Waterflood. Which change earned the barrels.

Wavepin watches a pipeline in seconds. Sweep reads a field over months. No shared physics; shares the tournament, ledger and honest-status pattern.

The causal engine (telecom harness, separate codebase): difference-in-differences, synthetic control with placebos, panel fixed and random effects, structural breaks, bootstrap, the finding contract (supported, projected, contested) and replay fingerprints; its compartment solver for display. pipeintel: ledger and tournament pattern. About 20–25 %. To fix before any run: synthetic-control scaling at barrel scale, and missing values read as zero. New: the capacitance-resistance family with producer coupling, constrained fit and per-pair identifiability, rate reconciliation, staggered and synthetic difference-in-differences, few-cluster bootstrap, a synthetic field generator with planted effects, full-physics truth runs on an open reservoir deck. The playbook engine: injection-change proposals (harness, not moat).

Three to four weeks, one person. Week 0: licence check on a released North Sea field's daily data. Week 1: capacitance-resistance fit with constraints; gain ranks, held-out rate error and identifiability per pair on that field. Weeks 2–3: a synthetic field (6 injectors, 12 producers) with worst-wells-first selection, staggered campaigns, interference and allocation error; the untouched open libraries beside naive before-after, decline refit and the fixed causal engine; one bias and coverage table.

Truth: a synthetic field with multi-tank physics, aquifer support, heterogeneous decline, producer interference, worst-wells-first selection, staggered campaigns, allocation error and planted uplift; full-physics runs of an open reservoir simulator with planted channels and events. Estimator counterfactual: a single-tank model without aquifer, a different model class. Against: naive before-after, decline-curve refit, 2 x 2 difference-in-differences, the untouched open capacitance-resistance library, the untouched open staggered difference-in-differences library, synthetic control with placebos, synthetic difference-in-differences, Bayesian structural time series. Scored: effect bias and error, attributable-barrels error, 90 % coverage, placebo false-positive rate, gain rank correlation against simulated flux, share of pairs flagged unidentifiable, held-out rate error. Real data: placebo runs on regulator well panels in untreated periods.

Loses to the untouched open libraries out of distribution: Sweep is dropped.