WeatherBlend

Multi-model forecast blending for Bonehill Rocks, Dartmoor

Dry-window models

Per-(station, window) P(N-hour dry block in 09–18 local). 3b LightGBM + 3p copula MC. Brier, lower better. Δ vs best single NWP — negative = blend wins.

Dry window — Bellever Dartmoor — 2-hour

Phase 3p · v2026-08-02_150329_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (14 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
v2026-07-26_150843_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144837_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144837_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145513_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145513_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144709_phase3p 14 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144709_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150944_phase3p 16 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150944_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-06-21_145445_phase3p 19 0.000
—		0.000
—		0.000
—		——
v2026-06-14_150640_phase3p
v2026-06-21_145445_phase3p		 24 0.000
—		0.000
—		0.002
—		——
v2026-06-14_150640_phase3p 22 0.035
—		0.021
—		0.003
—		——
v2026-06-14_150640_phase3p 10 0.075
—		0.056
—		0.012
—		——
v2026-06-10_080817_phase3p
v2026-06-11_054510_phase3p		 8 0.001
—		0.000
—		0.000
—		——

Dry window — Bellever Dartmoor — 3-hour

Phase 3b · v2026-08-02_145011 Δ -0.004 vs prev train

53-feature LightGBM per-(station, window). Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.037 has_dry_window_mf (0.078) -52.8%
+48h 0.040 has_dry_window_mf (0.099) -59.2%
+72h 0.044 has_dry_window_gfs (0.092) -51.8%
Verify history (29 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
v2026-07-26_145548 8 0.000
—		0.000
—		0.000
—		——
v2026-07-19_143826 22 0.000
—		0.000
—		0.004
—		——
v2026-07-19_143826 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_144401 22 0.000
—		0.000
—		0.000
—		——
v2026-07-12_144401 10 0.000
—		0.000
—		0.000
—		——
v2026-07-05_143708 14 0.000
—		0.000
—		0.000
—		——
v2026-07-05_143708 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_162100 16 0.031
—		0.015
—		0.000
—		——
v2026-06-28_162100 8 0.061
—		0.039
—		0.000
—		——
v2026-06-21_143941 19 0.001
—		0.001
—		0.000
—		——
v2026-06-14_145202
v2026-06-21_143941		 24 0.001
—		0.000
—		0.002
—		——
v2026-06-14_145202 22 0.023
—		0.037
—		0.002
—		——
v2026-06-14_145202 10 0.049
—		0.110
—		0.008
—		——
v2026-06-07_140739 26 0.060
—		0.047
—		0.074
—		——
v2026-06-07_140739 11 0.072
—		0.014
—		0.141
—		——
v2026-05-31_204435 23 0.305
—		0.219
—		0.168
—		——
v2026-05-31_204435 18 0.373
—		0.280
—		0.213
—		——
v2026-05-31_204435 6 0.676
—		0.306
—		0.101
—		——
v2026-05-24_131915 29 0.000
—		0.000
—		0.000
—		——
v2026-05-24_131915 14 0.000
—		0.000
—		0.000
—		——
v2026-05-14_052009 32 0.062
—		0.123
—		0.054
—		——
v2026-05-14_052009 14 0.141
—		0.345
—		0.206
—		——
v2026-04-29_120905 50 0.005
—		0.015
—		0.043
—		——
v2026-04-29_120905 57 0.006
—		0.002
—		0.097
—		——
v2026-04-29_120905 64 0.043
—		0.032
—		0.118
—		——
v2026-04-29_120905 54 0.049
—		0.039
—		0.127
—		——
v2026-04-29_120905 25 0.030
—		0.078
—		0.155
—		——
v2026-04-27_192657
v2026-04-28_161801
v2026-04-29_120905		 5 0.003
—		0.000
—		0.000
—		——
v2026-04-23_101107
v2026-04-27_192657		 12 0.000
—		0.000
—		0.000
—		——

Phase 3p · v2026-08-02_150329_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (20 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
v2026-07-26_150843_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144837_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144837_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145513_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145513_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144709_phase3p 14 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144709_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150944_phase3p 16 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150944_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-06-21_145445_phase3p 19 0.000
—		0.000
—		0.000
—		——
v2026-06-14_150640_phase3p
v2026-06-21_145445_phase3p		 24 0.000
—		0.000
—		0.008
—		——
v2026-06-14_150640_phase3p 22 0.067
—		0.058
—		0.013
—		——
v2026-06-14_150640_phase3p 10 0.143
—		0.144
—		0.055
—		——
v2026-06-10_080817_phase3p
v2026-06-11_054510_phase3p		 8 0.004
—		0.000
—		0.000
—		——
v2026-06-07_142135_phase3p 11 0.042
—		0.036
—		0.012
—		——
v2026-05-31_205722_phase3p 23 0.131
—		0.158
—		0.221
—		——
v2026-05-31_205722_phase3p 18 0.159
—		0.200
—		0.275
—		——
v2026-05-31_205722_phase3p 6 0.237
—		0.129
—		0.207
—		——
v2026-05-26_110451_phase3p 18 0.000
—		0.000
—		0.000
—		——
v2026-05-26_110451_phase3p 3 0.000
—		————

Dry window — Bellever Dartmoor — 4-hour

Phase 3b · v2026-08-02_145143 Δ -0.003 vs prev train

53-feature LightGBM per-(station, window). Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.048 has_dry_window_mf (0.128) -62.6%
+48h 0.065 has_dry_window_aifs (0.220) -70.4%
+72h 0.064 has_dry_window_jma (0.128) -49.8%
Verify history (29 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
v2026-07-26_145718 8 0.000
—		0.000
—		0.000
—		——
v2026-07-19_143937 22 0.000
—		0.002
—		0.006
—		——
v2026-07-19_143937 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_144518 22 0.000
—		0.001
—		0.013
—		——
v2026-07-12_144518 10 0.000
—		0.004
—		0.038
—		——
v2026-07-05_143817 14 0.000
—		0.000
—		0.000
—		——
v2026-07-05_143817 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_162222 16 0.035
—		0.030
—		0.001
—		——
v2026-06-28_162222 8 0.069
—		0.071
—		0.000
—		——
v2026-06-21_144059 19 0.005
—		0.011
—		0.008
—		——
v2026-06-14_145318
v2026-06-21_144059		 24 0.003
—		0.009
—		0.008
—		——
v2026-06-14_145318 22 0.024
—		0.136
—		0.009
—		——
v2026-06-14_145318 10 0.042
—		0.404
—		0.048
—		——
v2026-06-07_140851 26 0.053
—		0.051
—		0.062
—		——
v2026-06-07_140851 11 0.055
—		0.173
—		0.114
—		——
v2026-05-31_204541 23 0.190
—		0.121
—		0.088
—		——
v2026-05-31_204541 18 0.227
—		0.153
—		0.108
—		——
v2026-05-31_204541 6 0.543
—		0.104
—		0.073
—		——
v2026-05-24_132009 29 0.000
—		0.000
—		0.000
—		——
v2026-05-24_132009 14 0.000
—		0.000
—		0.000
—		——
v2026-05-14_052407 32 0.175
—		0.102
—		0.087
—		——
v2026-05-14_052407 14 0.401
—		0.286
—		0.332
—		——
v2026-04-29_121003 50 0.076
—		0.054
—		0.061
—		——
v2026-04-29_121003 57 0.002
—		0.021
—		0.101
—		——
v2026-04-29_121003 64 0.058
—		0.149
—		0.142
—		——
v2026-04-29_121003 54 0.068
—		0.185
—		0.161
—		——
v2026-04-29_121003 25 0.055
—		0.212
—		0.202
—		——
v2026-04-27_192749
v2026-04-28_161822
v2026-04-29_121003		 5 0.000
—		0.000
—		0.000
—		——
v2026-04-23_101150
v2026-04-27_192749		 12 0.000
—		0.000
—		0.000
—		——

Phase 3p · v2026-08-02_150329_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (20 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
v2026-07-26_150843_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144837_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144837_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145513_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145513_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144709_phase3p 14 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144709_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150944_phase3p 16 0.001
—		0.001
—		0.000
—		——
v2026-06-28_150944_phase3p 8 0.002
—		0.001
—		0.000
—		——
v2026-06-21_145445_phase3p 19 0.000
—		0.000
—		0.001
—		——
v2026-06-14_150640_phase3p
v2026-06-21_145445_phase3p		 24 0.000
—		0.000
—		0.020
—		——
v2026-06-14_150640_phase3p 22 0.103
—		0.103
—		0.034
—		——
v2026-06-14_150640_phase3p 10 0.214
—		0.242
—		0.119
—		——
v2026-06-10_080817_phase3p
v2026-06-11_054510_phase3p		 8 0.012
—		0.000
—		0.002
—		——
v2026-06-07_142135_phase3p 11 0.098
—		0.085
—		0.033
—		——
v2026-05-31_205722_phase3p 23 0.076
—		0.104
—		0.154
—		——
v2026-05-31_205722_phase3p 18 0.087
—		0.125
—		0.184
—		——
v2026-05-31_205722_phase3p 6 0.133
—		0.060
—		0.101
—		——
v2026-05-26_110451_phase3p 18 0.000
—		0.000
—		0.000
—		——
v2026-05-26_110451_phase3p 3 0.000
—		————

Dry window — Bellever Dartmoor — 5-hour

Phase 3p · v2026-08-02_150329_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (14 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
v2026-07-26_150843_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144837_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144837_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145513_phase3p 22 0.001
—		0.000
—		0.001
—		——
v2026-07-12_145513_phase3p 10 0.000
—		0.000
—		0.002
—		——
v2026-07-05_144709_phase3p 14 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144709_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150944_phase3p 16 0.004
—		0.003
—		0.000
—		——
v2026-06-28_150944_phase3p 8 0.008
—		0.007
—		0.000
—		——
v2026-06-21_145445_phase3p 19 0.000
—		0.002
—		0.003
—		——
v2026-06-14_150640_phase3p
v2026-06-21_145445_phase3p		 24 0.000
—		0.000
—		0.049
—		——
v2026-06-14_150640_phase3p 22 0.039
—		0.084
—		0.083
—		——
v2026-06-14_150640_phase3p 10 0.036
—		0.090
—		0.301
—		——
v2026-06-10_080817_phase3p
v2026-06-11_054510_phase3p		 8 0.034
—		0.000
—		0.006
—		——

Dry window — Bellever Dartmoor — 6-hour

Phase 3b · v2026-08-02_145311 Δ +0.003 vs prev train

53-feature LightGBM per-(station, window). Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.100 has_dry_window_aifs (0.284) -64.7%
+48h 0.109 has_dry_window_jma (0.220) -50.4%
+72h 0.132 has_dry_window_jma (0.234) -43.8%
Verify history (29 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
v2026-07-26_145844 8 0.000
—		0.001
—		0.002
—		——
v2026-07-19_144045 22 0.013
—		0.005
—		0.049
—		——
v2026-07-19_144045 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_144632 22 0.021
—		0.019
—		0.087
—		——
v2026-07-12_144632 10 0.000
—		0.005
—		0.415
—		——
v2026-07-05_143924 14 0.000
—		0.000
—		0.021
—		——
v2026-07-05_143924 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_162341 16 0.141
—		0.072
—		0.020
—		——
v2026-06-28_162341 8 0.272
—		0.179
—		0.000
—		——
v2026-06-21_144216 19 0.094
—		0.033
—		0.080
—		——
v2026-06-14_145432
v2026-06-21_144216		 24 0.084
—		0.001
—		0.046
—		——
v2026-06-14_145432 22 0.156
—		0.106
—		0.071
—		——
v2026-06-14_145432 10 0.118
—		0.121
—		0.253
—		——
v2026-06-07_141001 26 0.110
—		0.075
—		0.025
—		——
v2026-06-07_141001 11 0.252
—		0.234
—		0.081
—		——
v2026-05-31_204645 23 0.094
—		0.060
—		0.143
—		——
v2026-05-31_204645 18 0.074
—		0.062
—		0.118
—		——
v2026-05-31_204645 6 0.000
—		0.001
—		0.008
—		——
v2026-05-24_132101 29 0.012
—		0.021
—		0.007
—		——
v2026-05-24_132101 14 0.002
—		0.016
—		0.045
—		——
v2026-05-14_052612 32 0.241
—		0.134
—		0.160
—		——
v2026-05-14_052612 14 0.552
—		0.376
—		0.613
—		——
v2026-05-03_120517 50 0.128
—		0.049
—		0.203
—		——
v2026-05-03_120517 51 0.158
—		0.162
—		0.192
—		——
v2026-05-03_120517 31 0.110
—		0.288
—		0.311
—		——
v2026-05-03_120517 17 0.149
—		0.397
—		0.732
—		——
v2026-04-29_121058 25 0.079
—		0.031
—		0.125
—		——
v2026-04-27_192839
v2026-04-28_161844
v2026-04-29_121058		 5 0.013
—		0.002
—		0.002
—		——
v2026-04-23_101214
v2026-04-27_182428
v2026-04-27_192839		 11 0.001
—		0.000
—		0.001
—		——

Phase 3p · v2026-08-02_150329_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (20 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
v2026-07-26_150843_phase3p 8 0.000
—		0.001
—		0.001
—		——
v2026-07-19_144837_phase3p 22 0.001
—		0.001
—		0.002
—		——
v2026-07-19_144837_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145513_phase3p 22 0.002
—		0.001
—		0.004
—		——
v2026-07-12_145513_phase3p 10 0.000
—		0.001
—		0.003
—		——
v2026-07-05_144709_phase3p 14 0.000
—		0.000
—		0.001
—		——
v2026-07-05_144709_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150944_phase3p 16 0.011
—		0.006
—		0.001
—		——
v2026-06-28_150944_phase3p 8 0.023
—		0.013
—		0.000
—		——
v2026-06-21_145445_phase3p 19 0.001
—		0.005
—		0.006
—		——
v2026-06-14_150640_phase3p
v2026-06-21_145445_phase3p		 24 0.000
—		0.001
—		0.059
—		——
v2026-06-14_150640_phase3p 22 0.074
—		0.122
—		0.098
—		——
v2026-06-14_150640_phase3p 10 0.061
—		0.091
—		0.204
—		——
v2026-06-10_080817_phase3p
v2026-06-11_054510_phase3p		 8 0.059
—		0.001
—		0.016
—		——
v2026-06-07_142135_phase3p 11 0.270
—		0.303
—		0.368
—		——
v2026-05-31_205722_phase3p 23 0.062
—		0.057
—		0.116
—		——
v2026-05-31_205722_phase3p 18 0.053
—		0.044
—		0.113
—		——
v2026-05-31_205722_phase3p 6 0.032
—		0.011
—		0.012
—		——
v2026-05-26_110451_phase3p 18 0.001
—		0.001
—		0.002
—		——
v2026-05-26_110451_phase3p 3 0.000
—		————

Dry window — Bovey Tracey — 2-hour

Phase 3p · v2026-08-02_150331_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (14 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
v2026-07-26_150845_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144839_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144839_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145515_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145515_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144711_phase3p 14 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144711_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150945_phase3p 16 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150945_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-06-21_145447_phase3p 19 0.000
—		0.000
—		0.000
—		——
v2026-06-14_150642_phase3p
v2026-06-21_145447_phase3p		 24 0.000
—		0.000
—		0.000
—		——
v2026-06-14_150642_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-06-14_150642_phase3p 10 0.000
—		0.001
—		0.000
—		——
v2026-06-10_080819_phase3p
v2026-06-11_054512_phase3p		 8 0.000
—		0.000
—		0.000
—		——

Dry window — Bovey Tracey — 3-hour

Phase 3b · v2026-08-02_145440 Δ -0.007 vs prev train

53-feature LightGBM per-(station, window). Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.012 has_dry_window_icon (0.007) +66.4%
+48h 0.013 has_dry_window_mf (0.044) -71.1%
+72h 0.013 has_dry_window_mf (0.066) -80.2%
Verify history (26 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
v2026-07-26_150010 8 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144153 22 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144153 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_144746 22 0.000
—		0.000
—		0.000
—		——
v2026-07-12_144746 10 0.000
—		0.000
—		0.001
—		——
v2026-07-05_144030 14 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144030 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_162500 16 0.000
—		0.001
—		0.000
—		——
v2026-06-28_162500 8 0.000
—		0.002
—		0.000
—		——
v2026-06-21_144331 19 0.000
—		0.001
—		0.001
—		——
v2026-06-14_145544
v2026-06-21_144331		 24 0.000
—		0.001
—		0.000
—		——
v2026-06-14_145544 22 0.000
—		0.002
—		0.000
—		——
v2026-06-14_145544 10 0.000
—		0.006
—		0.001
—		——
v2026-06-07_141111 26 0.004
—		0.005
—		0.003
—		——
v2026-06-07_141111 11 0.000
—		0.006
—		0.001
—		——
v2026-05-31_204749 28 0.027
—		0.016
—		0.023
—		——
v2026-05-31_204749 24 0.031
—		0.020
—		0.030
—		——
v2026-05-31_204749 11 0.017
—		0.005
—		0.088
—		——
v2026-05-24_132152 29 0.000
—		0.000
—		0.000
—		——
v2026-05-24_132152 14 0.000
—		0.000
—		0.000
—		——
v2026-05-14_052815 32 0.002
—		0.008
—		0.002
—		——
v2026-05-14_052815 14 0.003
—		0.022
—		0.006
—		——
v2026-05-03_233816 50 0.000
—		0.000
—		0.003
—		——
v2026-05-03_233816 51 0.000
—		0.001
—		0.004
—		——
v2026-05-03_233816 24 0.001
—		0.001
—		0.010
—		——
v2026-05-03_233816 10 0.001
—		0.004
—		———

Phase 3p · v2026-08-02_150331_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (20 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
v2026-07-26_150845_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144839_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144839_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145515_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145515_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144711_phase3p 14 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144711_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150945_phase3p 16 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150945_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-06-21_145447_phase3p 19 0.000
—		0.000
—		0.000
—		——
v2026-06-14_150642_phase3p
v2026-06-21_145447_phase3p		 24 0.000
—		0.000
—		0.000
—		——
v2026-06-14_150642_phase3p 22 0.001
—		0.002
—		0.000
—		——
v2026-06-14_150642_phase3p 10 0.001
—		0.004
—		0.001
—		——
v2026-06-10_080819_phase3p
v2026-06-11_054512_phase3p		 8 0.000
—		0.000
—		0.000
—		——
v2026-06-07_142137_phase3p 11 0.002
—		0.002
—		0.001
—		——
v2026-05-31_205723_phase3p 28 0.053
—		0.049
—		0.018
—		——
v2026-05-31_205723_phase3p 24 0.061
—		0.062
—		0.023
—		——
v2026-05-31_205723_phase3p 11 0.036
—		0.096
—		0.092
—		——
v2026-05-26_110452_phase3p 18 0.000
—		0.000
—		0.000
—		——
v2026-05-26_110452_phase3p 3 0.000
—		————

Dry window — Bovey Tracey — 4-hour

Phase 3b · v2026-08-02_145607 Δ -0.006 vs prev train

53-feature LightGBM per-(station, window). Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.028 has_dry_window_mf (0.124) -77.2%
+48h 0.021 has_dry_window_mf (0.109) -80.9%
+72h 0.025 has_dry_window_gfs (0.109) -77.3%
Verify history (26 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
v2026-07-26_150135 8 0.000
—		0.000
—		0.001
—		——
v2026-07-19_144259 22 0.002
—		0.000
—		0.001
—		——
v2026-07-19_144259 10 0.003
—		0.000
—		0.001
—		——
v2026-07-12_144900 22 0.002
—		0.001
—		0.001
—		——
v2026-07-12_144900 10 0.002
—		0.001
—		0.002
—		——
v2026-07-05_144135 14 0.003
—		0.001
—		0.002
—		——
v2026-07-05_144135 10 0.003
—		0.001
—		0.002
—		——
v2026-06-28_162617 16 0.002
—		0.000
—		0.001
—		——
v2026-06-28_162617 8 0.003
—		0.000
—		0.000
—		——
v2026-06-21_144445 19 0.001
—		0.000
—		0.000
—		——
v2026-06-14_145655
v2026-06-21_144445		 24 0.001
—		0.000
—		0.001
—		——
v2026-06-14_145655 22 0.003
—		0.003
—		0.001
—		——
v2026-06-14_145655 10 0.005
—		0.008
—		0.002
—		——
v2026-06-07_141220 26 0.026
—		0.018
—		0.022
—		——
v2026-06-07_141220 11 0.018
—		0.021
—		0.028
—		——
v2026-05-31_204851 28 0.114
—		0.056
—		0.085
—		——
v2026-05-31_204851 24 0.133
—		0.071
—		0.108
—		——
v2026-05-31_204851 11 0.096
—		0.066
—		0.342
—		——
v2026-05-24_132243 29 0.001
—		0.001
—		0.002
—		——
v2026-05-24_132243 14 0.001
—		0.001
—		0.004
—		——
v2026-05-14_053030 32 0.001
—		0.009
—		0.003
—		——
v2026-05-14_053030 14 0.003
—		0.024
—		0.009
—		——
v2026-05-03_233924 50 0.004
—		0.003
—		0.013
—		——
v2026-05-03_233924 51 0.002
—		0.002
—		0.014
—		——
v2026-05-03_233924 24 0.003
—		0.004
—		0.031
—		——
v2026-05-03_233924 10 0.005
—		0.010
—		———

Phase 3p · v2026-08-02_150331_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (20 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
v2026-07-26_150845_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144839_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144839_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145515_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145515_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144711_phase3p 14 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144711_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150945_phase3p 16 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150945_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-06-21_145447_phase3p 19 0.000
—		0.000
—		0.000
—		——
v2026-06-14_150642_phase3p
v2026-06-21_145447_phase3p		 24 0.000
—		0.000
—		0.001
—		——
v2026-06-14_150642_phase3p 22 0.002
—		0.006
—		0.002
—		——
v2026-06-14_150642_phase3p 10 0.005
—		0.014
—		0.005
—		——
v2026-06-10_080819_phase3p
v2026-06-11_054512_phase3p		 8 0.000
—		0.000
—		0.000
—		——
v2026-06-07_142137_phase3p 11 0.010
—		0.007
—		0.003
—		——
v2026-05-31_205723_phase3p 28 0.127
—		0.098
—		0.048
—		——
v2026-05-31_205723_phase3p 24 0.148
—		0.122
—		0.060
—		——
v2026-05-31_205723_phase3p 11 0.125
—		0.184
—		0.199
—		——
v2026-05-26_110452_phase3p 18 0.000
—		0.000
—		0.000
—		——
v2026-05-26_110452_phase3p 3 0.000
—		————

Dry window — Bovey Tracey — 5-hour

Phase 3p · v2026-08-02_150331_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (14 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
v2026-07-26_150845_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144839_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144839_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145515_phase3p 22 0.001
—		0.000
—		0.001
—		——
v2026-07-12_145515_phase3p 10 0.000
—		0.000
—		0.003
—		——
v2026-07-05_144711_phase3p 14 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144711_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150945_phase3p 16 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150945_phase3p 8 0.001
—		0.001
—		0.000
—		——
v2026-06-21_145447_phase3p 19 0.000
—		0.000
—		0.001
—		——
v2026-06-14_150642_phase3p
v2026-06-21_145447_phase3p		 24 0.000
—		0.000
—		0.006
—		——
v2026-06-14_150642_phase3p 22 0.009
—		0.021
—		0.009
—		——
v2026-06-14_150642_phase3p 10 0.015
—		0.043
—		0.020
—		——
v2026-06-10_080819_phase3p
v2026-06-11_054512_phase3p		 8 0.001
—		0.000
—		0.002
—		——

Dry window — Bovey Tracey — 6-hour

Phase 3b · v2026-08-02_145734 Δ -0.013 vs prev train

53-feature LightGBM per-(station, window). Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.066 has_dry_window_mf (0.139) -52.2%
+48h 0.066 has_dry_window_ecmwf (0.292) -77.3%
+72h 0.077 has_dry_window_ecmwf (0.285) -73.0%
Verify history (26 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
v2026-07-26_150300 8 0.002
—		0.001
—		0.000
—		——
v2026-07-19_144406 22 0.014
—		0.006
—		0.005
—		——
v2026-07-19_144406 10 0.015
—		0.009
—		0.002
—		——
v2026-07-12_145014 22 0.018
—		0.019
—		0.004
—		——
v2026-07-12_145014 10 0.020
—		0.013
—		0.004
—		——
v2026-07-05_144241 14 0.023
—		0.012
—		0.002
—		——
v2026-07-05_144241 10 0.020
—		0.008
—		0.002
—		——
v2026-06-28_162735 16 0.022
—		0.007
—		0.002
—		——
v2026-06-28_162735 8 0.034
—		0.010
—		0.000
—		——
v2026-06-21_144600 19 0.027
—		0.005
—		0.000
—		——
v2026-06-14_145807
v2026-06-21_144600		 24 0.036
—		0.008
—		0.007
—		——
v2026-06-14_145807 22 0.014
—		0.013
—		0.002
—		——
v2026-06-14_145807 10 0.027
—		0.029
—		0.007
—		——
v2026-06-07_141329 26 0.114
—		0.055
—		0.085
—		——
v2026-06-07_141329 11 0.065
—		0.080
—		0.064
—		——
v2026-05-31_204955 28 0.106
—		0.161
—		0.194
—		——
v2026-05-31_204955 24 0.123
—		0.203
—		0.246
—		——
v2026-05-31_204955 11 0.091
—		0.162
—		0.010
—		——
v2026-05-24_132334 29 0.017
—		0.004
—		0.001
—		——
v2026-05-24_132334 14 0.033
—		0.008
—		0.001
—		——
v2026-05-14_053806 32 0.026
—		0.088
—		0.017
—		——
v2026-05-14_053806 14 0.058
—		0.245
—		0.064
—		——
v2026-05-03_234027 50 0.047
—		0.146
—		0.168
—		——
v2026-05-03_234027 51 0.012
—		0.018
—		0.017
—		——
v2026-05-03_234027 24 0.024
—		0.038
—		0.029
—		——
v2026-05-03_234027 10 0.050
—		0.103
—		———

Phase 3p · v2026-08-02_150331_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (20 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
v2026-07-26_150845_phase3p 8 0.000
—		0.001
—		0.001
—		——
v2026-07-19_144839_phase3p 22 0.001
—		0.000
—		0.001
—		——
v2026-07-19_144839_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145515_phase3p 22 0.003
—		0.002
—		0.007
—		——
v2026-07-12_145515_phase3p 10 0.001
—		0.001
—		0.008
—		——
v2026-07-05_144711_phase3p 14 0.000
—		0.000
—		0.002
—		——
v2026-07-05_144711_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150945_phase3p 16 0.002
—		0.001
—		0.000
—		——
v2026-06-28_150945_phase3p 8 0.004
—		0.002
—		0.000
—		——
v2026-06-21_145447_phase3p 19 0.000
—		0.002
—		0.003
—		——
v2026-06-14_150642_phase3p
v2026-06-21_145447_phase3p		 24 0.000
—		0.001
—		0.013
—		——
v2026-06-14_150642_phase3p 22 0.022
—		0.044
—		0.020
—		——
v2026-06-14_150642_phase3p 10 0.035
—		0.086
—		0.039
—		——
v2026-06-10_080819_phase3p
v2026-06-11_054512_phase3p		 8 0.003
—		0.001
—		0.006
—		——
v2026-06-07_142137_phase3p 11 0.066
—		0.053
—		0.028
—		——
v2026-05-31_205723_phase3p 28 0.185
—		0.150
—		0.192
—		——
v2026-05-31_205723_phase3p 24 0.213
—		0.182
—		0.236
—		——
v2026-05-31_205723_phase3p 11 0.236
—		0.115
—		0.067
—		——
v2026-05-26_110452_phase3p 18 0.000
—		0.001
—		0.001
—		——
v2026-05-26_110452_phase3p 3 0.000
—		————

Dry window — Dartmoor Nr Hexworthy — 2-hour

Phase 3p · v2026-08-02_150333_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (14 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
v2026-07-26_150846_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144840_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144840_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145516_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145516_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144712_phase3p 14 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144712_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150946_phase3p 16 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150946_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-06-21_145448_phase3p 19 0.000
—		0.000
—		0.000
—		——
v2026-06-14_150643_phase3p
v2026-06-21_145448_phase3p		 24 0.000
—		0.000
—		0.002
—		——
v2026-06-14_150643_phase3p 22 0.035
—		0.022
—		0.003
—		——
v2026-06-14_150643_phase3p 10 0.077
—		0.059
—		0.014
—		——
v2026-06-10_080820_phase3p
v2026-06-11_054514_phase3p		 8 0.001
—		0.000
—		0.000
—		——

Dry window — Dartmoor Nr Hexworthy — 3-hour

Phase 3b · v2026-08-02_145902 Δ -0.000 vs prev train

53-feature LightGBM per-(station, window). Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.076 has_dry_window_ecmwf (0.206) -62.9%
+48h 0.078 has_dry_window_mf (0.135) -42.3%
+72h 0.079 has_dry_window_ecmwf (0.248) -68.1%
Verify history (28 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
v2026-07-26_150425 8 0.001
—		0.001
—		0.003
—		——
v2026-07-19_144513 22 0.001
—		0.002
—		0.013
—		——
v2026-07-19_144513 10 0.000
—		0.002
—		0.003
—		——
v2026-07-12_145128 22 0.001
—		0.003
—		0.006
—		——
v2026-07-12_145128 10 0.000
—		0.004
—		0.004
—		——
v2026-07-05_144347 14 0.000
—		0.002
—		0.003
—		——
v2026-07-05_144347 10 0.000
—		0.002
—		0.003
—		——
v2026-06-28_162853 16 0.038
—		0.021
—		0.002
—		——
v2026-06-28_162853 8 0.075
—		0.050
—		0.001
—		——
v2026-06-21_144715 19 0.003
—		0.003
—		0.002
—		——
v2026-06-14_145920
v2026-06-21_144715		 24 0.000
—		0.002
—		0.011
—		——
v2026-06-14_145920 22 0.077
—		0.105
—		0.016
—		——
v2026-06-14_145920 10 0.143
—		0.307
—		0.063
—		——
v2026-06-07_141438 26 0.016
—		0.115
—		0.063
—		——
v2026-06-07_141438 11 0.028
—		0.108
—		0.055
—		——
v2026-05-31_205058 23 0.130
—		0.223
—		0.177
—		——
v2026-05-31_205058 18 0.158
—		0.283
—		0.221
—		——
v2026-05-31_205058 6 0.385
—		0.305
—		0.201
—		——
v2026-05-24_132426 29 0.000
—		0.002
—		0.003
—		——
v2026-05-24_132426 14 0.000
—		0.002
—		0.004
—		——
v2026-05-14_054446 32 0.400
—		0.233
—		0.188
—		——
v2026-05-14_054446 14 0.284
—		0.273
—		0.124
—		——
v2026-04-29_121459 50 0.081
—		0.156
—		0.153
—		——
v2026-04-29_121459 57 0.002
—		0.006
—		0.048
—		——
v2026-04-29_121459 64 0.018
—		0.032
—		0.048
—		——
v2026-04-29_121459 54 0.021
—		0.038
—		0.057
—		——
v2026-04-29_121459 25 0.005
—		0.070
—		0.061
—		——
v2026-04-29_121459 5 0.000
—		————

Phase 3p · v2026-08-02_150333_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (20 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
v2026-07-26_150846_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144840_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144840_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145516_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145516_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144712_phase3p 14 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144712_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150946_phase3p 16 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150946_phase3p 8 0.001
—		0.000
—		0.000
—		——
v2026-06-21_145448_phase3p 19 0.000
—		0.000
—		0.000
—		——
v2026-06-14_150643_phase3p
v2026-06-21_145448_phase3p		 24 0.000
—		0.000
—		0.009
—		——
v2026-06-14_150643_phase3p 22 0.071
—		0.062
—		0.015
—		——
v2026-06-14_150643_phase3p 10 0.151
—		0.157
—		0.065
—		——
v2026-06-10_080820_phase3p
v2026-06-11_054514_phase3p		 8 0.004
—		0.000
—		0.000
—		——
v2026-06-07_142138_phase3p 11 0.049
—		0.037
—		0.011
—		——
v2026-05-31_205725_phase3p 23 0.124
—		0.152
—		0.206
—		——
v2026-05-31_205725_phase3p 18 0.151
—		0.192
—		0.256
—		——
v2026-05-31_205725_phase3p 6 0.226
—		0.120
—		0.216
—		——
v2026-05-26_110453_phase3p 18 0.000
—		0.000
—		0.000
—		——
v2026-05-26_110453_phase3p 3 0.000
—		————

Dry window — Dartmoor Nr Hexworthy — 4-hour

Phase 3b · v2026-08-02_150030 Δ -0.006 vs prev train

53-feature LightGBM per-(station, window). Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.104 has_dry_window_aifs (0.213) -51.1%
+48h 0.107 has_dry_window_aifs (0.241) -55.4%
+72h 0.102 has_dry_window_aifs (0.213) -52.2%
Verify history (28 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
v2026-07-26_150550 8 0.002
—		0.006
—		0.002
—		——
v2026-07-19_144620 22 0.002
—		0.004
—		0.011
—		——
v2026-07-19_144620 10 0.001
—		0.001
—		0.002
—		——
v2026-07-12_145243 22 0.002
—		0.004
—		0.011
—		——
v2026-07-12_145243 10 0.002
—		0.007
—		0.023
—		——
v2026-07-05_144453 14 0.000
—		0.002
—		0.003
—		——
v2026-07-05_144453 10 0.000
—		0.002
—		0.003
—		——
v2026-06-28_163011 16 0.090
—		0.032
—		0.003
—		——
v2026-06-28_163011 8 0.174
—		0.078
—		0.001
—		——
v2026-06-21_144829 19 0.034
—		0.011
—		0.005
—		——
v2026-06-14_150033
v2026-06-21_144829		 24 0.042
—		0.007
—		0.208
—		——
v2026-06-14_150033 22 0.201
—		0.243
—		0.352
—		——
v2026-06-14_150033 10 0.040
—		0.099
—		0.590
—		——
v2026-06-07_141547 26 0.046
—		0.090
—		0.030
—		——
v2026-06-07_141547 11 0.073
—		0.312
—		0.034
—		——
v2026-05-31_205201 23 0.085
—		0.080
—		0.115
—		——
v2026-05-31_205201 18 0.101
—		0.100
—		0.140
—		——
v2026-05-31_205201 6 0.171
—		0.080
—		0.052
—		——
v2026-05-24_132517 29 0.001
—		0.001
—		0.001
—		——
v2026-05-24_132517 14 0.001
—		0.001
—		0.001
—		——
v2026-05-14_055125 32 0.432
—		0.246
—		0.241
—		——
v2026-05-14_055125 14 0.369
—		0.311
—		0.291
—		——
v2026-04-29_121554 50 0.234
—		0.175
—		0.174
—		——
v2026-04-29_121554 57 0.254
—		0.178
—		0.259
—		——
v2026-04-29_121554 64 0.152
—		0.150
—		0.225
—		——
v2026-04-29_121554 54 0.176
—		0.186
—		0.287
—		——
v2026-04-29_121554 25 0.090
—		0.097
—		0.044
—		——
v2026-04-29_121554 5 0.001
—		————

Phase 3p · v2026-08-02_150333_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (20 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
v2026-07-26_150846_phase3p 8 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144840_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-19_144840_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145516_phase3p 22 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145516_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144712_phase3p 14 0.000
—		0.000
—		0.000
—		——
v2026-07-05_144712_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-06-28_150946_phase3p 16 0.002
—		0.001
—		0.000
—		——
v2026-06-28_150946_phase3p 8 0.003
—		0.001
—		0.000
—		——
v2026-06-21_145448_phase3p 19 0.000
—		0.001
—		0.001
—		——
v2026-06-14_150643_phase3p
v2026-06-21_145448_phase3p		 24 0.000
—		0.000
—		0.207
—		——
v2026-06-14_150643_phase3p 22 0.199
—		0.269
—		0.355
—		——
v2026-06-14_150643_phase3p 10 0.038
—		0.113
—		0.400
—		——
v2026-06-10_080820_phase3p
v2026-06-11_054514_phase3p		 8 0.013
—		0.000
—		0.002
—		——
v2026-06-07_142138_phase3p 11 0.116
—		0.090
—		0.033
—		——
v2026-05-31_205725_phase3p 23 0.069
—		0.098
—		0.143
—		——
v2026-05-31_205725_phase3p 18 0.082
—		0.117
—		0.169
—		——
v2026-05-31_205725_phase3p 6 0.123
—		0.054
—		0.107
—		——
v2026-05-26_110453_phase3p 18 0.000
—		0.000
—		0.000
—		——
v2026-05-26_110453_phase3p 3 0.000
—		————

Dry window — Dartmoor Nr Hexworthy — 5-hour

Phase 3p · v2026-08-02_150333_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (14 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
v2026-07-26_150846_phase3p 8 0.000
—		0.001
—		0.000
—		——
v2026-07-19_144840_phase3p 22 0.000
—		0.001
—		0.001
—		——
v2026-07-19_144840_phase3p 10 0.000
—		0.000
—		0.000
—		——
v2026-07-12_145516_phase3p 22 0.001
—		0.001
—		0.002
—		——
v2026-07-12_145516_phase3p 10 0.001
—		0.001
—		0.002
—		——
v2026-07-05_144712_phase3p 14 0.142
—		0.000
—		0.001
—		——
v2026-07-05_144712_phase3p 10 0.198
—		0.000
—		0.000
—		——
v2026-06-28_150946_phase3p 16 0.243
—		0.266
—		0.376
—		——
v2026-06-28_150946_phase3p 8 0.011
—		0.007
—		0.000
—		——
v2026-06-21_145448_phase3p 19 0.001
—		0.003
—		0.003
—		——
v2026-06-14_150643_phase3p
v2026-06-21_145448_phase3p		 24 0.000
—		0.001
—		0.312
—		——
v2026-06-14_150643_phase3p 22 0.338
—		0.395
—		0.532
—		——
v2026-06-14_150643_phase3p 10 0.039
—		0.082
—		0.257
—		——
v2026-06-10_080820_phase3p
v2026-06-11_054514_phase3p		 8 0.521
—		0.954
—		0.838
—		——

Dry window — Dartmoor Nr Hexworthy — 6-hour

Phase 3b · v2026-08-02_150158 Δ -0.006 vs prev train

53-feature LightGBM per-(station, window). Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.167 has_dry_window_mf (0.241) -30.9%
+48h 0.165 has_dry_window_ecmwf (0.234) -29.4%
+72h 0.175 has_dry_window_ecmwf (0.248) -29.5%
Verify history (28 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
v2026-07-26_150714 8 0.002
—		0.006
—		0.002
—		——
v2026-07-19_144727 22 0.006
—		0.010
—		0.021
—		——
v2026-07-19_144727 10 0.005
—		0.005
—		0.003
—		——
v2026-07-12_145358 22 0.020
—		0.018
—		0.053
—		——
v2026-07-12_145358 10 0.022
—		0.034
—		0.166
—		——
v2026-07-05_144600 14 0.132
—		0.018
—		0.012
—		——
v2026-07-05_144600 10 0.182
—		0.011
—		0.012
—		——
v2026-06-28_163129 16 0.202
—		0.204
—		0.314
—		——
v2026-06-28_163129 8 0.042
—		0.011
—		0.011
—		——
v2026-06-21_144944 19 0.128
—		0.060
—		0.046
—		——
v2026-06-14_150146
v2026-06-21_144944		 24 0.068
—		0.027
—		0.288
—		——
v2026-06-14_150146 22 0.241
—		0.329
—		0.456
—		——
v2026-06-14_150146 10 0.141
—		0.116
—		0.094
—		——
v2026-06-07_141656 26 0.244
—		0.279
—		0.278
—		——
v2026-06-07_141656 11 0.227
—		0.123
—		0.107
—		——
v2026-05-31_205304 23 0.073
—		0.134
—		0.125
—		——
v2026-05-31_205304 18 0.049
—		0.070
—		0.093
—		——
v2026-05-31_205304 6 0.023
—		0.033
—		0.029
—		——
v2026-05-24_132609 29 0.019
—		0.008
—		0.024
—		——
v2026-05-24_132609 14 0.013
—		0.011
—		0.074
—		——
v2026-05-14_055801 32 0.328
—		0.138
—		0.164
—		——
v2026-05-14_055801 14 0.200
—		0.039
—		0.144
—		——
v2026-05-03_120600 50 0.120
—		0.111
—		0.197
—		——
v2026-05-03_120600 51 0.252
—		0.181
—		0.231
—		——
v2026-05-03_120600 31 0.411
—		0.290
—		0.415
—		——
v2026-05-03_120600 17 0.536
—		0.394
—		0.574
—		——
v2026-04-29_121648 25 0.405
—		0.403
—		0.170
—		——
v2026-04-29_121648 5 0.070
—		————

Phase 3p · v2026-08-02_150333_phase3p Δ +0.000 vs prev train

Gaussian copula MC over Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences. Captures within-day wet/dry autocorrelation. Trained 2026-08-02. Metric: Test Brier.

Lead Blend Best single Δ vs best
+24h 0.000 — (0.000 val) —
+48h 0.000 — (0.000 val) —
+72h 0.000 — (0.000 val) —
Verify history (20 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
v2026-07-26_150846_phase3p 8 0.001
—		0.003
—		0.002
—		——
v2026-07-19_144840_phase3p 22 0.002
—		0.002
—		0.004
—		——
v2026-07-19_144840_phase3p 10 0.001
—		0.001
—		0.000
—		——
v2026-07-12_145516_phase3p 22 0.004
—		0.004
—		0.008
—		——
v2026-07-12_145516_phase3p 10 0.002
—		0.002
—		0.005
—		——
v2026-07-05_144712_phase3p 14 0.141
—		0.000
—		0.002
—		——
v2026-07-05_144712_phase3p 10 0.196
—		0.000
—		0.001
—		——
v2026-06-28_150946_phase3p 16 0.382
—		0.345
—		0.363
—		——
v2026-06-28_150946_phase3p 8 0.302
—		0.216
—		0.000
—		——
v2026-06-21_145448_phase3p 19 0.002
—		0.008
—		0.007
—		——
v2026-06-14_150643_phase3p
v2026-06-21_145448_phase3p		 24 0.001
—		0.002
—		0.320
—		——
v2026-06-14_150643_phase3p 22 0.393
—		0.372
—		0.543
—		——
v2026-06-14_150643_phase3p 10 0.159
—		0.124
—		0.171
—		——
v2026-06-10_080820_phase3p
v2026-06-11_054514_phase3p		 8 0.532
—		0.924
—		0.758
—		——
v2026-06-07_142138_phase3p 11 0.217
—		0.241
—		0.379
—		——
v2026-05-31_205725_phase3p 23 0.168
—		0.198
—		0.165
—		——
v2026-05-31_205725_phase3p 18 0.085
—		0.131
—		0.099
—		——
v2026-05-31_205725_phase3p 6 0.029
—		0.010
—		0.012
—		——
v2026-05-26_110453_phase3p 18 0.001
—		0.001
—		0.002
—		——
v2026-05-26_110453_phase3p 3 0.000
—		————