[Polichinel]

Scaffolds the research program for how UCDP aggregate events are allocated in space and time.
Follows the catalog in reports/investigation_summary_event_concentration/ (#534) and responds to
#533 / #484.

No code, configuration or data changed. This is a staging area; on promote its design
graduates to a proposed ADR.

What's here

Ten documents in reports/2026-11-04_summary_event_allocation_dossier/: the README spine, an
840-paper library survey with a 19-item acquisition list, the option space, the harness audit, a
gated roadmap, a deliberately-empty analysis plan, a glossary, an append-only experiment log, and a
study of the legacy views3 GedLoader.

The harness audit: ~85% already exists

Found, not built: an OCP strategy registry (datafactory_provenance.registry.Registry) so a new
allocation strategy is registered rather than patched into the builder · profiles as the
default-off flag mechanism · the ADR-050 byte-identical fixture gates · a synthetic known-answer
path for cheap readout · sixteen post-mortems, so negatives have a home.

The one real gap: nothing in this repo evaluates allocation quality — no locked baseline, no
metric protocol, no ground truth — because allocation was never treated as a choice. That is the
build, and it is constructible: 343,865 events are well-located and well-dated, carrying
68.6% of all mass, so a coarsen-and-recover test has a known answer and needs no new data.

⚠️ Three divergences — governance, not research

All routed to roadmap P0 rather than resolved here.

1. ADR-049 is accepted and dormant. Dated 2026-06-28, it decided spatial distribution "on by
default", named cell 148759 and event 463131, and set a golden test that cell 148759 should drop
from 273,076 to its proportional share. The profile uses passthrough:

	ADR-049 said	measured 2026-11-04
cell 148759, all history	should drop from 273,076	274,358 — it grew
cell 150919, all history	79,841	80,064

Nothing is red because no test asserts it. That is C-365's lesson recurring, and 03 §C4 makes
writing that test a pre-flight blocker — expected red until the decision lands.

2. ADR-049's worked example misdescribes the code. It states event 463131 "is spread across 12
months by temporal distribution (correct)". It produces exactly one row; the
annual → date_end_only override catches it. So the spatial half is decided but dormant and the
temporal half is undecided and was believed decided.

3. production_parity does not reproduce the legacy country-month path. The legacy loader reads
different tables for its two aggregations — self.pg_ged (filtered) for priogrid-month,
self.ged (unfiltered) for country-month — so where_prec 4/6 deaths were present at country
level. Our grid_to_country_month() takes the grid as input, after the viewpoint filter, so
they are in neither: ~48,124 events, ~864,869 fatalities, 19.16% of all mass. Structural
rather than configurable, and an argument for making ADR-049 live, since distributed events would
reach country-month on their own. (Approximate — deduplicated by event id rather than replaying
full survivorship; the order is right, the number is provisional.)

A correction the legacy notebook forced

I had filed ceil_split's 0.91% inflation as an ADR-040 Invariant 1 violation to fix. It is
faithful parity: the legacy code applies np.ceil to best, high and low alike. Fixing it
is a parity break, not a bug fix, and belongs behind the same profile switch as everything else.
The legacy loader also truncates summary events spreading past the ingestion month — a second
inherited leak, pulling the other way.

Two more legacy findings that strengthen the design:

• An is_summary flag already existed and was dropped at the aggregation boundary — the same
  seam where ours dies. So Summary/aggregate events are concentrated by measurement, not by reality — and one of them is ~63% of our primary evaluation score #533's ask for a precision flag is a restoration of a legacy concept,
  not a new consumer demand.
• Every legacy column is named _nokgi — "no Known Geographical Imprecision". The old schema
  declared the exclusion in its column names and implied a _kgi counterpart nobody built. Our
  lr_sb_best says nothing. The legacy schema was more self-documenting on exactly this point.

The parity freeze is not the obstacle it looked like

ADR-049 §6 already established that parity is "achievable via configuration". So
production_parity stays frozen and byte-identical while a separate recommended profile carries the
changes, and the question becomes which profile the pipeline invokes — one line, reversible,
auditable — rather than whether to mutate a frozen contract.

Library: survey and acquisition list

840 papers, 3,807 claims. Strong on evaluation under domination by a few observations (31
scoring-rule papers incl. Lerch2017 on the forecaster's dilemma, Gneiting2011's threshold
decomposition, Bolin2023 SCRPS), point processes (22, incl. Brehmer2023 on scoring a point
process's intensity measure), and UCDP measurement error (Vesco2026: underreporting is non-linear
in event size, so the large aggregates are the most accurately-counted values we hold — which is
why option O4, dropping, should be rejected).

Verified absent, checked against all 840 titles rather than inferred from a failed search:
change-of-support/MAUP (0), temporal disaggregation/benchmarking (0), areal interpolation
(~0), coarsened-at-random/partial identification (3, all Tobit — a different problem).

09_library_acquisitions.md lists 19 items in 3 tiers with verified citations; three of the seven
Tier 1 items are open access. Notable: Tobler (1979) on pycnophylactic (mass-preserving)
interpolation is ADR-040 Invariant 1 as an existing method, and the R packages pycno and
TSdisaggregation can serve as oracles for the evaluation harness.

Caveats carried in the documents

1. Figures rest on a local viewpoint build dated 2026-06-01, data through 2026-04; re-measure
   against the deployed build before citing as current.
2. The coarsen-and-recover design has two declared confounds: the proxy population, and — from
   Weidmann (2015) — that UCDP-GED locational accuracy degrades sharply below district scale,
   so our "ground truth" carries its own error at ~55 km cells and the oracle is not a true ceiling.
3. No causal claim about any model's performance.

Verification

Full suite: 2,433 passed · 16 skipped · 17 xfailed · 0 failed · 0 errors.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EfeuULbPTqrx4qtNki5dy9