# PRE-REGISTRATION — humility, attempt 2: the divergence, tested correctly Written 29 September 2026, after attempt 1 failed, before this run. ## Why there is an attempt 2, stated plainly Attempt 1 (`PREREG_humility_is_not_faith_2026-09-29.md`) pre-registered: > cycles-to-land **diverges** as η → 0 (above 1000 cycles by η = 0.005) It measured **918.7**. The criterion **FAILED** and is recorded as failed. The number 1000 was picked by me with no derivation behind it, and 0.005 was picked to go with it. That is a badly set threshold, not a result, and attempt 1's verdict stands as FAILS regardless of what this run shows. The underlying quantity is analytic, not empirical: cycles to land = ln(tolerance / |b₀ − reading|) / ln(1 − η) As η → 0, ln(1 − η) → −η, so cycles → (a constant)/η. **The right test is the exponent, not a threshold.** ## Criterion, fixed before this run **HOLDS** if, fitting log(cycles) against log(η) over η ∈ [1e-5, 1e-1]: 1. the fitted exponent lies in **[−1.05, −0.95]** — an inverse law, and 2. the fit is tight: R² > 0.999, and 3. the ratio cycles·η approaches the constant ln(tolerance/gap) in absolute value, to within 1% at the small-η end. **WHAT KILLS IT:** an exponent outside that band, or a poor fit. Then cycles-to-land does not diverge as 1/η and my reading of the update law is wrong. **Prediction, recorded now:** HOLDS with exponent −1.000. This leg alone is re-run. **Leg 2 of attempt 1 is not re-run and is not re-reported** — it holds at 0.000e+00 spread and that result stands on its own pre-registration. ## Discipline Verdict assembled from measured quantities. Attempt 1's FAILS is published alongside this, not replaced by it.