Correction to the Admissibility Criterion: tau* is Monotone in R, and the Screen Pulls Against Adjudication Twice

The admissibility criterion for rival-constant tests asserts that tau* is "not monotone in the expected direction: demanding R >= 5 is harder to satisfy than R >= 3 at fixed precision." That statement is withdrawn here as false. The criterion places an upper bound sigma/x = 2, 3, 4, 5) make large R easier to satisfy, not harder. The tabulated values are correct; only the sentence interpreting them is wrong. The coefficients are generated by tau* = A (R-1)/(R+1) with A approximately 0.611, reproducing all four to within one part in a thousand. The factor (R-1)/(R+1) is the half-separation between R and unity relative to their mean, and A^-1 is approximately 1.64, a one-sided critical value. The reconstruction is reported as indicated rather than proved: the nearest competing form, sqrt(R-1), fits about six times worse, which on four values quoted to three decimals is not decisive. The monotonicity finding does not depend on it, since every form consistent with the table is increasing in R. The error was interpretive. R was treated as an evidential bar the analyst elects to raise; it is instead a property of the row, the ratio of the rival constant's deviation to the claimed constant's. Rivals lying five times further apart are genuinely easier to distinguish than rivals twice as far apart. Correcting the reading strengthens the original conclusion rather than weakening it. R is smallest where the claimed and rival constants both sit close to the measured value, and those are exactly the rows a proximity screen admits. It follows that the rows a programme selects as its strongest are the rows demanding the most precise instruments. The screen works against adjudication twice over: by magnitude through the delta-scaling, and by geometry through R. Unaffected: the inequality itself, the uncertainty propagation it rests on, the tabulated tau* values, the delta-scaling consequence, and every application at a specific R, including the 1.5% frequency precision requirement derived for prediction P74a.