Skip to main content

Experiment Surface Catalog

Document status: CURRENT SCIENTIFIC EXPERIMENT CATALOG

This page answers:

Which substantive FCV experiment surfaces can the current system support, and what still blocks them?

Calibration benchmarks remain separate. A reproduced report statistic or published-study pattern is an instrument test, not a new FCV hypothesis.

Status vocabulary

StatusMeaning
REAL CURRENT-ARTIFACT CHARACTERIZATION COMPLETEDCurrent hash-backed empirical chain has passed the declared gates and completed the bounded R0–R5 characterization wave.
REAL CALIBRATION COMPLETED — RECOVERED LANEHistorical/reconstructed experiment-calibration evidence exists.
PARTIAL — SCIENTIFIC USE NEEDEDEmpirical stack exists but exposure, roles, timing, uncertainty, or estimator use remains.
PARTIAL — MISSING EMPIRICAL MEASUREMENTSource facts exist but a required reusable measurement/linkage is absent.
BLOCKEDA scientific, annotation, source, or shared-infrastructure prerequisite is absent.

At a glance

Scientific surfaceEmpirical stateHarness/design stateCurrent status
China GeoGCDF commitments → ACLED violencecurrent governed treatment + certified outcome Gold existE0–E6 + R0–R5 characterization completedREAL CURRENT-ARTIFACT CHARACTERIZATION COMPLETED
World Bank projects → ACLED violenceWB source-native Silver exists; no current WB spatial/period measurementACLED side existsPARTIAL — MISSING EMPIRICAL MEASUREMENT
Pooled China + World Bank → ACLEDindependent sources; no canonical pooled productunion/multi-arm semantics undeclaredBLOCKED
Jobs-related investment → ACLEDannotation review infrastructure existsvalidated annotation/use rule incompleteBLOCKED
Afrobarometer respondent/EA exposurereusable substrate onlysource-specific scientific path absentBLOCKED
DHS household/cluster exposureHR measurement arm externally commissioned; GPS/GC substrate existscross-grain exposure + displacement/survey design remainPARTIAL — SCIENTIFIC USE NEEDED
Recovered WB → ACLED E2recovered area-period surfaces existreal historical calibration executedREAL CALIBRATION COMPLETED — RECOVERED LANE

Surface 1 — GeoGCDF commitments → ACLED violence

Status: REAL CURRENT-ARTIFACT CHARACTERIZATION COMPLETED

Empirical authority

The current reference uses:

  • GADM 4.1 ADM2 geography;
  • AidData GeoGCDF v3.0.1 commitment-area-period Gold;
  • coverage-certified ACLED area-period-native-event Gold.

GeoGCDF explicitly excludes 22 unresolved project geometries under resolution_policy = exclude_unresolved; no commitment-time projects were unresolved. ACLED structural zeros are licensed only within an explicit certified country/period surface. The experiment lattice is restricted to the 47 countries covered by the treatment MeasurementContract rather than extending treatment-zero semantics to the extra GADM country.

Reference design

periods: 2003–2004 ... 2013–2014
PRIMARY treatment: project_count > 0
STRESS treatment: positive_reported_amount_project_count > 0
outcome: ACLED Violence against civilians fatalities at t+1
pre-outcome: same VAC fatalities at t-1
estimator: OLS + pre-outcome + period FE + country FE
uncertainty: ADM2-clustered covariance

Real gate result

PRIMARY:

N = 38,520
units = 6,420
countries = 47
treated = 7,667
controls = 30,853
mixed-support periods = 6 / 6

All E0–E6 gates were GREEN. Outcome zero share was 0.9222; pretreatment |SMD| was 0.0040; prior-outcome placebo was 0.0047 outcome SD; the 0.20-SD injected positive control recovered 30 / 30.

The positive-reported-amount STRESS treatment also passed all gates.

Canonical estimate

effect = +0.550 fatalities
SE = 0.458
z ≈ 1.20
≈ +0.0176 outcome SD

This is not strong substantive evidence.

Closed R0–R5 characterization

The central question is no longer whether the surface can run or whether its small-effect credibility ladder still needs to be built. That ladder is now executed.

Detector resolution

~0.02 SD → weak detector region
~0.05 SD → usually observable
~0.10 SD → essentially certain under frozen synthetic worlds

Uncertainty

ADM2-cluster, finite-country-cluster, and wild-country-bootstrap procedures were characterized on identical known-truth worlds. They differ in null size/power without changing the point estimate; no procedure was selected because it produced a favorable observed-data SE.

Influence

No one country or high-influence ADM2 explains the canonical coefficient. Period dependence is materially stronger; omitting 2011–12 flips the fatalities sign.

Falsification

Deeper pre-treatment and future-treatment placebos remain small. The canonical coefficient is not especially extreme under a structured within-country complete-treatment-history permutation null (p ≈ 0.176, calibration interpretation only).

Sparse outcomes

The canonical remains fatalities OLS. Alternative predeclared representations are directionally positive and clearer on incidence/count margins:

OLS event count +0.0549 SD
LPM any VAC +2.04 percentage points
PPML event count IRR 1.348

These do not replace the canonical model.

Scientific interpretation for this surface

The instrument is healthier than the substantive fatalities evidence is strong. The actual canonical estimate lies in the same approximately 0.02 SD range that the detector itself says is difficult to distinguish reliably. Timing falsification is reassuring; temporal heterogeneity and the structured-null result constrain stronger interpretation.

The next useful work for FCV should therefore add a different empirical/calibration dimension rather than continue proliferating E2 models.

A supported-environment rerun of the exact numerical packet remains archival housekeeping because the original R0–R5 executions emitted a SciPy/NumPy compatibility warning.

Commitment remains a commitment-time measure rather than implementation onset. Project counts do not imply local spending. The completed characterization is instrument evidence, not causal validation.

Surface 2 — World Bank projects → ACLED violence

Status: PARTIAL — MISSING EMPIRICAL MEASUREMENT

World Bank Projects API Silver provides source-native project facts, dates, amounts, provenance, and source IDs.

The missing forward-looking piece is a trustworthy project-location → shared geography → shared period measurement comparable to the GeoGCDF path. Recovered WBad/WBkg evidence remains historical calibration rather than validation of a rebuilt current WB path.

This path should be advanced when it yields a genuinely independent donor comparison, not simply because another source can be processed.

Surface 3 — Pooled China + World Bank → ACLED

Status: BLOCKED

A pooled experiment would need independent contracted measurements, explicit overlap/reconciliation policy, preserved source-specific coverage, and downstream union/multi-arm semantics. Do not pool sources merely to increase apparent sample size.

Surface 4 — Jobs-related investment → ACLED

Status: BLOCKED

Annotation infrastructure can support versioned labels when backed by explicit evidence and review provenance. Geography, timing, ambiguity handling, comparison groups, and treatment derivation still belong in the experiment.

Surface 5 — Afrobarometer respondent / EA exposure

Status: BLOCKED

The generic survey substrate can represent respondent/EA grain, weights, temporal semantics, and uncertain geography, but current source-native Afrobarometer ingestion, semantic mappings, geography linkage, and named experiment use remain absent.

Surface 6 — DHS household / cluster exposure

Status: PARTIAL — SCIENTIFIC USE NEEDED

The DHS HR measurement arm has reproduced 8 / 8 official-report checks across Nigeria, Uganda, and Zambia. The historical Briggs measurement benchmark also independently recovered regional wealth-location quantities essentially to rounding precision.

The remaining frontier is downstream scientific use:

  1. choose a named outcome/control/subgroup role;
  2. bind households to cluster geography explicitly;
  3. name the investment/exposure measurement;
  4. declare a displacement-aware exposure uncertainty rule;
  5. declare survey/exposure timing;
  6. declare weight / PSU / strata use;
  7. gate linkage, support, displacement sensitivity, and semantic comparability;
  8. choose a survey-compatible estimator only after those gates.

Public DHS coordinates remain displaced measurements; reported-coordinate membership is not true-location authority.

Surface 7 — Recovered WB → ACLED E2

Status: REAL CALIBRATION COMPLETED — RECOVERED LANE

24,852 area-periods
4,142 GIDs
4 declared WBad/WBkg cells passed hard gates
old 0.20-SD injection recovery = 30/30 in every cell

This remains useful historical calibration but is no longer the strongest evidence that the current stack itself can execute a real experiment; the current GeoGCDF→ACLED gate and R0–R5 wave now provide that evidence.

Current experiment priority

1. archive the closed GeoGCDF→ACLED packet in a supported environment
2. advance DHS spatial exposure through explicit displacement/cross-grain design
3. build a current WB spatial-period measurement when an independent donor comparison adds value
4. keep pooled/multi-arm and extra estimator complexity downstream of explicit scientific demand

Briggs (2017) is now a closed calibration benchmark and belongs in the Calibration Benchmark Catalog rather than this substantive experiment catalog. Breckner–Sunde remains a future calibration benchmark contingent on truthful grid/month infrastructure.