# PRE-REGISTRATION — Drift's ladder, attempt 2: each leg judged by the right kind of error Written 29 September 2026, after attempt 1 failed, before this run. ## ⛔ Attempt 1 — FAILS, and the fault is the one I had just named `PREREG_drift_ladder_2026-09-29.md` required the fitted l = 1 rate to match (d−1)D to within **10⁻⁶ relative**, on the grounds that the geodesic step makes it exact by construction. **It is exact in expectation. I measured it from 8,000 sample paths.** The sample mean of ⟨c⟩ carries noise of order 1/√8000 ≈ 10⁻², so the fitted rate cannot possibly land within 10⁻⁶. Measured: 0.098695 and 0.102360 at d = 3. **The criterion was impossible and the verdict FAILS stands.** **That pre-registration contains the sentence "the tolerance comes from the quantity's own decay, not from a number chosen to pass. Said in advance, because seven criteria tonight were numbers picked without asking that question" — and then sets an impossible tolerance three paragraphs above it.** I wrote the warning and broke it in the same document. It is recorded here in those terms. **What attempt 1 did show, and it is the substantive part:** the l = 2 rate is **0.67% from 2dD at d = 5** with the finer step, and the residual shrinks at both dimensions as the step halves — 0.3217 → 0.0921 at d = 3 and 0.2943 → 0.0067 at d = 5. ## Corrected design: the right kind of error for each leg **An exactness-by-construction claim is algebra and must be checked as algebra.** The scheme sets `φ = arccos(exp(−(d−1)D·dt))`, so the thing to verify is that the code computes that — to machine precision, with no sampling in it. **A fitted rate is a statistic and must be judged against its own standard error**, which the fit itself provides. ## Criterion, fixed before the run 1. **The construction is implemented correctly.** `cos(φ) − exp(−(d−1)D·dt)` is below **10⁻¹⁵** at every d and dt used. Pure algebra, no sampling. **Labelled as a code check, not as evidence about the scheme.** 2. **The first harmonic lands within its own error.** The fitted l = 1 rate is within **3 standard errors of the fit** of (d−1)D, at d = 3 and d = 5 and both step sizes. At most **1 of 4** may exceed 3 SE. 3. **The second harmonic lands on the room's ladder within its own error.** At the finer step, the fitted l = 2 rate is within **3 standard errors of the fit** of **2dD**, at both d. **This is the real test: l = 2 is not built into the scheme.** 4. **It converges.** The relative residual on l = 2 is smaller at the finer step than at the coarser, at both d. *(Held in attempt 1 and re-reported.)* **WHAT KILLS IT:** leg 3. If l = 2 misses its own error bar at the finer step while leg 4 also fails, the scheme is not a faithful diffusion and the tool is still not built. **Prediction:** legs 1, 2 and 4 hold. **Leg 3 I am genuinely unsure of at d = 3**, where attempt 1 was 9.2% out and the second harmonic decays three times faster than the first, leaving few usable decades. If it misses, that is a statement about how much sampling the l = 2 rate needs, not about the room. ## Discipline Attempt 1's FAILS is published beside this, with the self-contradiction named. No prose conclusion in the script.