five rows in the 2025-11-19 block of flu_geo_exclusions.csv are dated 2024-11-19 (year typo) — linear, linearlog, climate_base, climate_geo_agged, climate_quantile_extrapolated receive no weights for that date #214

Severity (my assessment): high
Area: prod flu pipeline / ensemble weights (committed data)
Commit observed: current main (ccc2ee4); the file is untouched by ccc2ee4, so rows are identical at 5e35f53 and main
Accompanying patch: patches/exploration-tooling/fix-flu-weights-2025-11-19-year-typo.patch (candidate PR)

Summary

In the committed scripts/flu_geo_exclusions.csv, the block commented # Nov 19, data returns, switch back to seasonal forecasters (line 75) contains four rows dated 2025-11-19 followed by five rows dated 2024-11-19 (lines 81-85):

2025-11-19, all, mp, 0
2025-11-19, windowed_seasonal, all, 0.5
2025-11-19, windowed_seasonal_extra_sources, all, 3
2025-11-19, climate_linear, all, 0.001
2024-11-19, linear, all, 3
2024-11-19, linearlog, all, 0
2024-11-19, climate_base, all, 2
2024-11-19, climate_geo_agged, all, 0
2024-11-19, climate_quantile_extrapolated, all, 0

Everything about the context points to a copy-paste where the year wasn't updated:

• The neighboring complete blocks (Nov 26 at lines 65-73, Oct 15 at lines 89+) each contain exactly these five forecasters.
• The twin scripts/flu_nssp_geo_exclusions.csv has a complete 2025-11-19 block with the same shape.
• 2024-11-19 appears exactly 5 times in the file — all in this block — so there is no genuine 2024 block that these rows could belong to.

Mechanism

parse_prod_weights (R/utils.R:283, unchanged at main) filters to forecast_date == 2025-11-19, so the five typo'd rows are filtered out and those forecasters get no weight rows. In ensemble_climate_linear (R/forecasters/ensemble_linear_climate.R:59-98) their weight becomes NA, the renormalization mass = sum(weight) (computed without na.rm) becomes NA for every affected (geo, ahead, quantile) group, and the final sum(value, na.rm = TRUE) (line 98) zeroes the whole group — so the entire climate_linear ensemble output for 2025-11-19 is zeros. Nothing in the pipeline errors on the missing rows (a related, more general validation gap is covered in a separate issue I've drafted).

Maintainer-runnable repro

sed -n '75,85p' scripts/flu_geo_exclusions.csv
grep -c "2024-11-19" scripts/flu_geo_exclusions.csv   # → 5

In R:

w <- readr::read_csv("scripts/flu_geo_exclusions.csv", comment = "#", show_col_types = FALSE)
nrow(dplyr::filter(w, forecast_date == as.Date("2025-11-19")))   # → 4 (five forecasters missing)

Impact

For any run whose forecast date is 2025-11-19 — e.g., an evaluation-mode replay covering the 2025 season, which the evaluation window (scripts/flu_hosp_prod.R:74-86) includes — the climate_linear ensemble output is all zeros with no error anywhere, which also perturbs that week's score/report targets. I want to be careful about scope: I'm reporting what the committed file implies when the pipeline runs for that date; I don't have visibility into what was actually submitted at the time, and I'll leave that assessment to you.

Suggested fix

I've prepared a candidate patch (see accompanying diff) that changes the five rows' date from 2024-11-19 to 2025-11-19 and touches nothing else (weights unchanged). I verified it with both git apply --check and patch -p1 --dry-run against the current tree (the file is byte-identical between 5e35f53 and ccc2ee4, so it applies to both). Beyond fixing the rows themselves, I'd also suggest the fail-closed weight-coverage validation from the related issue so a missing/duplicate-date row can never silently zero an ensemble again.