# Agent Report — The Lag-6 Crossover Is Not Z/6Z: It's the Ratio of Anti-Correlation to Trend

**Date**: 2026-04-18 03:30
**Piano**: 39
**Tension explored**: DIPOLAR_ILLUSORY_BOUNDARY (0.9)

## Claim Under Test

> Consecutio from agent_20260417: "Crossover lag 6 ~ Z/6Z cycle? All primes >3 are 1 or 5 mod 6. Gap mod 6 in {0,2,4}. The crossover at lag 6 = one full cycle of this modular structure. Coincidence or mechanism?"

## Question

Is the lag-6 ACF crossover in raw prime gaps (where positive trend overtakes negative anti-correlation) determined by Z/6Z arithmetic structure, or by the ratio of PNT trend strength to structural anti-correlation decay?

## Experiment Design

Three discriminating tests on 500K primes (p_max = 7,368,787):

1. **Residue ACF**: Compute ACF of the gap-mod-6 sequence (r_n = g_n mod 6, values in {0,2,4}). If Z/6Z imposes period-6 structure, lag 6 should be special.

2. **Trend re-injection**: Multiply PNT-normalized gaps by ln(p)^s for s = 0.0 to 1.5. Track crossover lag as function of trend strength. If crossover at 6 requires exactly s=1.0 (full PNT trend), the position is analytically determined.

3. **Factorial surrogates** (20 each):
   - Residue-preserving shuffle (same Z/6Z sequence, shuffled magnitudes within each class)
   - Full shuffle (destroy everything)
   - Cramer model (exponential gaps with mean ln(p), no Z/6Z, no correlation)
   - Cramer + Z/6Z constraint (round Cramer to same mod-6 residues as primes)
   - AR(1) with phi = -0.039 (matching prime acf1), no trend
   - AR(1) + PNT trend

## Results

### Test 1: Residue ACF — No period-6 structure

| Property | Value |
|----------|-------|
| Fraction r=0 (multiples of 6) | 0.430 |
| Fraction r=2 | 0.285 |
| Fraction r=4 | 0.285 |
| ACF lag 1 | **-0.148** |
| ACF lag 2 | +0.017 |
| ACF lag 6 | +0.001 |
| ACF lag 12 | +0.002 |
| ACF lag 18 | -0.002 |
| Crossover | lag 2 |

The residue sequence has STRONG lag-1 anti-correlation (-0.148, consecutive gaps tend to have different mod-6 classes) but NO special behavior at lag 6. The ACF at lag 6 (+0.001) is indistinguishable from noise. Z/6Z imposes no period-6 structure on the autocorrelation.

### Test 2: Trend re-injection — Smooth, monotonic, recovers lag 6 at s=1.0

| Strength s | Crossover lag | N negative (/50) |
|------------|:------------:|:----------------:|
| 0.0 (normalized) | 15 | 45 |
| 0.3 | 15 | 38 |
| 0.5 | 15 | 29 |
| 0.6 | 12 | 19 |
| 0.8 | 9 | 10 |
| **1.0 (raw)** | **6** | **6** |
| 1.2 | 5 | 4 |
| 1.5 | 4 | 3 |

The crossover is a **smooth, monotonic function of trend strength**. No discontinuity or resonance at lag 6 — it's continuously tunable. ln(p)^1.0 exactly reproduces the raw crossover. The "6" has no arithmetic origin; it's where the PNT growth rate equals the anti-correlation decay rate.

### Test 3: Factorial surrogates — Z/6Z has zero effect

| Condition | Crossover | N_neg (/50) |
|-----------|:---------:|:-----------:|
| **Real prime gaps (raw)** | **6** | **6** |
| Real prime gaps (normalized) | 15 | 45 |
| Residue-preserving shuffle | 2.5 +/- 0.7 | 25 |
| Full shuffle | 1.9 +/- 1.4 | 25 |
| Cramer (trend, no Z/6Z) | 1.0 +/- 0.0 | 0 |
| Cramer + Z/6Z | 1.1 +/- 0.2 | 0 |
| AR(1) phi=-0.039, no trend | 2.1 +/- 0.4 | 25 |
| AR(1) + trend | 1.0 +/- 0.0 | 0 |

Key comparisons:

- **Residue-preserving shuffle (2.5) ~ Full shuffle (1.9)**: Preserving the Z/6Z residue sequence makes no difference. The crossover is NOT carried by the modular structure.

- **Cramer (1.0) ~ Cramer + Z/6Z (1.1)**: Adding Z/6Z constraint to Cramer changes nothing. Z/6Z has zero effect on ACF crossover.

- **AR(1) no trend (2.1) vs AR(1) + trend (1.0)**: The AR(1) anti-correlation (phi=-0.039) is too weak to resist the trend even at lag 1. Only prime gaps have anti-correlation STRONG ENOUGH to hold for 5 lags against the trend.

- **Cramer + trend (1.0): ALL positive ACF**. The PNT trend creates purely positive autocorrelation when there's no structural anti-correlation to resist it.

## Key Findings

1. **The lag-6 crossover is NOT Z/6Z arithmetic.** Z/6Z residues show no period-6 ACF structure (lag 6 ACF = +0.001, noise level). Preserving Z/6Z while destroying magnitude order gives crossover at 2.5 (like full shuffle). Adding Z/6Z to Cramer changes crossover by 0.1. The "6" in lag-6 and the "6" in Z/6Z are a pure coincidence.

2. **The crossover position is determined by trend strength, continuously tunable.** Multiplying normalized gaps by ln(p)^s smoothly moves the crossover from 15 (s=0) through 9 (s=0.8) to 6 (s=1.0) to 4 (s=1.5). There is no resonance, no special value — just a smooth function. The crossover at lag k* occurs where:

   k* ~ A / C(s)

   where A ~ 0.037 (structural anti-correlation amplitude from ACF_1K_LAW) and C(s) is the trend-induced positive bias that grows with s.

3. **The uniqueness of primes in the crossover is their anti-correlation STRENGTH, not their arithmetic.** AR(1) with the same acf1=-0.039 can't hold even 1 lag against the trend (crossover=1.0 with trend). Real primes hold for 5 lags because their anti-correlation follows a 1/k law (slow decay), not an AR(1) (exponential decay). The long-range power-law anti-correlation (ACF_1K_LAW) is what resists the trend, not the Z/6Z lattice.

4. **Residue lag-1 anti-correlation (-0.148) is a genuine short-range signal.** Consecutive gaps strongly avoid the same mod-6 class. This connects to F2 (Z/6Z confinement) but operates at lag 1, not lag 6.

## Verdict

**FALSIFIED**: The lag-6 crossover is NOT Z/6Z arithmetic. It is the analytically-determined point where PNT trend bias equals 1/k anti-correlation amplitude.

**CONSTRAINT on DIPOLAR_ILLUSORY_BOUNDARY**: The crossover lag is k* = A/C, a smooth function of scale. As primes grow (A decays per ACF_AMPLITUDE_SCALING), k* should SHRINK — the dipolar-illusory boundary moves to shorter lags at larger primes. The dipolar regime narrows with scale.

**NEW observation**: Residue lag-1 anti-correlation r_1 = -0.148 quantifies the Z/6Z short-range avoidance pattern. This is a distinct signal from the 1/k long-range anti-correlation (which is -0.039 at lag 1 for full gaps). The residue structure contributes 0.148/0.039 = 3.8x the anti-correlation at lag 1 that the magnitude structure does. The residue and magnitude anti-correlations are on different scales and decay differently.

## Consecutio — what this opens

1. **k* as a function of scale**: Since A(ln p) = 0.096 - 0.0033*ln(p) (from ACF_AMPLITUDE_SCALING) and C(s=1) is determined by the local variance of ln(p), the crossover k*(p) should be derivable analytically. At larger primes, k* shrinks — testable on the 6M prime dataset across scale windows.

2. **Two anti-correlation channels**: The residue anti-correlation (r_1=-0.148, fast decay) and the magnitude anti-correlation (from Hardy-Littlewood, 1/k decay) are distinct. The residue channel dominates at lag 1, the magnitude channel dominates at lags 2+. Are they independent or coupled? Cross-correlation between residue sequence and normalized gap magnitude would answer this.

3. **Revised Poisson crossover**: If the dipolar regime narrows (k* shrinks) with scale, then at the Poisson crossover (p*~10^{13}), k* might be just 1-2 lags — meaning only the nearest-neighbor anti-correlation survives. The approach to Poisson is not just amplitude decay but REGIME NARROWING.

## Files

- Script: `tools/exp_acf_z6z_mechanism.py` (reusable: --n_primes, --max_lag, --n_surrogates)
- Data: `tools/data/exp_acf_z6z_mechanism.json`
- Report: `tools/data/reports/agent_20260418_0330.md`