Abstract
<p>The What Works Clearinghouse (WWC) rates whether education programs have credible evidence of working, and those ratings feed the evidence tiers written into the Every Student Succeeds Act. Each rating rests on standardized computations: effect sizes, baseline equivalence, and corrections for clustering. I ask two narrow questions. Given only what the WWC releases publicly, how much of its evidence base can an outside researcher independently recompute, and does an open tool reproduce the WWC's own effect-size determination where it can? I answer with baselinr, an open R package that implements the WWC's formulas, applied to the official WWC study export of 13,300 findings. About two thirds of the evidence base is recomputable: 64.6 percent of findings report the group means, standard deviations, and sample sizes an effect size is built from, a share that varies by standards version (96 percent under the earliest, a low of 49 percent under version 4.1, 67 percent under the current version 5) rather than falling monotonically at the frontier. Where a finding is recomputable, baselinr reproduces the WWC's own effect size to within 0.005 for 83 percent of continuous findings (rank correlation 0.96), though agreement declines under the newer standards as reviews apply more design-based adjustments that the raw statistics do not carry. The effect-size verdict is stable to defensible standardization and small-sample choices, flipping for about 2 percent of findings. The clustering-and-significance determination, which turns on an assumed intraclass correlation, is out of scope. Because the public record carries no covariate-level baseline statistics, I validate the baseline-equivalence determination separately, as a check on the tool: it certifies a randomized baseline (Project STAR) as balanced and flags the socioeconomic imbalance of a Catholic-versus-public-school comparison, where family income and parental education fail equivalence. I release the tool and code with this paper.</p>