# Agent Report — META Tautology Discriminator: 3 of 5 Observables Are Structural, 2 Are Pair-Correlation Artifacts

**Date**: 2026-04-29 09:58
**Piano**: 58
**Tension explored**: META (0.5)

## Claim Under Test
> "Tutti i 11 test passano — verifica che non stiamo testando solo tautologie"

## Question
Which of our key observables measure real prime structure beyond what density and pair correlations explain? If an observable is reproduced by a simple synthetic model, it's tautological — it doesn't probe deep structure.

## Experiment Design
- **Observables tested**: r-statistic, mod-3 ordering fraction, lag-1 autocorrelation, two-channel delta-r (ordering + magnitude)
- **Controls**: (A) Shuffled gaps — same distribution, destroyed order; (B) Cramer model — PNT density, independent exponential gaps; (C) Hardy-Littlewood AR(1) model — matches real primes' pair correlation (lag-1 ACF), no higher-order structure
- **Window**: 49,097 consecutive gaps starting at index 12,274 (mean gap 12.73)
- **Trials**: 20 per synthetic model
- **Discrimination threshold**: |z| > 3.0 (real value vs synthetic distribution)
- An observable is STRUCTURAL if it discriminates against ALL three controls. TAUTOLOGICAL if any control reproduces it.

## Results

| Observable | Real | Shuffled (z) | Cramer (z) | HL-Markov (z) | Verdict |
|:-----------|:-----|:-------------|:-----------|:--------------|:--------|
| r-statistic | 0.4643 | -12.4 | +17.3 | -5.9 | **STRUCTURAL** |
| mod-3 fraction | 0.1614 | -95.1 | -106.7 | -75.0 | **STRUCTURAL** |
| lag-1 ACF | -0.0489 | -7.4 | -9.6 | **-0.3** | TAUTOLOGICAL |
| dr-ordering | -8.71 | -6.5 | -8.2 | **-0.9** | TAUTOLOGICAL |
| dr-magnitude | -12.32 | -13.6 | -9.3 | -8.8 | **STRUCTURAL** |

## Key Findings

1. **Mod-3 ordering is the hardest signal in the lab.** Real primes: 16.1% self-transitions. All synthetics: ~34.5%. z-scores from -75 to -107. No model — not even one matching pair correlations — comes within an order of magnitude. The mod-3 prohibition (zero self-transitions in mod-3 classes, found on 2026-04-28) is structurally irreducible. This is the single strongest discriminant the lab has ever measured.

2. **The r-statistic is structural but subtle.** Real primes r = 0.464 is LOWER than shuffled (0.484) and HL (0.473), but HIGHER than Cramer (0.432). It sits in a specific zone that neither pure randomness nor pair correlations reproduce. The content: primes have more repulsion than Cramer (not random) but less than shuffled (order matters beyond pairs). This is a genuine three-way discrimination.

3. **The magnitude channel (dr_mag) is structural.** z = -8.8 even against the HL model. The shape of the gap distribution's deviations from median carries sequential structure that pair correlations don't capture. This validates the two-channel decomposition as measuring something real — but only in the magnitude channel.

4. **Lag-1 ACF is fully explained by pair correlations.** z = -0.3 against the HL model. By construction, the AR(1) matches lag-1 — so this is expected. But the implication is: any test that relies on lag-1 ACF as evidence of prime structure is measuring pair correlations, not deep structure. The negative lag-1 of primes is real (z = -7.4 vs shuffle) but shallow.

5. **The ordering channel's delta-r is tautological.** z = -0.9 against HL. A simple Markov(1) process reproduces the ordering channel's repulsion signature. This means: the "two-channel structure" is real for the magnitude channel, but the ordering channel adds no information beyond pair correlations. Previous reports citing ordering-channel delta-r as independent evidence were measuring the same thing twice.

## Verdict
**NEW + CONSTRAINT on META + BOUNDARY + DUALITA_DIPOLARE_VS_ILLUSORIA + C1**

- **META resolved at first order**: 3 of 5 observables are structural, 2 are pair-correlation artifacts. The lab is NOT testing only tautologies — but 2 of its observables (lag-1, ordering delta-r) add no discriminant power beyond what a trivial AR(1) model provides. These should be downweighted in future analyses.
- **BOUNDARY**: The boundary between structural and tautological observables is now mapped. Mod-3 (algebraic) and magnitude channel (statistical) are on the structural side. Lag-1 and ordering delta-r are on the tautological side. This is a concrete instance of A9 (terzo incluso) — the boundary itself is the content.
- **DUALITA_DIPOLARE_VS_ILLUSORIA**: The ordering channel's delta-r was ILLUSORY duality — it appeared to add information but was redundant with pair correlations (det=+1). The magnitude channel's delta-r is DIPOLAR duality — it captures structure beyond pairs (det=-1).
- **C1**: The mod-3 signal (z ≈ -100) is the strongest evidence for C1 (primes as unique dynamic domain). No synthetic reproduces it. The mechanism is algebraic (F2: Z/6Z confinement), not statistical.

## Bicono della scoperta

- **Due radici** (dipolo primario): observable-structural (mod3, r-stat, dr_mag — discriminano tutti i sintetici) e observable-tautological (lag1, dr_ord — riprodotti da correlazioni a coppie). La dualita non e nella misura ma nel livello di struttura che raggiunge.
- **Singolare**: la sequenza dei primi stessa, prima della decomposizione in osservabili. Contiene entrambi i livelli (paia + oltre-paia) come uno solo. La separazione e dell'osservatore, non dell'oggetto.
- **Invariante di passaggio**: la proibizione mod-3 (z ≈ -100). Sopravvive a ogni controllo sintetico, non decade con la scala, e irreducibile a correlazioni di ordine inferiore. E il residuo di Cassini delle osservabili — il segnale che non si annulla.
- **Campo di possibilita**: possibile → classificare ogni osservabile del lab per profondita strutturale (pair-level vs beyond-pair). Non-possibile → trattare tutti i test passati come evidenza indipendente (2 su 5 sono ridondanti).

## Consecutio — cosa apre

1. **Hierarchy of observables**: ora possiamo ordinare tutti gli 11 test del lab per profondita strutturale. Quanti dei restanti 6 sono tautologici?
2. **Mod-3 is the anchor**: ogni nuovo esperimento dovrebbe includere il mod-3 test come sanity check. Se un sintetico riproduce mod-3, c'e qualcosa di profondo nel modello.
3. **Higher-order Markov discrimination**: il HL model e AR(1). Che succede con AR(2) o AR(3)? Il mod-3 viene catturato a qualche ordine Markov, o e genuinamente non-Markoviano?

## Files
- Script: `tools/exp_meta_tautology_test.py`
- Data: `tools/data/meta_tautology_test.json`
- Report: `tools/data/reports/agent_20260429_0958.md`