WeatherBlend

Multi-model forecast blending for Bonehill Rocks, Dartmoor

Rain models

Per-station P(wet ≥ 0.1 mm/h) — 3a (lean) + 3c (rich), Brier, lower better. Plus per-station rainfall amount (mm/h) — 3f distributional, CRPS, lower better. Δ vs best single NWP — negative = blend wins.

Precipitation — Bellever Dartmoor

Phase 3o · v2026-06-15_161744_phase3o Δ -0.000 vs prev train

Rich + orographic + upper-air P(wet), pooled across the 4 Bonehill gauges (110 features = rich 59 + 9 terrain + 42 multi-level pressure). Bonehill rich champion; the decorrelated UA-route member of the 3-way 4b blend. Trained 2026-06-15. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.115 precip_ukmo (0.210) -45.2%
+48h 0.129 precip_icon (0.253) -48.9%
+72h 0.136 precip_icon (0.269) -49.5%
+96h 0.152 precip_icon (0.281) -45.8%
+120h 0.165 precip_gfs (0.313) -47.2%
Verify history (7 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_133956_phase3o
v2026-06-10_171114_phase3o		 304 0.086
jma: 0.164		0.088
jma: 0.082		0.044
jma: 0.036		0.013
gfs: 0.000		0.106
ecmwf: 0.134		0.086
ecmwf: 0.023
v2026-05-31_202616_phase3o
v2026-06-07_133956_phase3o
v2026-06-10_171114_phase3o		 600 0.404
gfs: 0.400		0.197
icon: 0.237		0.249
icon: 0.318		0.316
gfs: 0.200		0.217
gem: 0.355		0.216
icon: 0.290
v2026-05-31_202616_phase3o 636 —0.166
ecmwf: 0.200		0.173
jma: 0.244		0.206
ukmo: 0.285		0.240
gfs: 0.360		0.220
icon: 0.275
v2026-05-31_202616_phase3o 494 —0.197
ecmwf: 0.219		0.207
jma: 0.168		0.257
jma: 0.296		0.277
aifs: 0.462		0.360
aifs: 0.250
v2026-05-26_104548_phase3o
v2026-05-31_202616_phase3o		 464 —0.209
aifs: 0.148		0.283
jma: 0.134		0.510
aifs: 0.200		0.200
ecmwf: 0.244		0.269
ecmwf: 0.265
v2026-05-26_104548_phase3o 366 —0.011
mf: 0.016		0.030
ecmwf: 0.030		0.034
icon: 0.036		0.097
gfs: 0.100		—
v2026-05-26_104548_phase3o 51 —0.000
ukmo: 0.000		————

Phase 3c · v2026-06-15_160709_phase3c Δ -0.002 vs prev train

Rich P(wet) classifier, 101 features — rich surface (humidity/dew-depression/cloud/EA persistence) + the multi-level pressure (upper-air) block (850/700/500 hPa temps, heights, winds, RH850 across 4 exact NWPs, lead-matched by a leak-free backward ASOF). UA added in-place 2026-06-02 (−3.7%→−6.8% Brier across 24→72h). Trained 2026-06-15. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.096 precip_mf (0.189) -49.1%
+48h 0.112 precip_ecmwf (0.229) -51.2%
+72h 0.118 precip_ecmwf (0.250) -52.6%
+96h 0.137 precip_aifs (0.297) -53.8%
+120h 0.150 precip_ecmwf (0.284) -47.3%
Verify history (16 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_132416_phase3c
v2026-06-10_165447_phase3c		 304 0.100
jma: 0.164		0.068
jma: 0.082		0.029
jma: 0.036		0.009
gfs: 0.000		0.157
ecmwf: 0.134		0.076
ecmwf: 0.023
v2026-05-31_201848_phase3c
v2026-06-07_132416_phase3c
v2026-06-10_165447_phase3c		 600 0.385
gfs: 0.400		0.178
icon: 0.237		0.244
icon: 0.318		0.305
gfs: 0.200		0.218
gem: 0.355		0.206
icon: 0.290
v2026-05-31_201848_phase3c 636 —0.196
ecmwf: 0.200		0.190
jma: 0.244		0.195
ukmo: 0.285		0.226
gfs: 0.360		0.193
icon: 0.275
v2026-05-31_201848_phase3c 494 —0.240
ecmwf: 0.219		0.240
jma: 0.168		0.246
jma: 0.296		0.264
aifs: 0.462		0.287
aifs: 0.250
v2026-05-24_130922_phase3c
v2026-05-31_201848_phase3c		 726 —0.265
aifs: 0.148		0.362
jma: 0.134		0.474
aifs: 0.200		0.116
ecmwf: 0.169		0.134
ecmwf: 0.200
v2026-05-24_130922_phase3c 612 —0.007
icon: 0.015		0.017
mf: 0.021		0.012
mf: 0.011		0.035
ukmo: 0.036		0.043
ecmwf: 0.045
v2026-05-17_174340_phase3c
v2026-05-24_130922_phase3c		 708 —0.001
gem: 0.000		0.002
icon: 0.000		0.001
gfs: 0.000		0.012
ukmo: 0.021		0.003
gfs: 0.000
v2026-05-17_174340_phase3c 584 —0.134
mf: 0.199		0.092
ukmo: 0.115		0.078
mf: 0.108		0.042
ukmo: 0.094		0.012
gfs: 0.000
v2026-04-28_232840_phase3c
v2026-05-17_174340_phase3c		 1205 —0.208
icon: 0.296		0.288
ukmo: 0.356		0.370
icon: 0.333		0.225
ukmo: 0.337		0.262
gfs: 0.373
v2026-04-28_232840_phase3c 1214 —0.102
ukmo: 0.119		0.091
jma: 0.116		0.083
mf: 0.136		0.137
icon: 0.256		0.145
gem: 0.250
v2026-04-28_232840_phase3c 1374 —0.036
ukmo: 0.044		0.024
ukmo: 0.117		0.034
gem: 0.142		0.046
gem: 0.205		0.043
gem: 0.116
v2026-04-28_232840_phase3c 1298 —0.076
ukmo: 0.140		0.068
icon: 0.225		0.047
icon: 0.189		0.037
gem: 0.179		0.032
gem: 0.048
v2026-04-28_232840_phase3c 1003 —0.081
ukmo: 0.158		0.083
icon: 0.277		0.052
icon: 0.180		0.028
gem: 0.199		0.018
gem: 0.053
v2026-04-28_232840_phase3c 289 —0.044
gfs: 0.031		0.043
mf: 0.037		0.006
gfs: 0.000		——
v2026-04-23_154405_phase3c
v2026-04-28_161625_phase3c
v2026-04-28_232840_phase3c		 11 —0.022
gfs: 0.000		0.002
gfs: 0.000		0.001
icon: 0.000		——
v2026-04-23_154405_phase3c 11 —0.000
gfs: 0.000		0.000
ecmwf: 0.000		0.001
ecmwf: 0.000		——

Phase 3a · v2026-06-14_134821 Δ -0.004 vs prev train

Lean P(wet) classifier, 27 features. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.106 precip_mf (0.192) -45.0%
+48h 0.122 precip_ecmwf (0.233) -47.7%
+72h 0.128 precip_ecmwf (0.253) -49.4%
+96h 0.140 precip_aifs (0.298) -53.0%
+120h 0.154 precip_ecmwf (0.287) -46.4%
Verify history (16 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +24h+48h+72h+96h+120h
v2026-06-07_131821 560 0.131
icon: 0.263		0.141
jma: 0.227		0.114
jma: 0.156		0.102
ecmwf: 0.134		0.066
ecmwf: 0.023
v2026-05-31_201322
v2026-06-07_131821		 600 0.191
icon: 0.237		0.264
icon: 0.318		0.317
gfs: 0.200		0.218
gem: 0.355		0.214
icon: 0.290
v2026-05-31_201322 636 0.184
ecmwf: 0.200		0.208
jma: 0.244		0.185
ecmwf: 0.296		0.224
gfs: 0.360		0.211
icon: 0.275
v2026-05-31_201322 494 0.212
ecmwf: 0.219		0.249
jma: 0.168		0.227
jma: 0.296		0.252
aifs: 0.462		0.323
aifs: 0.250
v2026-05-24_130501
v2026-05-31_201322		 726 0.268
aifs: 0.148		0.357
jma: 0.134		0.558
aifs: 0.200		0.117
ecmwf: 0.169		0.140
ecmwf: 0.200
v2026-05-24_130501 612 0.007
icon: 0.015		0.017
mf: 0.021		0.013
mf: 0.011		0.035
aifs: 0.036		0.045
ecmwf: 0.045
v2026-05-17_173909
v2026-05-24_130501		 708 0.001
gem: 0.000		0.002
icon: 0.000		0.001
gfs: 0.000		0.010
gfs: 0.049		0.006
gfs: 0.000
v2026-05-17_173909 584 0.125
mf: 0.199		0.091
mf: 0.134		0.079
mf: 0.108		0.038
gfs: 0.119		0.027
gfs: 0.000
v2026-04-28_232709
v2026-05-17_173909		 1205 0.204
icon: 0.296		0.273
ecmwf: 0.407		0.323
icon: 0.333		0.232
gem: 0.348		0.259
gfs: 0.373
v2026-04-28_232709 1214 0.086
jma: 0.124		0.093
jma: 0.116		0.085
mf: 0.136		0.134
icon: 0.256		0.146
gem: 0.250
v2026-04-28_232709 1374 0.032
mf: 0.088		0.025
icon: 0.124		0.036
gem: 0.142		0.036
gem: 0.205		0.045
gem: 0.116
v2026-04-28_232709 1298 0.080
mf: 0.173		0.069
icon: 0.225		0.046
icon: 0.189		0.025
gem: 0.179		0.034
gem: 0.048
v2026-04-28_232709 1003 0.086
mf: 0.163		0.085
icon: 0.277		0.050
icon: 0.180		0.018
gem: 0.199		0.018
gem: 0.053
v2026-04-28_232709 289 0.040
gfs: 0.031		0.062
mf: 0.037		0.003
gfs: 0.000		——
v2026-04-27_175435
v2026-04-28_161456
v2026-04-28_232709		 2 0.072
gfs: 0.000		0.003
gfs: 0.000		0.002
gfs: 0.000		——
v2026-04-23_071842
v2026-04-27_175435		 10 0.000
gfs: 0.000		0.017
gfs: 0.000		0.002
ecmwf: 0.000		——

Phase 3d · v2026-06-16_090419_phase3d Δ -0.002 vs prev train

Exact-runtime P(wet) classifier. Trains on raw S3 cycles (GFS + IFS oper + AIFS required, MO Global + UKV optional) instead of Open-Meteo offset_day, with rigorous (RunTime, ValidTime, Lead) provenance per row. UKV pulled per-V-hour with target-lead-aware tuples. Trained 2026-06-16. Metric: Test Brier.

Lead Blend Best single Δ vs best
+12h 0.090 precip_moglobal (0.148) -39.1%
+24h 0.099 precip_moglobal (0.170) -41.5%
+48h 0.110 precip_moglobal (0.158) -30.7%
+72h 0.107 precip_gfs (0.265) -59.6%
+96h 0.165 precip_moglobal (0.144) +14.4%
+120h 0.171 precip_gfs (0.310) -44.8%
Verify history (13 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-05-08_061521_phase3d
v2026-06-07_140152_phase3d
v2026-06-14_144555_phase3d		 223 0.001
—		0.156
—		0.210
—		0.127
—		0.219
—		0.276
—
v2026-05-08_061521_phase3d
v2026-06-07_140152_phase3d		 234 0.144
—		0.176
—		0.202
—		0.271
—		0.331
—		0.369
—
v2026-05-08_061521_phase3d
v2026-05-31_203938_phase3d
v2026-06-07_140152_phase3d		 253 0.046
—		0.021
—		0.185
—		0.235
—		0.225
—		0.286
—
v2026-05-08_061521_phase3d
v2026-05-31_203938_phase3d
v2026-06-07_140152_phase3d		 253 0.007
—		0.261
—		0.218
—		0.273
—		0.209
—		0.275
—
v2026-05-08_061521_phase3d
v2026-05-31_203938_phase3d		 224 0.208
—		0.315
—		0.258
—		0.435
—		0.109
—		0.131
—
v2026-05-08_061521_phase3d
v2026-05-26_105519_phase3d
v2026-05-31_203938_phase3d		 133 0.001
—		0.011
—		0.004
—		0.010
—		0.005
—		0.015
—
v2026-05-08_061521_phase3d
v2026-05-24_131622_phase3d
v2026-05-26_105519_phase3d		 75 0.003
—		0.012
—		0.003
—		0.002
—		0.021
—		0.068
—
v2026-05-08_061521_phase3d
v2026-05-17_175449_phase3d
v2026-05-24_131622_phase3d		 36 0.004
—		0.024
—		0.161
—		0.246
—		0.175
—		0.178
—
v2026-05-08_061237_phase3d
v2026-05-08_061521_phase3d
v2026-05-17_175449_phase3d		 50 0.212
—		0.188
—		0.043
—		0.200
—		0.280
—		0.350
—
v2026-05-08_061237_phase3d
v2026-05-08_061521_phase3d		 49 0.054
—		0.042
—		0.106
—		0.142
—		0.270
—		0.414
—
v2026-05-08_061237_phase3d
v2026-05-08_061521_phase3d		 38 0.003
—		0.006
—		0.033
—		0.044
—		0.205
—		0.347
—
v2026-05-08_061237_phase3d 5 0.002
—		0.006
—		————
v2026-05-05_190733_phase3d 2 0.001
—		0.005
—		————

Phase 4a · v2026-06-14_123009_phase4a Δ -0.002 vs prev train

Phase 4a blender. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.101 (per-cell BART) (NaN val) —
+48h 0.115 (per-cell BART) (NaN val) —
+72h 0.122 (per-cell BART) (NaN val) —
+96h 0.135 (per-cell BART) (NaN val) —
+120h 0.148 (per-cell BART) (NaN val) —
Verify history (11 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_122520_phase4a 126 0.124
—		0.123
—		0.164
—		0.141
—		0.234
—		0.098
—
v2026-05-31_122003_phase4a
v2026-06-07_122520_phase4a		 150 0.418
—		0.157
—		0.232
—		0.373
—		0.210
—		0.195
—
v2026-05-31_122003_phase4a 160 —0.154
—		0.180
—		0.179
—		0.229
—		0.221
—
v2026-05-31_122003_phase4a 128 —0.177
—		0.212
—		0.208
—		0.294
—		0.260
—
v2026-05-24_122715_phase4a
v2026-05-31_122003_phase4a		 151 —0.193
—		0.234
—		0.352
—		0.125
—		0.141
—
v2026-05-24_122715_phase4a 125 —0.008
—		0.010
—		0.013
—		0.019
—		0.031
—
v2026-05-17_165023_phase4a
v2026-05-24_122715_phase4a		 135 —0.000
—		0.000
—		0.000
—		0.044
—		0.053
—
v2026-05-17_165023_phase4a 93 —0.088
—		0.108
—		0.112
—		0.193
—		0.209
—
v2026-05-17_165023_phase4a 21 —0.211
—		0.251
—		0.279
—		0.310
—		0.328
—
v2026-05-10_220841_phase4a 66 —0.175
—		0.221
—		0.222
—		0.192
—		0.203
—
v2026-05-10_220841_phase4a 15 —0.088
—		0.151
—		0.101
—		0.115
—		0.120
—

Phase 4b · v2026-06-14_150943_phase4b Δ -0.002 vs prev train

Phase 4b blender. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.097 p_3c (0.098) -0.6%
+48h 0.111 p_3o (0.112) -1.0%
+72h 0.116 p_3o (0.116) -0.3%
+96h 0.132 p_3o (0.132) 0.0%
+120h 0.145 p_3o (0.147) -1.2%
Verify history (11 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_142500_phase4b 536 0.094
jma: 0.173		0.122
icon: 0.263		0.132
jma: 0.236		0.126
jma: 0.162		0.139
ecmwf: 0.141		0.077
ecmwf: 0.023
v2026-05-31_210032_phase4b
v2026-06-07_142500_phase4b		 600 0.428
gfs: 0.444		0.178
icon: 0.227		0.247
icon: 0.318		0.325
gfs: 0.200		0.215
gem: 0.355		0.204
icon: 0.290
v2026-05-31_210032_phase4b 636 —0.168
ecmwf: 0.200		0.170
jma: 0.244		0.189
ukmo: 0.285		0.231
gfs: 0.360		0.210
icon: 0.275
v2026-05-31_210032_phase4b 494 —0.201
ecmwf: 0.219		0.207
jma: 0.168		0.232
jma: 0.296		0.270
aifs: 0.462		0.309
aifs: 0.250
v2026-05-26_110622_phase4b
v2026-05-31_210032_phase4b		 464 —0.219
aifs: 0.148		0.294
jma: 0.134		0.426
aifs: 0.200		0.188
ecmwf: 0.244		0.243
ecmwf: 0.265
v2026-05-24_140756_phase4b
v2026-05-26_110622_phase4b		 366 —0.011
mf: 0.016		0.030
ecmwf: 0.030		0.034
icon: 0.036		0.096
gfs: 0.100		0.043
ecmwf: 0.045
v2026-05-17_185156_phase4b
v2026-05-24_140756_phase4b
v2026-05-26_110622_phase4b		 637 —0.000
ukmo: 0.000		0.000
icon: 0.000		0.000
gfs: 0.000		0.028
ukmo: 0.028		0.031
gfs: 0.030
v2026-05-17_185156_phase4b 471 —0.106
ecmwf: 0.163		0.060
ukmo: 0.078		0.088
mf: 0.126		0.178
ukmo: 0.205		0.193
gfs: 0.289
v2026-05-17_185156_phase4b 53 —0.205
icon: 0.170		0.230
icon: 0.273		0.305
icon: 0.375		0.294
ukmo: 0.396		0.324
ukmo: 0.396
v2026-05-12_202905_phase4b 89 —0.254
ukmo: 0.159		0.332
jma: 0.056		0.003
icon: 0.000		0.003
gfs: 0.000		—
v2026-05-12_202905_phase4b 6 —0.618
mf: 0.000		————

Phase 3oni · v2026-06-14_143409_phase3oni

Phase 3oni blender. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.116 precip_ukmo (0.211) -45.1%
+48h 0.131 precip_icon (0.254) -48.4%
+72h 0.136 precip_mf (0.261) -47.9%
+96h 0.149 precip_icon (0.281) -46.9%
+120h 0.164 precip_gfs (0.314) -47.7%
Verify history (no runs yet)

No verify rows yet for phase 3oni. Next cycle: Mon/Thu 09:30 UTC, then 5d ERA5 latency.

Precipitation — Bovey Tracey

Phase 3o · v2026-06-15_161744_phase3o Δ +0.000 vs prev train

Rich + orographic + upper-air P(wet), pooled across the 4 Bonehill gauges (110 features = rich 59 + 9 terrain + 42 multi-level pressure). Bonehill rich champion; the decorrelated UA-route member of the 3-way 4b blend. Trained 2026-06-15. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.088 precip_icon (0.162) -45.7%
+48h 0.101 precip_icon (0.190) -46.7%
+72h 0.110 precip_icon (0.209) -47.2%
+96h 0.123 precip_icon (0.228) -45.9%
+120h 0.134 precip_gfs (0.244) -45.0%
Verify history (7 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_133956_phase3o
v2026-06-10_171114_phase3o		 304 0.060
icon: 0.079		0.018
icon: 0.019		0.001
gfs: 0.000		0.001
gfs: 0.000		0.036
gfs: 0.036		0.004
gfs: 0.000
v2026-05-31_202616_phase3o
v2026-06-07_133956_phase3o
v2026-06-10_171114_phase3o		 616 0.012
gfs: 0.000		0.069
icon: 0.062		0.046
icon: 0.045		0.003
ecmwf: 0.000		0.124
gem: 0.177		0.170
icon: 0.184
v2026-05-31_202616_phase3o 646 —0.094
icon: 0.178		0.104
icon: 0.223		0.095
icon: 0.144		0.141
gem: 0.199		0.227
icon: 0.213
v2026-05-31_202616_phase3o 504 —0.114
icon: 0.214		0.130
icon: 0.268		0.114
icon: 0.158		0.174
icon: 0.207		0.483
ukmo: 0.167
v2026-05-26_104548_phase3o
v2026-05-31_202616_phase3o		 458 —0.095
icon: 0.180		0.097
gfs: 0.167		0.086
ecmwf: 0.000		0.120
gfs: 0.159		0.152
ukmo: 0.269
v2026-05-26_104548_phase3o 376 —0.000
mf: 0.005		0.000
ecmwf: 0.000		0.000
icon: 0.000		0.000
gfs: 0.000		—
v2026-05-26_104548_phase3o 57 —0.000
ukmo: 0.000		————

Phase 3c · v2026-06-15_160947_phase3c Δ -0.001 vs prev train

Rich P(wet) classifier, 101 features — rich surface (humidity/dew-depression/cloud/EA persistence) + the multi-level pressure (upper-air) block (850/700/500 hPa temps, heights, winds, RH850 across 4 exact NWPs, lead-matched by a leak-free backward ASOF). UA added in-place 2026-06-02 (−3.7%→−6.8% Brier across 24→72h). Trained 2026-06-15. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.064 precip_icon (0.128) -49.8%
+48h 0.074 precip_mf (0.149) -50.4%
+72h 0.082 precip_mf (0.173) -52.3%
+96h 0.090 precip_icon (0.174) -48.1%
+120h 0.098 precip_gfs (0.184) -46.7%
Verify history (13 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_132814_phase3c
v2026-06-10_165856_phase3c		 304 0.062
icon: 0.079		0.016
icon: 0.019		0.002
gfs: 0.000		0.000
gfs: 0.000		0.039
gfs: 0.036		0.002
gfs: 0.000
v2026-05-31_202040_phase3c
v2026-06-07_132814_phase3c
v2026-06-10_165856_phase3c		 616 0.007
gfs: 0.000		0.064
icon: 0.062		0.041
icon: 0.045		0.002
ecmwf: 0.000		0.100
gem: 0.177		0.144
icon: 0.184
v2026-05-31_202040_phase3c 646 —0.095
icon: 0.178		0.100
icon: 0.223		0.088
icon: 0.144		0.114
gem: 0.199		0.193
icon: 0.213
v2026-05-31_202040_phase3c 504 —0.118
icon: 0.214		0.127
icon: 0.268		0.102
icon: 0.158		0.139
icon: 0.207		0.385
ukmo: 0.167
v2026-05-24_131051_phase3c
v2026-05-31_202040_phase3c		 722 —0.097
icon: 0.180		0.080
gfs: 0.167		0.121
ecmwf: 0.000		0.069
gfs: 0.136		0.077
icon: 0.190
v2026-05-24_131051_phase3c 640 —0.000
icon: 0.008		0.000
mf: 0.006		0.000
mf: 0.000		0.000
ukmo: 0.000		0.001
ecmwf: 0.000
v2026-05-17_174525_phase3c
v2026-05-24_131051_phase3c		 724 —0.000
gem: 0.000		0.000
icon: 0.000		0.000
gfs: 0.000		0.001
ukmo: 0.012		0.001
gfs: 0.000
v2026-05-17_174525_phase3c 592 —0.026
icon: 0.078		0.013
ukmo: 0.039		0.006
mf: 0.016		0.004
ukmo: 0.056		0.004
gfs: 0.000
v2026-05-03_233604_phase3c
v2026-05-17_174525_phase3c		 1214 —0.094
gfs: 0.213		0.091
ukmo: 0.266		0.174
gfs: 0.100		0.083
icon: 0.192		0.088
gfs: 0.197
v2026-05-03_233604_phase3c 1214 —0.038
icon: 0.091		0.035
icon: 0.077		0.031
icon: 0.099		0.051
icon: 0.137		0.064
gfs: 0.226
v2026-05-03_233604_phase3c 1088 —0.019
ukmo: 0.053		0.012
icon: 0.063		0.013
mf: 0.082		0.027
icon: 0.147		0.042
gem: 0.206
v2026-05-03_233604_phase3c 496 —0.026
ukmo: 0.054		0.019
icon: 0.104		0.021
ukmo: 0.135		0.042
ukmo: 0.119		0.111
gem: 0.000
v2026-05-03_233604_phase3c 200 —0.007
ukmo: 0.025		0.023
gfs: 0.090		———

Phase 3a · v2026-06-14_135027 Δ -0.008 vs prev train

Lean P(wet) classifier, 27 features. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.069 precip_icon (0.132) -47.6%
+48h 0.078 precip_mf (0.151) -48.3%
+72h 0.088 precip_gfs (0.168) -47.4%
+96h 0.092 precip_icon (0.178) -48.3%
+120h 0.101 precip_gfs (0.188) -46.4%
Verify history (13 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +24h+48h+72h+96h+120h
v2026-06-07_132019 568 0.051
icon: 0.060		0.044
icon: 0.050		0.046
mf: 0.064		0.038
gfs: 0.036		0.004
gfs: 0.000
v2026-05-31_201511
v2026-06-07_132019		 616 0.066
icon: 0.062		0.042
icon: 0.045		0.002
ecmwf: 0.000		0.098
gem: 0.177		0.146
icon: 0.184
v2026-05-31_201511 646 0.092
icon: 0.178		0.110
icon: 0.223		0.096
icon: 0.144		0.104
gem: 0.199		0.196
icon: 0.213
v2026-05-31_201511 504 0.115
icon: 0.214		0.139
icon: 0.268		0.115
icon: 0.158		0.124
icon: 0.207		0.385
aifs: 0.167
v2026-05-24_130628
v2026-05-31_201511		 722 0.091
icon: 0.180		0.100
gfs: 0.167		0.126
ecmwf: 0.000		0.076
gfs: 0.136		0.083
icon: 0.190
v2026-05-24_130628 640 0.000
icon: 0.008		0.001
mf: 0.006		0.001
mf: 0.000		0.001
aifs: 0.000		0.002
ecmwf: 0.000
v2026-05-03_233357
v2026-05-24_130628		 1037 0.000
gem: 0.000		0.000
icon: 0.000		0.000
gfs: 0.000		0.030
gfs: 0.087		0.048
gfs: 0.048
v2026-05-03_233357 1116 0.046
icon: 0.135		0.058
gfs: 0.175		0.054
mf: 0.175		0.081
icon: 0.225		0.090
gfs: 0.179
v2026-05-03_233357 1214 0.066
icon: 0.153		0.071
icon: 0.161		0.060
icon: 0.144		0.089
icon: 0.192		0.097
gfs: 0.197
v2026-05-03_233357 1214 0.037
icon: 0.091		0.036
icon: 0.077		0.031
icon: 0.099		0.052
icon: 0.137		0.057
gfs: 0.226
v2026-05-03_233357 1088 0.018
gfs: 0.066		0.013
icon: 0.063		0.014
mf: 0.082		0.027
icon: 0.147		0.030
gem: 0.206
v2026-05-03_233357 496 0.024
gfs: 0.095		0.022
icon: 0.104		0.021
gem: 0.143		0.040
gem: 0.206		0.082
gem: 0.000
v2026-05-03_233357 200 0.015
icon: 0.100		0.035
gfs: 0.090		———

Phase 3d · v2026-06-16_090557_phase3d Δ -0.002 vs prev train

Exact-runtime P(wet) classifier. Trains on raw S3 cycles (GFS + IFS oper + AIFS required, MO Global + UKV optional) instead of Open-Meteo offset_day, with rigorous (RunTime, ValidTime, Lead) provenance per row. UKV pulled per-V-hour with target-lead-aware tuples. Trained 2026-06-16. Metric: Test Brier.

Lead Blend Best single Δ vs best
+12h 0.059 precip_moglobal (0.116) -49.1%
+24h 0.063 precip_moglobal (0.133) -52.3%
+48h 0.062 precip_moglobal (0.124) -49.8%
+72h 0.067 precip_moglobal (0.130) -48.5%
+96h 0.110 precip_moglobal (0.126) -12.8%
+120h 0.115 precip_moglobal (0.196) -41.2%
Verify history (13 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-05-08_062649_phase3d
v2026-06-07_140321_phase3d
v2026-06-14_144729_phase3d		 223 0.000
—		0.059
—		0.066
—		0.077
—		0.129
—		0.145
—
v2026-05-08_062649_phase3d
v2026-06-07_140321_phase3d		 234 0.020
—		0.040
—		0.021
—		0.001
—		0.210
—		0.233
—
v2026-05-08_062649_phase3d
v2026-05-31_204054_phase3d
v2026-06-07_140321_phase3d		 244 0.002
—		0.001
—		0.102
—		0.152
—		0.150
—		0.188
—
v2026-05-08_062649_phase3d
v2026-05-31_204054_phase3d
v2026-06-07_140321_phase3d		 244 0.011
—		0.149
—		0.103
—		0.174
—		0.133
—		0.172
—
v2026-05-08_062649_phase3d
v2026-05-31_204054_phase3d		 222 0.104
—		0.188
—		0.066
—		0.026
—		0.059
—		0.095
—
v2026-05-08_062649_phase3d
v2026-05-26_105611_phase3d
v2026-05-31_204054_phase3d		 140 0.000
—		0.004
—		0.001
—		0.002
—		0.003
—		0.009
—
v2026-05-08_062649_phase3d
v2026-05-24_131720_phase3d
v2026-05-26_105611_phase3d		 82 0.003
—		0.006
—		0.001
—		0.001
—		0.011
—		0.033
—
v2026-05-08_062649_phase3d
v2026-05-17_175600_phase3d
v2026-05-24_131720_phase3d		 39 0.000
—		0.014
—		0.013
—		0.003
—		0.049
—		0.068
—
v2026-05-08_062649_phase3d
v2026-05-17_175600_phase3d		 53 0.084
—		0.074
—		0.051
—		0.007
—		0.102
—		0.090
—
v2026-05-08_062428_phase3d
v2026-05-08_062649_phase3d		 49 0.032
—		0.018
—		0.059
—		0.053
—		0.126
—		0.090
—
v2026-05-08_062428_phase3d
v2026-05-08_062649_phase3d		 38 0.001
—		0.001
—		0.055
—		0.052
—		0.206
—		0.365
—
v2026-05-08_062428_phase3d 5 0.001
—		0.001
—		————
v2026-05-05_190928_phase3d 2 0.003
—		0.009
—		————

Phase 4a · v2026-06-14_123009_phase4a Δ -0.007 vs prev train

Phase 4a blender. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.066 (per-cell BART) (NaN val) —
+48h 0.077 (per-cell BART) (NaN val) —
+72h 0.082 (per-cell BART) (NaN val) —
+96h 0.088 (per-cell BART) (NaN val) —
+120h 0.095 (per-cell BART) (NaN val) —
Verify history (11 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_122520_phase4a 130 0.036
—		0.059
—		0.042
—		0.042
—		0.067
—		0.000
—
v2026-05-31_122003_phase4a
v2026-06-07_122520_phase4a		 154 0.004
—		0.072
—		0.053
—		0.001
—		0.096
—		0.099
—
v2026-05-31_122003_phase4a 160 —0.100
—		0.099
—		0.111
—		0.111
—		0.113
—
v2026-05-31_122003_phase4a 128 —0.121
—		0.121
—		0.129
—		0.145
—		0.159
—
v2026-05-24_122715_phase4a
v2026-05-31_122003_phase4a		 152 —0.113
—		0.099
—		0.054
—		0.081
—		0.093
—
v2026-05-24_122715_phase4a 130 —0.000
—		0.000
—		0.000
—		0.001
—		0.000
—
v2026-05-17_165023_phase4a
v2026-05-24_122715_phase4a		 141 —0.000
—		0.000
—		0.000
—		0.006
—		0.014
—
v2026-05-17_165023_phase4a 94 —0.008
—		0.022
—		0.034
—		0.030
—		0.059
—
v2026-05-17_165023_phase4a 22 —0.032
—		0.090
—		0.093
—		0.061
—		0.111
—
v2026-05-10_220841_phase4a 66 —0.107
—		0.127
—		0.102
—		0.090
—		0.084
—
v2026-05-10_220841_phase4a 15 —0.142
—		0.188
—		0.156
—		0.165
—		0.168
—

Phase 4b · v2026-06-14_150943_phase4b Δ -0.007 vs prev train

Phase 4b blender. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.065 p_3c (0.065) -0.7%
+48h 0.075 p_3c (0.076) -1.4%
+72h 0.081 p_4a (0.082) -1.8%
+96h 0.088 p_4a (0.088) +0.6%
+120h 0.096 p_4a (0.095) +0.7%
Verify history (11 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_142500_phase4b 544 0.061
icon: 0.083		0.052
icon: 0.063		0.036
icon: 0.043		0.048
mf: 0.069		0.038
gfs: 0.038		0.002
gfs: 0.000
v2026-05-31_210032_phase4b
v2026-06-07_142500_phase4b		 616 0.008
gfs: 0.000		0.069
icon: 0.065		0.042
icon: 0.045		0.002
ecmwf: 0.000		0.106
gem: 0.177		0.140
icon: 0.184
v2026-05-31_210032_phase4b 646 —0.091
icon: 0.178		0.102
icon: 0.223		0.093
icon: 0.144		0.120
gem: 0.199		0.185
icon: 0.213
v2026-05-31_210032_phase4b 504 —0.112
icon: 0.214		0.128
icon: 0.268		0.109
icon: 0.158		0.150
icon: 0.207		0.380
ukmo: 0.167
v2026-05-26_110622_phase4b
v2026-05-31_210032_phase4b		 458 —0.094
icon: 0.180		0.085
gfs: 0.167		0.069
ecmwf: 0.000		0.133
gfs: 0.159		0.152
ukmo: 0.269
v2026-05-24_140756_phase4b
v2026-05-26_110622_phase4b		 376 —0.000
mf: 0.005		0.000
ecmwf: 0.000		0.000
icon: 0.000		0.000
gfs: 0.000		0.000
ecmwf: 0.000
v2026-05-17_185156_phase4b
v2026-05-24_140756_phase4b
v2026-05-26_110622_phase4b		 659 —0.000
ukmo: 0.000		0.000
icon: 0.000		0.000
gfs: 0.000		0.006
ukmo: 0.018		0.009
gfs: 0.019
v2026-05-17_185156_phase4b 474 —0.006
icon: 0.025		0.013
ukmo: 0.015		0.029
mf: 0.081		0.038
ukmo: 0.102		0.071
gfs: 0.175
v2026-05-17_185156_phase4b 56 —0.049
icon: 0.107		0.106
ukmo: 0.125		0.125
gfs: 0.315		0.088
ukmo: 0.241		0.154
ukmo: 0.241
v2026-05-12_202905_phase4b 89 —0.144
gem: 0.000		0.110
gfs: 0.112		0.001
icon: 0.000		0.000
gfs: 0.000		—
v2026-05-12_202905_phase4b 6 —0.769
mf: 0.000		————

Phase 3oni · v2026-06-14_143409_phase3oni

Phase 3oni blender. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.088 precip_icon (0.163) -45.9%
+48h 0.102 precip_icon (0.191) -46.4%
+72h 0.111 precip_icon (0.210) -46.9%
+96h 0.121 precip_icon (0.228) -46.8%
+120h 0.133 precip_gfs (0.245) -45.4%
Verify history (no runs yet)

No verify rows yet for phase 3oni. Next cycle: Mon/Thu 09:30 UTC, then 5d ERA5 latency.

Precipitation — Dartmoor Nr Hexworthy

Phase 3o · v2026-06-15_161744_phase3o Δ -0.000 vs prev train

Rich + orographic + upper-air P(wet), pooled across the 4 Bonehill gauges (110 features = rich 59 + 9 terrain + 42 multi-level pressure). Bonehill rich champion; the decorrelated UA-route member of the 3-way 4b blend. Trained 2026-06-15. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.131 precip_ukmo (0.228) -42.5%
+48h 0.145 precip_icon (0.273) -46.8%
+72h 0.152 precip_gem (0.292) -47.9%
+96h 0.170 precip_icon (0.295) -42.3%
+120h 0.181 precip_gfs (0.336) -45.9%
Verify history (7 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_133956_phase3o
v2026-06-10_171114_phase3o		 304 0.128
jma: 0.184		0.162
jma: 0.154		0.139
jma: 0.131		0.108
gfs: 0.100		0.220
jma: 0.284		0.191
ecmwf: 0.159
v2026-05-31_202616_phase3o
v2026-06-07_133956_phase3o
v2026-06-10_171114_phase3o		 594 0.326
gfs: 0.400		0.216
icon: 0.290		0.239
gfs: 0.341		0.300
gfs: 0.200		0.222
icon: 0.337		0.240
icon: 0.323
v2026-05-31_202616_phase3o 632 —0.159
ecmwf: 0.231		0.184
ecmwf: 0.250		0.207
ecmwf: 0.255		0.223
gfs: 0.297		0.228
icon: 0.310
v2026-05-31_202616_phase3o 490 —0.180
ecmwf: 0.243		0.207
jma: 0.223		0.243
ecmwf: 0.273		0.239
gfs: 0.356		0.345
aifs: 0.193
v2026-05-26_104548_phase3o
v2026-05-31_202616_phase3o		 464 —0.198
aifs: 0.255		0.261
jma: 0.244		0.452
aifs: 0.300		0.180
ecmwf: 0.239		0.243
ecmwf: 0.320
v2026-05-26_104548_phase3o 366 —0.009
mf: 0.014		0.000
ecmwf: 0.000		0.001
icon: 0.000		0.001
gfs: 0.000		—
v2026-05-26_104548_phase3o 51 —0.058
ukmo: 0.059		————

Phase 3c · v2026-06-15_161225_phase3c Δ -0.001 vs prev train

Rich P(wet) classifier, 101 features — rich surface (humidity/dew-depression/cloud/EA persistence) + the multi-level pressure (upper-air) block (850/700/500 hPa temps, heights, winds, RH850 across 4 exact NWPs, lead-matched by a leak-free backward ASOF). UA added in-place 2026-06-02 (−3.7%→−6.8% Brier across 24→72h). Trained 2026-06-15. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.125 precip_ecmwf (0.221) -43.5%
+48h 0.138 precip_ecmwf (0.241) -42.6%
+72h 0.147 precip_aifs (0.275) -46.5%
+96h 0.164 precip_aifs (0.304) -46.0%
+120h 0.171 precip_ecmwf (0.305) -43.9%
Verify history (16 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_133207_phase3c
v2026-06-10_170301_phase3c		 304 0.147
jma: 0.184		0.183
jma: 0.154		0.127
jma: 0.131		0.101
gfs: 0.100		0.219
jma: 0.284		0.209
ecmwf: 0.159
v2026-05-31_202231_phase3c
v2026-06-07_133207_phase3c
v2026-06-10_170301_phase3c		 594 0.314
gfs: 0.400		0.222
icon: 0.290		0.247
gfs: 0.341		0.290
gfs: 0.200		0.211
icon: 0.337		0.212
icon: 0.323
v2026-05-31_202231_phase3c 632 —0.182
ecmwf: 0.231		0.199
ecmwf: 0.250		0.235
ecmwf: 0.255		0.197
gfs: 0.297		0.190
icon: 0.310
v2026-05-31_202231_phase3c 490 —0.208
ecmwf: 0.243		0.227
jma: 0.223		0.283
ecmwf: 0.273		0.226
gfs: 0.356		0.246
aifs: 0.193
v2026-05-24_131221_phase3c
v2026-05-31_202231_phase3c		 726 —0.205
aifs: 0.255		0.289
jma: 0.244		0.466
aifs: 0.300		0.104
ecmwf: 0.160		0.131
ecmwf: 0.224
v2026-05-24_131221_phase3c 612 —0.014
icon: 0.021		0.011
mf: 0.015		0.010
mf: 0.005		0.005
ukmo: 0.000		0.005
ecmwf: 0.000
v2026-05-17_174711_phase3c
v2026-05-24_131221_phase3c		 708 —0.030
gem: 0.029		0.035
icon: 0.036		0.056
gfs: 0.059		0.049
ukmo: 0.064		0.022
gfs: 0.018
v2026-05-17_174711_phase3c 584 —0.251
aifs: 0.308		0.208
ecmwf: 0.247		0.184
mf: 0.236		0.135
ukmo: 0.200		0.064
gfs: 0.047
v2026-04-28_232945_phase3c
v2026-05-17_174711_phase3c		 1205 —0.216
aifs: 0.276		0.311
ecmwf: 0.237		0.343
icon: 0.222		0.254
gem: 0.368		0.272
ecmwf: 0.374
v2026-04-28_232945_phase3c 1214 —0.147
mf: 0.175		0.132
jma: 0.166		0.129
mf: 0.163		0.175
icon: 0.280		0.172
gem: 0.263
v2026-04-28_232945_phase3c 1374 —0.091
ukmo: 0.102		0.085
ukmo: 0.169		0.095
icon: 0.185		0.108
gem: 0.247		0.110
gem: 0.180
v2026-04-28_232945_phase3c 1298 —0.151
ukmo: 0.222		0.140
icon: 0.292		0.126
icon: 0.249		0.121
gem: 0.238		0.117
gem: 0.145
v2026-04-28_232945_phase3c 1003 —0.148
ukmo: 0.233		0.151
gem: 0.334		0.135
icon: 0.254		0.118
gem: 0.247		0.112
ukmo: 0.118
v2026-04-28_232945_phase3c 289 —0.091
gfs: 0.100		0.148
mf: 0.198		0.004
gfs: 0.000		——
v2026-04-23_154459_phase3c
v2026-04-28_161729_phase3c
v2026-04-28_232945_phase3c		 11 —0.027
gfs: 0.000		0.003
gfs: 0.000		0.001
icon: 0.000		——
v2026-04-23_154459_phase3c 11 —0.000
gfs: 0.000		0.001
ecmwf: 0.000		0.001
ecmwf: 0.000		——

Phase 3a · v2026-06-14_135233 Δ +0.002 vs prev train

Lean P(wet) classifier, 27 features. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.133 precip_ecmwf (0.222) -40.0%
+48h 0.148 precip_ecmwf (0.244) -39.2%
+72h 0.154 precip_aifs (0.276) -44.3%
+96h 0.167 precip_aifs (0.306) -45.4%
+120h 0.176 precip_ecmwf (0.307) -42.5%
Verify history (16 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +24h+48h+72h+96h+120h
v2026-06-07_132217 560 0.196
jma: 0.295		0.187
jma: 0.239		0.213
jma: 0.217		0.233
jma: 0.284		0.210
ecmwf: 0.159
v2026-05-31_201659
v2026-06-07_132217		 594 0.225
icon: 0.290		0.250
gfs: 0.341		0.275
gfs: 0.200		0.212
icon: 0.337		0.216
icon: 0.323
v2026-05-31_201659 632 0.170
ecmwf: 0.231		0.189
ecmwf: 0.250		0.185
ecmwf: 0.255		0.195
gfs: 0.297		0.207
icon: 0.310
v2026-05-31_201659 490 0.184
ecmwf: 0.243		0.211
jma: 0.223		0.211
ecmwf: 0.273		0.195
gfs: 0.356		0.272
aifs: 0.193
v2026-05-24_130755
v2026-05-31_201659		 726 0.217
aifs: 0.255		0.279
jma: 0.244		0.497
aifs: 0.300		0.100
ecmwf: 0.160		0.125
ecmwf: 0.224
v2026-05-24_130755 612 0.014
icon: 0.021		0.012
mf: 0.015		0.009
mf: 0.005		0.005
aifs: 0.000		0.003
ecmwf: 0.000
v2026-04-28_232809
v2026-05-24_130755		 1010 0.030
gem: 0.029		0.036
icon: 0.036		0.055
gfs: 0.059		0.115
aifs: 0.174		0.130
gfs: 0.181
v2026-04-28_232809 1107 0.253
ecmwf: 0.294		0.237
ecmwf: 0.252		0.246
ecmwf: 0.324		0.280
gfs: 0.379		0.306
ecmwf: 0.410
v2026-04-28_232809 1205 0.204
jma: 0.289		0.200
ecmwf: 0.240		0.207
icon: 0.276		0.256
gem: 0.368		0.274
ecmwf: 0.374
v2026-04-28_232809 1214 0.134
mf: 0.175		0.137
jma: 0.166		0.132
mf: 0.163		0.172
icon: 0.280		0.170
gem: 0.263
v2026-04-28_232809 1374 0.087
mf: 0.127		0.082
icon: 0.174		0.092
icon: 0.185		0.103
gem: 0.247		0.107
gem: 0.180
v2026-04-28_232809 1298 0.150
mf: 0.247		0.142
icon: 0.292		0.120
icon: 0.249		0.114
gem: 0.238		0.123
gem: 0.145
v2026-04-28_232809 1003 0.151
mf: 0.250		0.155
gem: 0.334		0.127
icon: 0.254		0.114
gem: 0.247		0.116
gem: 0.148
v2026-04-28_232809 289 0.090
gfs: 0.100		0.156
mf: 0.198		0.002
gfs: 0.000		——
v2026-04-26_184501
v2026-04-28_161556
v2026-04-28_232809		 4 0.034
gfs: 0.000		0.001
gfs: 0.000		0.004
ecmwf: 0.000		——
v2026-04-23_163848
v2026-04-26_085202
v2026-04-26_184501		 4 0.002
gfs: 0.000		0.007
gfs: 0.000		0.000
gfs: 0.000		——

Phase 3d · v2026-06-16_090731_phase3d Δ +0.002 vs prev train

Exact-runtime P(wet) classifier. Trains on raw S3 cycles (GFS + IFS oper + AIFS required, MO Global + UKV optional) instead of Open-Meteo offset_day, with rigorous (RunTime, ValidTime, Lead) provenance per row. UKV pulled per-V-hour with target-lead-aware tuples. Trained 2026-06-16. Metric: Test Brier.

Lead Blend Best single Δ vs best
+12h 0.145 precip_aifs (0.304) -52.4%
+24h 0.159 precip_moglobal (0.230) -30.5%
+48h 0.154 precip_gfs (0.332) -53.5%
+72h 0.163 precip_gfs (0.319) -48.8%
+96h 0.195 precip_gfs (0.327) -40.3%
+120h 0.207 precip_gfs (0.359) -42.4%
Verify history (13 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-05-08_061830_phase3d
v2026-06-07_140447_phase3d
v2026-06-14_144900_phase3d		 223 0.027
—		0.179
—		0.199
—		0.175
—		0.207
—		0.285
—
v2026-05-08_061830_phase3d
v2026-06-07_140447_phase3d		 234 0.188
—		0.213
—		0.190
—		0.174
—		0.255
—		0.319
—
v2026-05-08_061830_phase3d
v2026-05-31_204207_phase3d
v2026-06-07_140447_phase3d		 253 0.037
—		0.144
—		0.182
—		0.240
—		0.189
—		0.232
—
v2026-05-08_061830_phase3d
v2026-05-31_204207_phase3d
v2026-06-07_140447_phase3d		 253 0.004
—		0.247
—		0.207
—		0.281
—		0.176
—		0.221
—
v2026-05-08_061830_phase3d
v2026-05-31_204207_phase3d		 224 0.183
—		0.286
—		0.268
—		0.421
—		0.132
—		0.134
—
v2026-05-08_061830_phase3d
v2026-05-26_105702_phase3d
v2026-05-31_204207_phase3d		 133 0.002
—		0.074
—		0.039
—		0.016
—		0.055
—		0.069
—
v2026-05-08_061830_phase3d
v2026-05-24_131817_phase3d
v2026-05-26_105702_phase3d		 75 0.128
—		0.157
—		0.235
—		0.204
—		0.101
—		0.164
—
v2026-05-08_061830_phase3d
v2026-05-17_175709_phase3d
v2026-05-24_131817_phase3d		 36 0.002
—		0.021
—		0.215
—		0.350
—		0.359
—		0.294
—
v2026-05-08_061616_phase3d
v2026-05-08_061830_phase3d
v2026-05-17_175709_phase3d		 50 0.101
—		0.161
—		0.043
—		0.202
—		0.334
—		0.295
—
v2026-05-08_061616_phase3d
v2026-05-08_061830_phase3d		 49 0.089
—		0.085
—		0.120
—		0.145
—		0.339
—		0.336
—
v2026-05-08_061616_phase3d
v2026-05-08_061830_phase3d		 38 0.002
—		0.003
—		0.032
—		0.053
—		0.203
—		0.351
—
v2026-05-08_061616_phase3d 5 0.002
—		0.003
—		————
v2026-05-05_191123_phase3d 2 0.010
—		0.002
—		————

Phase 4a · v2026-06-14_123009_phase4a Δ +0.002 vs prev train

Phase 4a blender. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.129 (per-cell BART) (NaN val) —
+48h 0.139 (per-cell BART) (NaN val) —
+72h 0.149 (per-cell BART) (NaN val) —
+96h 0.162 (per-cell BART) (NaN val) —
+120h 0.173 (per-cell BART) (NaN val) —
Verify history (11 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_122520_phase4a 126 0.171
—		0.176
—		0.194
—		0.209
—		0.296
—		0.271
—
v2026-05-31_122003_phase4a
v2026-06-07_122520_phase4a		 149 0.369
—		0.208
—		0.238
—		0.334
—		0.228
—		0.207
—
v2026-05-31_122003_phase4a 159 —0.151
—		0.181
—		0.184
—		0.228
—		0.192
—
v2026-05-31_122003_phase4a 127 —0.165
—		0.196
—		0.203
—		0.284
—		0.239
—
v2026-05-24_122715_phase4a
v2026-05-31_122003_phase4a		 151 —0.158
—		0.218
—		0.325
—		0.111
—		0.127
—
v2026-05-24_122715_phase4a 125 —0.008
—		0.010
—		0.013
—		0.003
—		0.002
—
v2026-05-17_165023_phase4a
v2026-05-24_122715_phase4a		 135 —0.015
—		0.024
—		0.056
—		0.072
—		0.088
—
v2026-05-17_165023_phase4a 93 —0.205
—		0.200
—		0.176
—		0.235
—		0.265
—
v2026-05-17_165023_phase4a 21 —0.219
—		0.216
—		0.253
—		0.282
—		0.307
—
v2026-05-10_220841_phase4a 66 —0.333
—		0.354
—		0.370
—		0.370
—		0.398
—
v2026-05-10_220841_phase4a 15 —0.121
—		0.172
—		0.134
—		0.147
—		0.148
—

Phase 4b · v2026-06-14_150943_phase4b Δ +0.002 vs prev train

Phase 4b blender. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.126 p_3c (0.127) -0.4%
+48h 0.137 p_3o (0.138) -0.6%
+72h 0.144 p_3o (0.142) +1.3%
+96h 0.159 p_3o (0.159) -0.1%
+120h 0.169 p_3o (0.170) -0.8%
Verify history (11 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_142500_phase4b 536 0.140
jma: 0.194		0.175
jma: 0.291		0.184
jma: 0.245		0.199
jma: 0.221		0.232
jma: 0.293		0.209
ecmwf: 0.159
v2026-05-31_210032_phase4b
v2026-06-07_142500_phase4b		 594 0.357
gfs: 0.444		0.212
icon: 0.284		0.242
gfs: 0.341		0.308
gfs: 0.200		0.216
icon: 0.337		0.221
icon: 0.323
v2026-05-31_210032_phase4b 632 —0.163
ecmwf: 0.231		0.181
ecmwf: 0.250		0.186
ecmwf: 0.255		0.211
gfs: 0.297		0.211
icon: 0.310
v2026-05-31_210032_phase4b 490 —0.184
ecmwf: 0.243		0.202
jma: 0.223		0.215
ecmwf: 0.273		0.235
gfs: 0.356		0.302
aifs: 0.193
v2026-05-26_110622_phase4b
v2026-05-31_210032_phase4b		 464 —0.193
aifs: 0.255		0.260
jma: 0.244		0.383
aifs: 0.300		0.172
ecmwf: 0.239		0.222
ecmwf: 0.320
v2026-05-24_140756_phase4b
v2026-05-26_110622_phase4b		 366 —0.009
mf: 0.014		0.000
ecmwf: 0.000		0.001
icon: 0.000		0.001
gfs: 0.000		0.002
ecmwf: 0.000
v2026-05-17_185156_phase4b
v2026-05-24_140756_phase4b
v2026-05-26_110622_phase4b		 637 —0.058
ukmo: 0.059		0.035
icon: 0.036		0.057
gfs: 0.059		0.057
ukmo: 0.068		0.058
gfs: 0.064
v2026-05-17_185156_phase4b 471 —0.245
aifs: 0.312		0.163
ukmo: 0.190		0.144
mf: 0.191		0.233
ukmo: 0.303		0.281
gfs: 0.395
v2026-05-17_185156_phase4b 53 —0.200
icon: 0.226		0.206
icon: 0.273		0.299
icon: 0.375		0.280
ukmo: 0.396		0.349
ukmo: 0.396
v2026-05-12_202905_phase4b 89 —0.468
ukmo: 0.270		0.568
ecmwf: 0.315		0.482
gfs: 0.333		0.418
gfs: 0.438		—
v2026-05-12_202905_phase4b 6 —0.711
mf: 0.000		————

Phase 3oni · v2026-06-14_143409_phase3oni

Phase 3oni blender. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.132 precip_ukmo (0.228) -42.2%
+48h 0.146 precip_icon (0.274) -46.6%
+72h 0.153 precip_gem (0.293) -47.9%
+96h 0.167 precip_icon (0.295) -43.5%
+120h 0.182 precip_gfs (0.337) -45.9%
Verify history (no runs yet)

No verify rows yet for phase 3oni. Next cycle: Mon/Thu 09:30 UTC, then 5d ERA5 latency.

Precipitation — Princetown

Phase 3o · v2026-06-15_161744_phase3o Δ +0.000 vs prev train

Rich + orographic + upper-air P(wet), pooled across the 4 Bonehill gauges (110 features = rich 59 + 9 terrain + 42 multi-level pressure). Bonehill rich champion; the decorrelated UA-route member of the 3-way 4b blend. Trained 2026-06-15. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.121 precip_ukmo (0.209) -42.0%
+48h 0.135 precip_icon (0.235) -42.6%
+72h 0.142 precip_icon (0.253) -44.0%
+96h 0.157 precip_icon (0.271) -42.2%
+120h 0.167 precip_gfs (0.303) -44.9%
Verify history (7 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_133956_phase3o
v2026-06-10_171114_phase3o		 304 0.076
jma: 0.128		0.061
jma: 0.048		0.049
jma: 0.024		0.012
gfs: 0.000		0.145
jma: 0.165		0.112
ecmwf: 0.045
v2026-05-31_202616_phase3o
v2026-06-07_133956_phase3o
v2026-06-10_171114_phase3o		 584 0.321
gfs: 0.400		0.204
icon: 0.270		0.227
gfs: 0.318		0.302
gfs: 0.400		0.207
gem: 0.325		0.215
icon: 0.316
v2026-05-31_202616_phase3o 648 —0.159
mf: 0.231		0.174
ecmwf: 0.229		0.195
ecmwf: 0.271		0.211
gfs: 0.327		0.222
icon: 0.288
v2026-05-31_202616_phase3o 506 —0.184
mf: 0.281		0.199
ecmwf: 0.268		0.233
ecmwf: 0.296		0.232
icon: 0.394		0.351
aifs: 0.238
v2026-05-26_104548_phase3o
v2026-05-31_202616_phase3o		 472 —0.194
mf: 0.313		0.235
ecmwf: 0.321		0.332
icon: 0.300		0.165
ecmwf: 0.208		0.219
ecmwf: 0.314
v2026-05-26_104548_phase3o 376 —0.000
mf: 0.005		0.001
ecmwf: 0.000		0.001
icon: 0.000		0.001
gfs: 0.000		—
v2026-05-26_104548_phase3o 57 —0.000
ukmo: 0.000		————

Phase 3c · v2026-06-15_161504_phase3c Δ -0.002 vs prev train

Rich P(wet) classifier, 101 features — rich surface (humidity/dew-depression/cloud/EA persistence) + the multi-level pressure (upper-air) block (850/700/500 hPa temps, heights, winds, RH850 across 4 exact NWPs, lead-matched by a leak-free backward ASOF). UA added in-place 2026-06-02 (−3.7%→−6.8% Brier across 24→72h). Trained 2026-06-15. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.105 precip_mf (0.202) -47.9%
+48h 0.117 precip_mf (0.216) -45.8%
+72h 0.123 precip_mf (0.240) -48.7%
+96h 0.142 precip_icon (0.258) -45.0%
+120h 0.155 precip_ecmwf (0.289) -46.4%
Verify history (6 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_133602_phase3c
v2026-06-10_170708_phase3c		 304 0.102
jma: 0.128		0.074
jma: 0.048		0.032
jma: 0.024		0.012
gfs: 0.000		0.167
jma: 0.165		0.098
ecmwf: 0.045
v2026-05-31_202423_phase3c
v2026-06-07_133602_phase3c
v2026-06-10_170708_phase3c		 584 0.297
gfs: 0.400		0.200
icon: 0.270		0.245
gfs: 0.318		0.305
gfs: 0.400		0.208
gem: 0.325		0.208
icon: 0.316
v2026-05-31_202423_phase3c 648 —0.195
mf: 0.231		0.193
ecmwf: 0.229		0.232
ecmwf: 0.271		0.197
gfs: 0.327		0.204
icon: 0.288
v2026-05-31_202423_phase3c 506 —0.223
mf: 0.281		0.228
ecmwf: 0.268		0.280
ecmwf: 0.296		0.213
icon: 0.394		0.296
aifs: 0.238
v2026-05-31_202423_phase3c 208 —0.202
mf: 0.313		0.262
ecmwf: 0.321		0.384
icon: 0.300		——
v2026-04-23_154433_phase3c 11 —0.000
gfs: 0.000		0.000
ecmwf: 0.000		0.001
ecmwf: 0.000		——

Phase 3d · v2026-06-16_090904_phase3d Δ -0.006 vs prev train

Exact-runtime P(wet) classifier. Trains on raw S3 cycles (GFS + IFS oper + AIFS required, MO Global + UKV optional) instead of Open-Meteo offset_day, with rigorous (RunTime, ValidTime, Lead) provenance per row. UKV pulled per-V-hour with target-lead-aware tuples. Trained 2026-06-16. Metric: Test Brier.

Lead Blend Best single Δ vs best
+12h 0.090 precip_moglobal (0.171) -47.6%
+24h 0.104 precip_moglobal (0.177) -41.0%
+48h 0.112 precip_gfs (0.307) -63.6%
+72h 0.111 precip_moglobal (0.145) -23.6%
Verify history (7 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_140613_phase3d
v2026-06-14_145030_phase3d		 142 0.001
—		0.196
—		0.261
—		0.221
—		——
v2026-06-07_140613_phase3d 118 0.105
—		0.147
—		0.131
—		0.069
—		——
v2026-05-31_204321_phase3d
v2026-06-07_140613_phase3d		 148 0.033
—		0.069
—		0.213
—		0.250
—		——
v2026-05-31_204321_phase3d
v2026-06-07_140613_phase3d		 143 0.008
—		0.278
—		0.247
—		0.301
—		——
v2026-05-31_204321_phase3d 119 0.274
—		0.378
—		0.342
—		0.535
—		——
v2026-05-26_105752_phase3d
v2026-05-31_204321_phase3d		 110 0.001
—		0.012
—		0.005
—		0.012
—		——
v2026-05-26_105752_phase3d 82 0.005
—		0.005
—		0.004
—		———

Phase 4a · v2026-06-14_123009_phase4a Δ -0.000 vs prev train

Phase 4a blender. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.107 (per-cell BART) (NaN val) —
+48h 0.119 (per-cell BART) (NaN val) —
+72h 0.129 (per-cell BART) (NaN val) —
+96h 0.141 (per-cell BART) (NaN val) —
+120h 0.154 (per-cell BART) (NaN val) —
Verify history (5 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_122520_phase4a 122 0.118
—		0.146
—		0.145
—		0.140
—		0.253
—		0.141
—
v2026-05-31_122003_phase4a
v2026-06-07_122520_phase4a		 146 0.323
—		0.195
—		0.219
—		0.341
—		0.212
—		0.188
—
v2026-05-31_122003_phase4a 162 —0.151
—		0.170
—		0.181
—		0.211
—		0.176
—
v2026-05-31_122003_phase4a 130 —0.173
—		0.197
—		0.208
—		0.259
—		0.194
—
v2026-05-31_122003_phase4a 58 —0.162
—		0.198
—		0.285
—		——

Phase 4b · v2026-06-14_150943_phase4b Δ -0.001 vs prev train

Phase 4b blender. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.105 p_3c (0.106) -1.1%
+48h 0.117 p_3o (0.119) -1.3%
+72h 0.122 p_3o (0.121) +0.6%
+96h 0.137 p_3o (0.137) 0.0%
+120h 0.150 p_3o (0.150) -0.1%
Verify history (5 runs)

Mon + Thu rolling Brier. Per-lead cells turn red when the rolling metric breaches the lead-specific drift threshold; check the verify report for the per-cell breakdown. Version column names the trained model — a fresh champion takes ~5-9d to show.

Run (UTC) Version N +12h+24h+48h+72h+96h+120h
v2026-06-07_142500_phase4b 528 0.091
jma: 0.135		0.132
ukmo: 0.241		0.132
jma: 0.205		0.142
jma: 0.155		0.165
jma: 0.174		0.105
ecmwf: 0.045
v2026-05-31_210032_phase4b
v2026-06-07_142500_phase4b		 584 0.335
gfs: 0.444		0.196
icon: 0.262		0.232
gfs: 0.318		0.315
gfs: 0.400		0.204
gem: 0.325		0.202
icon: 0.316
v2026-05-31_210032_phase4b 648 —0.164
mf: 0.231		0.174
ecmwf: 0.229		0.184
ecmwf: 0.271		0.201
gfs: 0.327		0.197
icon: 0.288
v2026-05-31_210032_phase4b 506 —0.189
mf: 0.281		0.203
ecmwf: 0.268		0.215
ecmwf: 0.296		0.221
icon: 0.394		0.287
aifs: 0.238
v2026-05-31_210032_phase4b 208 —0.191
mf: 0.313		0.238
ecmwf: 0.321		0.305
icon: 0.300		——

Phase 3oni · v2026-06-14_143409_phase3oni

Phase 3oni blender. Trained 2026-06-14. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.122 precip_ukmo (0.209) -41.8%
+48h 0.136 precip_icon (0.235) -42.3%
+72h 0.142 precip_icon (0.253) -44.1%
+96h 0.154 precip_icon (0.271) -43.3%
+120h 0.167 precip_gfs (0.304) -45.0%
Verify history (no runs yet)

No verify rows yet for phase 3oni. Next cycle: Mon/Thu 09:30 UTC, then 5d ERA5 latency.