Proof

Scored against what actually happened. Misses included.

Every number below was computed after the fact, from independent evidence, without tuning anything to it. Where AURORA is weak, it says so.

Roads under water: forecast against Sentinel-1 radar

Montha, IMD Bulletin No. 21 run. Radar flood maps from the first Sentinel-1 passes after landfall (78 h, 102 h after it) were compared with every road segment AURORA had forecast to close. 3,65,573 segments were observed; 206 showed water.

How to read this

The first Sentinel-1 pass came 78 and 102 hours after landfall, when much of the water had drained, and it found water at only 206 observed road segments. This is a weak test of a landfall-day forecast, and AURORA's false-alarm rate against it is very high.

Flagging the same number of segments, AURORA caught more flooded segments than the lowest-ground baseline (17 against 16) and more than the nearest-to-track baseline (17 against 0).

Skill scores per district and method
District	Method	Hit rate (POD)	False alarms (FAR)	CSI	Hits · misses · false alarms
Kakinada	AURORA (≥50% of futures)	0.00	1.00	0.00	0 · 8 · 1707
	IMD track alone	0.75	1.00	0.00	6 · 2 · 11113
	Baseline: nearest the track	0.00	1.00	0.00	0 · 8 · 1707
	Baseline: lowest ground	0.13	1.00	0.00	1 · 7 · 1706
Konaseema	AURORA (≥50% of futures)	0.71	0.99	0.01	5 · 2 · 904
	IMD track alone	0.00	1.00	0.00	0 · 7 · 2047
	Baseline: nearest the track	0.00	1.00	0.00	0 · 7 · 909
	Baseline: lowest ground	0.00	1.00	0.00	0 · 7 · 909
West Godavari	AURORA (≥50% of futures)	0.23	1.00	0.00	3 · 10 · 3420
	IMD track alone	0.46	1.00	0.00	6 · 7 · 9123
	Baseline: nearest the track	0.00	1.00	0.00	0 · 13 · 3423
	Baseline: lowest ground	0.08	1.00	0.00	1 · 12 · 3422
East Godavari	AURORA (≥50% of futures)	0.00	1.00	0.00	0 · 7 · 1496
	IMD track alone	0.29	1.00	0.00	2 · 5 · 6207
	Baseline: nearest the track	0.00	1.00	0.00	0 · 7 · 1496
	Baseline: lowest ground	0.00	1.00	0.00	0 · 7 · 1496
Eluru	AURORA (≥50% of futures)	0.15	1.00	0.00	7 · 40 · 2374
	IMD track alone	0.28	1.00	0.00	13 · 34 · 12906
	Baseline: nearest the track	0.00	1.00	0.00	0 · 47 · 2381
	Baseline: lowest ground	0.19	1.00	0.00	9 · 38 · 2372
Krishna district	AURORA (≥50% of futures)	0.02	1.00	0.00	2 · 122 · 818
	IMD track alone	0.56	1.00	0.00	70 · 54 · 18934
	Baseline: nearest the track	0.01	1.00	0.00	1 · 123 · 819
	Baseline: lowest ground	0.02	1.00	0.00	3 · 121 · 817

observed flooded: midpoint flooded, or at least 2 of 5 points along the edge
predicted closed: closed by the observation time in at least 50% of weighted members
after horizon: passes after the 85 h run horizon are compared with the state at the horizon
Sentinel-1 sees standing water at the pass time; it can miss the flood peak between passes.
Radar under-detects flooding in built-up areas and under tree cover, so hits there are undercounted.
A flooded pixel at a road is evidence of water near the road, not proof the road was impassable.
The model closes roads for rain flooding and storm surge only, and never reopens them.

Contains modified Copernicus Sentinel data [2025]

Reading IMD bulletins: Gemini against hand-checked labels

Each bulletin was read on the deployed service. Agreement is field by field with a reading entered by hand from the PDF; those labels are drafts awaiting a second person's confirmation.

Montha · National Bulletin No. 19

16 pages · gemini-3.7-flash · bulletin-v2

48/49

fields agree with the labels; differs: rainfall categories and coverage

✓ passMatches the reading schema (all required fields present)
✓ passPositions inside the North Indian Ocean (11 positions)
✓ passLead times increase (0, 6, 12, 18, 24, 36, 48, 60, 72, 84)
✓ passFix-to-fix speed at most 60 km/h (fastest 16 km/h)
✓ passWinds consistent with the stated category (all rows consistent)
✓ passEvery numeric element has a quote (33/33 covered)
✓ passQuotes found verbatim in the PDF (78/78 found in the text layer)
✓ passTrack and winds match an independent table parser (10/10 table rows matched)
– n/a+24 h position within 150 km of the ensemble median (checked when a storm run uses this reading)

Montha · National Bulletin No. 21

17 pages · gemini-3.7-flash · bulletin-v2

48/49

fields agree with the labels; differs: rainfall categories and coverage

✓ passMatches the reading schema (all required fields present)
✓ passPositions inside the North Indian Ocean (11 positions)
✓ passLead times increase (0, 6, 12, 18, 24, 36, 48, 60, 72, 84)
✓ passFix-to-fix speed at most 60 km/h (fastest 16 km/h)
✓ passWinds consistent with the stated category (all rows consistent)
✓ passEvery numeric element has a quote (35/35 covered)
✕ failQuotes found verbatim in the PDF (76/79 found in the text layer)
✓ passTrack and winds match an independent table parser (10/10 table rows matched)
– n/a+24 h position within 150 km of the ensemble median (checked when a storm run uses this reading)

Live · IMD National Bulletin No. 4 (2026, BOB/07/2026)

8 pages · gemini-3.7-flash · bulletin-v2

–

no labels (live bulletin)

✓ passMatches the reading schema (all required fields present)
✓ passPositions inside the North Indian Ocean (7 positions)
✓ passLead times increase (0, 6, 12, 18, 24, 36)
✓ passFix-to-fix speed at most 60 km/h (fastest 22 km/h)
✓ passWinds consistent with the stated category (all rows consistent)
✓ passEvery numeric element has a quote (13/13 covered)
✓ passQuotes found verbatim in the PDF (41/41 found in the text layer)
– n/aTrack and winds match an independent table parser (no national-bulletin table found)
– n/a+24 h position within 150 km of the ensemble median (checked when a storm run uses this reading)

Not validated yet

Power: VIIRS night-lights outage scoring is not done; power results are not shown.
Storm surge: the screening model is not yet compared with IMD's surge guidance and observations.
Closures reported in state situation reports are not yet matched to the forecast.
Parameters still marked prior (uncalibrated): rain_k_m_per_mm, rain_r0_mm, rain_h_max_m, rain_coverage_fractions, rain_patch_h3_res6, road_formation_allowance, close_depth_0.3m, surge_decay_0.083m_per_km, surge_sigma_2rm, surge_timing_rm_plus_50km, holland_b_1.5, district_buffer_30km, settlement_min_pop_25, publication_delays, source_balanced_weights.