# PRE-REGISTRATION — crowd, attempt 2: the algebra, and a simulation judged by its own error Written 29 September 2026, after attempt 1 failed, before this run. ## Why there is an attempt 2 Attempt 1 (`PREREG_crowd_strength_is_crowd_error_2026-09-29.md`) required the simulated ⟨|ā|²⟩ to match `ρ + (1−ρ)/N` within **0.002**. Worst measured **0.0039**, at N = 10, ρ = 0.10. **FAILS, and that verdict stands.** **That is the fourth time tonight I set a numerical threshold without first asking what range or precision the quantity actually has.** 0.002 was not derived from anything — the Monte Carlo standard error at that sample size is larger than the threshold I demanded, so the test could fail on noise alone and tells us nothing about the claim. ## The right test ⟨|ā|²⟩ has a closed form. With `aᵢ = √ρ·u + √(1−ρ)·eᵢ`, u and eᵢ independent unit vectors: ā = √ρ·u + √(1−ρ)·ē, ē = (1/N)Σeᵢ |ā|² = ρ|u|² + 2√(ρ(1−ρ))·(u·ē) + (1−ρ)|ē|² ⟨|u|²⟩ = 1, ⟨u·ē⟩ = 0, ⟨|ē|²⟩ = 1/N ⇒ ⟨|ā|²⟩ = ρ + (1−ρ)/N exactly So the claim is **algebra**, and a simulation can only confirm that the code implements the algebra. It must therefore be judged against **its own sampling error**, not a number I invent. ## Criterion, fixed before the run 1. **Within its own error.** At every N ∈ {10, 100, 1000} and ρ ∈ {0, 0.1, 0.25, 0.5, 0.9, 1}, the measured mean lies within **3 standard errors** of `ρ + (1−ρ)/N`, where the standard error is computed from the run's own sample variance. Across 18 cells, at most **1** may exceed 3 SE (by chance roughly 0.05 expected). 2. **The error shrinks as it must.** Quadrupling the trials at one cell (N = 10, ρ = 0.1) halves the standard error, within 20%. 3. **Coherence's endpoints.** At ρ = 1 the resultant is exactly N; at ρ = 0 it is √N within 3 SE. **WHAT KILLS IT:** more than one cell outside 3 SE, or an error that does not shrink as 1/√trials — either would mean the code is not implementing the algebra. **Prediction:** holds. This is a test of my code, not of the world. ## ⛔ And the honest limit on leg 3, restated Attempt 1's leg 3 came out 1.000000 at every cell **by construction**: setting ρ_error = ρ_aim makes the two expressions literally identical, so their ratio is 1 as algebra. **That is not evidence.** The content is entirely in the identification — whether the reference a crowd shares really does carry the error it shares — and **nothing here measures that.** It is recorded as an assumption, and any write-up must say so. ## Discipline Attempt 1's FAILS is published beside this.