ASFM Blind 87/91 meta-analysis
================================

Locked blind grids included: 87^3 and 91^3
Node-detection parameters: quantile=0.995, min_distance=0.12, top_nodes=28
Each grid: 28 nodes, 378 all-pair evaluations, 6 seeds per pair, 12000 events per seed

Combined pairs: 756
Bell-stable pairs: 98/756 (12.963%)
Tsirelson exceedance pairs: 0/756
Max mean |S|: 2.758714643
Max mean+3sigma: 2.776575626
Min Tsirelson margin: 0.051851499

Grid 87^3:
  Bell-stable pairs: 46/378 (12.169%)
  Tsirelson exceedance pairs: 0/378
  Max mean |S|: 2.758714643
  Max mean+3sigma: 2.776575626
  chi correlation all pairs: r=0.889861, r2=0.791853
  chi correlation bell-stable: r=0.700333, r2=0.490467, n=46

Grid 91^3:
  Bell-stable pairs: 52/378 (13.757%)
  Tsirelson exceedance pairs: 0/378
  Max mean |S|: 2.683855784
  Max mean+3sigma: 2.693652583
  chi correlation all pairs: r=0.910272, r2=0.828594
  chi correlation bell-stable: r=0.739809, r2=0.547317, n=52

Combined chi correlation all pairs: r=0.900633, r2=0.811139, n=756
Combined chi correlation bell-stable: r=0.694819, r2=0.482774, n=98

Interpretation:
The locked blind-grid result replicates the class-level finding: unseen lattices produce many Bell-violating and Tsirelson-safe residual-field resonance pairs under fixed node-detection parameters. The near-Tsirelson strength is not universal across grids: Blind87 contains stronger high-S pairs than Blind91. Direct Noether coherence remains a strong global classifier, while high-S calibration is still incomplete and resonance-class dependent.