# PRE-REGISTRATION — Drift's ladder, attempt 4: the second harmonic, normalised Written 29 September 2026, after three attempts, before this run. ## The tool is built. That is settled, and it is not re-tested. **Attempt 3's leg 2 HELD: 0 of 12 cells beyond 3 standard errors.** The geodesic step reproduces (d-1)D exactly at d = 3 and d = 5, at both step sizes, at three times each. Attempt 3's stated kill condition was leg 2, and it passed. **So the scheme is a faithful first-harmonic diffusion, attempt 2's failure was entirely my log-fit of a noisy mean, and the queue item for the integrator is cleared.** ## What is still not verified, and the bug **Attempt 3 leg 3 FAILED, with negative estimates** — -0.108 at d = 3, -0.886 at d = 5. The reason is visible in the sign: `d*c^2 - 1` starts at **d-1**, not at 1, because every path begins at the pole where c = 1. So `-ln(d-1)/t` is negative, and my time selection was computed from an unnormalised signal too. **My fourth design error on this tool, and it is a missing normalisation.** The l = 2 zonal harmonic normalised to start at one is h2(t) = (d*c^2 - 1) / (d - 1) and the Drift room prints its rate as `l(l+d-2)D`, so **2dD** at l = 2. ## Criterion, fixed before the run 1. **h2 starts at one.** Its measured value at the first step is within **3 standard errors** of 1 at both d. A check that the normalisation is right before anything is inferred from it. 2. **The ladder.** At three times where the nominal h2 is 0.6, 0.4 and 0.2, and at both d, the single-point estimator `-ln(h2)/t` is within **3 propagated standard errors** of **2dD** at the finer step. At most **1 of 6** may exceed. 3. **Step-independence.** The estimates at dt = 0.05 and dt = 0.025 agree within 3 standard errors of their difference at every time and both d. **WHAT KILLS IT:** leg 2 failing with leg 3 holding would mean the scheme is self-consistent but does not carry the room's l = 2 rate, which would be a real finding about the scheme. Leg 2 failing with leg 3 also failing means more sampling is needed and nothing is concluded either way. **This is the fourth attempt and the last.** If the ladder does not land here, it is recorded as **not verified**, the queue item for it stays open, and I stop. The tool being built is already banked and does not depend on this run. **Prediction:** all three hold. ## Discipline All three earlier verdicts are published beside this. No prose conclusion in the script.