Model performance bridge: 11–15 August 2026

Executive summary

The missing interval between the two published analyses can be scored. The retained archive supplied 185 complete deterministic runs across five resorts, and the observation archive covers the full five-day window.

It was not another substantial weather event. The five verification gauges recorded only 1.0 mm combined: Mount Buller 0.6 mm, Falls Creek 0.2 mm and Mount Hotham 0.2 mm, with Perisher and Thredbo Top dry. This makes the period useful chiefly as a false-alarm rainfall check. ECMWF AIFS produced the smallest precipitation error at 6.0 mm MAE, followed by JMA at 7.1 mm and ECMWF IFS at 8.0 mm. Every model was wet-biased.

Phase and snow-level figures are retained for exploration but should not be used to declare a thermal winner. The observed 80% snow fraction represents only 0.8 mm of inferred snow, and the precipitation-weighted snow-level summary rests on that same very small total.

Deterministic precipitation

Ranked by mean absolute error against the observed totals. Because every model overforecast the near-dry period, MAE and positive bias are identical after rounding.

The comparison is not perfectly sample-matched. The only lead bucket shared by every available model was lead day 0. Restricting to that bucket preserves the same leading pair: AIFS 6.2 mm, IFS 6.4 mm, JMA and GFS 6.9 mm. Only one initialisation time was present for every model at every site, which is too sparse for an exactly matched ranking.

Thermal and phase context

The ten-minute gauge feed recorded precipitation at three of the five verification stations. The pooled inferred snow fraction was 80%, but that is 0.8 mm of snow from a 1.0 mm total—not an event-sized phase sample. Ninety-five model runs were wet enough to define a phase bracket; 38 contained the observed fraction and 57 sat entirely above it. None sat entirely below it.

The inferred precipitation-weighted Victorian snow level was 1,636 m across 26 precipitating regression timestamps. AIFS had the smallest raw snow-level MAE at 84 m, followed by ICON at 90 m and UKMO at 105 m. These values are presented as context only: tiny rainfall increments decide the weighting, so the apparent separation between models is too fragile to promote into a rank.

Ensembles

Each ensemble system contributed only five runs. None of their 10th–90th percentile precipitation ranges contained the observed near-zero total, and every ensemble median was too wet. This is directionally consistent with the deterministic false alarms, but the samples are marked insufficient for ranking.

How this completes the record

The retained explorer now has three adjacent primary periods:

1. 9–10 August: a warm rain-on-snow event.
2. 11–15 August: this near-dry bridge, useful for false alarms.
3. 16–22 August: a whole-week check containing a later three-day event.

The 19–21 August event-only analysis is a nested lens inside the third period, not a fourth consecutive window. Keeping that distinction visible is more informative than reducing precipitation, phase and snow level to a single composite score with arbitrary weights.

Reproducibility and limitations

npm run verify:event -- --start 2026-08-11 --end 2026-08-16 --slug bridge-2026-08-11 --precip-source-contains 10-minute --exclude 95908 --exclude-precip 95901

The run used the default earliest-archive selection and snapshotted the observations. Thredbo Village was excluded because its observation cadence is too sparse; Baw Baw remained available to the thermal regression but was excluded from precipitation, matching the later weekly analysis. ARPEGE had no archived runs spanning the full five-day window. Relative percentage precipitation error is undefined at dry gauges, so those gauges remain in millimetre MAE and bias but are excluded from that secondary percentage field.