# PRE-REGISTRATION — does an instrument that grades itself over-trust itself, and by how much? Written 29 September 2026, before the run. **Drift + Trust**, rank 8, sharing L and σ. ## The two rooms **Drift** (Debye's rotational diffusion, 1929): `⟨cos θ⟩` decays as `e^(−(d−1)Dt)`, and its own text says *"turning it back does not undo the drift: only checking it against something outside restores it."* **Trust** (Gauss's inverse-variance weighting): `w ∝ 1/σ²`. ## The problem neither room states **Gauss's rule assumes unbiasedness.** A drifted instrument is not noisy — it is **precise about the wrong thing**. Its axis has moved off true, but its readings still cluster tightly about its own axis. So an instrument that estimates σ from **its own recent readings** measures only the drift accumulated inside that window, not the drift accumulated since it was last checked against something outside. It then claims a weight its accuracy does not earn. ## The quantities With λ = (d−1)D and alignment `c(t) = e^(−λt)`, the variance of a reading about **truth** is σ²_true(t) = 1 − e^(−2λt) while the scatter an instrument sees in **its own** readings over a window τ is σ²_self(τ) = 1 − e^(−2λτ) independent of t **Over-trust factor** = w_self / w_true = σ²_true(t) / σ²_self(τ), which for large t tends to `1 / (1 − e^(−2λτ))`. **The prediction that follows and is not in either room: a LONGER self-assessment window trusts you LESS, and is right to.** A short window is confident because it has not lived long enough to see its own drift. ## Criterion, fixed before the run 1. **The alignment decays at the room's own rate.** A simulated random walk on the sphere gives a fitted decay rate for ⟨cos θ⟩ matching `(d−1)D` to within **3 standard errors** of the fit, at d = 3 and d = 5. 2. **Self-scatter depends on the window, not on the age.** σ²_self measured over a window τ agrees with `1 − e^(−2λτ)` within 3 standard errors, and measurements taken at t = 1/λ and t = 8/λ agree with each other within 3 standard errors of their difference. **This is the leg that makes the claim non-trivial: if self-scatter grows with age, the instrument catches its own drift and there is no over-trust.** 3. **The over-trust factor is σ²_true(t)/σ²_self(τ)** and approaches `1/(1 − e^(−2λτ))` at large t, within **2%** at t = 8/λ, for τλ in {0.01, 0.05, 0.2, 1.0}. Simulation legs are judged against **their own standard errors**; leg 3's 2% is the asymptote's own truncation, `e^(−16) = 1.1×10⁻⁷`, so 2% is loose. Stated in advance, because six criteria tonight were numbers chosen without asking what precision the quantity has. **WHAT KILLS IT:** leg 2 failing. Then drift shows up in an instrument's own scatter, 1/σ² sees it, and there is no over-trust to report. **Prediction:** all three hold. ## ⛔ What will not be established That any real instrument or mind grades itself over a short window. **The claim is conditional:** IF reliability is judged from recent self-agreement, THEN the over-trust factor is the stated one. Nothing here measures how anyone actually grades themselves. ## Discipline No prose conclusion in the script. Tables printed by the run. Hashes read from the file.