Objective 4-domain nest placement for WRF radar data assimilation A case study over Macaé / INEA radar, 21–22 January 2024 · draft for WRF User Meeting 2026

Novelty vs prior art. Ingredients-based forecasting (Doswell et al. 1996), composite parameters (Thompson et al. 2004, 2012), standardised climatological anomalies (Hart & Grumm 2001), EFI for convection (Tsonevsky et al. 2018), and WoFS-style relocatable domains (Skinner et al.) are all individually standard. The combination — composite z-scores of many indices across many forecasts and times, swept for a fixed-size box, with a climatology-referenced algorithmic go/no-go — is what, to our knowledge, is not yet published.

This page was re-issued after a full multi-agent audit of every generation step (radar conversion, DA cycling, MET verification, aggregation and reporting). The pre-audit state is preserved in git (tag published-v1-leandromachadocruz; pipeline tag pre-audit-2026-06). The central conclusion — radar 3D-Var materially improves the precipitation forecast inside radar coverage at convection-permitting resolution — survives all corrections. What changed:

• E1 — matched-pair double counting. Point-Stat ran two interpolation methods at once and the aggregator pooled both, doubling every pair/hit count (e.g. dawn d03 "2796 pairs" → honest 1402). Ratio scores barely move: POD at ≥2.5 mm/h is 0.04 → 0.26 (was reported 0.04 → 0.27), CSI 0.04 → 0.20 (was 0.21).
• E2 — Alerta Rio gauges shifted +1 h. The "HBV" label in the source files is legacy; timestamps are civil UTC−3 (Brazil abolished DST in 2019). All 33 Rio gauges were re-ingested; hourly scores change <0.5 pp.
• E3 — MODE centroid distances were inflated 9× on d03 (a hardcoded 9 km/grid conversion applied to the 1 km grid). Absolute values are now in true km (e.g. dawn-mean 23.5 → 19.4 km); the relative improvement (−17%, −45% at peak hours) is unchanged.
• E4 — innovation "OmA" now shows the final analysis fit (2nd outer loop) instead of the loop-1 linearised residual. The reflectivity impact range becomes +10%…+38.5% (was +15%…+45%); every cycle remains positive.
• E5 — the d04 "non-degradation" claim is withdrawn. The published tables already showed degradation outside radar coverage (hourly POD drops at all thresholds; 21-h MAE 18.7 → 23.6 mm; ΔBSS < 0): the text now states it. Also: only the morning d04 is the out-of-coverage Angra sibling; the afternoon d04 sits ~126 km from the radar, inside coverage.
• E6 — HiRA re-aggregated. The old code silently dropped no-event hours and averaged per-hour ROC AUCs; with the correct pooled aggregation the d03 BSS advantage disappears, but the radar DA shows a genuine discrimination gain (d03 ROC AUC 0.52 → 0.64 at ≥2.5 mm/h, w=15). The "+1.00σ / top 16%" climatological label relies on a Gaussian assumption and sits 0.0002 above the |z|≥1 cut — read it as "favourable", not as a precise percentile.
• E7 — placement provenance. The morning picker previously published a d04 at (18,77) that was not the executed nest; this page now shows the geometry actually run, d04 at (17,70). The published morning box was pinned to a prior run's position (not the sweep optimum) — disclosed on the picker page. The observed-reflectivity regrid also gained a true nearest-neighbour lookup (obs geolocation error was up to ~4.6 km on d02).
• E8 — afternoon window re-run with a fair control. The original afternoon control was a GFS 21/00Z cold start (zero hydrometeors), mixing spin-up with the DA signal; a fair-lineage control (same GFS 20/12Z pre-forecast as the radar branch, no DA) was re-run after the audit. Verdict: the afternoon window shows no DA benefit — light-rain POD rises via over-forecast while CSI and the 21-h accumulation degrade (d02 MAE 11.3 → 36.4 mm). The morning-gain / afternoon-loss contrast (information age of the assimilated volumes) is itself a result; this page's verification section reports the morning pair, which was always fair.

Methodology, evidence and before/after for all 104 audit findings: AUDITORIA_RADAR_DA_2026-06.md in the pipeline repository.