This console runs the manuscript's pipeline on its frozen public geography: 57 Malaria Atlas Project survey sites in Uganda. Panels marked Live recompute from raw site data in your browser (a JavaScript port of the deposited R code; how closely it agrees with the deposited run is stated in the footer). Panels marked Deposited display Monte-Carlo results from the archived R runs. The trial is simulated. The geography, the survey counts and the connectivity are real; the arms, the intervention and its effect are not. As in the paper's main tables, every figure here assumes no between-site variance beyond binomial sampling; the paper's Appendix D adds between-site and between-cluster heterogeneity (intracluster correlation 0.02 to 0.10), under which a binomial analysis is too narrow even without contamination and the design's precision gains shrink.
Sites are sized by number examined and colored by baseline prevalence. Edges are the top gravity weights Gij = massi·massj/d² (the contamination kernel). Building the pools scores 6,000 stratified-balanced allocations exactly as PROPOSED_estimand_and_calibration_MAP.R does: covariate balance + cross-arm mobility leakage L(a) = Σ Gij·1[ai≠aj]. SR keeps all; CCR keeps the bottom quartile of balance; CC keeps the top 40 by leakage alone (contamination control with no balance constraint); CRR keeps the top 40 by combined rank.
| design | leakage (live) | deposited | Δ | imbalance (live) | deposited | Δ |
|---|---|---|---|---|---|---|
| build the pools to populate | ||||||
All 331 deduplicated candidate partitions (the eight method-by-resolution combinations, of five clustering methods and K ∈ {8,12,16}, that survive singleton rejection) with their ten standardized metrics are embedded from PROPOSED_design_selection_MAP.csv. The composite is recomputed live in your browser as the deposited weighted sum of signed metric scores; the weights are embedded to four decimals, so it agrees with the deposited composite to within 0.0001. Drag weights to see how rankings move; the featured winner (Spectral, K=8) holds rank 1 under the paper's CRITIC–entropy fusion weights (the bootstrap stability leg is the same for every metric and cancels; Limitation 8). The composite is a selection statistic, not a test.
| # | method | K | composite | optimality /100 |
|---|
The full data-generating model is ported to JavaScript: latent log-odds ηi = (1−c·si)·η₀ + c·si·(gravity-weighted cross-arm neighbor mean), with saturating boundary exposure si = 1−exp(−Ei/κi); outcomes y ~ Binomial(N, expit η). True effect δ = ln(0.5) ≈ −0.693 (treatment halves infection odds). Pick a design pool (built live in Panel 1), an estimator, and contamination strength c, then watch bias, coverage and RMSE accumulate replicate by replicate with Monte-Carlo standard errors. CC is the contamination-only baseline: it ranks on leakage alone with no balance constraint, so compare its bias against CRR's to see what the balance criterion buys (and check Panel 1's table for what it costs in covariate imbalance).
| bias | MCSE | coverage | RMSE | n done | |
|---|---|---|---|---|---|
| run the simulation to populate | |||||
δ = −0.6931 (true effect). Bias is mean estimate minus δ; coverage = fraction of 95% CIs containing δ.
Monte-Carlo results from the archived R runs (1,000–1,500 replicates per cell), embedded verbatim. These are the manuscript's Tables 13 to 16 plus the allocation-unit exposure bridge.
The same estimand, two-criterion balance-plus-leakage allocation and three estimators as Panels 1 and 3 (not the ten-metric suite or the design-selection engine of Panel 2), exercised on structurally different systems: the 2017–18 Yemen cholera outbreak by governorate (waterborne disease, human-mobility spread, symmetric gravity kernel) and early-2020 COVID-19 across Louisiana's parishes, a directional stress test over real reported incidence (cumulative confirmed cases), not a model of COVID transmission. In both, geography/population/baseline incidence are real; assignment and effect are simulated (hypothetical per-unit survey N = 800). The COVID testbed uses the manuscript's directed production-attraction flux Ki→j = SiRj/d², so it can compare a direction-blind design (CRR_sym) against a direction-aware one (CRR_dir). Each testbed also carries the CC contamination-only baseline (top 40 by symmetric leakage alone, no balance constraint), from PROPOSED_cc_baseline_cholera.R / PROPOSED_cc_baseline_covid.R, with its deposited diagnostics and OC rows embedded verbatim.
| design | leakage (live) | deposited | imbalance (live) | deposited |
|---|
| bias | coverage | RMSE | n |
|---|
| design | estimator | bias | cover | typeI | power |
|---|
With 20 governorates (about 10 per arm) the naive model-based test keeps mild size inflation (Type-I about 0.08 to 0.11), as the manuscript reports. Only the spillover-adjusted estimator has low power: 0.15 to 0.35 for SR, CCR and CRR (the manuscript's Table 9) and 0.10 for the CC baseline.
| design | net flux imbalance (live) | deposited | total cross-arm flux (live) | deposited |
|---|
The key contrast: CRR_dir crushes net directional imbalance but its TOTAL cross-arm flux stays high: total flux is what drives bias.
| bias | coverage | RMSE | n |
|---|
| design | estimator | bias | cover | typeI | power |
|---|
Negative result: the direction-aware design (CRR_dir) does not beat the direction-blind one on bias: at c=0.35, CRR_dir naive bias 0.293 vs CRR_sym 0.272.
The design reduces cross-arm exposure and the measured residual bias. It does not deliver freedom from SUTVA violation. The design does not restore nominal coverage for the naive or the boundary-censored estimator. All figures assume no between-site variance (paper, Appendix D). The composite ranking is a selection statistic, not a test; no p-value is reported for it. The spillover-adjusted estimator needs the mobility kernel, and its good behavior does not survive a kernel the analysis does not assume. The Type-I columns come from model-based tests; randomization inference is recommended for the primary test and is not evaluated in the manuscript. In the Panel 5 testbeds the survey counts are a hypothetical N = 800 per unit, not real counts.