator was fitted on the inner out-of-fold scores. Fitting on in-sample scores would be optimistic in exactly the direction being measured. Artifacts: results/roofclf_weighting_calibration.json, results/roofclf_weighting_calibration_folds.csv.

Spectral coherence, and where transferability lives (2026-09-19)

Two more attempts at the features the classifier leans hardest on, both using data already on disk.

Variant	AUC	Within size band	Median rate error	Paired vs shipped
Baseline (shipped)	0.8574	0.8206	0.0308	--
Medoid composite	0.8415	0.7896	0.0280	-0.0074, 9 of 30
Plus context features	0.8540	0.8166	0.0291	-0.0003, 14 of 30
Context features ONLY	0.8595	0.8189	0.0569	-0.0011, 14 of 30
Medoid plus context	0.8546	0.7982	0.0287	-0.0070, 9 of 30

The band-wise median really does invent a spectrum, and fixing it does not pay. annual_composite reduces with median(dim="time") over a Dataset, so each band's median is taken independently: a pixel's B02 can come from January and its B11 from March, and the result is a spectrum no scene observed. That is harmless for one band and not harmless for a RATIO, which is what the model weights most (swir_vis_ratio -3.92, ndbi +3.65, blue_red_ratio -1.56). A medoid -- the single date per pixel whose whole spectrum is closest to the median -- is coherent by construction, and verified to equal exactly one real date where the band-wise median equals none. It scores -0.0074 AUC and -0.0102 within size band. The trade is visible in the construction: a medoid is ONE observation, so it forfeits the twelve-scene median's noise averaging, and that costs more than incoherent ratios do. It does improve the rate error slightly (0.0308 to 0.0280), which is consistent with the diagnosis rather than with the medoid being simply worse.

The result worth keeping is the context-only row. local_zscore re-centres a building against its own quadrat or cell, and ships applied to brightness alone, one feature of fifteen. Extended to every spectral feature, added alongside the absolute ones it changes nothing (-0.0003, 14 of 30). But replacing them entirely ties the shipped model (-0.0011, 14 of 30, p = 1.00) while nearly doubling the rate error, 0.0308 to 0.0569.

That separates two things this project had confounded:

• Ranking is relative. Which building in a cell carries PV is fully answerable from within-cell contrast; the absolute reflectance level adds nothing to it.
• Calibration is absolute. Z-scoring removes the level that says one quadrat is more PV-dense than another, and the aggregate collapses.

For scaling that is a useful thing to know in both directions. The ranking half -- the leads product, and the precision-thresholded population the coverage ratio is fitted on -- should transfer to a new country robustly, because it is immune to the cross-quadrat level shifts that atmospheric and seasonal differences produce. The aggregate half depends on absolute reflectance being comparable between places, which is exactly why removing atmospheric correction in the L1C test tripled the fold-to-fold spread without touching the median.

A footnote that turned out to be a defect. brightness_zscore re-centres a building's brightness against its own quadrat or cell and has shipped as an available block, plus_local_contrast, since 2026-08-09. It had never actually been measured: local_zscore used np.median rather than np.nanmedian, and it runs inside building_table BEFORE the fill/edge rows are dropped, so a single building whose footprint has no valid pixel turned the entire quadrat's z-scores into NaN. That was 21% of rows and 5 of 30 quadrats entirely NaN -- among them sialkot and sukkur, the two quadrats already recorded here as carrying composite fill. One NaN poisons a logistic fit, so every fold trained on those quadrats returned NaN and pivot_table silently dropped the block from the ablation output. Fixed 2026-09-19 (NaN-safe, and bit-identical on clean input), after which the block measures +0.0000 AUC, 13 of 30 folds and is rejected on its merits rather than by accident.

Neither is adopted. The context block is kept as roofclf.CONTEXT_FEATURES / add_context_features, the medoid as preprocess.medoid_composite and building_table(preprocess="medoid"). One caveat on the medoid comparison: its table has 36 more rows than the baseline (123,903 against 123,867), because a scene stack covers the quadrat geobox plus margin rather than the boundary bbox. At 0.03% that cannot carry the result, but the two are not row-identical the way the earlier preprocessing comparisons were.

Artifacts: results/roofclf_medoid_context.json, results/roofclf_medoid_context_*_folds.csv.

The spectral SNR budget, and why the domain is exhausted (2026-09-19)

After a run of null preprocessing results it is worth asking what the spectral signal actually IS, rather than trying another representation of it. Detection here is a mixture, y = f.r_PV + (1 - f).r_bg, so the signal is f.(r_PV - r_bg) and the question is what each term is worth. Measured on the 30-quadrat table, in B08, the highest-contrast band:

Term	Value
Panel-minus-roof contrast at full cover	0.0579 reflectance = 1.34 background sd
Median fill fraction on a PV roof	0.309
Typical available signal	0.41 sd
Background sd between PV-free roofs	0.0431 reflectance
Sentinel-2 radiometric noise at this level	~0.001-0.002 reflectance

The mixture model holds tightly: B08 falls linearly with fill fraction (B08 = 0.3142 - 0.0579 x f), and the separation grows with it exactly as predicted -- d' of -0.25, -0.36, -0.54, -0.83 across fill-fraction bands averaging 0.08, 0.15, 0.29 and 0.65.

The noise is not the sensor, and this rules out a whole class of ideas by arithmetic. Roof-to-roof heterogeneity is twenty to forty times Sentinel-2's radiometric noise. Every intervention aimed at measurement quality -- more scenes, a better atmospheric correction, denoising, spectral coherence -- is tuning a term worth about 2% of the variance. That is the retrospective explanation for temporal statistics, L1C, the medoid and sharpening all landing at zero.

The linear spectral limit is already reached. The Mahalanobis separation across all ten bands is d' = 1.226, an implied AUC of 0.807, while spectral_only measures 0.8333 in leave-one-quadrat-out -- the shipped model is at or past the Gaussian linear bound, exceeding it because the hand-made ratios are nonlinear in the raw bands. Whitening buys 1.95x over the best single band (0.629, B08) and the model already has it. With 79.3% of band variance in the first principal component and a maximum inter-band correlation of 0.984, there are roughly two effective dimensions here. No new index, matched filter or whitening of these bands can add meaningful signal. (The bound is in-sample and assumes Gaussian classes, so read it as an order of magnitude, not a decimal.)

That leaves shrinking the denominator, which fails for an interesting reason. Since the noise IS roof diversity, estimating a background per building from its nearest neighbours should raise SNR directly, and roofing material is spatially clustered so the neighbours are the right reference. It does cut the noise: the PV-free standard deviation falls 23 to 43% across bands. It cuts the signal faster. Median d' falls to 0.36x, B11 even changes sign, and the multivariate separation drops 1.226 to 0.904.

The cause is a property this project already documents: PV adoption is spatially clustered, which is why nn_median_m exists. A PV roof's neighbours disproportionately carry PV, so the local median is itself pulled toward the panel spectrum and subtracting it cancels the contrast. Sweeping the neighbourhood scale shows there is no escape:

Neighbours	Median radius	Mahalanobis d'	Against raw
10	31 m	0.717	0.58x
50	78 m	0.904	0.74x
200	169 m	0.980	0.80x
800	369 m	1.022	0.83x
none	--	1.226	1.00x

Monotone, and asymptotic toward not conditioning at all. This also explains why the quadrat-scale version (context features) merely TIED rather than helped or hurt: at 1-4 km² the group is large enough that adoption clustering does not dominate its median, so it neither cancels signal nor removes much heterogeneity.

Conclusion: the spectral domain is exhausted for this sensor. The remaining term in the budget is fill fraction, which no amount of processing changes -- it is set by pixel size against array size, and it is the same quantity that makes France fail and Pakistan work. Kept as roofclf.add_neighbour_features for anyone who wants to re-measure. Panels are strongly polarising, which would be a nearly background-free channel, but no free satellite measures it.

PV-free neighbours only (2026-10-23), also rejected. The obvious repair is to build the background from neighbours that carry no PV, which removes the clustering that cancels the contrast. Measured on the area-weighted table (raw d' 1.564 there, against 1.226 in the pixel-centre measurement above), 50 nearest buildings in the same quadrat: neighbours chosen by label keep 0.90x of the raw separation against the all-neighbour 0.75x, and choosing them by out-of-fold score below the deployment threshold, which needs no labels, keeps 0.94x. It still does not help. The PV-free spread falls 42% in every variant, but even with label-chosen neighbours the visible and SWIR bands keep only 0 to 0.4x of their separation (B11 changes sign) while B06 to B8A keep all of it: most of the visible and SWIR contrast is neighbourhood-level, which roof stock adopts PV, rather than panel physics. Dividing by the neighbours' spread, a local z-score, is worse than plain differencing. Added to the shipped features, every variant moves within-size AUC by -0.0001 to -0.0016 per fold, better in 10 to 13 of 29 folds, none significant; on their own they reach 0.769 to 0.774 against 0.819. The score-chosen version carries a second-order leak (a neighbour's score comes from a model that saw the held-out quadrat), which can only flatter it. Script: scripts/roofclf_pvfree_neighbour_contrast.py; artifacts results/roofclf_pvfree_neighbour_contrast_sep_k50.csv and _folds_k50.csv.

Sun geometry: neither glint nor high sun raises the contrast (2026-09-19)

Two follow-ups to the SNR budget, asking whether some sun-sensor geometry gives a better look at the same panels.

Glint is not a scene-level lever. A glass-fronted panel is specular, so on a date when the sun-panel-sensor geometry approaches the mirror condition it should stand out. Scored over 239 scenes in 22 quadrats, using each scene's own sun and view angles and the 192 poses fitted from real Pakistani installations rather than an assumed one, near-specular scenes do show 34% more contrast: d' -0.549 in the nearest quartile of misalignment (1.17 deg) against -0.409 in the farthest (8.54 deg).

It is not glint. Misalignment and solar zenith are collinear at +0.475 by construction, and partialling out the sun kills the geometry term while the sun survives:

	Spearman	p
d' vs misalignment	+0.101	0.12
d' vs solar zenith	+0.204	0.0015
d' vs misalignment, controlling for solar zenith	+0.025	0.70
d' vs solar zenith, controlling for misalignment	+0.161	0.013

The sign was wrong for glint from the start: PV gets DARKER relative to the roof near specular, not brighter. And glint could not work as a scene selector even if it were real, because 93.7% of scenes already have some pose within 10 deg of specular (median 3.77 deg) -- the condition almost never discriminates between observations. What this does NOT test is per-target glint on individual installations, which is how glint.py uses it and where it does work: no scene has more than half the pose population glinting, so a narrow lobe on a few panels is diluted by pixel-averaging across a quadrat.

And a high-sun window does not help either, which refutes a prediction made here. The winter fit was d' = -1.296 + 0.0145 x sun_zenith, which extrapolated to roughly double the contrast at summer sun angles and suggested the dry-season compose window might be costing signal. Measured against pre-monsoon May-June scenes over 10 quadrats:

	Pre-monsoon	Dry season
Median d'	-0.601	-0.672
Roof brightness	3,064 DN	2,532 DN

Gain 0.98x, better in 5 of 10 quadrats, Wilcoxon p = 0.77. The first half of the mechanism happened exactly as predicted -- roofs really are 21% brighter under high sun -- and the contrast did not move, because d' is a NORMALISED quantity: illumination scales the panel, the roof and the between-roof spread together, so numerator and denominator grow in step and the ratio is invariant to first order.

That means the within-window correlation the prediction rested on was not illumination scaling at all; more likely seasonal surface change (moisture, dust, vegetation) tracking date across a narrow 37-59 deg span. The compose window is not costing contrast, and acting on the extrapolation would have meant recompositing a country onto a worse epoch for no gain. Artifacts: results/glint_geometry_snr.json, results/summer_window_contrast.json and their per-scene CSVs.

The level half of brightness does not transfer (2026-09-20)

absolute = cell_mean + deviation, and roofclf sees only absolute, which confounds "this building is bright" with "this cell is bright". Handing it the cell mean adds one degree of freedom per feature. The motivation was the measured split reported above: within-cell z-scores alone tie the shipped ranking while doubling the rate error, so ranking is relative and calibration is absolute. The level half looked like the unexploited part.

Registered before measuring: a cell-constant feature adds the same number to every building's logit inside a held-out quadrat, so it cannot reorder them, and any AUC movement can only come from refitting the other coefficients. That half held -- the single cell-mean moved AUC by +0.0001.

The substantive half failed, and significantly.

Block	AUC	Within size band	Rate error	Rate ratio
Baseline	0.8574	0.8206	0.0308	1.061
Plus brightness cell-mean	0.8571	0.8220	0.0309	1.056
Plus all 15 cell-means	0.8494	0.8229	0.0430	1.012

Calibration gets worse: the per-quadrat adoption rate is closer to truth in only 7 of 29 quadrats, sign test p = 0.008, Wilcoxon p = 0.002. The 15-feature version pushes median rate error 0.0308 to 0.0430 while pulling the ratio to 1.012, which is the signature of centring the level and widening its spread.

The reason is this register's own finding, turned against the idea. A cell-constant feature is fitted across 29 quadrats and then extrapolated to an unseen one, so the model learns a BETWEEN-quadrat relationship between brightness and adoption rate -- and between-quadrat absolute rates are exactly what does not transfer. Adding the level as a feature does not give the model the level; it gives it a spurious slope to extrapolate along. Kept as roofclf.CELL_LEVEL_FEATURES / add_cell_level_features for re-measurement.

Snow in Germany: the opportunity is real, the contrast is not concludable (2026-09-20)

Snow-cover contrast was rejected for France on opportunity, 0.096 usable scenes per commune-winter outside the Alps, and that entry named the one variant worth testing: Germany, which has more reliable lowland snow. Tested over the 8 densest PV cells in the German OSM rooftop layer, three winters, with SCL class 11 deliberately KEPT (the shipped mask drops classes outside 4-7 and would discard every observation under test) and snow measured over building footprints rather than over the cell.

The opportunity objection is answered. 6 snow scenes out of 58 low-cloud scenes, a 10.3% rate, against France's 0.096 per commune-winter. Observations of snow-covered roofs genuinely exist in Germany.

The contrast is not. Six scenes, Mann-Whitney p = 0.39, and bimodal:

Cell	Snow on roofs	Roof DN	PV DN	d'
137_515	61%	2,035	1,146	-2.40
137_515	100%	3,806	2,405	-1.72
149_513	20%	1,647	829	-1.03
149_513	100%	4,333	1,702	-1.74
60_508	100%	3,250	8,143	+3.84
60_508	84%	3,524	5,588	+2.48

Four scenes behave as the physics predicts. The two that inverted are explained by array size rather than by snow: cell 60_508's mapped arrays have a median area of 126 m², about 1.3 pixels at 10 m, against 338 and 320 m² in the two well-behaved cells. At roughly 1.2 PV pixels per array the mask is mostly snow-covered ROOF, which reads brighter than a roof average that includes shadowed and wet surfaces. The project's central limitation, array size against pixel size, reappears here amplified rather than relieved.

The headline ratio of 15.9x is not usable either: it divides by a bare-season baseline of -0.087, i.e. German OSM rooftop PV barely separates from PV-free roofs in bare winter at all, which is consistent with a control contaminated by the 96.4% of German rooftop PV that OSM does not map. A near-zero denominator makes any ratio large.

What would settle it is more cells and a restriction to arrays large enough to own a pixel. Recorded as partial rather than shipped or rejected, because the half that killed the French version is genuinely answered and the half that matters is untested at this sample size. Artifacts: results/germany_snow_contrast.json, results/germany_snow_contrast_scenes.csv.

The composite reducer is the biggest lever in this register (2026-09-20)

Chasing Google's Open Buildings 2.5D Temporal, which fuses up to 32 Sentinel-2 acquisitions to reach an effective ~4 m, produced the largest single improvement measured anywhere in this register -- and it turned out to have nothing to do with super-resolution.

annual_composite reduces the scene stack with a per-pixel median. Replacing it with a mean gains +0.0188 AUC within size band at 10 m, with no change to resolution, features or model.

That figure is smaller than first reported here, and the correction matters. The variants below reduce the saved SCENE STACKS while the baseline reduces the COMPOSITE FILES, and the two were fetched at different times through different read paths. A control run afterwards -- the MEDIAN of the same stack, which should score zero if the two paths are equivalent -- scores +0.0098 within size band (21 of 30, p = 0.024). So roughly a third of the original +0.0286 was the path, not the estimator:

Effect	AUC	Within size band
Path (stack median against composite median)	+0.0054	+0.0098
Estimator (mean against median, same stack)	+0.0122	+0.0188
Estimator, trimmed mean	--	+0.0102

The estimator effect is real and significant either way, and the path effect is worth its own look: the baseline reads through composite_index.read_window, which mosaics and can carry the documented cell-edge fill frame, while a stack is a clean geobox read. If that is the cause it is a free win independent of any reducer.

The decomposition that isolates it:

Variant	AUC	Within size band	Folds better
Baseline, median at 10 m (composite file)	0.8575	0.8206	--
Median at 10 m (same stack) -- PATH CONTROL	0.8628	+0.0098	21 of 30
Mean at 10 m	0.8756	+0.0286 (+0.0188 net of path)	25 of 30
Mean at 5 m	0.8789	+0.0313	25 of 30
Multi-frame fusion at 5 m	0.8767	+0.0273	26 of 30
Area-weighted zonal means at 10 m	0.8588	+0.0113	25 of 30
Area-weighted zonal means, as SHIPPED (nodata-masked, no rows dropped)	0.8682	+0.0119	24 of 29
Median at 5 m, nearest	0.8623	+0.0058	23 of 30
SEN2SR single-image at 2.5 m	0.8402	-0.0367	6 of 30

So 91% of the effect is the estimator, about 9% is the finer grid, and the fusion itself is NEGATIVE: applying measured sub-pixel shifts scores below not applying them.

Why the mean wins is not mysterious once separated out. SCL masking already removes cloud, shadow and snow, so the median's robustness is largely redundant, and at twelve samples a median carries about 1.57x the variance of a mean. Less feature noise, better separation, and it helps most for the sub-pixel footprints whose single pixel is noisiest.

The route there was three wrong mechanisms in a row, which is worth recording. The gain first looked like multi-frame fusion; a zero-shift control beat the fused version, so it looked like resolution; plain upsampling recovered only a fifth of it, so it looked like footprint sampling; exact area-weighted zonal means recovered a third. Only the mean explains it, and tmean_up2 reproduces the shift-and-add control to four decimals, which is the consistency check that the accumulator had been computing a mean all along.

Why fusion cannot work here, measured rather than assumed. Multi-frame super-resolution needs frames that sample the ground at different sub-pixel phases. Measured across 298 frame pairs in 30 quadrats, the median inter-acquisition shift is 0.122 px (1.22 m), with phase offsets at roughly a quarter of the uniform-spread ideal and only about 3 frames in 12 usefully displaced. Re-measured from native granules with no reprojection or resampling, to rule out our own loading path as the cause: 0.132 px (1.32 m), essentially identical. The diversity really is that small. A synthetic check puts the consequence precisely: with ideal phases shift-and-add recovers +0.166 of correlation against an upsample, and at the diversity Sentinel-2 actually provides, +0.015 -- about 9% of the achievable gain.

And single-image super-resolution actively hurts. SEN2SRLite (CC0-1.0, runs in 0.31 GB on a GTX 1060) produces a faithful 2.5 m raster -- downsampled it correlates 0.9976 with the original and it genuinely adds high-frequency energy, 0.688 to 0.918. It still costs -0.0367 AUC within size band, 6 of 30 folds, because it cannot add information: it redistributes 10 m content using learned priors, and texture uncorrelated with PV dilutes real contrast. That is the 2026-07 rejection reproduced with a much better model, and it answers the hallucination objection empirically rather than by assertion.

Building height, from GlobalBuildingAtlas, is redundant. Height alone separates PV from PV-free roofs at d' = +0.651 (median 6.74 m against 4.03 m), comparable to the best single spectral band. On top of the model it is worth +0.0006 within size band, 6 of 9 folds, p = 0.51, and height-plus-size alone scores 0.173 BELOW baseline: tall buildings are large buildings, and log_roof_area already carries it. Measured on the 5 quadrats where GBA covers more than half our footprints, a biased subsample. GBA heights are CC BY-NC 4.0 in any case, which would have blocked deployment.

The mean's weakness is real, and a trimmed mean is the answer. The median is there to reject residual cloud that SCL missed, and a mean cannot: with one unmasked cloud pixel among twelve frames, a clean value near 100 reads 100.7 under the median, 100.2 under a 20%-trimmed mean, and 924.4 under a plain mean. Measured on the quadrats, the trimmed mean keeps most of the gain:

Reducer	Within size band	Folds better	One unmasked cloud
Median (current)	--	--	100.7
Trimmed mean (drop top and bottom decile)	+0.0200 (+0.0102 net of path)	25 of 30	100.2
Mean	+0.0286 (+0.0188 net of path)	25 of 30	924.4

So the trimmed mean is the one to ship: 70% of the gain, and robustness indistinguishable from the median on the failure mode the median exists for.

Wired as imagery.REDUCERS / annual_composite(reducer=...) / compose --reducer, with the default left at "median" and run_compose inheriting an AOI's existing reducer, exactly as for the resampling change and for the same reason: a model calibrated on medians and scored on means is a domain shift. Every existing AOI -- Pakistan, Germany, France, Nigeria, Zambia -- inherits "median". Shipping it nationally means recompositing, which costs roughly 2 TB of transfer per country, so it is free for the next country and expensive for the existing ones.

Two more noise reductions, and they stack. Weighting each 10 m pixel by the fraction of the footprint covering it -- instead of rasterize's all-or-nothing pixel-centre test, which gives about half of VIDA footprints zero pixels and a representative-point fallback -- is worth +0.0113 within size band on its own (25 of 30 folds, p = 0.0001). Combined with the trimmed mean it reaches +0.0305 raw, +0.0207 net of the path effect, against +0.0215 predicted by adding them: additive to within noise, because they attack different noise sources.

SHIPPED 2026-09-20, and it is the only part of this section that costs nothing. The reducer and the frame count both need a country recomposited; area weighting is a read-side change to roofclf and needs no new imagery at all, so it was promoted to the default (roofclf.AREA_WEIGHTED_ZONAL, roofclf.area_weighted_zonal_mean, applied in BOTH building_table and score_buildings_national, with --no-area-weighted on roof-classifier and roofclf-score-national reproducing every earlier figure).

The shipped version is not the experimental code, so it was re-measured rather than carried over. Two things had to change first, neither of which shows up in an interior quadrat:

• The nodata mask. The experiment tested np.isfinite, right for a NaN-masked scene stack and wrong for a composite window, where fill is 0.0. Unmasked, a cell-edge fill pixel enters the weighted mean as genuine near-zero reflectance, and PV is dark: that is the mechanism that once made 45.6% of every nationally flagged building a cell-edge artefact. The shipped version applies the same all-bands-equal-fill test zonal_mean_max uses.
• The fallback. A footprint with no subpixel of its own got zero weight and came back NaN, silently dropping 0.16 to 0.47% of buildings per quadrat -- the smallest ones, i.e. the population this module exists for. The shipped version keeps zonal_mean_max's value for those rows, so the scored population is unchanged by construction. 0.06 to 0.8% of buildings take that path.

Re-measured on all 30 quadrats, both arms on identical rows, no rows lost:

Metric	Pixel-centre	Area-weighted	Delta	Folds better	Sign p	Wilcoxon p
AUC	0.8591	0.8682	+0.0085	28 of 29	0.0000	0.0000
Within size band	0.8093	0.8193	+0.0119	24 of 29	0.0005	0.0017

and in the units the atlas consumes, at an identical precision of 0.500:

Arm	Flagged	Precision	Recall
Pixel-centre	21,904	0.500	0.6386
Area-weighted	22,580	0.500	0.6583

+2.0 points of recall, 3.1% relative, for no imagery. The shipped version slightly exceeds the experimental +0.0113, which is the expected direction: it stopped dropping the smallest buildings. Acceptance test: scripts/verify_area_weighted_default.py, result results/roofclf_area_weighted_default.json.

Pakistan was refitted on 2026-09-20 (30 quadrats, parcel label; pre-refit state kept at data/roofclf_PRE_areaweighted_20260920/). Against the 2026-08-20 production fit:

	2026-08-20	2026-09-20
Buildings	123,867	123,898
Median fold AUC	0.8574	0.8682
Median within size band	0.8206	0.8193
Deployment threshold	0.2515	0.2446
Flagged at precision 0.500	21,580	22,580
Recall at precision 0.500	0.6292	0.6583

The within-size line reads down and that is a difference of medians, not a regression. Two things changed between those summaries -- area weighting and the grid-snap read fix -- and neither summary is a paired comparison: different tables, different fold sets, and a median of fold medians is not additive. Measured properly on identical rows the two are +0.0119 (24 of 29 folds, p = 0.0017) and +0.0032 (18 of 29, p = 0.27) respectively; the read-path figure reproduces the +0.0032 already recorded for it, and its own unpaired medians move 0.8206 to 0.8092 while its paired delta is positive. This register has retracted a claim to this exact trap once already ("the 0.8593 that made it look best was a difference of medians"), so the paired test is the one to read. The deployment row is computed on pooled out-of-fold scores rather than a median, which is why it moves unambiguously.

The calibration and the national scoring must share one zonal convention, and check_scoring_matches_calibration now refuses a mismatched pair by name rather than by coefficient hash. National rescoring was run the same day off this fit; the capacity chain and the atlas are a separate, owner-gated step because they move the published headline.

Frame count helps, weakly, and only through variance. Full-year stacks (~36 scenes against the dry-season 12) separate frame count from season by drawing a fixed random 12 out of the SAME year stack, so the two differ only in N:

Trimmed mean over	Within size band	Net of path
12 dry-season frames	+0.0200	+0.0102
12 frames of the year	+0.0247	+0.0149
~36 frames of the year	+0.0295	+0.0197

The frame-count effect is +0.0048 within size band for tripling N, and the path effect cancels in that comparison so it is clean. That is what 1/sqrt(N) predicts: three times the frames cuts the standard error by 42%. So Google's 32-acquisition budget does transfer to this problem -- through estimator variance, not through the shifted viewpoints its design exploits -- and it costs three times the download, which against roughly 2 TB of recompositing per country is a poor trade at national scale. Note also that the full-year season mix is not itself harmful: at matched N = 12 the year frames slightly outscore the dry-season ones, so a wider window is a viable way to reach more frames if the bandwidth is ever there.

The whole package, net of the path effect: about +0.026 AUC within size band, from a trimmed mean, area-weighted zonal means and more frames -- all of it noise reduction, none of it new information, and none of it super-resolution.

What it is worth to the PRODUCT, which is not AUC. run_roof_classifier picks a threshold targeting precision 0.5 on pooled out-of-fold scores, and sub400_capacity fits its coverage ratio and area recall on the population that threshold flags. A ranking gain that leaves that population unchanged changes no capacity figure, so the deployment step was re-run for each variant:

Variant	AUC	Flagged	Precision	Recall
Baseline	0.8601	21,592	0.500	0.6295
Trimmed mean	0.8771	23,148	0.500	0.6751
Plus area-weighted zonal	0.8810	23,382	0.500	0.6822
Plus the full year's frames	0.8851	23,794	0.500	0.6937

At identical precision, recall goes 0.6295 to 0.6937: +6.4 points, or 10.2% more of the true PV roofs captured. That is the number to quote, because it is the population the capacity chain is fitted on. The two free changes alone -- trimmed mean and area-weighted zonal -- deliver +5.3 points of it; the remaining +1.1 costs three times the download.

One threshold is already the right one (2026-09-20)

If the deployment threshold is what the capacity chain consumes, the obvious next lever is the threshold itself. run_roof_classifier picks ONE cut targeting precision 0.5 on pooled out-of-fold scores. One cut is recall-optimal only if the score means the same thing everywhere, and this register has repeatedly found it does not: rate_ratio spans 0.2 to 5x across quadrats, gradient boosting beat the linear model on per-quadrat rate error while ranking worse, and no global monotone map recovered that gain. So: cut per stratum instead, on the two stratifiers the capacity chain already uses, building-size band and quadrat density. Every threshold is fitted on the OTHER quadrats' out-of-fold scores, the same nesting the model itself uses, and every comparison is against a global cut RE-TUNED to whatever precision the stratified rule actually reached, so a recall gain bought by dropping precision cannot read as a gain.

Equal precision in every stratum is much worse than one cut, and not marginally:

Rule	Flagged	Precision	Recall	Net of matched global	Quadrats better	Sign p
One global cut	21,199	0.5012	0.6195	--	--	--
Per size band	20,180	0.5004	0.5889	-0.0298	7 of 26	0.029
Per density stratum	19,074	0.5023	0.5586	-0.0579	3 of 23	0.0005
Per size and density	19,029	0.4885	0.5420	-0.0919	4 of 25	0.0009

The mechanism is that equal AVERAGE precision is the wrong objective. At a fixed pooled precision the recall-maximising allocation equalises the MARGINAL precision -- the precision of the last building admitted -- not the stratum's average. A stratum whose base rate is 3.6% (the 0-50 m² band, 50,502 of 123,867 buildings) has to be cut very deep before its average precision reaches 0.5, and its recall collapses from 0.191 to 0.059 while the bands that were already easy barely move.

So the theoretically-correct version was run too: recalibrate the score to a per-stratum posterior with isotonic regression fitted on the training folds, then apply ONE global cut to the recalibrated score, which equalises the margin by construction. This is not the monotone recalibration already rejected here -- that map was global, and a per-stratum map is not monotone in the original score.

It wins pooled and dies paired. Per size band: recall 0.6392 at precision 0.4947 against the matched global cut's 0.6260, so +1.3 points pooled -- and a median per-quadrat difference of exactly +0.0000, 12 of 24 quadrats better, sign test p = 1.00. The pooled gain is a few large quadrats, which is the same failure the register logged when 55 PV-dense German municipalities read as an 18.2% error against a representative 33.4%. Recalibrating per density stratum loses outright (-0.0482 net), and per size and density together loses (-0.0110 net).

The roof is its own control (2026-09-20)

Everything above works on ONE epoch, and the SNR budget says that is the binding constraint: the noise limiting roofclf is roof-to-roof heterogeneity, 0.0431 reflectance, 20 to 40x the sensor noise, and the linear spectral limit on a single-epoch zonal mean is essentially reached. Every intervention aimed at the MEASUREMENT -- L1C, 20 m sharpening, spatial unmixing, single-image and multi-frame super-resolution, medoid compositing -- has failed, which is exactly what that budget predicts.

There is one term that cancels roof-to-roof heterogeneity outright, and it is not spectral: the same roof, earlier. Pakistan's rooftop boom is post-2022, so a 2019/20 dry-season composite is a pre-installation look at most of today's arrays, and current - 2019 differences away everything about the roof that did not change. Thirty-one quadrat-clipped scene stacks were pulled for 2019-11-01 to 2020-03-15 (6 to 12 frames, median 7, 223 MB total) and reduced the SAME way as the current stack, so a difference cannot pick up a reducer or a read-path change.

Block	AUC	Within size band	vs baseline	Folds better	Sign p	Wilcoxon p
Baseline	0.8574	0.8205	--	--	--	--
Plus the 2019 difference	0.8632	0.8388	+0.0234	25 of 29	0.0001	0.0002
Plus a split-half difference (placebo)	0.8551	0.8182	-0.0009	14 of 29	1.00	0.70
Plus the 2019 LEVELS, undifferenced	0.8590	0.8294	+0.0036	20 of 29	0.061	0.26
Plus both differences	0.8695	0.8392	+0.0222	25 of 29	0.0001	0.008

Measured against the shipped read path this is the largest single feature block in this register, and roughly what the entire noise-reduction package above is worth, from one extra historical composite rather than a national recomposite. That framing does not survive its own control -- see "The two levers are not additive" below, which is the part to read before acting on any of this.

Three controls, all registered before measuring, and all of them matter:

The placebo is exactly null. Differencing the first half of the CURRENT dry-season stack against its second half gives the same band count, the same features, the same imagery and the same reduction, with no installation possible in between: -0.0009, 14 of 29 folds, p = 1.00. So what helps is the EPOCH, not having two estimates of the same roof.

The difference beats the levels, which is the L2 prior doing work. For an unpenalised linear model, adding the 2019 levels spans the same space as adding the differences and must tie. Under L2 it does not, and it does not: +0.0036 (p = 0.061) against +0.0234. The difference parameterisation is the one that encodes "the roof cancels", and encoding it is worth 6x more than handing the model the same information unshaped.

It is PV, not construction. A building that did not exist in 2019, or was re-roofed, also differences large, and VIDA footprints come from recent imagery so new construction is in the layer. Two reads separate them (scripts/epoch_difference_mechanism.py):

• Sign. PV modules are dark; bare ground turning into a roof is not. PV-bearing roofs darkened relative to PV-free ones by a median 0.0082 reflectance, darker in 28 of 29 quadrats, sign test p < 0.0001.
• Magnitude. Construction would live in the extreme tail of the difference. It does not: the top decile of |d19_brightness| has a LOWER PV base rate than the rest (0.130 against 0.140), and the gain is in the bulk (+0.0165, 25 of 28 folds) not the tail (+0.0041, 12 of 19).

The fitted coefficients are also self-consistent with the static model rather than with a generic change detector: the largest is d19_b11 at +1.88, and b11_mean is the static model's largest coefficient at +4.33. The block is reading "this roof moved TOWARDS the PV spectral signature", not "this roof moved".

In the units the atlas consumes (the same deployment step the reducer package was priced in): at an identical precision of 0.500 the flagged population goes 21,596 to 22,704 buildings and recall goes 0.6298 to 0.6621, +3.2 points, 5.1% relative. For comparison the whole noise-reduction package is +6.4 points, and that one needs the existing national imagery recomposited; this one needs a second epoch alongside it.

THE TWO LEVERS ARE NOT ADDITIVE, AND THIS IS THE RESULT THAT MATTERS FOR DEPLOYMENT. Everything above measures the epoch difference against the SHIPPED read path: a median composite read through CompositeIndex.read_window. The noise-reduction package (trimmed mean, area-weighted zonal means) replaces exactly that, and it was fitted on the same 30 quadrats. Re-running the whole ablation on the trimzonal table instead:

Base	Baseline within size band	Plus the 2019 difference	Gain	Folds better	Sign p
Shipped read path	0.8205	0.8388	+0.0234	25 of 29	0.0001
Noise-reduced (trimzonal)	0.8370	0.8510	+0.0047	20 of 28	0.036

Four fifths of the epoch gain is already captured by the noise reduction. It is still there, still significant, and the placebo is still null on the better base (-0.0017, 10 of 29) -- but a project that ships the reducer package should expect well under +0.01, not +0.023, from adding a second epoch on top.

The old epoch was frame-starved, and it is still not saturated. Those 2019/20 stacks carry a median 7 frames against the current 12. Re-pulling them over a wider window (2019-10-01 to 2020-04-30, --max-items 24 --max-cloud 50) more than doubles that to a median 16, and on the same noise-reduced base the gain rises and sharpens:

Old epoch	Median frames	Gain on trimzonal	Folds better	Sign p	Wilcoxon p
2019-11-01 to 2020-03-15	7	+0.0047	20 of 28	0.036	0.011
2019-10-01 to 2020-04-30, cap 24	16	+0.0067	23 of 29	0.002	0.0002

So the honest deployment number is +0.0067 and rising, not +0.0047, and the curve has not flattened -- a deeper old epoch is the cheapest remaining way to improve this, since it costs archive queries rather than a second national composite. One thing does change with the cleaner old epoch: the undifferenced 2019 LEVELS catch up almost completely (+0.0063, 24 of 29, p = 0.0005 against the difference's +0.0067). The "the roof cancels" prior was worth 6x on a noisy old epoch and is worth almost nothing on a clean one, which is what a regularisation prior should do: it buys most where the data is weakest.

The two are not measuring the same thing mechanically -- one cleans the current estimate, the other cancels the roof -- but they raise the same underlying quantity, the contrast between panel and roof relative to the noise on it, and once the measurement is clean there is less left for the difference to recover. The honest summary is that this register has found one large improvement available to roofclf, reachable two different ways, and not two large improvements. Best measured combination: trimzonal plus the 2019 difference at 0.8510 within size band against the shipped 0.8205, and the noise reduction is the cheaper half of it.

Caveats. The 2019 arm carries a median 7 frames against the current 12, so the difference is noisier on the old side than it needs to be; more 2019/20 scenes would likely widen the gain rather than close it. Rule-1 labels are epoch-relative, so an array installed between the mapping imagery and the composite is an unlabelled positive in both arms. Deploying this nationally means a second epoch of imagery for the whole country, which is the same cost, roughly 2 TB per country, as the reducer change -- but unlike the reducer it buys new information rather than less noise, and it does not require recompositing what already exists. Scripts: scripts/compose_scene_stacks.py --window 2019-11-01:2020-03-15 --subdir stacks_2019, scripts/run_epoch_difference_ablation.py. Results: results/roofclf_epoch_difference.json, results/roofclf_epoch_mechanism.json.

Most footprints are one pixel (2026-09-20)

Every spectral feature roofclf ships is a zonal MEAN over the footprint. That is the right statistic only if a flagged roof is uniformly covered, and it is not: an array usually occupies part of a roof, so the mean is a mixture of panel and roof in an unknown proportion. This register's SNR budget says the roof half is the dominant noise term (roof-to-roof heterogeneity 0.0431 reflectance, 20 to 40x the sensor noise) and that the linear spectral limit ON THE MEAN is essentially reached. Order statistics are not on that axis. They are a different summary of the same pixels, not a cleaner estimate of the same summary, so the limit does not bind them.

Two hypotheses, registered before measuring. EXTREMES: the darkest pixel in a footprint approximates the panel endmember, so min and max are less diluted than the mean, and this should help MORE on larger footprints. SPREAD: a partially covered roof is internally heterogeneous and a uniform one is not, so the within-footprint standard deviation is direct evidence of partial cover. Both on four quantities (brightness, SPPI, SWIR/visible, blue/red), both free -- the window is already read and rasterized. Plus the control that matters here: the pixel COUNT alone, because a size proxy hiding inside either block would look like a win.

Block	AUC	Within size band	vs baseline	Folds better	Sign p
Baseline	0.8591	0.8093	--	--	--
Plus pixel count (control)	0.8579	0.8092	-0.0015	11 of 29	0.265
Plus extremes	0.8579	0.8074	-0.0006	12 of 29	0.458
Plus spread	0.8583	0.8080	-0.0012	11 of 28	0.345
Plus both	0.8575	0.8086	-0.0007	11 of 29	0.265

Nothing, and the diagnostic is the point: 72.4% of the 123,898 buildings in these 30 quadrats resolve to a SINGLE 10 m pixel. Median pixel count is 1 overall and 2 for a PV-bearing roof. For three quarters of the population min, max and mean are the same number and the standard deviation is exactly zero, so there is no within-footprint distribution to read. Splitting each fold's AUC by pixel count confirms it is not merely diluted: on the multi-pixel subset the blocks are still flat (spread +0.0009, 15 of 28 folds, p = 0.85; extremes -0.0003, 13 of 28, p = 0.85), so the registered size-conditional prediction fails on its own terms too.

This bounds a whole class of ideas, not just this one. Anything that needs structure INSIDE a footprint -- texture, panel-versus-roof segmentation, within-roof unmixing, a shape prior on the array -- has no substrate in three quarters of the population that roofclf exists to cover. It also explains retrospectively why area-weighted zonal means work (+0.0113 within size band, above): that change attacks the same fact from the other side, fixing how a footprint reads a pixel it only PARTLY covers rather than trying to read structure within one. Script: scripts/run_dispersion_ablation.py, feature block roofclf.DISPERSION_FEATURES, result results/roofclf_dispersion_baseline.json.

Verdict: one global threshold is already the right allocation, and the score is closer to calibrated across size bands than the per-quadrat rate spread suggests. Those are different claims about the same model -- rate error is a level error within a quadrat, and what a threshold needs is that the score ORDER the same way across strata, which it does. Nothing changes in the pipeline. Script: scripts/run_stratified_threshold.py, result results/roofclf_stratified_threshold.json.

Artifacts: results/roofclf_preprocess_ablation.json, results/subpixel_shifts.json, results/native_shifts.json, results/roofclf_gba_height.json.

Keeping more than the median composite (2026-09-19)

The composite reduces about twelve cloud-masked scenes to their per-pixel median (imagery.annual_composite). A median is the maximum-robustness central estimator, so it deletes the tail by construction, and the tail is where a specular surface at a fixed tilt should put its information. The scenes are already downloaded and then discarded, so the cheapest possible version of "use the time dimension" is to keep more of that distribution and ask whether it helps.

compose --stats now writes a temporal_stats sidecar beside each composite: per-pixel p10, p50, p90 and standard deviation for all ten bands, plus a valid-observation count. No extra network traffic, which matters because this stage is bandwidth-bound. What roofclf gets per building is the bright tail (p90 - p50, a free proxy for the glint that otherwise needs bespoke per-target scene pulls), the dark tail (p50 - p10, which should separate dark-and-static PV from dark-and-moving shadow), the temporal spread, and n_obs, without which spread is unreadable: std = 0 means "stable" at twelve looks and "one look" at one.

The measurement is a leave-one-quadrat-out ablation on the 30 production quadrats, and it reproduces the shipped fit exactly (123,867 buildings, 17,150 positives, median fold AUC 0.8574, 0.8206 within size band, min fold 0.4964), so the block is the only variable.

Feature block	AUC	Within size band
Size only	0.7441	0.5773
Size + temporal	0.7311	0.5868
Median composite only	0.8333	--
Size + spectral (shipped)	0.8574	0.8206
Shipped + temporal (compact, 6 columns)	0.8579	0.8232
Shipped + temporal (full, 31 columns)	0.8576	0.8262

It buys nothing. Paired per fold, the compact block moves AUC by a median +0.0003 (IQR -0.0013 to +0.0016), improving 16 of 30 folds, sign test p=0.57; within size band +0.0006 (IQR -0.0018 to +0.0033), 16 of 30, p=0.71. Widening from 6 columns to 31 does not rescue it, so the compact grouping is not what threw the signal away.

The temporal_only row is the part worth keeping, because it distinguishes "adds nothing new" from "says nothing". On its own the block sits at 0.5868 within size band against a 0.5773 size-only baseline: a trace above chance, not nothing, but nothing that survives having the median composite in the model. So the distribution's higher moments are close to redundant with its centre for this task, at this resolution.

Two things make this a fair test rather than a plumbing failure. The sidecar stores its own p50, and that band is bit-identical to composite_0.tif's median on 100% of valid pixels, so the temporal features and the existing reflectance features describe the same twelve scenes and no epoch confound is possible. And the block has real dynamic range rather than being flat: for B02 the median pixel spans 129 DN between p50 and p90, 183 DN between p10 and p50, with a standard deviation of 148 DN, against roof reflectances of order 1,000 to 3,000 DN.

The likeliest reason it fails is the same sensor ceiling that the 400 m² floor and the French transfer result keep returning: a sub-400 m² array is a minority of a 100 m² pixel, so whatever anisotropy it has is diluted by a roof that does not share it, in every statistic equally. Twelve dry-season scenes also give little specular opportunity, which is the ceiling glint already documents from the other direction.

Kept behind flags (compose --stats, roof-classifier --temporal-features), off by default, because re-measuring costs one afternoon and no retraining. Cost if it were ever wanted nationally: about 4x the per-cell composite disk, roughly 215 GB for Pakistan's 4,473 cells. Raw numbers: results/roofclf_temporal_ablation.csv (median per block) and results/roofclf_temporal_ablation_folds.csv (per fold, which is what the paired test above is computed from).

Temporal features for roofclf

Two candidate features were tested against the quadrats and neither was kept. Epoch jump, the change in the same classifier's score between two imagery epochs, measured at 0.8736 to 0.8736 AUC in its probability form (exactly zero effect) and 0.8736 to 0.8608 in its cheaper reflectance-delta form, with one quadrat failing outright. Step change, per-building aggregation of the pixel-level time-series signal below, did not survive a within-size-band control. An older design note proposed using the epoch comparison as a recall rescue rather than a feature; that specific plumbing was never built, and the measurement above is the reason it was not pursued.

Retraining on known failures

Two retrains aimed at documented failure modes, both negative in an instructive way.

The fraction-head quadrat-supervised retrain beat the production checkpoint in 13 of 13 quadrats in-sample and collapsed the scale dispersion from 11.95x to 2.48x, which looks decisive until the held-out quadrat comes back at 0.461 against an in-sample band of 1.878 to 3.684. A paired spatial block bootstrap puts the held-out AUC gain at -0.009 to +0.061, one-sided p 0.062. One quadrat and 797 labelled pixels is the binding constraint, so more holdout folds, not more analysis, is what would settle it.

The roofclf hard-negative retrain took the one concretely locatable confirmed false positive on record (six very bright roofs that both detectors flag) and folded them in as negatives. Six rows against 104,423 moved their own scores by 0.0001 to 0.003 and changed nothing else. Oversampling them 20x, 100x and 500x does eventually suppress them, but median_fold_auc falls 0.8824 to 0.8811 to 0.8726 to 0.839 on the way, so the model is learning to distrust one tiny neighbourhood of feature space at the expense of everything else. No factor tested threads that needle. The fix this actually recommends is mining more examples of the same bright-roof pattern nationally, for which no roofclf-side mining tool exists yet.

SPPI alone as Germany's sub-400 estimator (2026-09-18)

Germany's published sub-400 m² half uses no classifier: it prices roof area in the 200-400 m² band at a register-fitted constant, because that beat both roofclf and total roof area. The obvious thing left untried was the other detector. SPPI is a zero-training spectral index, it is already computed for every scored German building, and it is the corroborating half of Pakistan's AND-gate, so if any classifier were going to help here it is the one that needs no labels.

It does not. All four estimators below were run through scripts/validate_sub400_against_mastr.py on identical data, the same cells, the same municipalities and the same five-fold cross-validation, with only the weighting changed:

estimator	Gemeinden	median municipal error	best stratified	Spearman	slope
roof area only (published)	10,589	31.2%	28.8%	0.942	0.888
SPPI percentile rank	10,588	33.7%	31.0%	0.922	0.845
roofclf probability	10,005	36.8%	34.9%	0.855	0.333
SPPI top decile	10,209	49.9%	48.8%	0.824	0.498

Two formulations, because SPPI is a spectral index in roughly [-2.6, 0.07] rather than a probability and cannot be multiplied in raw. Percentile rank weights each roof by where its SPPI falls in the national distribution, the analogue of probability weighting. Top decile credits only roofs above the 90^th percentile, the analogue of thresholding and the way SPPI is used in Pakistan's AND-gate.

The interesting half of the result is that SPPI ranks German roofs better than the trained classifier does, on both error and slope, while needing no labels at all. The decisive half is that neither beats doing nothing. This is the same finding as the roof-area baseline, from a third direction: where PV is near-ubiquitous, capacity is close to proportional to roof area by construction, so an instrument that concentrates credit on a subset discards more real capacity than it correctly withholds. Thresholding is worst precisely because it discards the most.

The baselines here read 31.2% and 36.8% where Germany's page records 35.4% and 48.4%, because the grid was rebuilt without the building-density filter (4,871 cells against 4,657) after those were measured. The ordering is unchanged and the within-run comparison is what the table is for.

A NaN trap worth recording, because it produced a confident wrong answer rather than an error. The first top-decile run reported zero municipalities and zero capacity. 2,660 of 6.2M buildings, 0.043%, carry no SPPI value; np.quantile returns NaN if any input is NaN, and x >= NaN is False for every row, so the threshold flagged nothing and the harness dutifully reported zeros all the way through. The harness is now NaN-safe and logs how many buildings lack a value.

A sub-400 m² level without quadrats, from OSM and a past epoch (2026-10-20)

Countries whose only high-resolution imagery is years old cannot get Rule-1 quadrats (Box 17's gate), and without quadrats they get no sub-400 m² figure at all. The classifier can be borrowed, since ranking transfers; the level cannot, because the coverage ratio carries precision and precision carries the local adoption rate. The idea tested here measured the level from two things such a country still has (earthpv.transfer_calibration), as an adjusted count, pi = (f - FPR) / (TPR - FPR) per roof-size bin:

• TPR from OSM positives: the detection rate among buildings carrying mapped PV anywhere in the country. Incomplete mapping is fine, since only the rate among them is used.
• FPR from the past: the flag rate of the same roofs in a pre-boom Sentinel-2 epoch, scored by the same model, as an upper bound on today's false-positive rate.

It was validated against Pakistan's 30 quadrats with no quadrat information at all: the classifier from the other quadrats, positives from 8,584 OSM-mapped sub-400 m² roofs outside every quadrat, negatives from composite_1 (2021-10 to 2022-01). Population: sub-400 m² roofs.

Level from	Median error per quadrat	Pooled ratio	Spearman
Quadrats, leave-one-out (flagged roof area x ratio)	35%	1.00	0.92
Quadrats, leave-one-out (roof area x constant, the null)	54%	0.98	0.90
OSM TPR + 2021 FPR	318%	-3.71	-0.48
OSM TPR + 2021 FPR, epochs quantile-matched	113%	-0.07	-0.25
OSM TPR + the TRUE current FPR (an oracle)	46%	1.26	0.96
A strict threshold, no FPR correction (p >= 0.55)	83%	1.08	0.91
200 random flagged roofs reviewed yes/no, OSM TPR	56%	1.27 (90%: 1.09-1.47)	0.93

The past is not a negative set, and the reason is the classifier. The same roofs flag more often in 2021 than today (0.236 against 0.160, while today's true FPR is 0.093), so the adjusted count goes negative. Matching the old epoch to today per band (median, median+IQR, quantile) only makes the two flag rates equal (0.156 to 0.164 against 0.160), and the estimate collapses to zero. The control that explains both: roofclf's 2021 score predicts TODAY's PV at AUC 0.832, against 0.870 for today's score (0.848 once the epochs are matched), and Pakistan had imported only 14% of today's panels by January 2022 (Ember). The classifier recognises the kind of roof that adopts, not the panels on it, which is also why the epoch difference adds information but not much. The same finding rules out the softer version of the idea, composing Sentinel-2 at an old mapping epoch so a stale-imagery quadrat becomes Rule-1 again: the flagged set barely changes between epochs, so such a quadrat would measure the old epoch's PV per flagged roof, which in a boom country is several-fold below today's. That last step is inferred, not measured.

The positives half works. With the true FPR plugged in, OSM-derived TPR and array size track the quadrats at Spearman 0.96. OSM positives are the easy ones (TPR 0.52-0.93 against 0.23-0.85 on quadrat PV; mean array 1.3-1.6x larger), and the two biases largely cancel, to +26%. A yes/no review of random flagged roofs is the cheapest human input that supplies the missing FPR (simulated at +27%, with 200 reviews holding the national total within about 15% of that), but each reviewed roof still needs recent imagery, which is the constraint this started from.

Nigeria, the motivating case, and India (2026-10-22): does Pakistan's classifier still rank PV roofs abroad? PU AUC on OSM-mapped sub-400 m² arrays against every other sub-400 m² roof in the same cells (scripts/transfer_calibration_pu_auc.py; 90% intervals from a cell-clustered bootstrap with negatives capped per cell):

Country	Positives (cells)	PU AUC	Within size bins	Flag rate, positives / unlabelled
Pakistan (own model, quadrats excluded)	4,385 (133)	0.930 (0.926-0.949)	0.887-0.924	0.750 / 0.049
India	498 (148)	0.778 (0.758-0.847)	0.690-0.807	0.504 / 0.079
Nigeria	60 (25)	0.661 (0.555-0.783)	0.628-0.833	0.367 / 0.115

Nigeria's first figure, 0.902 on 357 positives, was an artefact, and so was India's first, 0.673 on 47,911. The script labelled every footprint touched by any non-ground OSM polygon, and VIDA traces a park's panel rows as buildings, so a large array or a park mistagged rooftop turned into hundreds of "PV roofs" that were not the target population. Positives now come only from arrays under 400 m², and footprints touching a larger PV polygon are dropped from both classes; Pakistan, whose mapped rooftop PV is mostly small, barely moves (0.934 -> 0.930), which is why the original rule looked sound. The ranking transfers to India only partly: well above chance, well below Pakistan, and the mapped Indian arrays are not smaller than Pakistan's (OSM sub-400 m² median 74.8 m² against 75.4), so this is not France's sub-pixel regime. The level cannot be borrowed either way (above), so neither country gets a sub-400 m² half: India needs its own Rule-1 quadrats (eight sampled candidates are in data/labels/india/candidate_quadrats/), which would also measure the borrowed model properly and allow a refit. Scored cells: data/roofclf_transfer/{nigeria,india}/prob.

A side finding: composite_1 carries no BOA offset (scenes before baseline 04.00, 2022-01-25) while composite_0 carries +1000 DN; each layer's own darkest 0.5% sits 730 to 860 DN apart in every band (median over 60 cells). roofclf.epoch_feature_frame(dn_offset=) harmonizes it; see open question 10 ("Growth as a product") for what it means for the growth atlas, and for the 2026-10-22 finding that the offset differs by cell. Scripts: scripts/transfer_calibration_pakistan_check.py, scripts/transfer_calibration_followups.py, scripts/transfer_calibration_pu_auc.py. Results: results/transfer_calibration/.

AlphaEarth embeddings: propensity, not panels (2026-10-22)

Google's Satellite Embedding dataset (AlphaEarth Foundations v2.1, CC-BY-4.0) gives every 10 m pixel a 64-number, unit-length summary of one calendar year of Sentinel-2, Landsat, Sentinel-1 and other inputs, from 2017. It is the cheapest new input available: no training and no GPU, and the Source Cooperative mirror (tge-labs/aef, anonymous S3) needs no Earth Engine account. Each building got the area-weighted mean of the 64 channels over its footprint, with the same preprocess.coverage_matrix weights the shipped Sentinel-2 features use. Production 30-quadrat table, leave-one-quadrat-out, paired per fold against the shipped model. Every variant was declared in the script before measuring; the harness reproduces the shipped model exactly (0.8682 AUC, recall 0.658 at precision 0.5).

Block	AUC	Within size band	vs baseline	Folds better	Sign p	Recall at precision 0.5	Rate error
Baseline (shipped)	0.8682	0.8194	--	--	--	0.658	0.0286
Plus AEF 2025 (primary)	0.8682	0.8214	+0.0032	18 of 29	0.26	0.642	0.0346
Plus AEF 2024	0.8675	0.8220	+0.0050	19 of 29	0.14	0.673	0.0372
Plus AEF 2021, pre-boom control	0.8741	0.8241	+0.0047	20 of 29	0.061	0.691	0.0292
Plus AEF 2025 and its change since 2021	0.8673	0.8295	+0.0066	21 of 29	0.024	0.672	0.0264
Plus AEF 2025 over a 70 m box, not the footprint	0.8701	0.8182	-0.0007	14 of 29	1.00	0.608	0.0384
Size plus AEF 2025, no Sentinel-2 block	0.7475	0.6424	-0.1691	1 of 29	< 0.001	0.030	0.0354
Gradient boosting, plus AEF 2025 against without	0.8517	0.7908	+0.0012	16 of 29	0.71	0.647	0.0292

The primary comparison is null. -0.0003 AUC (14 of 29 folds, p = 1.00), +0.0032 within size band (p = 0.26), and the two quantities the product consumes both get worse: recall at the deployment precision falls 0.658 to 0.642 (quadrat bootstrap 95%: -0.066 to +0.034) and the median per-quadrat adoption-rate error rises 0.0286 to 0.0346. A non-linear head does not change that (+0.0012 within size band for gradient boosting).

The pre-boom control explains the rest. Pakistan had imported 14% of today's panels by January 2022, so a 2021 embedding cannot see most mapped arrays; it scores at least as well as 2025 (2025 against 2021: -0.0016 within size band, 12 of 29 folds; recall -0.049, 95% -0.092 to -0.021). The one block that clears a sign test, 2025 plus the change since 2021, is 2021 plus 2025 in a linear model, and against the 2021 levels alone it adds nothing (+0.0016, 18 of 29, p = 0.26; Wilcoxon p = 0.13 against baseline). The direct check agrees: between 2021 and 2025 the embedding of a footprint that gained PV moved exactly as far as one that did not (median cosine 0.9607 against 0.9604). What the embedding carries is the kind of roof and neighbourhood that adopts, which is what roofclf's own 2021 score already showed the Sentinel-2 block carries.

It cannot replace the Sentinel-2 block. Size plus the embedding reaches 0.6424 within size band against 0.8194 for size plus reflectance, and flags almost nothing at precision 0.5. The embedding is spatially smooth at roof scale: a footprint's mean embedding sits at cosine 0.992 from the mean over the 70 m box around it, so for a median 68 m2 Pakistani footprint it mostly describes the block. The footprint is still marginally more informative than the box (+0.0029 within size band, 22 of 29, p = 0.008), but neither beats the baseline.

Large arrays are visible, small ones are not (scripts/aef_pixel_change_check.py). On the AEF grid, pixels at least 75% covered by mapped PV moved further 2021 to 2025 than PV-free roof pixels more than 50 m from any PV (pooled AUC 0.653). That is carried by the industrial quadrats (Multan, Rahim Yar Khan, Sanghar, SITE and the Karachi coast, 0.78 to 0.82); in the residential Islamabad and Lahore quadrats PV pixels moved LESS than distant roofs (0.30 to 0.49), where new construction dominates the change. Roof pixels within 20 m of PV moved less than distant ones (0.43), so the change does not smear into neighbours. An annual AEF change map might corroborate large arrays or ground-mount, which is a separate question from roofclf and was not tested.

Closed for roofclf. Scripts: scripts/run_aef_embedding_ablation.py, scripts/aef_pixel_change_check.py. Results: results/roofclf_aef_embedding.json and _folds.csv, results/roofclf_aef_pixel_change.json. Embedding windows for the 30 quadrats (2021, 2024, 2025) are cached under /home/tobi/earthpv_data/aef/.

AlphaEarth change finds unmapped ground-mount (2026-10-22)

The roofclf test closed the embedding for small roofs but left one lead: pixels under large mapped arrays do move between years. The place that could use it is the ground-mount precision table. pakistan_candidate_precision.yaml sets p_unmapped = 0.0 in every ground bin, so every segmentation detection off a roof with no OSM feature within 100 m (about 4,300 of 5,200) is booked as false, because nothing independent had ever looked at them.

Data. Scattered objects are unaffordable from the COG mirror (one band block per read, ~38 MB per object-year), so this reads the Source Cooperative Zarr mosaic (tge-labs/aef-mosaic) by HTTP range request, ~2.5 MB per 256 x 256 chunk (scripts/aef_mosaic.py, no new dependency; it matches the COG values at Multan at median cosine 0.997). 2,215 chunks per year, 2017, 2021 and 2025.

Populations (scripts/aef_large_array_change.py, declared and seeded): every candidate with OSM within 100 m (real), every rooftop candidate without, a size-proportional random 1,000 of the 4,326 unmapped ground ones, and 500 per placement of the OSM arrays >= 400 m2 that no candidate is within 100 m of (missed). Controls: each ground object's own polygon displaced 300 to 1,200 m, and two PV-free VIDA roofs >= 400 m2 within 2 km for each rooftop one (contaminated by unmapped PV, so rooftop AUCs are lower bounds). Scores: raw change 1 - cos(e2017, e2025), and two logistic models fitted real against controls with 1-degree cells held out, one on the change vector alone (learned) and one adding the 2025 levels (learned_level).

Real against controls, AUC	Raw change	Learned, change only	Learned, change and levels
Ground candidates (876)	0.718	0.964	0.990
Rooftop candidates (1,126)	0.624	0.884	0.918
Ground arrays segmentation missed (500)	0.594	0.879	0.941
Rooftop arrays segmentation missed (500)	0.610	0.824	0.855

The direction of change carries the signal, not its size: raw change is weak, a model of which way an embedding moves when an array goes in is strong, from 0.83 below 500 m2 to 0.98 at 5,000 to 50,000 m2 for ground. At the controls' 5% false-positive rate the change-only model recovers 54% of the ground arrays segmentation missed and 40% of the rooftop ones. That rate is against the land around them, not against look-alikes, so this is a corroborating score for an existing candidate set, not a detector.

The unmapped real share. A threshold mixture (pi = (P_unmapped - P_controls) / (P_real - P_controls)) puts it at 0.19 (1,000-5,000 m2), 0.29 (5,000-50,000) and 0.13 (> 50,000) on the change-only score, about zero below 1,000 m2. That assumes a false positive changes like the land around it, so it was checked by a blind review (scripts/aef_render_review_tiles.py): 60 shuffled Esri World Imagery panels under random codes, verdicts written before the key was opened.

Stratum (score above or below the controls' 95th percentile)	PV	No PV	Unsure
Mapped arrays, a check on the reviewer	6	0	0
Displaced controls	0	6	0
Unmapped, above	7	12	5
Unmapped, below	1	23	0

The assumption fails, in the expected direction: the twelve false positives above threshold are motorways, riverbeds, dark rock, forest, an open-pit mine and a factory yard, which score higher than random land. Weighted back by stratum (31% of unmapped ground candidates >= 1,000 m2 sit above), the real share is 0.12 to 0.18 (unsure counted as no, then yes), below the mixture but far from 0.0. The review was a check on the method, by Claude from Esri imagery; it is not a calibration, and the strata were drawn on learned_level.

Two leads came from the review and are post hoc, on 43 decided objects with 8 PV, so they are recorded, not adopted. The change-only score separates them far better (AUC 0.975 against 0.857 with levels, 0.750 for raw change): forest, rock and mountain are static, and only levels let them look like a solar park. And timing separates the rest: 63% of real ground arrays changed mostly after 2021 against 39% of controls, and 75% of the reviewed unmapped PV did, while the motorways, rivers, mine and yard changed mostly before.

What it would mean, not applied. At the table's current recall values, a ground p_unmapped of 0.12 to 0.18 in the bins >= 1,000 m2 adds very roughly 0.44 to 0.66 GWp to segmentation's ground estimate (1,135.6 MWp). Re-deriving that YAML is an owner decision and interacts with the pending same-placement recall fix (two errors that were cancelling), which pushes the other way. The next step is a real review: scripts/aef_ground_review_sample.py has drawn 120 unmapped ground candidates, 40 per tercile of the change-only score, each with its Esri capture date (81 of 120 from 2025 or later), into results/aef_ground_review/review.geojson for JOSM.

Results: results/aef_large_array_change.json (and _scores.csv, _population.parquet), results/aef_ground_review/. Embedding chunks are cached under /home/tobi/earthpv_data/aef/mosaic_chunks/ (16.6 GB).

AlphaEarth against segmentation: a different detector, not a better one (2026-10-22)

The owner's review settles the ground-mount share. The 120 unmapped ground candidates of results/aef_ground_review/review.geojson were reviewed in JOSM on Esri imagery and every validated array mapped into OSM; scripts/aef_review_from_osm.py reads the verdicts back (two agreeing Overpass answers, data base checked to postdate the edits). 67 solar features are new since the July snapshot. Weighted back by tercile of the change-only score, the real share of unmapped ground candidates is 0.26 under the precision table's own 100 m match convention and 0.13 if the new feature must touch the candidate (0.27 and 0.14 at >= 1,000 m2; undecidable old-imagery candidates excluded). Per tercile, strict: high 10 of 29, mid 1 of 23, low 0 of 35. The change score ranks these verdicts at AUC 0.84 (0.89 strict); raw change at 0.53. Claude's blind review had said 0.12 to 0.18, i.e. it was applying the strict rule. Either way the table's 0.0 is measurably too low; at its current recall values 0.14 to 0.27 is very roughly +0.5 to +1.0 GWp of ground-mount. Not applied.

Wall to wall (scripts/aef_vs_segmentation.py --national-negatives). A per-pixel logistic over the frozen embedding (2025 levels and the change since 2017, positives from mapped arrays >= 400 m2, negatives from around them and from 60 random national cells, never within 2 km of an evaluation area), against the production candidate set, on 20 uniformly random density cells, 20 random PV-rich cells and the Rule-1 quadrats. The first run drew negatives only around known PV and fired 300 to 1,300 times per 1,000 km2 of ordinary land; random-land negatives fixed that, which is the difference between a usable and an unusable detector here.

At tau 0.99 (declared: the first AEF threshold whose random-land burden is below segmentation's)	Segmentation	AlphaEarth
Random cells: unmapped detections per 1,000 km2	8.5	5.7
PV-rich cells: precision, mapped lower bound	0.755	0.606
PV-rich cells: precision with the reviewed share of unmapped	0.83-0.88	0.65-0.70
PV-rich cells: recall, ground / rooftop	0.68 / 0.43	0.47 / 0.48
Quadrats, exhaustive: precision	0.952	0.969
Quadrats: recall, ground / rooftop	0.63 / 0.35	0.59 / 0.51

The reviewed share comes from a second blind review mixing 72 unmapped detections from both instruments (scripts/aef_vs_seg_review.py): in the PV-rich cells 7 of 24 of segmentation's were PV against 3 of 24 of AlphaEarth's, so the precision gap is not an artefact of OSM having been mapped from segmentation's leads.

What it adds is different arrays. Arrays found by AlphaEarth and missed by segmentation lift union recall from 0.35 to 0.60 for rooftop and from 0.63 to 0.81 for ground in the quadrats, where its additions are 94% real; in the PV-rich cells from 0.43 to 0.63 and 0.68 to 0.76, where they are about 0.4 real. The gain sits at 400 to 5,000 m2 (quadrats: 0.31 to 0.47 at 400 to 1,000 m2, 0.56 to 0.83 at 1,000 to 5,000 m2). Its false positives are landscape change, roads, rivers, mines, construction, where segmentation's are spectral look-alikes.

Verdict: not a replacement and not a third capacity instrument; two roles. A precision corroborator for segmentation's unmapped candidates, calibrated on the owner's review, and a leads generator for the human-validation workflow, aimed at urban and industrial cells. Its newest year is calendar 2025, so it cannot see the newest arrays, and a national wall-to-wall pass is ~0.5 TB from the Zarr mosaic, or a server-side job on Earth Engine.

Results: results/aef_ground_review/owner_review_summary.json, owner_verdicts.csv; results/aef_vs_segmentation_natneg.csv (and the first run without random-land negatives, aef_vs_segmentation.csv); results/aef_vs_seg_review/.

The capacity chain is a quantification problem (2026-10-22)

A sweep of the levers this register had not yet tried, on the 30 production quadrats with the same paired leave-one-quadrat-out design as above. Every block is scored three ways: ranking (AUC within roof-size band), recall at the precision-0.5 deployment threshold, and the error on each held-out quadrat's true PV area when it is priced the way the capacity chain prices it. The harness recomputes the shipped band means from the composites and matches the production table exactly before adding anything (scripts/roofclf_untapped_extract.py, scripts/roofclf_untapped_ablation.py; results/roofclf_untapped.json and _pass2.json).

The classifier: nothing new ships.

Block	Within size band vs shipped	Folds better	Verdict
Footprint weights convolved with a Gaussian PSF, sigma 0.4 / 0.65 / 1.0 px	+0.0005 / -0.0022 / -0.0144	19 / 11 / 7 of 29	rejected: the area weights already fit
Footprint buffered 1 / 2 / 4 m before area weighting	+0.0002 / -0.0012 / -0.0105	15 / 12 / 6	rejected
VIDA micro-context: counts within 25-100 m, nearest neighbour, relative size, built-up share	-0.0049	10	rejected
The ten nearest buildings' mean features (a spatial lag)	-0.0010	13	rejected
Spectral features x log roof area; all 45 pairwise band indices	-0.0003; -0.0002	14; 14	rejected
L2 0.01 / 100 / 1,000	-0.0009 / -0.0046 / -0.0263	7 / 9 / 7	the shipped L2 = 1 stands
Equal training weight per quadrat	-0.0011	8	rejected
Train without positives under 10 m2 of PV; without negatives within 5 m of a positive	-0.0009; +0.0002	4; 15	rejected
Yard block; segmentation and fraction rasters (re-measured on the current fit)	+0.0004; +0.0002	16; 16	still rejected
Mean logit of the linear model and gradient boosting	+0.0003	16	recall +1.4 points (95% 0.2 to 4.0), nothing else
Difference against composite_1 (2021-10 to 2022-01), per-window median match	+0.0024	20 (p = 0.061)	partial, below

A second epoch is already on disk for Pakistan. composite_1 exists for 4,467 of 4,473 cells, so the download the epoch-difference result priced at a national recomposite is not needed here. Its difference is the one positive block above, and on the capacity metric it helps thresholded pricing (median error 0.457 to 0.367, 20 of 29 quadrats, p = 0.061) but not probability-weighted pricing (19 of 29, p = 0.26), so it is not deployed. Two cautions came with it. composite_1's radiometric scale differs by cell: the 122 cells built on 2026-07-05, which hold 24 of the 30 quadrats, already sit on composite_0's scale, so the documented flat +1000 DN offset makes them 0.1 too bright, and the version using it scores worse (+0.0014). And the 2021 epoch alone reaches 0.7956 within size band against 0.8194, the propensity already recorded.

Gradient boosting, priced end to end. The model-class test left open whether its better per-quadrat adoption rate survives the capacity chain. It does not: per-quadrat PV-area error 0.409 against 0.457 thresholded (12 of 29 quadrats closer, p = 0.46), and 0.245 against 0.242 probability-weighted.

The lever is the aggregation. The capacity chain counts a flagged roof in full and an unflagged one not at all, then rescales by how that count behaved in the quadrats. Rebuilt in the chain's own structure and held out one quadrat at a time (scripts/roofclf_soft_capacity_check.py, results/roofclf_soft_capacity.json), median error on true sub-400 m² PV area:

Pricing, same scores	Median error	IQR	Pooled ratio
No classifier (roof area x ratio)	0.518	0.23-1.84	1.085
Flagged roof area (published)	0.386	0.16-0.83	1.017
Probability-weighted roof area	0.094	0.07-0.67	0.998

Probability weighting is closer in 20 of 29 quadrats (Wilcoxon p = 0.0035), and 18 to 23 of 29 for four other scores and the roof-only label; the flagged version does not beat the null at the median (16 of 29). The mechanism is the sparse end: below 300 bldg/km² the mean score matches the true adoption rate (0.020 against 0.020) while only 0.2-0.3% of roofs cross the threshold, so the flagged version books 12% of the mapped PV there and the weighted one 92%. Through the real chain this moved Pakistan's Best estimate from 24,438 to 34,589 MWp, most of it in cells of 100-1,000 bldg/km², where calibration is thinnest, so it shipped opt-in (--weighting probability) with thresholding as the default.

Adopted 2026-10-25 (owner decision): probability weighting is now the default of both capacity commands, and the published Pakistan atlas uses it. Best estimate 24,438.3 to 34,588.8 MWp (90% 29,308-41,736); sub-400 m² roofclf 13,262.6 to 24,206.3, the ≥ 400 m² replacement 8,256.7 to 7,503.2; the roofclf-AND-SPPI floor stays thresholded by design, so Verified is unchanged at 6,994.1 MWp. 80% of the +10.2 GWp lands in cells of 100-1,000 bldg/km², and against the independent rooftop-solar reference the geography agrees better (Spearman 0.868 to 0.882; 0.773 to 0.861 on hexagons where this project is non-zero). A --weighting threshold rerun the same day reproduced the previous page exactly. The thresholded count stays available as --weighting threshold. Measured on the quadrats but not in sparse national cells: roofclf lists what it trusts and the two measurements that would revert it.

Estimators built for prior shift fail. Treating the capacity chain as the quantification problem it is, the standard estimators for a population whose base rate differs from training were tried on the same scores (scripts/roofclf_quantification.py): EM prior adjustment (Saerens et al. 2002), adjusted counts by TPR and FPR, and their probabilistic version. All three reach a median error near 1.0, collapsing most quadrats to zero. Their assumption is that only the base rate differs between quadrats while the score distribution within each class stays put, and these quadrats differ in roof stock as well.

Temporal unmixing on a full-year stack is dropped from the open list. Its best measured gain was +0.0015 within size band in the most dynamic quadrats, and this sweep puts the binding error in the aggregation, not in another thousandth of ranking.

A finer token grid: THOR at 8 px against TerraMind (2026-10-24)

TerraMind's patch embedding is fixed at 16 px, so at 10 m every token is 160 m across and a 400 m² array is 1/64 of one. ESA and IBM's comparison of TerraMind and THOR (arXiv 2607.18504) found patch size and decoder explain more variance than which model is used, with compact minority classes gaining most from small patches. THOR (ESA Phi-lab and NR, Apache-2.0) accepts any patch size, so the test was whether a finer grid recovers the rooftop arrays of 400-5,000 m² that segmentation misses.

Design. Three THOR-Tiny runs on v14's corpus with v14's recipe (loss, learning rate, schedule, patience, neck tail, UNet decoder), so v14 itself is the baseline and was not retrained. THOR tokenises each GSD group separately, which needed a control that matches TerraMind's grid band for band:

Run	10 m group (B02-B04, B08)	20 m group (B05-B07, B8A, B11, B12)	Tokens	Train cost per batch of 8
v14 (TerraMind-Tiny)	160 m	160 m	196	274 ms
THOR ps16	160 m	320 m	245	133 ms
THOR ps16m (matched grid)	160 m	160 m	392	182 ms
THOR ps8	80 m	160 m	980	614 ms

Scored by scripts/compare_backbones.py on every AOI's own val chips: the six holdouts of the training corpus plus five countries none of the models saw (Poland, Vietnam, India, Morocco, Egypt). Recall is per installation, paired on the same (chip, installation) and tested with an exact McNemar. Each THOR model is also compared at its own lowest threshold whose pooled pixel precision reaches v14's at the production 0.3, because a model that marks more pixels finds more installations: a 50-step smoke model scored recall 0.97 against v14's 0.75 at pixel IoU 0.02.

Result: no pooled recall gain. At matched precision, all 11 AOIs pooled:

	v14	THOR ps16m	THOR ps8
Rooftop 400-5,000 m2 recall (n = 4,633)	0.574	0.573 (p = 0.97)	0.585 (p = 0.08)
Ground >= 400 m2 recall (n = 10,722)	0.933	0.924 (p < 0.001)	0.928 (p = 0.005)
Pixel precision	0.733	0.776	0.782
Pixel IoU	0.633	0.645	0.656

On the matched grid THOR ties TerraMind, and halving the 10 m group's token adds +1.2 points of rooftop recall over it (+496/-441 installations, p = 0.08). The ps16 run with 320 m SWIR tokens lost in every AOI but Morocco (rooftop 0.29 against 0.57): coarse SWIR tokens cost far more than finer visible tokens gain, the same band ranking roofclf's coefficients show. What the finer grid does buy is outline: pixel precision +5 points and IoU +2.3 over v14, consistent at both saved epochs and in 9 of 11 AOIs, which matters to an area-based capacity estimate more than to installation recall.

The regional effects are mostly selection noise. ps8 against ps16m is +18 points on French rooftops and +18 on Pakistani ones, but -9 on South African and -15 on German (54 installations). Within v14's own run, the second-best checkpoint (epoch 49, val mIoU 0.805 against 0.811) has pooled rooftop recall 0.621 against the shipped epoch 41's 0.565, +9 points in South Africa, at pixel precision 0.717 against 0.736. So one run and one selected epoch per model cannot resolve a few points either way. The France gain is the one that held at both saved epochs of both runs (0.57 and 0.54 against 0.42 and 0.43).

Verdict: not adopted. A finer grid does not recover the missed rooftop arrays at a level one run can show, at 2.2 times v14's training cost per batch (614 against 274 ms) and three times THOR ps16's inference cost per window (45 against 14 ms). 4 px was not run: it does not fit at batch 8 on the GTX 1060, an estimated 5 s per batch puts a run at 8-12 days, and the ps8 over ps16m increment it would extend is +1.2 points. Two leads came out of it:

• Checkpoint selection. val/mIoU is dominated by large ground plants' pixels and is a poor selector of rooftop recall. A selection rule on held-out installation recall, from a split that is not the holdout, or averaging the last epochs' weights, is untested and free.
• Pascal attention, and it applies to production. On the GTX 1060 PyTorch's default SDPA kernel (memory-efficient) is about 15 times slower in fp16 than the math kernel (269 against 18 ms per layer at 980 tokens). TerraMind calls the same SDPA: forcing the math kernel takes TerraMind-Tiny from 275 to 97 ms per training batch of 8, with logits within 0.003 and 99.9% of pixels unchanged. Applied to training 2026-10-24 (earthpv.attention; a 50-step smoke run 17 against 36 s). Not to TerraMind inference, which runs one window at a time and gains only 17% (25 against 29 ms per window), while 0.03% of a raster's pixels move by 1-2 uint8 levels (7 of 51,120 detected pixels crossed 0.3 on six South African cells): every published raster stays reproducible byte for byte instead.

Code: src/earthpv/thor_backbone.py (THOR from a third_party/ checkout, input standardisation from scripts/chip_band_stats.py, compile disabled and the math kernel forced on pre-Volta GPUs), configs configs/thor_pv_ps{16,16m,8}_south_africa.yaml, runner scripts/run_backbone_compare.sh. infer uses a stride of 102 for tokens finer than 16 px, since 104 is a multiple of 8. Results: results/backbone_compare/thor16_vs_v14/, thor8_vs_v14_lowthr/, thor16m_thor8_vs_v14/, epoch_noise/.

OlmoEarth embeddings: ahead in Pakistan, behind in South Africa (2026-10-25)

Ai2's OlmoEarth v1.3 annual embeddings (OlmoEarth-v1.3-Base over a calendar year of Sentinel-1, Sentinel-2 and Landsat; 128 dimensions, int8, 10 m on each UTM zone's grid; released 2026-10-22 on Source Cooperative) went through the AlphaEarth head-to-head unchanged: the same evaluation areas, truth, negatives, logistic probe and thresholds, so the embedding is the only thing that differs. One deviation was forced: OlmoEarth covers 2017-2022 for only part of Pakistan (zone 43 from 2017, Multan's zone 42 from 2022, Quetta from 2023), so change is measured from 2023 for both embeddings, and AlphaEarth was re-run on that baseline. That costs AlphaEarth little (PV-rich recall 0.44 against 0.48 from 2017, quadrats 0.52 against 0.54).

Operating points. The declared rule (the first threshold whose random-land burden falls below segmentation's 8.5 unmapped detections per 1,000 km²) picks tau 0.99 for AlphaEarth and 0.995 for OlmoEarth's 64-dimension Matryoshka prefix. The full 128 dimensions never get below it (15.1 at 0.995, the highest threshold tested), so they have no declared point and are shown at 0.995 for reference.

	Segmentation	AlphaEarth	OlmoEarth 64-d	OlmoEarth 128-d
Random cells: unmapped detections per 1,000 km2	8.5	5.7	6.1	15.1
Random cells: of those, PV in Claude's blind review	2-5 of 12	1-4 of 12	7-11 of 13	not reviewed
Random cells: of those, PV in the owner's blind review	not reviewed	not reviewed	10-12 of 13	not reviewed
PV-rich cells: precision, mapped lower bound	0.755	0.679	0.745	0.660
PV-rich cells: precision with Claude's reviewed share	0.84-0.88	0.71-0.73	0.85-0.86	not reviewed
PV-rich cells: precision with the owner's reviewed share	0.86-1.00	0.73-0.89	0.85-0.98	not reviewed
PV-rich cells: recall, ground / rooftop	0.68 / 0.43	0.46 / 0.40	0.83 / 0.67	0.92 / 0.82
Quadrats, exhaustive: precision	0.952	0.971	0.959	0.941
Quadrats: recall, ground / rooftop	0.63 / 0.35	0.53 / 0.51	0.92 / 0.65	0.99 / 0.69
Quadrats: rooftop arrays a detection overlaps	0.12	0.34	0.53	0.59

AlphaEarth is the 2023 baseline throughout except its random-cell review, which is the earlier review of the 2017 baseline. At the same false-alarm budget OlmoEarth 64-d matches segmentation's precision and finds far more arrays: quadrat rooftop recall 0.65 against 0.35, PV-rich ground 0.83 against 0.68. Segmentation adds almost nothing on top of it (the union reaches 0.65 rooftop and 0.95 ground in the quadrats), where AlphaEarth's union with segmentation reached 0.60 and 0.81. In the quadrats 51 arrays are found by OlmoEarth and by neither of the other two, against 6 found only by AlphaEarth. On random land its unmapped detections are mostly real: in Claude's review 7 of 13 are clearly PV and 4 more are unsure, so its false alarms are 0.9-2.8 per 1,000 km² against segmentation's 5.0-7.1; the owner's review puts OlmoEarth's at 0.5-1.4 (segmentation's random-cell detections were not in the owner's sample).

Checked against four ways it could be an artefact:

• Loose matching. A truth array counts as found when a detection lies within 100 m, which rewards blobs painted over built-up land. It does not here. OlmoEarth 64-d covers 0.24% of the PV-rich cells' land against segmentation's 0.35%, 80% of its detected area lies on mapped PV (segmentation 84%), and requiring overlap instead of proximity keeps the ranking: quadrat rooftop 0.53 against 0.12 for segmentation and 0.34 for AlphaEarth, as tabled.
• Pixel leakage. Training tiles touching a 0.02-degree buffer around any evaluation area are skipped, the rule AlphaEarth used.
• OSM labels in pretraining. OlmoEarth's pretraining decodes 30 rasterised OSM layers (as targets only; inference sees imagery alone), among them power_plant, i.e. every power=plant outline including solar parks, and building. Its training tiles were also chosen from 120 OSM categories that include solar plants. So part of its plant-outline recall may be exposure to these very labels: quadrat plant outlines are overlapped at 0.69 against segmentation's 0.22. Solar power=generator was never a target, and rooftop generators still go from 0.12 to 0.53 in the quadrats, so the rooftop gain cannot come from it.
• The reviewer. The 61 panels (scripts/olmoearth_vs_seg_review.py, blind and mixed) were judged twice, by Claude and then by the owner (scripts/build_review_page.py). Fresh draws of segmentation and AlphaEarth read in Claude's review as its earlier review did (5 PV and 3 unsure of 12 against 7 and 4 of 24; 1 and 1 of 12 against 3 and 1 of 24), so the instruments were read alike. The two reviewers never disagree about PV: every panel Claude called PV the owner did too, plus four more. They differ on the rest: of the 29 panels Claude called not PV, the owner marked 22 unsure. So the owner's lower ends match Claude's (OlmoEarth 0.85, segmentation 0.86 in the PV-rich cells) and the upper ends are wider. In the owner's review OlmoEarth's PV-rich detections are 10 PV, 12 unsure and 2 not PV of 24, segmentation's 5, 7 and 0 of 12, AlphaEarth's 2, 6 and 4 of 12; on random land OlmoEarth's are 10, 2 and 1 of 13.

Plausible reasons, untested: a year of monthly multi-sensor input against one dry-season median, and the Sentinel-2 grid itself rather than AlphaEarth's resampled lat/lon grid.

The repeat in South Africa reverses it. scripts/olmoearth_vs_segmentation_country.py ran the same 64-d probe, recipe and operating-point rule against v14's candidates, with the PV-rich cells drawn from the Western Cape, segmentation's val holdout, and the probe barred from that region too, so neither instrument trained there: 20 random national cells and 20 of the 78 holdout cells holding two or more mapped arrays (the country has no quadrats). The rule again picks tau 0.995. Both Claude and the owner reviewed the 96 blind panels; they agree on 78.

	Segmentation (v14)	OlmoEarth 64-d
Random cells: unmapped detections per 1,000 km2	19.3	15.2
Random cells: of 24 unmapped, PV in Claude's review	1-6	1-5
Random cells: of 24 unmapped, PV in the owner's review	2	1-2
PV-rich holdout cells: precision, mapped lower bound	0.287	0.314
PV-rich holdout cells: precision with Claude's reviewed share	0.50-0.55	0.49-0.54
PV-rich holdout cells: precision with the owner's reviewed share	0.53	0.54
PV-rich holdout cells: recall, all / rooftop / ground	0.43 / 0.42 / 0.52	0.22 / 0.22 / 0.20

Precision ties again, at half the recall. OlmoEarth matches segmentation's recall only at tau 0.98 (0.44), at twice its random-land burden (37.3), and adds little in union (0.43 to 0.49). Both instruments are far less precise than in Pakistan: the reviewed false alarms are mostly roofs without panels, fields and sheds, plus ten electrical substations, all of them OlmoEarth's. power_substation is one of OlmoEarth's pretraining OSM targets beside power_plant, which may be why a probe trained on PV lands on them. As a corroborator it still separates: segmentation candidates it confirms are mapped at 0.47, the rest at 0.26.

Local training data does not explain it. In Pakistan both instruments trained on labels from around the evaluated cells (the probe outside a 0.02-degree buffer); in South Africa neither did. A second run (--train-in-holdout) adds the 759 Western Cape training tiles outside that buffer to the first run's 2,500 and national negatives, drawn again from the same seed, with segmentation still held out. The rule then picks tau 0.99, and the probe's recall in the holdout cells moves from 0.22 to 0.26 (rooftop 0.24, against segmentation's 0.42), its mapped lower bound from 0.314 to 0.244. At every fixed tau its recall is slightly lower than without the local tiles. So the gap to Pakistan is the country (its arrays, roofs and confusers), not missing local labels.

Verdict: a corroborator worth knowing about, not a detector to adopt.

• It does not transfer. In South Africa a pixel logistic on the frozen embedding finds half of what segmentation finds at the same precision, even with labels from the evaluated region, and it brings its own false alarm.
• The licence. The OlmoEarth Artifact License forbids military and extractive uses, requires those restrictions to travel with distributed derivatives, and requires citing Ai2. Detections derived from it cannot go into the ODbL atlas or the Apache-2.0 model without an owner decision on how the terms combine. Leads that only point mappers at imagery are the least entangled use.
• Cost. At 64 dimensions and two years it is 1.1 MB per km²: about 1 TB for Pakistan, about 0.5 TB for the building-populated cells. That is roughly a day per 0.5 TB with nothing else on the link; this run fetched 49 GB in 4.5 hours beside a national compose.
• One probe. Only a pixel logistic regression on frozen embeddings was measured, and nationally change is available only from 2023.

In Pakistan, as a corroborator, it confirms 238 of segmentation's 290 PV-rich candidates, at a mapped precision of 0.815, against 0.481 for the 52 it does not. AlphaEarth confirmed 134, at 0.836 against 0.686, so OlmoEarth separates segmentation's real candidates from its false ones better.

Code: scripts/olmoearth_mosaic.py (Zarr v3 shard reader over range requests, no new dependency), scripts/olmoearth_vs_segmentation.py, scripts/olmoearth_vs_seg_review.py, runner scripts/run_olmoearth_compare.sh; scripts/aef_vs_segmentation.py --base-year 2023. Results: results/olmoearth_vs_segmentation.csv (with _unions.csv and .json), results/aef_vs_segmentation_natneg_base2023.csv, results/olmoearth_vs_seg_review/ (the owner's verdicts in summary_owner.json and blind_review_joined_owner.csv, Claude's in summary.json and claude_verdicts.csv). South Africa: scripts/olmoearth_vs_segmentation_country.py (--review, --join), runner scripts/run_olmoearth_country.sh; results/olmoearth_vs_segmentation_south_africa.csv (with _unions.csv and .json; the second run beside it with _holdout_trained) and results/olmoearth_vs_seg_review_south_africa/ (the owner's verdicts in summary_verdicts.json, Claude's in claude_verdicts.csv; the owner's review page is blind/review.html, from scripts/build_review_page.py).

One segmentation model for every country? v18 pooled (2026-10-30)

Every localized checkpoint is v5's corpus plus ONE country, so no model had seen more than one of Poland, India, Vietnam, Egypt, Morocco, Yemen, Nigeria and Zambia, and Germany's chips came from 14 MGRS tiles. v18 pools them: v16's corpus row for row, each of those eight countries' train chips repeated up to 8 times towards about 7,000 rows, and 7,205 German chips cut from the national compose outside the 14 tiles (69,364 rows, 64,967 train; provenance in data/chips/combined_pooled_v18/corpus.json). The recipe is v16's unchanged: one run, best val/mIoU 0.8191 at epoch 18, early-stopped after epoch 26.

Scored against v14 at matched precision. scripts/compare_backbones.py on 14 holdouts, paired McNemar per installation. Its first pass could not match: v18's pixel precision at its lowest scored threshold (0.790 at 0.2) was already above v14's at production (0.755 at 0.3), so its "matched" row was still a more conservative model. The re-score (results/backbone_compare/v18_pooled_vs_v14_lowthr/) adds thresholds down to 0.0003. Pooled over the six matching AOIs, v18 reaches v14's precision at 0.01 (0.764), its epoch 23 at 0.3 (0.758). Over all 14 holdouts:

Checkpoint (threshold)	>= 400 m2	Rooftop 400-5,000 m2	Ground >= 400 m2	Pixel IoU (6 AOIs)
v14 (0.3)	0.816	0.573	0.916	0.622
v18 epoch 18 (0.01)	0.783 (+256/-790)	0.474 (+144/-609)	0.910 (+108/-177)	0.590
v18 epoch 23 (0.3)	0.787 (+219/-688)	0.492 (+137/-517)	0.908 (+79/-167)	0.615

Every row is p < 1e-4 against v14. v18 at epoch 18 never reaches v14's pixel recall at any threshold (0.739 at 0.0003 against 0.780 at 0.3): this is not only a shifted operating point.

Where it loses, matched to v14's own precision in each region:

• South Africa (4,700 installations): rooftop 0.48 (epoch 18) and 0.53 (epoch 23) against 0.62, 333 and 212 rooftop arrays lost net; ground 0.945 against 0.950.
• France (2,217): rooftop 0.37 and 0.31 against 0.42; ground 0.56 and 0.58 against 0.66.
• Poland (4,334): rooftop 0.53 against 0.64 but not matched, since v14 is far less precise there (0.69) than v18 at any threshold. Poland's own v15 makes the same trade (79 of 154 rooftop arrays against v14's 98, results/release_model_benchmark.csv). Ground up to +0.8 points.
• Egypt (23): 13 (epoch 18) and 6 (epoch 23) against 1, the one clear win, on v18's own training data.
• Germany, Pakistan (both holdouts), Punjab, Gujarat, India, Vietnam, Morocco, Zambia and Yemen: no significant difference, except two small losses (German ground, 7 of 399 at epoch 18; pakistan_fresh, 11 net of 314 at epoch 23).

So on the seven holdouts v14 never trained on and v18 did (India, Poland, Vietnam, Morocco, Egypt, Yemen, Zambia), v18 wins one, and v14 transfers as well as v18 does with home data.

Why, as far as one run shows. The losses sit in the two largest holdouts, South Africa and France, which both corpora contain and whose share v18 halved (South Africa 7,092 of 30,458 train rows in v14, 23%, and of 64,967 in v18, 11%). Early stopping on pooled val/mIoU ended v18 at about 19 passes over each South African chip against v14's 42, and epoch 23 is better than epoch 18 on South African rooftops (0.54 against 0.47 at 0.3) at the same val/mIoU (0.8188 against 0.8191). Some of the gap may therefore be dilution and under-training rather than harmful labels. Two harmful-label candidates are untested: the German national chips (OSM holds about 3.6% of Germany's rooftop units, so unmapped roofs train as background) and Yemen's 233 chips repeated 8 times (labels segmentation does not find at all). last.ckpt is epoch 23; epochs 24 to 26 were not saved.

Verdict: rejected as trained. v18 replaces neither the released v14 nor any country's checkpoint. What would separate the explanations: v18 trained on to v14's epoch count without early stopping (resume from last.ckpt, about 11 GPU hours); v18 as a starting point for one country's fine-tune; and a run without the German national chips and the Yemen repeats.

Which training labels teach rooftop PV as background? A per-source audit (2026-10-30)

A chip burns every mapped installation as 1 and everything else as 0, so a source that maps few of its PV roofs teaches the rest as "not PV". v18's entry left that untested. scripts/label_audit.py measures it per source without training anything. The released v14, at threshold 0.5, is run over up to 1,500 chips per source, keeping only chips at least one chip width (2,240 m) from every v14 training chip, where v14 has learned the labels' gaps. Every confident blob of 400 m² or more that lies more than 30 m from any label or ignore region is counted. One with at least a quarter of its area on a VIDA footprint is a roof blob. That is 10,473 chips from 15 sources. Punjab drops out: all 274 of its chips lie inside v14's training footprint.

A roof blob is either unmapped PV or a false positive. Two checks separate them: Germany's register, and Esri imagery of 12 random roof blobs per source (review; crops and the shared review page in data/label_audit/review/). The verdicts so far are Claude's (results/label_audit/review_verdicts_assistant.csv), not the owner's.

Source	Share of v18 train rows	Roof blobs per 100 chips	PV on Esri (of decided)	Unmapped roof PV per 100 chips	Per mapped rooftop array
Vietnam	10.7%	53.3	8 of 11	38.9	~10
South Africa	10.9%	35.5	8 of 11	25.9	0.27
Germany, national chips	11.1%	18.1	8 of 8 (register: at least 52%)	9.4-18.1	0.7-1.3 (register: ~16)
Gujarat	1.0%	23.4	5 of 9	13.1
India	8.0%	19.7	5 of 9	11.0	0.50
Nigeria	1.6%	8.7	9 of 10	7.8
France	26.3%	6.8	7 of 8	6.0	0.10
Germany, 14 tiles	4.7%	10.9	4 of 9	4.8	0.66
Pakistan, fresh chips	8.6%	11.0	3 of 7	4.7	0.22
Morocco	0.6%	13.2	2 of 9	2.9
Egypt	1.3%	5.0	2 of 5	2.0
Poland	11.8%	1.8	0 of 7	0	0
Yemen	2.9%	0.4	0 of 1	0
Zambia	0.3%	0		0
Namibia	not in v18	12.0	none decidable (Esri 2019)

"Unmapped roof PV" is roof blobs times the PV share among decided crops. The last column divides it by mapped installations per chip times the rooftop share of mapped installations of 400 m² or more in that source's holdout. That share is approximate for the chips, and it is left blank where no holdout share was surveyed.

Germany's register settles the national chips without a model. The 1,500 audited national chips hold 3,988 registered rooftop units of 72 kWp or more (about 400 m² of module) commissioned before the compose window. 83% of them sit on pixels the chip trains as background, 8% on a label and 6% within 30 m of one. That is 2.2 per chip, against 1.3 mapped installations of any kind, or about 16 per mapped rooftop array. The 14-tile germany source reads the same (82% on background). v14's roof blobs there lie within 30 m of a registered unit of 30 kWp or more 61% of the time. Undetected, unlabelled footprints of the same size in the same chips do so 19% of the time (mastr-null, land-use matched as in scripts/mastr_p_unmapped.py), so at least 52% of the blobs are registered PV.

What it shows:

• The hypothesis holds for Vietnam and the German national chips. Together they are 21.8% of v18's train rows. Their labels hold almost no rooftop arrays (3% and 11% of mapped installations), while their roofs carry many.
• The rooftop teachers have the same gap, but small: 0.1 to 0.3 unmapped PV roofs per mapped rooftop array in France, South Africa and Pakistan, and 0.5 in India.
• Poland's labels are clean of it. After France it is v18's largest source (11.8%), but its few roof blobs are mostly not PV (0 of 7 decided, on Esri captures from 2020 to 2024). Whatever it cost v18 is dilution.
• Morocco's roof blobs are mostly roofs with a grid of skylights (7 of 9), a false-positive class worth hard negatives.
• Yemen has the opposite problem: v14 finds 2 of the 2,454 labels in its chips, so Yemen holds labels no detector here can see.

Limits. These are floors: v14 finds about 57% of rooftop arrays, and blobs within 30 m of a label are not counted. Twelve crops per source put about ±25 points on each PV share. Esri predates the composite in Namibia (2019), Poland (2020 to 2024) and parts of Pakistan.

Verdict. The harmful-label explanation of v18's rooftop loss is supported for two sources and refuted for Poland. It does not rule out the dilution and under-training explanations in v18's entry. A second pooled corpus will burn unlabelled building footprints as ignore in Vietnam and in the German national chips, plus an ignore disc around every unmapped register unit of 30 kWp or more in Germany. India and Gujarat probably need the same; Poland keeps its labels as they are.

Feeding TerraMind the input it was pretrained on: v21 (2026-10-30)

TerraMind was pretrained on Sentinel-2 L2A standardised per band with fixed statistics, but TerraTorch 1.2.8's TerraMind ViT backbones (terramind_v1_tiny and its siblings) do not apply them; only its generation and TiM builders do. Every segmentation checkpoint up to v20 was fed reflectance (DN / 10000), so each band's spread was 5 to 8 times narrower than in pretraining. TiM variants were not an alternative: they refuse a band subset and need 12 bands, where the composites have 10.

The probe said it mattered. A frozen-feature probe on v14's corpus (logistic regression on the four encoder layers the neck selects, one prediction per 16 px token, 1,487 training and 754 held-out chips) measured what the pretrained encoder can read before any fine-tuning:

Encoder input	Held-out token average precision
DN / 10000 (every checkpoint so far)	0.469
TerraMind's own standardisation of raw DN	0.737
The same, 1,000 DN BOA offset removed first	0.696
The corpus's own statistics (what thor_backbone applies)	0.452
Randomly initialised weights, DN / 10000 and standardised	0.231, 0.259

It improved in all six training AOIs (France 0.115 to 0.553, South Africa 0.426 to 0.718).

The test. v21 is v14 with one change: backbone: earthpv_terramind, which standardises each band with TerraMind's statistics inside the model. Same pretrained weights, bands, neck, decoder, loss, learning rate (1e-4, constant), patience and corpus (data/chips/combined_south_africa). It was scored as THOR was (above): 11 holdouts, paired McNemar per installation, each checkpoint at its own threshold reaching v14's pooled pixel precision (0.755 over the six training AOIs). Since one selected epoch moves rooftop recall by several points, both of v21's saved epochs were scored against both of v14's.

v21 converges twice as fast. Its best val/mIoU, 0.8132 at epoch 20, is above v14's 0.8109 at epoch 41, and early stopping ended it after epoch 28 against v14's 49.

It does not find more installations. All 11 AOIs pooled, at matched precision:

	v14 epoch 41 (shipped)	v14 epoch 49	v21 epoch 20	v21 epoch 27
Threshold	0.3	0.8	0.5	0.1
Rooftop 400-5,000 m2 recall (n = 4,634)	0.574	0.610 (p < 0.001)	0.575 (+317/-312, p = 0.87)	0.576 (p = 0.67)
Ground >= 400 m2 recall (n = 10,725)	0.933	0.929 (p = 0.003)	0.926 (+75/-142, p < 0.001)	0.925 (p < 0.001)
Pixel precision	0.733	0.760	0.754	0.763
Pixel IoU	0.633	0.644	0.650	0.654

v14 epoch 49 does not quite reach the target at the highest scored threshold (0.754 at 0.8). The pixel IoU gain comes mostly from Morocco, where the shipped epoch alone is imprecise (0.68 against 0.91 to 0.96 for the other three).

The regional moves are mostly epoch noise. v21 gains rooftop arrays in South Africa (0.66 and 0.65 against 0.62, p < 0.002) and Punjab (0.69 and 0.71 against 0.66) and loses ground plants in Poland (0.960 against 0.970), but v14's own epoch 49 makes all three moves, each by more (South Africa 0.69, Punjab 0.71, Poland 0.955). The one move that is v21's own, at both saved epochs, is a loss in France: rooftop 0.35 and 0.38 against 0.42 (+63/-165, p < 0.001), ground 0.62 and 0.60 against 0.66, where v14 epoch 49 differs from epoch 41 by about one point either way. Elsewhere the differences are small: Pakistan's ground plants rise at both v21 epochs (0.81 against 0.77) as they do at v14 epoch 49, German ground falls by 7 of 399 at epoch 27 only (p = 0.02), and Gujarat, Vietnam, India, Morocco and Egypt show none.

Verdict: rejected as a model change. The probe measured a frozen encoder. Fine-tuning every layer at 1e-4 lets the encoder adapt to whatever scale it is fed, and here it adapts fully; the input only shortens the way there. v21 replaces neither the released v14 nor any country's checkpoint, and no published number changes. The standardisation could still matter where the pretrained features have to carry more: with a lower encoder learning rate, a frozen encoder, or a larger backbone fine-tuned briefly. That is untested, and worth trying only together with a learning-rate schedule (constant 1e-4 with early stopping is the recipe every checkpoint used). The faster convergence is real but small in hours: 20 fewer epochs at about 9 minutes each on the GTX 1060, so v21 trained in 4.5 hours.

Code: src/earthpv/terramind_backbone.py (earthpv_terramind: the statistics as buffers, in the units the model receives, so they travel with the checkpoint and datamodule, infer and evaluate are unchanged; every existing checkpoint loads and predicts exactly as before), config configs/terramind_pv_v21_tmnorm_south_africa.yaml, runner scripts/run_tmnorm_compare.sh. Checkpoints in /home/tobi/earthpv_data/models/v21_tmnorm_south_africa/. Results: results/backbone_compare/v21_tmnorm_vs_v14/.

Training on arrays down to 100 m²: v24 (2026-11-01)

The owner asked whether training segmentation on arrays from 100 m² up, instead of 400 m², would find more of Pakistan's small rooftop installations, and find them better. chips.MIN_PV_AREA burns every array below 400 m² as ignore, so no checkpoint has ever had a gradient from one.

The test. v24 is v16 with one change. In Pakistan's training chips (pakistan_fresh and punjab), the 4,048 mapped arrays of 100 to 400 m² in the 2026-10-25 pull are burned as PV, all-touched as the large ones are, in 713 of 3,064 chips (3.9% more PV pixels). Arrays below 100 m² stay ignore, and so does every array touching a registered calibration area (2,633), so the areas are a clean test for both models. Corpus, images, splits and recipe are v16's; val and buffer masks are unchanged, so val/mIoU picks epochs on the same terms. Both of v24's saved epochs (26, its best val/mIoU, and 33) were scored against both of v16's (30, shipped, and 38), on two sets neither was trained on:

• 44 calibration areas, Rule-1 complete (every visible panel mapped; Jhang f2, corroborated only, is left out), on production rasters: 2,614 rooftop arrays of 100 to 400 m² and 12,633 below 100 m².
• v16's Pakistani holdout chips (the Multan and Punjab val chips and their buffer rings, clear of every training chip): 289 rooftop arrays of 100 to 400 m². The map is incomplete there, so precision is a floor.

Each checkpoint is read at the lowest threshold reaching v16's pixel precision at 0.3 on that set, with an exact McNemar test per installation against v16 epoch 30 (scripts/small_array_eval.py).

It does not find more small arrays. Calibration areas, at matched pixel precision (0.576):

	v16 epoch 30 (shipped)	v16 epoch 38	v24 epoch 26	v24 epoch 33
Threshold	0.3	0.15	0.6	0.9
Rooftop 100-400 m2 recall (n = 2,614)	0.0375	0.0375 (+32/-32)	0.0318 (+15/-30, p = 0.04)	0.0360 (+19/-23, p = 0.64)
Rooftop below 100 m2 recall (n = 12,633)	0.0012	0.0010	0.0016	0.0013
Ground 100-400 m2 recall (n = 221)	0.104	0.095	0.086	0.109
Rooftop 400 m2 and up recall (n = 512)	0.607	0.594	0.582 (p = 0.047)	0.617
Outlines of 400 m2 or less touching mapped PV	21 of 34	18 of 29	15 of 25	10 of 35

On the holdout chips (matched pixel precision 0.569), rooftop 100-400 m² recall is 0.059 for v16 epoch 30, 0.024 for its epoch 38 (+1/-11, p = 0.006), 0.083 for v24 epoch 26 (+10/-3, p = 0.09) and 0.059 for v24 epoch 33 (+7/-7). That one hint of a gain rests on 13 arrays and one epoch, and the calibration areas, nine times the sample, show the same epoch losing.

At the production threshold it looks better, misleadingly. At 0.3, v24 epoch 33 finds 0.052 of the areas' small rooftop arrays (+52/-14) and epoch 26 finds 0.043 (+30/-15). But both flag more pixels at every threshold: 27,925 and 35,564 in the areas at 0.3, against v16's 24,824 (12% and 43% more). Raised to v16's precision, the gain is gone. What the labels taught is to fire more readily, not to see a 100 m² array in a 10 m pixel.

Arrays of 400 m² and up. On Pakistan's three holdouts, as v16 was shipped (compare_backbones.py, pakistan_fresh, pakistan, punjab), v24 epoch 26 finds more at matched precision: 0.756 against 0.740 of 1,318 installations (+52/-31, p = 0.03), mostly ground-mount (0.790 against 0.768, +27/-13). Its epoch 33 ties (0.735, +48/-54), and in the calibration areas epoch 26 finds fewer large rooftop arrays than v16 (table). v16's own epoch 38 reaches 0.691 on the same holdouts (+26/-90), so the choice of epoch moves this number more than the change does.

Verdict: rejected. At equal precision v24 finds the same 3 to 4 in 100 rooftop arrays of 100 to 400 m² as v16, or fewer, and nothing below 100 m². Small rooftop PV stays roofclf's job; v16 stays Pakistan's checkpoint and no published number changes. The one positive, the 400 m² holdouts at epoch 26, does not survive the second epoch or the calibration areas, so it is no case for a swap. Only the labels changed here: a model built for small objects (finer tokens, a loss weighted toward them) was not part of this test.

Code: scripts/lower_floor_corpus.py (re-burns a corpus's training masks from a lower floor, calibration areas held out), scripts/small_array_eval.py (both test sets, every threshold from one inference pass, bins down to 0-100 m²), config configs/terramind_pv_v24_pakistan_floor100.yaml, runner scripts/run_floor100_v24.sh. Checkpoints in /home/tobi/earthpv_data/models/v24_pakistan_floor100/. Results: results/small_array_eval/v24_floor100_vs_v16/ and results/backbone_compare/v24_floor100_vs_v16/.

The partial result worth watching

Time-series step detection below the floor

At a median installation area near 50 m², the trained model reads 0.000 probability on 99.8% of true footprints and glint validates zero of 1,021. Static appearance is exhausted. But appearance in time is a different signal: a panel installed in 2023 is a step change in a dense per-pixel Sentinel-2 series, even if no single scene shows it.

scripts/pv_step_signal.py removes common-mode atmosphere against reference pixels, guards co-registration by phase correlation, learns the PV installation change vector spectrally rather than assuming a fixed index, deseasonalises with annual harmonics and per-orbit offsets, and scans for the best breakpoint per pixel.

The good part: Area under the ROC curve of 0.875 and 0.74 on held-out halves, against 0.50 for the model on the same footprints. That is the first non-zero discrimination anyone here has achieved below 500 m².

The part that failed: Converting that into a city-scale unmapped-capacity number was rejected by its own control. The method's estimate of unmapped area per built-up pixel sat inside the false-area floor measured on two PV-free cropland control cubes, so nothing in its totals block is quotable as capacity. What survives is the ranking: step-leads are defensible as a lead ordering.

Two landmines are documented for anyone continuing: a propensity confound (the households that install panels differ systematically from those that do not, in ways visible from space) and duplicated Sentinel-2 baseline products that silently double-count dates.

Deeper write-ups

These pages hold the full tables and derivations behind the rows above. Each one carries a status banner stating where it now stands relative to the current pipeline, because several of them were written before the thing they describe was replaced.

Write-up	Subject
Calibration boxes	The running log of every ground-truth area, how each was mapped and what it changed
roofclf national deployment and temporal features	The first national scoring run, and the epoch-jump and step-change feature tests
roofclf cell-edge false positives	Two compounding bugs that made 45.6% of national flags artifacts, and both fixes
Quadrat-supervised fraction retrain	The in-sample win, the holdout that did not support it, and the block bootstrap
Fraction-head hard-negative retrain	A regime-specific win: better on large arrays, worse on dense small rooftops
SPPI spectral index evaluation	What SPPI can and cannot do, and why it is used only as a second opinion
Boom-window stacking	Why the 2021-versus-current retrain is inconclusive rather than negative
Standard-pose matched filter	Assessed against real pose data and not recommended
Glint PSF and matched filtering	The measured Sentinel-2 glint PSF, why filtering on it loses to the aperture statistic, and the verified-negative false-spike rate
Glint spike-rate estimator	Glint as a statistical density estimator, and the floor it hits
Glint tile-batched coverage	The 22x speedup, and the silent scene-loss bug it uncovered
Glint-validated training labels	A proposal to feed corroborated detections back as labels
Quadrats as training data	The proposal that became roofclf
Calibration imagery dating	Why ground-truth completeness is relative to an imagery date, and what closing that costs