[Polichinel]

Closes nothing. Documents the measurement behind #533 and #484; both stay open pending a policy
decision that is not this PR's to take.

What this is

views-hydranet reported that one cell-month held 62% of a scored horizon's truth mass, collapsing
that score's effective sample size to 2.58 cells out of 842,634 — deleting a single data point
moved the score by 83×. This PR catalogs the whole class that cell belongs to.

Measurement only. No code, configuration or data changed.

Headline figures

Events landing in a cell-month their own UCDP precision fields disown	21,766 (5.8%), carrying 21.95% of all fatality mass
The severe class — spatially and temporally unresolved	457 events (0.12%), carrying 11.22%
Top ten cell-months	431,802 fatalities = 12.4% of the dataset
One 0.5° cell (148759, 13.25N 39.25E, Tigray/Amhara)	270,688 = 7.79% of every fatality recorded 1989–2026

Two mechanisms, not one

The first pass at #533 found only the temporal one. The spatial one is nearly as large and was
missed:

• Temporal — date_end_only assigns an annual-source event to its date_end month "regardless
  of whether the event spans multiple months" (its own docstring). A year-long aggregate lands on
  one month. 3,315 events, 17.79% of mass.
• Spatial — passthrough returns an event unchanged whatever its where_prec, and
  production_parity excludes only where_prec 4 and 6. So 5 (country-level) and 7 (unclear) are
  pinned to whatever single cell their coordinates happen to name. 18,908 events, 15.37% of mass.

Only 181 events in the entire dataset are ever spread across months.

The figure with the largest operational consequence

Concentrated share of total fatality mass:

period	share
2010–14	5.5%
2015–19	8.6%
2020–24	48.8%

2020–2024 is the window training and evaluation splits draw on. Nearly half its fatality mass is in
the wrong place, the wrong month, or both.

Why the catalog is tiered rather than a top-N list

Two shapes need different remedies. Ethiopia is few enormous events (142 events, 7.96% of all
mass). Ukraine is 1,893 small ones. Nagorno-Karabakh is 159 events across 78 months of persistent
low-grade concentration. A magnitude threshold catches the first and misses the other two —
independent support for views-hydranet's decision not to winsorise.

Also found

• The largest instance had not been reported anywhere. Event 463137, 121,848 fatalities,
  Aug–Oct 2022, date_prec 5, where_prec 5, on one cell in October 2022 — 94.7% of all
  state-based fatalities recorded on earth that month. It is 8% larger than the event in Summary/aggregate events are concentrated by measurement, not by reality — and one of them is ~63% of our primary evaluation score #533,
  which could not see it because the scored window ends at month 504.
• A correction to Summary/aggregate events are concentrated by measurement, not by reality — and one of them is ~63% of our primary evaluation score #533. The anchor event is Ethiopia/Tigray, not Darfur — event 463131,
  the Tigray War's whole-year total (2021-01-01→2021-12-31, "Tigray, Afar, Amhara") on one cell,
  beside three cell-mates totalling 27. The Darfur inference came by analogy with Check two very large lr_os_best values in priogrid 149451 (Darfur, 2025-10 and 2025-12) — possible summary/aggregate assignment #484. Check two very large lr_os_best values in priogrid 149451 (Darfur, 2025-10 and 2025-12) — possible summary/aggregate assignment #484's cell
  is separately confirmed and is genuinely North Darfur.
• A non-conservation, recorded and not acted on. Mass over rows exceeds mass over events by
  31,672 (0.91%) — ceil_split rounds up per month, as its docstring documents. Totals are not
  conserved today, which matters for any remedy claiming to conserve them. even_split already
  exists, is tested, and does conserve.

Negative results

• date_prec == 5 with span_months == 1: zero events. Had any existed they would have been
  invisible to a span-based test.
• where_prec 4 and 6 are correctly excluded, as configured.
• Survivorship is working — 372,058 unique ids over 372,303 rows; the excess is month spreading.
• No UCDP data-quality fault. Every event carries honest precision metadata, and it is retained
  through the builder. The information needed to handle these correctly is already in the store.

What is deliberately not here

source_distribution_map={"annual": "date_end_only"} is the production-parity contract and sits
formally with Håvard; changing it would move every published fatality figure. This PR proposes no
policy and changes no behaviour. It records what already exists so the decision is cheap when it
is taken.

Files

• reports/investigation_summary_event_concentration/findings.md — the catalog
• scripts/investigate_summary_event_concentration.py — prints every figure in it and writes the
  two CSVs; ruff and mypy --strict clean
• .gitignore — the two CSVs are not committed: 6.7 MB of derived data regenerated by one
  command from a viewpoint build superseded every refresh. A committed snapshot would only be a
  stale one nobody could tell was stale.

Caveats carried in the document

1. Basis is a local viewpoint build dated 2026-06-01, data through 2026-04. The deployed build
   is newer; re-run the script before citing any figure as current.
2. Leverage is proxied by best, not measured on model scores. Sound for a count-unit CRPS, and
   views-hydranet reports the direct figure is the same order — but it is a proxy, and the effective
   sample sizes are computed on truth.
3. No causal claim about any model's performance. The instrument is degenerate on these slices;
   that is not the same as saying a particular reading was wrong.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EfeuULbPTqrx4qtNki5dy9