Experiment complete. Summary:

**Five observables, each placing primes on a GUE-Poisson scale (τ ∈ [0,1]):**

- **3/5 are order-invariant** (Δτ = 0.000 vs shuffle) — they measure gap distribution, not ordering. META partially confirmed: 60% of typical observables are tautological under shuffle.

- **2/5 form a dipole**: spacing_ratio pushed toward Poisson (Δτ = −0.12), lag1_acf pushed toward GUE (Δτ = +0.20). The same phenomenon (consecutive gap anticorrelation) looks like opposite things depending on which observable you ask.

- **Ordering creates 2x more coherence** (τ std 0.09 vs 0.19 for shuffle). The ordering doesn't add noise — it makes observables agree more.

- **All τ drift toward Poisson with scale** (confirming Brody flow from previous run).

The boundary between GUE and Poisson is not a point on a 1D axis. It's a 2D structure with a dipolar ordering signature. Report at `tools/data/reports/agent_20260430_1905.md`, tool at `tools/exp_boundary_coherence.py`, new tension DIPOLAR_ORDERING added to seme.