# Agent Report — The Spectral Landscape: Primes Are Not a Mixture
**Date**: 2026-04-06 10:30
**Piano**: 39
**Tension explored**: BOUNDARY (0.75) + METRIC_TENSOR (0.9) + GAP_ANTICORR (0.75)

## Claim Under Test
> The boundary between GUE and Poisson is "the third included" (A9). Is this boundary populated by multiple domains, or are primes special?

## Question
Do multiple domains occupy the spectral boundary, or is the prime niche unique? If Berry-Robnik mixtures fill the same <r> range, what DISCRIMINATES primes from mixtures?

## Experiment Design
- 17 domains (GUE, GOE, GSE, Poisson, power-law, picket fence, clock jitter, primes, semi-Poisson, Berry-Robnik x3, Anderson 1D, Harper phi/rational, quadratic residues)
- N=5000 spacings per domain, 20 shuffled surrogates for z-scores
- Three observables: <r> (ratio statistic), Brody beta, gap_acf1 (lag-1 autocorrelation)
- Scale trajectory: 6 windows across 5.76M primes (p=8K to p=100M)

## Results

### Spectral Landscape (17 domains)
| Domain | <r> | acf1 | z_acf1 | Class |
|:-------|:----|:-----|:-------|:------|
| Harper_phi | 0.320 | -0.003 | -0.6 | POISSON |
| power_law_1.5 | 0.377 | -0.003 | -0.1 | POISSON |
| Anderson_1D | 0.385 | -0.029 | -2.1 | POISSON |
| Poisson | 0.390 | +0.004 | +0.3 | POISSON |
| Berry-Robnik 0.3 | 0.443 | +0.027 | +2.1 | BOUNDARY |
| **primes** | **0.481** | **-0.101** | **-7.5** | **BOUNDARY** |
| Berry-Robnik 0.5 | 0.485 | +0.009 | +0.6 | BOUNDARY |
| semi-Poisson | 0.507 | +0.010 | +0.8 | GOE-zone |
| GOE | 0.533 | -0.280 | -16.9 | GOE |
| Berry-Robnik 0.7 | 0.546 | -0.014 | -1.5 | GOE-zone |
| zeta zeros | 0.596 | -0.319 | -18.3 | GUE |
| GUE | 0.602 | -0.278 | -20.8 | GUE |
| quad. residues | 0.613 | -0.046 | -4.0 | GUE |
| GSE | 0.672 | -0.319 | -21.6 | RIGID |
| clock jitter | 0.724 | +0.018 | +1.2 | RIGID |
| picket fence | 0.989 | +0.017 | +1.2 | RIGID |
| Harper rational | 0.996 | +0.002 | +0.6 | RIGID |

### The discriminant: <r> vs acf1
Two mechanisms produce intermediate <r>:
- **MIXTURE** (Berry-Robnik): superpose chaotic + regular levels. Any <r> achievable. acf1 ~ 0.
- **INTRINSIC** (primes): gaps anti-correlated by nature. acf1 = -0.10, z = -7.5.

Berry-Robnik 0.5 has <r>=0.485, nearly identical to primes (0.481). But its acf1=+0.009 (no ordering). The <r> statistic alone CANNOT distinguish these fundamentally different spectral types.

### Prime trajectory (scale dependence)
| Scale (p range) | <r> | acf1 | <r>_shuf | delta_order |
|:----------------|:----|:-----|:---------|:------------|
| 8K-612K | 0.465 | -0.071 | 0.483 | -0.018 |
| 612K-2.8M | 0.455 | -0.067 | 0.472 | -0.017 |
| 2.8M-7.4M | 0.452 | -0.058 | 0.468 | -0.015 |
| 7.4M-24M | 0.449 | -0.054 | 0.463 | -0.014 |
| 24M-50M | 0.447 | -0.053 | 0.460 | -0.013 |
| 50M-100M | 0.445 | -0.051 | 0.458 | -0.013 |

Slopes: d<r>/d(ln p) = -0.0035, d(acf1)/d(ln p) = +0.0039.
Both drift toward Poisson origin (0.386, 0).

Decomposition: <r>_prime = <r>_distribution + delta_ordering
- <r>_distribution (shuffled) drifts toward Poisson from gap distribution (PNT)
- delta_ordering = <r>_prime - <r>_shuffled, always negative, SHRINKS with scale

## Key Findings

1. **Primes are NOT a Berry-Robnik mixture.** Berry-Robnik systems achieve intermediate <r> by mixing chaotic and regular levels. Their gap autocorrelation is zero — no ordering. Primes achieve intermediate <r> through INTRINSIC anti-correlation (big-small alternation, Hardy-Littlewood). Same <r>, different physics.

2. **The spectral landscape is 2D, not 1D.** <r> alone classifies 4 zones (Poisson/boundary/GOE-GUE/rigid). Adding acf1 splits the boundary zone: mixtures (acf1~0) vs intrinsically ordered (acf1<<0). Primes are the ONLY tested domain at intermediate position on BOTH axes.

3. **The prime trajectory converges to Poisson from a unique direction.** In (<r>, acf1) space, primes move from (0.465, -0.071) toward (0.445, -0.051). Both coordinates drift toward the Poisson origin. The ordering effect (delta_ordering) shrinks: from -0.018 at small scale to -0.013 at large scale.

4. **The "third included" is confirmed but refined.** The boundary is not an <r> value — it's a REGION in (<r>, acf1) space. Mixtures fill this region horizontally (any <r>, acf1=0). Primes cut through it diagonally (intermediate <r> AND intermediate acf1). The third included is not a point on a line — it's a trajectory through a plane.

5. **Quadratic residues are GUE-like.** <r>=0.613, confirming Katz-Sarnak for quadratic L-functions. But their acf1=-0.046 is MUCH weaker than GUE (-0.28), suggesting incomplete level repulsion — another boundary domain on a different axis.

## Verdict
**NEW** — The spectral landscape is 2-dimensional: (<r>, acf1). Primes occupy a unique niche: intrinsically anti-correlated at intermediate repulsion. No mixture model reproduces this. The discriminant is the gap autocorrelation, not the ratio statistic.

New tension: **SPECTRAL_NICHE** — primes are the only known domain at (intermediate <r>, significantly negative acf1). The 2D classification opens: are there other intrinsically-ordered boundary domains?

Constraint: **<r> alone is insufficient** to characterize spectral statistics. Any study claiming "primes are GUE-like" or "primes are between GOE and Poisson" based solely on <r> is missing the ordering dimension.

## Consecutio — What This Opens
1. The 2D plane (<r>, acf1) can be parameterized. What is the natural coordinate system? Is there a one-parameter family connecting Poisson to GUE that passes through the prime point?
2. Quadratic residues are GUE in <r> but weakly ordered (acf1=-0.05). Are there other "weak-GUE" domains?
3. The delta_ordering shrinks with scale. At what scale does it vanish? This gives a SECOND Poisson horizon (the ordering horizon) independent of the <r> horizon.

## Files
- Script: tools/exp_spectral_landscape.py (reusable for any domain)
- Data: tools/data/exp_spectral_landscape.json
- Report: this file