# PRE-REGISTRATION — over-trust, attempt 2: every leg measured, and the regime named Written 29 September 2026, after attempt 1 failed, before this run. ## Attempt 1 — FAILS, three separate faults, all mine `PREREG_drifted_overtrust_2026-09-29.md`. 1. **Leg 1 failed at d = 5**: fitted rate −0.186668 against −0.200000, z = 58.67. My step — isotropic Gaussian noise then renormalise onto the sphere — has an effective rate that departs from (d−1)D as d grows. d = 3 was fine at z = 1.93. **My integrator, not the room.** 2. **Leg 2 compared two different quantities.** I measured scatter about the *window's own mean* and held it against `1 − e^(−2λτ)`, which is not that. And I included **τλ = 1.0**, which is not a short window at all, when the claim is explicitly about an instrument grading itself over a *short* one. That is the same fault as the veil in entry 31: a claim tested outside its own regime. 3. **⛔ Leg 3 "passed" and it was vacuous.** It compared `σ²_true/σ²_self` against `1/σ²_self` — two closed forms, no measurement, differing only by a truncation of e⁻¹⁶. **I wrote a leg that could not fail.** That is worse than a badly chosen threshold, because a threshold can at least be wrong; this manufactured a pass. It is recorded here so it is not forgotten. ## Corrected design **Everything is measured from the simulation. No leg compares a closed form to a closed form.** - `σ²_true(t)` = mean over paths of `1 − (u(t)·L)²`, the squared perpendicular error from truth. - `σ²_self(t, τ)` = mean over paths of the scatter of `u` about its own mean inside the window [t, t+τ]. - over-trust factor = `σ²_true(t) / σ²_self(t, τ)`, both measured. **The regime is named in advance:** the claim concerns a **short** self-assessment window, τλ ≪ 1. Windows comparable to the decorrelation time necessarily see the drift, which is not a defect but the point — attempt 1's τλ = 1.0 was outside the claim, not evidence against it. This run uses **τλ ∈ {0.01, 0.02, 0.05}** and reports τλ = 0.5 and 1.0 separately as the contrast. **d = 3 only**, with the integrator verified by halving the step. Attempt 1 showed my sphere step is not trustworthy at larger d, and fixing that is a different job from this claim. ## Criterion, fixed before the run 1. **The integrator is trustworthy.** Halving dt changes the fitted decay rate by less than **1%**, and the fitted rate agrees with (d−1)D within **3 standard errors** at d = 3. Reported before leg 2. 2. **Self-scatter is blind to age, in the short-window regime.** For each τλ in {0.01, 0.02, 0.05}, `σ²_self` at t = 1/λ and at t = 8/λ agree within **3 standard errors of their difference**. Zero of three may disagree. **This is the leg that can kill the claim**: if self-scatter grows with age, 1/σ² catches the drift and there is no over-trust. 3. **Error from truth does grow.** `σ²_true(8/λ)` exceeds `σ²_true(1/λ)` by more than 3 standard errors of the difference, so the two quantities in leg 2 and leg 3 are genuinely different. 4. **The over-trust factor grows with age.** For each short τλ, the measured factor at t = 8/λ exceeds the factor at t = 1/λ by at least **1.5×**. **Reported, not scored:** the same factors at τλ = 0.5 and 1.0, where the window is no longer short. **WHAT KILLS IT:** leg 2 or leg 3. Leg 2 failing means the instrument sees its own drift. Leg 3 failing means there is nothing to over-trust about. **Prediction:** all four hold in the short-window regime, and the contrast cases at τλ = 0.5 and 1.0 show the factor collapsing toward 1. ## ⛔ What will not be established That any real instrument or mind grades itself over a short window. **The claim is conditional** and the write-up says so. ## Discipline Attempt 1's FAILS is published beside this, with all three faults named. No prose conclusion in the script, and **no leg that compares one closed form to another.**