# PRE-REGISTRATION — over-trust, attempt 3: the window, not the age Written 29 September 2026, after two failures, before this run. ## What the first two attempts established, and what they refuted **Attempt 1 — FAILS.** Integrator off at d = 5 (z = 58.67); leg 2 compared scatter-about-the-window-mean to a formula for a different quantity and included τλ = 1.0, which is not a short window; **leg 3 was vacuous**, comparing two closed forms with no measurement in it. **Attempt 2 — FAILS.** Leg 1 (integrator) failed: halving the step moved the fitted rate by **4.85%** and it sat **39 standard errors** from (d−1)D. Legs 2 and 4 failed. **⛔ And attempt 2's leg 4 refuted my own prediction.** I predicted the over-trust factor grows with age. It does not: 1.153, 1.001, 1.063 at τλ of 0.01, 0.02, 0.05. The reason is visible in the data — `σ²_true` is **0.630 at t = 1/λ and 0.666 at 8/λ**, already at its saturation value of 2/3 for d = 3. **I sampled two times that were both late.** That is the third time in this pipe that I chose sample points without asking where the variation lives. **What the numbers did show:** the factor is large and set by the window — 201× at τλ = 0.01, 39.7× at 0.05, 4.67× at 0.5, 2.71× at 1.0. ## The corrected claim — narrower, and it is what the data supports **How much an instrument over-trusts itself is set by how short its self-assessment window is, not by how long it has been drifting.** Once drift saturates, age carries no further information; the over-trust is fixed by the checking interval alone. ## ⛔ What this attempt will NOT claim, and why **Nothing about Drift's decay rate.** My sphere step — Gaussian noise then renormalise — moved 4.85% on halving and sits 39 SE from (d−1)D, so **it cannot be used to verify the room's law**, and no leg here depends on the rate being right. Verifying Drift's own exponent needs a faithful rotational-diffusion integrator and is a separate job, filed as separate work. Everything below concerns only quantities measured from whatever process this walk actually is. ## Criterion, fixed before the run 1. **Saturation.** `σ²_true` at t = 4/λ and t = 8/λ agree within **3 standard errors of their difference**, establishing that both are past the growth phase. 2. **Age-independence once saturated.** For each τλ in {0.01, 0.02, 0.05, 0.1}, the over-trust factor at t = 4/λ and t = 8/λ agree within **3 standard errors**. Zero of four may disagree. 3. **Monotone in the window.** Across τλ in {0.01, 0.02, 0.05, 0.1, 0.25, 0.5, 1.0} the factor at t = 8/λ **decreases at every step**, with no reversals. 4. **The size is worth stating.** The factor at τλ = 0.01 exceeds the factor at τλ = 1.0 by more than **20×**. **WHAT KILLS IT:** leg 2 failing would mean age matters after all and the corrected claim is wrong too. Leg 3 failing would mean the window does not set it either, and the whole reading is wrong. **Prediction:** all four hold. **Recorded plainly: I have been wrong twice in this pipe, and once my prediction was refuted by my own run. This criterion is written so a third wrong reading would show.** ## Discipline Both earlier FAILS are published beside this. No prose conclusion in the script, and no leg that compares one closed form to another.