# PRE-REGISTRATION — does an unaimed crowd cancel, or only look like it? Written 29 September 2026, before the run. **Coherence + The Switch**, rank 8, sharing L and a. ## The two rooms **Coherence** (Rayleigh's random walk): `N shared · √N unshared (√(πN)/2 in the plane)`. Minds that share a reference add up as N; minds that do not add up as a random walk. **The Switch**: `sign(a·L̂) → growth or load`. *"The same input becomes motion or stored load depending only on which way you face."* ## What neither room says **Coherence counts the vector sum `|Σaᵢ|`.** The Switch says that is not what the ledger sees: a mind facing away does not *cancel* a mind facing toward — it contributes **load** while the other contributes **growth**. So the growth channel is Σ max(0, aᵢ·L̂) which is a sum of non-negative terms and **cannot cancel at all.** For uniformly aimed minds it should grow as **N**, not √N, while the vector sum still grows as √N. **The √N of Rayleigh is produced entirely by cancellation, and routing destroys cancellation.** ## The exact per-mind means, to compare against For a uniform unit vector, c = a·L̂: - **d = 3:** c is uniform on [−1, 1], so `E[max(0,c)] = ∫₀¹ c/2 dc = 1/4` exactly. - **d = 2:** c = cos θ with θ uniform, so `E[max(0,c)] = (1/2π)∫_{−π/2}^{π/2} cos θ dθ = 1/π`. ## Criterion, fixed before the run 1. **The vector sum is Rayleigh's.** The fitted exponent of `⟨|Σaᵢ|⟩` against N over N = 10 … 10,000 is within **3 standard errors of the fit** of 0.5, at d = 2 and d = 3. And in the plane the constant matches **√π/2 = 0.886227** to within 3 standard errors of the measured mean at the largest N. 2. **The growth channel is linear.** The fitted exponent of `⟨Σ max(0, aᵢ·L̂)⟩` against N is within 3 standard errors of **1.0**, at both d. 3. **The per-mind growth is the exact mean.** `⟨Σ max(0,c)⟩ / N` is within 3 standard errors of **1/4** at d = 3 and **1/π** at d = 2, at every N tested. 4. **Growth and load are equal.** `|⟨Σmax(0,c)⟩ − ⟨Σmax(0,−c)⟩| / N` is within 3 standard errors of zero at every N and both d. Every leg is judged against **its own sampling error**, and the two per-mind means are exact algebra written down before the run. **No tolerance is invented.** Said because roughly ten criteria last night were numbers fixed without asking what precision the quantity has, and one leg could not fail at all. 5. **Reported, not scored:** the ratio `Σ max(0,c) / |Σaᵢ|` against N. If legs 1 and 2 hold it must grow as **√N**, so the routed channel outruns the resultant without bound. **WHAT KILLS IT:** leg 2 failing. If the growth channel is not linear in N then routing does not remove the cancellation and the reading is wrong. **Prediction:** all four hold. Legs 3 and 4 are algebra checked by sampling; **legs 1 and 2 together are the finding**, and it is a statement about the two rooms, not about the world. ## Discipline No prose conclusion in the script. Tables printed by the run. Hashes read from the file.