# PRE-REGISTRATION — "The more he depends on you, the worse he checks you" **Batter's box item. Status before tonight: on deck. Its note: _"Unknown. The first-order effect is zero; a second-order effect is left."_ The owed next step: _"Restate it as a second-order effect and look for data sized to see one."_** Written 2026-09-29, **before any number exists**. --- ## The claim under test A checker whose own position depends on the party being checked checks them worse. The earlier finding was that the first-order effect vanishes, which was read as the claim failing. **This run asks whether that is the right reading, or whether the effect is real and simply quadratic.** ## The model, fixed before running Standard signal detection, which is what checking is. - The checked party is sound with probability 0.5. The checker sees a noisy signal **x**, drawn from N(0,1) if unsound and N(d', 1) if sound; **d' = 2** is the checker's raw discriminability, held fixed throughout. - The checker picks a threshold **t** and reports "sound" when x > t. - **Dependence** is a number **q ∈ [0, 1]**: the checker carries an extra private cost **q** for every report of "unsound", because a bad verdict costs them something they depend on. - The checker is **not corrupt**. They minimise their own expected cost honestly, given that private cost. The question is what honest optimisation under dependence does to the check. ## What is measured 1. **t\*(q)** — the threshold they choose. 2. **J(q) = sensitivity + specificity − 1** (Youden's J), the check's actual quality, which the house already uses. 3. **ΔJ(q) = J(0) − J(q)** — how much checking ability is lost. ## The pass criterion, fixed now - **The claim SURVIVES, restated as second-order**, if ΔJ(q) is **quadratic in q near zero** — the fitted exponent on log ΔJ against log q lies in **[1.7, 2.3]** — and ΔJ is strictly increasing in q. - **The claim FAILS** if ΔJ is flat (ΔJ < 0.001 at q = 0.5) or if the exponent is **below 1.3**, which would make it first-order after all and contradict the earlier finding. - Anything else is **INCONCLUSIVE**. ## What is predicted, and why By the envelope theorem, a small perturbation of the objective at an interior optimum costs nothing at first order, so **ΔJ should go as q²**. The prediction is therefore exponent ≈ 2, and the practical consequence, if it holds, is that **small dependence is nearly free and large dependence is disproportionately damaging** — the damage grows faster than the dependence does. ## Method Analytic where possible: t\*(q) from the first-order condition on the Gaussian mixture, J from the normal CDF. A numerical sweep of q ∈ {0.01 … 1.0} confirms it. No simulation noise, so no seeds and no intervals are needed; the curve is exact to the precision of the normal CDF. ## Deviations Any change to the above is written here, with its reason, **before** the run it affects.