Skip to main content

ACLED Violence Measurement

Product status: IMPLEMENTED / CURRENT CONTRACT-BACKED PRODUCT
Real current-artifact experiment status: NOT YET RECORDED AS CANONICAL
Producing repository: fcv-empirical-data

What it is​

A source-specific ACLED measurement path that preserves source events first, then relates them explicitly to shared geography and time before producing a sparse analytical measurement.

registered ACLED snapshot
→ source-native Silver events
→ event ↔ geography membership
→ event ↔ PeriodIndex membership
→ sparse Gold area × period × native event type

Source authority​

The authoritative input is a registered external ACLED export represented by SourceSnapshotRef, including source/release/snapshot identity and file hashes.

Historical recovered ACLED aggregates are parity and calibration evidence; they are not the authority for the rebuilt source rows.

Natural grain​

Silver: one supplied ACLED source row, with a stable event_row_id plus the native source event ID kept separately.

Gold:

geo_uid × period_id × native_event_type

The Gold table is intentionally sparse.

Geography​

Spatial assignment is produced through spatial-data-foundation, not embedded as an irreversible source rewrite.

Membership states preserve distinctions such as:

  • uniquely matched;
  • outside geography;
  • boundary / multiple candidates;
  • missing coordinates;
  • invalid coordinates.

The first Gold policy includes only uniquely matched events. Ambiguous events are not duplicated or tie-broken silently.

Time and coverage​

Period assignment uses the shared PeriodIndex and a declared PeriodScheme. The legacy-compatible T2_y2001 scheme is a parameter, not source semantics.

The current Gold coverage contract uses:

absent_row_semantics = unknown

Therefore a missing area-period-event-type row is not automatically zero violence.

Observed minimum/maximum event dates are descriptive support; they do not by themselves prove complete source reporting over the interval.

Important measured facts​

The normalized Silver envelope retains at least:

  • event date;
  • coordinates;
  • fatalities;
  • native event and sub-event type;
  • geography precision;
  • time precision;
  • country identifiers where supplied;
  • source release and snapshot identity;
  • every original source column under a source__ prefix.

The Gold measurement includes:

  • event_count;
  • fatal_event_count;
  • fatalities;
  • known/missing fatality event counts;
  • record_present = true for represented aggregate rows.

Zero-fatality source events are preserved. No default GEO_PRECISION == 1 filter is applied upstream.

Durable contract artifacts​

The path persists or references:

  • SourceSnapshotRef;
  • Silver and Gold DatasetRef objects;
  • geography/period lineage inputs;
  • MeasurementContract;
  • CoverageContract;
  • RunManifest;
  • QA and parity artifacts;
  • content hashes and code revision where supplied.

Known limitations / do not infer​

Do not infer that:

  • an absent Gold row means zero fatalities;
  • ACLED is automatically the outcome for every experiment;
  • Violence against civilians is the only legitimate taxonomy selection;
  • geography ambiguity has been solved by convenience;
  • source precision is a scientific inclusion rule;
  • current ACLED is harmonized with UCDP;
  • synthetic CI proves real ACLED/GADM acceptance.

Experiment history​

The experiment harness has a current contracted projection path that can explicitly choose, for example:

native_event_type = Violence against civilians
value = fatalities
role = outcome / pre-outcome
timing = +1 / -1 period

That path is synthetic-accepted, but the harness documentation explicitly states that a real current-artifact contracted E1/E2 run has not yet been recorded as canonical.

Recovered real-data E1/E2 calibration exists on the historical/reconstructed lane and should remain labeled as such.

Technical references​