# PRE-REGISTRATION — Drift's ladder, attempt 3: no log-fit at all Written 29 September 2026, after two failures, before this run. ## The two failures, and the bug they exposed **Attempt 1 — FAILS.** Demanded 10⁻⁶ relative agreement from a Monte Carlo estimate, in a file whose own text warned against choosing tolerances that way. Impossible criterion. **Attempt 2 — FAILS.** Judged against the fit's own standard error, and **all four l = 1 cells missed by 7 to 41 standard errors** — 0.100721 and 0.103375 against 0.100000, with a fit SE of 7.8×10⁻⁵. That is not noise. **It is bias, and halving the step made it worse, which is the tell.** **⛔ The bug is my estimator.** I fit `log⟨c⟩` against t down to ⟨c⟩ > 0.02. At that tail the sampling error of ⟨c⟩ is about 0.003, and since `E[log X] < log E[X]`, the tail points are pulled **down** and the fitted decay looks **faster**. A finer step supplies more tail points and so more bias. **The scheme was never the problem in attempt 2; the way I measured it was.** ## The corrected estimator, and why it needs no fit The geodesic step gives `⟨c(t)⟩ = exp(−λt)` **exactly in expectation**, with λ = (d−1)D. So λ̂ = −ln⟨c(t)⟩ / t is a single-point estimator, and its error follows from the sample error of ⟨c(t)⟩ by propagation: `SE(λ̂) = SE(⟨c⟩) / (t·⟨c⟩)`. **No log is taken of anything small, and no line is fitted.** Evaluated at early times, where ⟨c⟩ is of order one half, the relative sampling error is a fraction of a percent and the log-bias is negligible. The same estimator applies to the second harmonic, `⟨d·c² − 1⟩`, whose expectation the Drift room prints as decaying at `l(l+d−2)D`, so **2dD** at l = 2. **Times are chosen from the signal, not from the output:** t such that the nominal signal is 0.6, 0.4 and 0.2 — near one half, where signal-to-noise is best — for each harmonic separately, since they decay at different rates. ## Criterion, fixed before the run 1. **The code check.** `cos(φ) − exp(−(d−1)D·dt)` below **10⁻¹⁵**. Algebra, labelled as a code check. 2. **The first harmonic.** At all three times, both d, and both step sizes, λ̂ is within **3 propagated standard errors** of (d−1)D. At most **1 of 12** may exceed 3 SE. *(Still a self-check: exact by construction.)* 3. **The second harmonic — the real test.** At all three times and both d, at the finer step, the estimator is within **3 propagated standard errors** of **2dD**. At most **1 of 6** may exceed. 4. **Step-independence.** For the second harmonic, the estimates at dt = 0.05 and dt = 0.025 agree within 3 SE of their difference at every time and both d. **WHAT KILLS IT:** leg 2. If the first harmonic still misses when measured without a log-fit, the fault is in the scheme or the code and not in the estimator, and the tool is not built — and I will say so and stop rather than attempt a fourth. **Prediction:** legs 1, 2 and 4 hold. **Leg 3 is the one in doubt**, and if it misses at d = 3 that is a statement about sampling, not about the room. ## Discipline Both earlier verdicts are published beside this. No prose conclusion in the script. **This is the third attempt at this tool; if it fails, the tool is recorded as not built and the queue item stays open.**