When the Counterfactual Is Adjudicated: A Grid-Curtailment Benchmark Separating Structure from One-Sided Estimation

Methods that recover an outcome suppressed by an intervening policy are often evaluated against a simulator written by the evaluator, and their ranking changes with the simulator. On one power grid, the suppressed outcome is adjudicated instead. Whenever the Brazilian system operator restricts a wind or solar plant, it publishes a half-hourly estimate of what the plant would have generated. The estimate is computed from wind-speed or irradiance measurements at the plant, under the procedure used to settle compensation payments. It is not a second meter reading, but the process that produces it predates this work and uses those measurements, which the benchmark withholds from every method. This paper builds a benchmark from the published record. The benchmark covers 18 wind clusters over 40,896 hours and 22 solar clusters over 18,984 hours, and it enforces five invariants as build-time assertions. Its evaluation protocol declares what each method may read, splits the record by date, and scores recovery on the hours the operator restricts separately from false corrections on the hours it does not. At a configuration fixed before any test score is seen, a structured latent decomposition models the censoring in its likelihood and floors its predictions at realised output. On restricted hours it beats a least-squares ratio estimator on solar but loses on wind. A ratio estimator that instead fits the 0.75 quantile of realised output, with calendar inputs and no latent structure, beats the decomposition on censored hours on all 15 seeds of both technologies (on seeds 1 to 5, 0.1237 against 0.1327 on solar and 0.1388 against 0.1713 on wind). On solar it also cuts the error on unrestricted hours from 0.2483 to 0.0198. With or without calendar terms in its inputs, a least-squares fit loses to the decomposition on every solar seed, and a 0.75-quantile fit with the same inputs beats it on at least 14 of 15. For those inputs, the change of loss alone reverses the ordering. The quantile grid is fixed in advance and reported in full, and this arm is its best cell on solar. A validation block inside the training window picks cells that still beat the decomposition on 14 of 15 solar seeds and all 15 wind seeds. Over all clusters, the solar advantage of the decomposition over least squares therefore does not show a benefit from its latent structure. On the solar clusters live throughout the record, the decomposition gains more and keeps a lead of about 0.02 over the quantile fit on restricted hours, so one-sided estimation reproduces only part of that gain. Scoring the unrestricted hours separately is what exposes the false corrections of the decomposition.