# Agent Report — Two Kinds of GUE: Distribution-Level vs Ordering-Level Classification

**Date**: 2026-04-24 03:30
**Piano**: 48
**Tension explored**: META (0.5) + BOUNDARY (0.8)

## Claim Under Test
> The GUE/Poisson classification of 13 domains is treated as a structural finding. But is it a property of sequential correlations (genuine) or of the gap distribution shape alone (tautological)?

## Question
If I shuffle the gap sequence of each domain (destroying ordering, preserving distribution), does the GUE/Poisson classification survive?

## Experiment Design
- **Metric**: r-statistic = mean(min(g_i, g_{i+1}) / max(g_i, g_{i+1})) on consecutive gap pairs
- **Null baseline**: 1000 random permutations of the same gap sequence
- **z-score**: (r_original - r_shuffled_mean) / r_shuffled_std
- **Scope**: 10 domains (primes, GUE matrices, Poisson, logistic, Fibonacci spectrum, Ising 2D, percolation, Brownian, coupled oscillators, cellular automata)
- **Reference**: R_GUE = 0.5307, R_Poisson = 0.3863

## Results

| Domain | N | r_orig | r_shuf | z | Class | Shuf_Class | Verdict |
|--------|---|--------|--------|---|-------|------------|---------|
| primes | 100K | 0.4622 | 0.4813 | -26.6 | GUE | GUE | STRUCTURAL |
| gue | 5K | 0.5995 | 0.6351 | -14.7 | GUE | GUE | STRUCTURAL |
| fibonacci | 609 | 0.4782 | 0.4073 | +8.0 | GUE | **Poisson** | STRUCTURAL + FLIP |
| coupled_osc | 427 | 0.8775 | 0.4146 | +43.5 | GUE | **Poisson** | STRUCTURAL + FLIP |
| percolation | 510 | 0.6551 | 0.4508 | +16.1 | GUE | **Poisson** | STRUCTURAL + FLIP |
| logistic | 100K | 0.3887 | 0.3423 | +61.6 | Poisson | Poisson | STRUCTURAL |
| poisson | 100K | 0.3847 | 0.3856 | -1.1 | Poisson | Poisson | TAUTOLOGICAL |
| brownian | 650 | 0.3148 | 0.3324 | -1.6 | Poisson | Poisson | TAUTOLOGICAL |
| ising_2d | 95 | 0.9858 | 0.9859 | -0.1 | GUE | GUE | TAUTOLOGICAL |
| cell_auto | 105 | 0.8446 | 0.8455 | -0.1 | GUE | GUE | TAUTOLOGICAL |

**Totals**: 6/10 structural (|z|>3), 4/10 tautological (|z|<3), **3/10 class flips on shuffle**.

## Key Findings

1. **Two kinds of GUE.** The GUE-classified domains split into two fundamentally different categories:
   - **Distribution-GUE** (primes, GUE matrices): the gap distribution itself is GUE-like. Shuffling preserves the classification. Sequential ordering adds fine structure (shifts r downward by 0.02-0.04) but doesn't create the classification.
   - **Ordering-GUE** (fibonacci, coupled oscillators, percolation): the gap distribution is Poisson. The GUE classification exists ONLY because of sequential ordering. Destroy the order and they collapse to Poisson. r-shifts are massive: +0.07 to +0.46.

2. **Primes are distribution-GUE.** Shuffled prime gaps still give r=0.4813 (GUE side). The ordering pushes r DOWN by 0.019 (z=-26.6), adding extra gap repulsion beyond what the distribution predicts. This is the opposite sign from ordering-GUE domains (where ordering pushes r UP).

3. **The sign of delta_r is a discriminant.** Distribution-GUE domains have delta_r < 0 (ordering increases repulsion). Ordering-GUE domains have delta_r > 0 (ordering creates attraction/clustering that looks like level repulsion in the r-statistic). The sign tells you which mechanism drives the classification.

4. **3/8 GUE domains are ordering-GUE.** The BOUNDARY claim "8 GUE, 5 Poisson" conflates two distinct mechanisms. The refined picture: 2 distribution-GUE (primes, GUE), 3 ordering-GUE (fibonacci, coupled_osc, percolation), 2 small-N ambiguous (ising, cell_auto), 3 Poisson (poisson, brownian, logistic). The logistic map is Poisson at distribution level but has massive ordering structure (z=+61.6) that doesn't flip the class.

5. **META constraint confirmed.** A binary GUE/Poisson test that doesn't include a shuffle control conflates the two mechanisms. Testing "is r closer to 0.53 or 0.39?" is necessary but insufficient — it doesn't distinguish whether the ordering or the distribution is the source.

## Verdict
**NEW + CONSTRAINT on BOUNDARY + META**

The BOUNDARY claim must be refined: the 8 GUE domains are not homogeneous. Two distinct mechanisms generate GUE statistics. The boundary between GUE and Poisson has two layers: distribution-level and ordering-level. The sign of delta_r = r_original - r_shuffled discriminates which layer operates.

## Bicono della scoperta

- **Due radici** (dipolo primario): Distribution-GUE (il repulsore intrinseco, det=-1 nella forma dei gap) / Ordering-GUE (il repulsore emergente, det=-1 nella sequenza dei gap). Invertite: il primo nasce dalla distribuzione e l'ordine lo affina; il secondo nasce dall'ordine e senza esso collassa.
- **Singolare** (1-che-e-tutto): la r-statistic prima della decomposizione in distribuzione + ordine. Non esiste come ente autonomo — e sovrapposizione dei due canali (distribuzione e sequenza), come la decomposizione two-channel trovata nei run precedenti.
- **Invariante di passaggio**: il segno di delta_r. Sopravvive al passaggio del vertice: delta_r < 0 = repulsione intrinseca (primes, GUE), delta_r > 0 = repulsione emergente (fibonacci, percolation, coupled_osc). Il segno e stabile, non dipende dalla scala.
- **Campo di possibilita**: qui diventa possibile discriminare PERCHE un dominio e GUE (distribuzione vs ordine) — non solo CHE e GUE. Qui diventa non-possibile trattare tutti i domini GUE come omogenei: i test che non separano i due meccanismi (shuffle) sono incompleti.

## Consecutio
La decomposizione two-channel (magnitude/residue) dei run precedenti si sovrappone a questa: il canale magnitudine corrisponde alla distribuzione (chi sei senza ordine), il canale residuo corrisponde all'ordine (cosa l'ordine aggiunge). I 3 domini che flippano sono quelli dove il canale residuo E' il segnale. Per i primi, il canale magnitudine domina ma il residuo aggiunge repulsione (delta_r < 0, z=-26.6). Prossima domanda: i domini ordering-GUE (fibonacci, percolation) hanno la stessa struttura Markov-3 trovata nei primi, o il loro meccanismo di ordine e diverso?

## Files
- Script: `tools/exp_boundary_shuffle_audit.py`
- Data: `tools/data/boundary_shuffle_audit.json`
- Report: `tools/data/reports/agent_20260424_0330.md`