halemarshal/b41-bangkok

B41 (25 Mt) over Bangkok: weapons-effects model, renders and thermal-shadow analysis

Package built 2026-11-02; Asok cut corrected 2026-11-03 (round 3); round 4 on 2026-11-03: single-deck viaduct, δ = asin(R/D) everywhere, the street-point and 5 km cells re-solved with the fireball disc, MP4 re-encoded. Physics only: blast, thermal, fire, fallout, EMP, crater, and the thermal-shadow analysis at three sight-line points plus the Asok street cut. No human depiction.

Location record

• GZ: 13.746000, 100.557000 (b41_bangkok.py line 39). Nearest named road: Soi Sukhumvit 13 Yaek 1-1, 46.6 m.
• Street point: 13.746044, 100.557404, on Soi Sukhumvit 13 Yaek 1-1. It is 43.96 m from GZ, which bears 263.643° from it.
• 5 km point: 13.751056, 100.603407. It is 5,050.0 m from GZ and stands on land, 2.4 m from a mapped water edge.
• Far point: 13.762012, 100.703963. It is 15,992.2 m from GZ; φ is 29.1699°, derived at the point.
• Asok cut (camera B): 13.735666, 100.562157, 1.7 m above Sukhumvit Rd, 9.3 m SSW of the viaduct midline, heading 306.118°, pitch +16°, HFOV 90°. Ground standoff from GZ 1,271.48 m in the render's local frame (1,272.15 m WGS84 geodesic); GZ bears 333.996° from it. At HOB 8,927 m: φ = 81.8936°, slant D = 9,016.79 m, δ = asin(3.0 km / D) = 19.4335° (atan gave 18.4029°), φ + δ = 101.3271° (past vertical; tan(φ + δ) = −4.9922), Q_ff = 790.8329 cal/cm².

Contents

video/

• b41_bangkok_render_v3_1080p60.mp4: current MP4 (1080p, 60 fps, 135.5 s, 8,130 frames, H.264), encoded from the v3 scrubber on 2026-11-03. Cut 1b shows the single-deck viaduct, the disc-integrated shadows and the corrected label; the end card is the round-4 one.
• b41_bangkok_render_v3.html: the self-contained scrubber the MP4 was encoded from, with an optional total-fluence overlay for cut 1b. Frame rate, camera geometry, timeline and every other cut are as in v2.
• withdrawn/: the v1 and v2 MP4s and scrubbers. Cut 1b in them draws the viaduct as two 9.2 m boxes (15.9 m) with a point-source binary shadow mask; the scrubbers' label text is neutralised, the MP4 frames are not.

figures/current/

• asok_still_v3.png: Asok cut at T+11.00 s, single-deck geometry, total-fluence overlay (log scale, cyan ≈38 cal/cm² charring line), ground gauge across the viaduct's shadow-side band (0 / 38 / 432 / 760 cal/cm²) and the label "BTS viaduct thermal shadow: penumbra, graded but still charring: 0 → 760 cal/cm² across a 7.8 m ramp; above the ≈38 charring line beyond the inner 0.7 m; no hard edge at this range: fireball disc δ 19.4°; umbra only under the deck (direct beam; scatter not modelled)." asok_still_v3_plain.png is the same frame without the overlay.
• end_card_v3.png: end card: shadows penumbral, δ = 19.43° at the cut, φ + δ = 101.3°, viaduct smear 7.8 m (shadow side) and 7.4 m (GZ side), no hard shadow at this range. Frame rate, camera geometry and imagery credit lines unchanged.
• field4_0_street.png, field4_1_5km.png, field4_2_15km.png: one figure per sight-line point, facades pulled at each coordinate, re-solved in round 4 by integration over the fireball disc (256 directions per 1° × 1 cm surface element). The 15.9 km cell remains superseded by far5. field4_provenance.png: the provenance table (unchanged figure; see corrections for its step-6 row).
• far5_AB.png, far5_C.png, far5_provenance.png: the far cell, re-run with δ = asin(R/D).
• b41_bangkok_effects_v2.png: effects figure (Kingery-Bulmash blast, glass, fire network, fallout).

figures/withdrawn/: earlier figures, stamped WITHDRAWN.

results/

• results.json, sensitivity.json, run_log*.txt, fields.npz: model outputs and the sensitivity sweep.
• provenance_all_elements.csv, radial_walk_stretches.csv, solve3_report.txt, ob_sample.log, corrected_captions.txt.
• Asok cut: asok_cut.json (pinned geometry), asok_casters.json (bounds, band fluences, spreads, deck variants, convergence), asok_caster_table.csv (two blocks, 4 dp, header notes), asok_deck_profiles.json, asok_convergence.txt. The per-pixel field (asok_field_s1.npz, ~30 MB) is not shipped; step 15 regenerates it.
• Sight-line cells: street_disc.txt (onsets, ladders, convergence, from tools/street_disc.py).
• mp4_check.txt: frame count, OCR and end-card comparison for the v3 MP4, with the v2 MP4 as a positive control.

data/: derived inputs needed to re-run the code, plus street_disc.pkl. The large third-party sources are not included; see Sources.

code/: model (b41_bangkok.py, kb_blast.py), tools/ (extraction, walk solve, DSM calibration, Open Buildings sampling, far-cell penumbra, sight-line disc solve, Asok penumbra) and render/ (pre-render, compositors, capture, figure scripts).

Round 4 corrections (2026-11-03)

Street point: the premise was wrong for ordinary stock. field4_0_street.png read "NO SHADOW, closed by geometry" from a point-source test: the nearest element on the centre ray needed 1,881 m. With the disc (δ = 19.64°, φ + δ = 109.4°, the disc straddles the zenith, so every azimuth counts), MS:132203311:1278754 (UNKNOWN; GLO-30 lower bound 5.1 m; nearest edge 3.7 m) shades the figure's feet if H ≥ 9.8 m, against 1,881 m: 192× lower. Round 3's infinite-edge estimate of 27.1 m was also too high, because the footprint's nearest edge is off the centre ray. A three-storey shophouse next door is enough; this is ordinary low-rise stock, not towers. The cell is OPEN, not closed. No known-height element reaches. The second element on the ray, way/339617803, needs ≥ 38.0 m (point-source 4,180 m); its calibrated estimate is ≈30 m, so it does not reach the figure and is carried as undetermined, not as a near miss.

• Burn class on the figure changed. Point-source gave source-facing 4.0 (1st°), 45° flank 2.8, tangential 0.7, rear 0. The disc gives 56–60 cal/cm² on every side, crown 786: above the ≈38 charring line all round.
• If MS:132203311:1278754 is at its calibrated GLO-30 estimate (5.1 m / 0.2405 = 21.3 m): penumbra, graded but still charring (weakest element 39.1 cal/cm²). Part of the figure drops below the charring line once that element passes 22.4 m, so at the estimate the figure is barely charring, not safe.

5 km point. The cell is decided by an element off the centre ray: MS:132203311:258923 (UNKNOWN, 24.7 m away) shades the feet if H ≥ 30.5 m. The on-ray MS:132203311:100687 needs 53.4 m (point-source 100.3 m). The figure's 90° flank takes 38 cal/cm² unshaded (point-source 0). Charring line crossed on the figure once MS:132203311:258923 exceeds 36.2 m.

Asok deck geometry fixed. OSM maps each BTS track as its own way (6.70 m apart at the cut) and render_B2 built a 9.2 m box on each, drawing a 15.9 m deck. render/prerender.py bts_structure now builds one 9.2 m deck on the midline between paired track centrelines, one pier row every 30 m, station box within 75 m of the junction. Camera B was re-rendered with it (same camera; the pipeline reproduced the v2 base to a mean 0.23/255 with the old geometry first). The still, table, scrubber and MP4 all use the corrected geometry. The withdrawn two-box row stays in asok_casters.json for the record.

• The table's row set follows the geometry. Casters are listed when their penumbra lands on ground visible in the frame (≥ 100 px). The narrower deck exposed more street and added osm_bld#157822 and #157827; that, with the asin switch, moved the vertical casters' base fluence from 163–262 to 169–288 cal/cm². A later deck or camera change can add or drop rows the same way without the table being wrong.

δ = asin(R/D). A sphere of radius R at distance D subtends asin(R/D); the package used atan(R/D). Switched in tools/asok_geom.py, tools/street_disc.py and tools/penumbra.py. Deltas, same geometry: Asok δ 18.40° → 19.43°; deck smear 7.36 → 7.78 m (+0.42 m; 7.80 m on the rendered deck); GZ side 7.00 → 7.44 m; umbra 2.80 → 2.42 m; charring line 0.64 → 0.68 m in; mid-band 434.3 → 433.7; far bound 762.4 → 760.1 cal/cm². Vertical casters: far bounds +4.4–4.6% (Interchange 21 80.43 → 84.15 m), base fluence +6–10 cal/cm². Far cell: δ 9.303° → 9.428°, onset 2.833 → 2.814 m, ankle at the Open Buildings height 80.1 → 80.2 cal/cm²; far5_provenance.png unchanged.

• Scaling checks under the switch. Asok deck: the charring crossing sits 0.64 m into a 7.36 m band (8.70%) with atan and 0.68 m into 7.80 m (8.72%) with asin, so it scales with the band. Street point (a separate cell, 44 m from GZ; its 22.4 m is a building height, not a distance into the Asok band): the onset falls 10.39 → 9.79 m but the charring crossing rises 21.57 → 22.35 m, because the wider disc puts more of its low edge on the figure's vertical flanks (weakest unshaded flank 53.1 → 56.2 cal/cm²), so it takes more building to push them below the line. At the 21.3 m estimate the figure stays above the line under both conventions (38.7 atan, 39.1 asin).

The 30 cal/cm² figure. Round 2's README said the viaduct shadow was "a penumbra receiving about 30 cal/cm²". Nothing in the package derives it: no script, log, report or caption computes a fluence for the strip; the v2 compositor drew it as zero. The only nearby number is the fluence at the old label's anchor pixel on the withdrawn two-box deck (21.9 cal/cm² with atan, 34.5 with asin): an anchor on the ramp, not a band statistic. Dropped.

Calibration error in round 3, mine. Round 3 wrote "≈52 m calibrated" for way/339617803. The provenance table divides the GLO-30 reading by the 10–30 m median ratio, 7.2 / 0.2405 = 30 m; 52 m used the all-buildings median. With the disc onset at 38.0 m, the 30 m estimate does not reach the feet; the element stays undetermined.

Sampling budget, per regime. Images: 256 disc directions per pixel (Asok frame) or per surface element (field4 cylinders). Numbers: 4,096 directions for every table and caption value, within 1.1% of 16,384 across the Asok band (1.5 cal/cm² at worst). The 1.3% bar at 128 directions holds in the median (0.76% on the street-point figure, 0.88% on the Asok band) and fails only where the visible sliver of the disc is thin: 1.39% at the worst grazing element of the street figure, up to 16.6% at the Asok band's inner edge, where fluence is small (14.8 cal/cm², 1.9% of Q_ff). That is sampling error on small values, not a table error; results/asok_convergence.txt and results/street_disc.txt have the full statistics.

Round 3 corrections (2026-11-03, Asok cut)

• Cut 1b used a point-source binary shadow map. Fluence is now Q_ff × (1/N) Σ vis_k · max(n·d_k, 0) over the disc seen from each pixel, from ray casts against the scene the render builds. No mask, no blur.
• Past vertical (φ + δ > 90°): the code tests φ + δ ≥ 90°, sets the near bound of every vertical caster to its base and logs the raw (negative) infinite-edge value. Vertical casters have no umbra: fluence at the base is 169–288 cal/cm² (round 4 values). The elevated deck has a true umbra (direct beam only). See the table header for why the two classes are kept apart.

Round 2 corrections (2026-11-02)

• Burn classes: every field figure before that set labelled classes one step too severe; fluence values were correct.
• Street point: it is not GZ and not on the soi it was earlier labelled with.
• GLO-30 fails calibration as a building-height source (median 0.139× of tagged height, 664 buildings; 0.030× for towers ≥ 60 m); used only as a lower bound. Open Buildings 2.5D Temporal v1 passes (median 0.977×, Pearson r 0.822) and over-reads low-rise buildings against levels × 3.2.

Model limits to read with the results

• Building heights. 434,201 of 459,831 OSM footprints (94.4%) carry no height or levels tag; their heights, the A–E bands drawn in cut 1 and the Asok casters marked UNKNOWN use type/area defaults. The casualty model's band stock is an assumed radial split, not these footprints, so the fragility sweep (medians × 0.7–1.3; 12:00 prompt deaths 8.01–9.56 M) covers building response, not a misassigned stock.
• Shelter split. The fallout PF mix (urban 32/38/20/9/1% at PF 2/4/10/40/1000, no basements) is an assumption written into the code, with no survey behind it; its effective urban PF is 3.6. The v1 mix (basements allowed, PF 6.5) gives 3.59 M fallout deaths against 5.62 M. Sweep #1, shelter PF × 0.5–2 (the largest swing, 72%), moves the base to 3.49–7.51 M with the split held fixed. Together they span 1.36–7.51 M, a 5.5× range the sweep alone does not show.
• Surface burst. At 12:00 the surface burst kills fewer people promptly (6.70 M against 8.69 M) and more in total once 7-day fallout is counted (6.70 + 5.62 = 12.32 M against 8.69 M, wet-season base case). The ranking holds across sweep #1 on the base split and flips only at the protective corner (basement split with PF × 2: 8.05 M).

Reproduce

Needs Python 3 with numpy, scipy, shapely, pyproj, rasterio, matplotlib, pyosmium and s2sphere; trimesh, embreex and mapbox_earcut for the Asok solve; Node, Playwright and ffmpeg for stills and video. The scripts use absolute paths:

0. Layout: mkdir -p /home/claude/model/out && cp -r code/* data video /home/claude/model/ && cd /home/claude/model. Run every step from /home/claude/model (dsm_calib.py and ob_lowrise.py use relative data/ paths).
1. Re-fetch the raw sources into /home/claude/ext/: osm/Bangkok.osm.pbf, osm/ms/132203311.csv.gz, dsm/Copernicus_DSM_COG_10_N13_00_E100_00_DEM.tif, and the Open Buildings manifests ob/m_2016.json … ob/m_2023.json.
2. python3 -I tools/corridor_extract.py (Bangkok.osm.pbf → data/corridor.pkl)
3. python3 -I tools/dsm_calib.py (corridor.pkl, GLO-30 tile → data/dsm_calib.pkl); before solve3, which loads it.
4. python3 -I tools/solve3.py (corridor.pkl, dsm_calib.pkl, GLO-30, MS footprints → data/solve3.pkl, out/provenance_all_elements.csv, out/radial_walk_stretches.csv, out/solve3_report.txt)
5. python3 -I tools/ob_sample.py (corridor.pkl, solve3.pkl, dsm_calib.pkl, ob/m_2016…2023.json, HTTPS range reads from the public bucket → data/ob_sample.pkl). Set GDAL_CURL_CA_BUNDLE behind a TLS-intercepting proxy.
6. python3 -I tools/ob_lowrise.py (corridor.pkl, ob/m_2023.json, bucket reads → data/ob_lowrise.pkl)
7. python3 -I tools/step123.py (corridor.pkl, solve3.pkl → data/step123.pkl)
8. python3 -I tools/penumbra.py (solve3.pkl, ob_sample.pkl → data/penumbra.pkl)
9. python3 -I tools/street_disc.py (solve3.pkl → data/street_disc.pkl; log → results/street_disc.txt)
10. python3 -I render/fluence_field4.py (solve3.pkl, corridor.pkl, street_disc.pkl → out/field4_0_street.png, out/field4_1_5km.png, out/field4_2_15km.png)
11. python3 -I render/provenance_table4.py (solve3.pkl, out/provenance_all_elements.csv → out/field4_provenance.png)
12. python3 -I render/far_cell.py (penumbra.pkl, ob_sample.pkl, ob_lowrise.pkl, step123.pkl, solve3.pkl → out/far5_AB.png, out/far5_C.png, out/far5_provenance.png)

For step 11 without step 4, copy results/provenance_all_elements.csv to out/ first.

What rebuilds from this package, checked 2026-11-03 from a fresh copy:

Output	Rebuilds from the package?	Check
results.json, sensitivity.json, fields.npz, b41_bangkok_effects_v2.png (b41_bangkok.py)	Yes, from data/fire_network.npz and the model code	byte-identical
field4_*.png, field4_provenance.png, far5_*.png (steps 8–12)	Yes, from the shipped derived data in data/	byte-identical
Asok still, plain still, end card, caster table and JSON, v3 scrubber (steps 13–18)	Yes, plus Esri tiles fetched over the network (not shipped; the imagery can change)	byte-identical with tiles fetched 2026-11-03
v3 MP4 (step 19)	Yes, from the v3 scrubber	frames checked by OCR and against the end card; not byte-checked
Derived data: corridor.pkl, dsm_calib.pkl, solve3.pkl, ob_sample.pkl, ob_lowrise.pkl, step123.pkl, osm_bld.npz, osm_bld_meta.json, osm_rail.json, fire_network.npz	No. Steps 2–7, tools/osm_buildings.py and tools/fire_network.py need the raw sources in step 1, which are not shipped	not reproducible from the package

The OSM input was the BBBike Bangkok extract at replication 2026-10-31T00:00Z. BBBike extracts are rebuilt continuously, so a download today will not reproduce the derived data byte for byte; that needs an OSM history extract cut at that timestamp. Open Buildings is read from a remote bucket. Every figure therefore rebuilds from the shipped derived data, but the derived data themselves cannot be rebuilt from the package.

Asok cut (needs data/osm_bld.npz, data/osm_bld_meta.json, data/osm_rail.json, the v2 scrubber and Esri tiles):

13. python3 -I render/fetch_tiles.py B (Esri World Imagery z14 within 8 km of camera B, z17, z19 → /home/claude/ext/tiles)
14. python3 -I render/prerender2.py B (→ render/build/B_base.jpg, B_data.png, scene_cams.json; single-deck viaduct)
15. python3 -I tools/scene_from_html.py video/withdrawn/b41_bangkok_render_v2.html out/asok, then python3 -I tools/asok_penumbra.py out/asok/D.json out/asok (→ asok_field_s1.npz, cut.json)
16. python3 -I tools/asok_casters.py out/asok/D.json out/asok/asok_field_s1.npz out/asok, then python3 -I tools/asok_table.py out/asok/asok_casters.json out/asok/asok_caster_table.csv
17. python3 -I render/build_html3.py video/withdrawn/b41_bangkok_render_v2.html out/asok/asok_field_s1.npz out/asok/asok_casters.json out/b41_bangkok_render_v3.html render/build/B_base.jpg render/build/B_data.png
18. node render/still3.js out/b41_bangkok_render_v3.html out --flu asok_still_v3=73.5, then node render/still3.js out/b41_bangkok_render_v3.html out end_card_v3=132 asok_still_v3_plain=73.5
19. node render/capture.js out/b41_bangkok_render_v3.html out/b41_bangkok_render_v3_1080p60.mp4 0 8130 (all 8,130 frames; without the frame range capture.js stops at 133.5 s, inside the end card)

Model constants carried in every figure

• B41 25 Mt airburst, HOB 8,927 m at η = 0.5.
• Thermal per G&D eq 7.96.2, with f = 0.43, V = 10 km and pulse t_max = 2.311 s.
• Burn thresholds: 3rd°-100% at 12.75 cal/cm². The charring line at about 38 cal/cm² (38.24) is 3× that value, an extrapolation past G&D's validated curves.
• Fireball: radius 3.0 km, uniform radiance; δ = asin(R / D) from round 4 (19.43° at the Asok cut, 19.64° at the street point, 17.01° at 5 km, 9.43° at 15.9 km). Direct beam only; scattered sky radiation is not modelled.
• The ±15% floor-height caveat applies to levels × 3.2 heights.

Sources and licences

• OpenStreetMap via the BBBike Bangkok extract (replication 2026-10-31T00:00Z). © OpenStreetMap contributors, ODbL.
• Microsoft Global ML Building Footprints, quadkey 132203311. ODbL.
• Copernicus DEM GLO-30, tile N13 E100: https://copernicus-dem-30m.s3.amazonaws.com/ (Copernicus DEM licence).
• Google Open Buildings 2.5D Temporal v1, from the public GCS bucket open-buildings-temporal-data. CC BY 4.0 or ODbL v1.0; contains modified Copernicus Sentinel-2 data.
• Esri World Imagery, used as the basemap in the renders under Esri's terms of use (camera-B tiles re-fetched 2026-11-03).
• Glasstone & Dolan, The Effects of Nuclear Weapons (1977).
• Kingery-Bulmash blast fits as given in UFC 3-340-02.
• BTS viaduct cross-check (width and girder height, not used as data): railway-technology.com, "Bangkok Mass Transit System (BTS) Skytrain".