AWARE-Geo: Assumption-Aware Reliability Evaluation for Geographic Causal Measurement
This repository is an evidence-controlled research benchmark for deciding when a geo-causal estimator should be selected, blended, or rejected using only information available before final treatment outcomes are inspected. The primary domain is randomized geographic incrementality; observational staggered-panel evidence is separate. AWARE-Geo is a controller, not a new universal treatment-effect estimator.
The corrected frozen protocol AWARE-GEO-e52a8660a97ef174 completed 500 final
worlds. SIMULATED: H1 clean noninferiority passed (RMSE ratio 1.00 versus
fixed DiD); H2 passed with a small paired normalized-regret difference of
-0.00210 (bootstrap 95% CI -0.00347 to -0.00111). Frozen Type-I, coverage,
clean-abstention, and truth-ledger gates passed. AWARE chose DiD in 483/500
worlds, so the evidence supports a narrow promotion of the gated decision
procedure, not broad learned estimator-selection superiority.
Protocol 001 is preserved as invalid because a post-run adversarial review
found truth-derived decision materiality. It cannot support claims. Protocol
002's GeoLift result remains historically gated; the separate post-audit
GEOLIFT-COMPAT-001 extension qualified unmodified GeoLift 2.7.5 against an
exact earlier unmodified augsynth commit, without changing the frozen result.
- Randomized geo experiments: the primary known-truth simulation track, including clean, null, single-stress, and compound-OOD worlds.
- Observational staggered panels: a separately governed secondary track for heterogeneous timing, dynamic effects, anticipation, and parallel-trends failures. It is not mixed into the randomized leaderboard.
- Industry tools: executable Meridian GeoX design/analysis qualification and an exact-commit GeoLift compatibility overlay. Tool-native design metrics are not treated as interchangeable or as an accuracy leaderboard.
- Econometric baselines: DiD, synthetic-control-family, staggered-DiD, and structural time-series implementations are qualified or transparently gated.
The completed exploratory extensions intentionally preserve results that limit the headline finding. S04 structural-break and S13 compound-OOD coverage was 0.567 and 0.467. Predicted AWARE risk weakly ranked realized failure (Spearman 0.024; catastrophic-loss AUROC 0.362), and fixed DiD had slightly lower CVaR95 decision regret than AWARE. Longer seasonal histories improved simulated coverage but did not restore nominal coverage. Separate MDE/power and formal assignment-versus-dose compliance studies provide operational evidence without retroactively becoming Protocol 002 confirmation.
The original scientific runtime was WSL2 Ubuntu with one physical NVIDIA GPU. GPU claims are limited to the measured JAX/GeoX/XGBoost workloads; this is not multi-GPU or distributed evidence. Core verification is CPU-safe:
uv sync --locked --extra dev
uv run python scripts/verify_evidence.py --read-only
uv run python scripts/verify_post_audit_suite.pyThe canonical results already exist. python -m geocausal.cli pipeline --resume reports the legal resume state and refuses to rerun completed
protected evaluation. Hosted CI verifies software, package, claims, manifests,
and immutable evidence contracts; it does not rerun protected worlds, heavy
Monte Carlo, CUDA, or external R estimator stacks.
Start with reports/FINAL_INDUSTRY_ANALYTICS_SUMMARY.md, then see
EXECUTION_PLAN.md, docs/ESTIMAND_CONTRACT.md,
docs/CLAIM_BOUNDARIES.md, and reports/REPRODUCTION_GUIDE.md. Public
visibility does not make third-party data or dependencies part of this
project's MIT grant; their original licenses and source manifests continue to
govern them.