# PRE-REGISTRATION — routed crowd, attempt 2: the multiplicity allowance, derived Written 29 September 2026, after attempt 1 failed, before this run. ## ⛔ Attempt 1 — FAILS, on a criterion that was miscalibrated by design `PREREG_routed_crowd_2026-09-29.md` required **zero of 14 cells** beyond 3 standard errors, on legs 3 and 4. **One cell exceeded on each: 3.11 SE.** The verdict FAILS stands. **That criterion could not have been right.** With 14 cells and a two-sided 3-SE threshold, the per-cell false-exceedance rate is p = 0.0027, so the expected count is **14p = 0.038** and P(at least one) = 1 − (1 − p)^14 = 0.0372 P(at least two) = 0.0007 **"Zero of 14" therefore fails about one run in twenty-seven with everything correct.** I had this right in entry 33 — *"at most 1 of 6 may exceed 3 SE, by chance roughly 0.05 expected"* — and then wrote zero here. **That is the eleventh time in two days I fixed a number without asking what it implies about the measurement.** ## What attempt 1 established and does not need re-running - **Leg 1 held:** the vector sum's fitted exponent is **0.5000** at d = 2 and **0.5010** at d = 3, and the plane constant matched √π/2 within 3 SE. Rayleigh's √N. - **Leg 2 held:** the growth channel's fitted exponent is **0.999380** and **0.999263**. Linear in N. - The ratio growth ÷ resultant ran 1.15 → 3.63 → 11.2 → **36.7** at d = 2 over N = 10 … 10,000. **Those are the finding, and they are not re-run here.** This attempt re-tests only legs 3 and 4, whose failure was a criterion artefact. ## Criterion, fixed before the run Both legs are judged at 3 standard errors per cell, with the allowance taken from the binomial above: 1. **Per-mind growth.** `⟨Σ max(0,c)⟩ / N` against the exact mean — **1/4** at d = 3, **1/π** at d = 2 — at all 7 crowd sizes and both dimensions. **At most 1 of 14 may exceed 3 SE** (false-failure rate 0.07%). 2. **Growth equals load.** `|⟨Σmax(0,c)⟩ − ⟨Σmax(0,−c)⟩| / N` against zero, same 14 cells, **at most 1 may exceed.** 3. **And the fix must not simply be more luck.** Both legs are re-run at **four times** the trials of attempt 1. With four times the sampling, the exceeding cell's z should fall by roughly a factor of two **if it was noise**; if it stays above 3 the deviation is real and the exact mean is wrong for this sampler. Leg 3 is the part that makes this more than a widened goalpost: **a genuine deviation survives more sampling, and noise does not.** **WHAT KILLS IT:** the same cell exceeding 3 SE again at four times the trials. Then the per-mind mean is not 1/π or not 1/4 for these draws, and the algebra or the sampler is wrong. **Prediction:** legs 1 and 2 hold and leg 3 shows the z falling. **If the cell survives, that is the more interesting outcome and it gets reported as a defect in my stated exact means.** ## Discipline Attempt 1's FAILS is published beside this. No prose conclusion in the script.