Check two very large lr_os_best values in priogrid 149451 (Darfur, 2025-10 and 2025-12) — possible summary/aggregate assignment

[Polichinel]

Summary

While reading forecast maps against validation truth, two lr_os_best values turned up in
priogrid 149451 that are 25–30× larger than anything else in the file. Both look like they
could be summary / aggregate assignments — a large unattributed fatality total placed on one
cell-month — rather than per-cell counts. Flagging for a check on the datafactory side; I have not
changed anything.

Cell: priogrid 149451 — row 208, col 411 → 13.75°N, 25.25°E, North Darfur, Sudan
(El Fasher area), c_id 161.

The two values

month_id	date	lr_os_best
550	2025-10	27,413
551	2025-11	2
552	2025-12	32,505

Why they look like summary events rather than per-cell counts

1. Scale break. The next largest lr_os_best cell-month anywhere in the validation file
   (months 505–552, africa_me_legacy, 13,110 cells) is 1,100. These two are 25× and 30×
   that.

   552  149451   32505
   550  149451   27413
   527  148725    1100     <- next largest, 25x smaller
   540  131459     937
   543  180432     792
   

2. The cell's own history is two orders of magnitude smaller. Months 527–549 in this same
   cell run 0–53, with a median of 2:

   527..549: 0 0 0 0 0 36 11 0 2 28 0 0 20 7 1 13 23 26 0 2 3 42 53
   550: 27413      551: 2      552: 32505
   

   A 2 sitting between 27,413 and 32,505 is the part that reads as an artefact of assignment
   rather than a three-month atrocity pattern.

3. sb and ns are both exactly 0 in this cell-month, so whatever produced the os figure
   did not co-occur with recorded state-based or non-state activity in the same cell.

What I am not claiming

The underlying event is almost certainly real — El Fasher fell in October 2025 and mass killings
were widely reported, so a very large os total for that place and period is expected. The
question is narrower and purely about the data pipeline:

• Is this the UCDP/source aggregate for a wider offensive being geocoded to a single grid cell
  because the source event has one coordinate?
• If so, is that the intended behaviour, or should such events carry a precision/aggregation flag
  that downstream consumers can see?
• Is the split across 550 and 552 with 2 in between an artefact of how a multi-month event is
  date-assigned?

Why it matters downstream

views-hydranet scores against these values. One cell carrying 32,505 in a field whose next
largest value is 1,100 dominates any magnitude metric for that month, so a model is graded largely
on whether it hit one cell. Knowing whether this is a true per-cell count or an aggregate changes
how we treat it — a real count is a modelling target, an aggregate is arguably something to flag
or down-weight in evaluation.

No action requested beyond a look; @Polichinel is checking from the datafactory side.

Found: 2026-09-08, reading reports/2026-09-06_ensemble_roster_dossier/results/v2_truth_validation.parquet
(views-hydranet), built from the datafactory africa_me_legacy queryset.