SF × SIZE GRID RESULTS
======================================================================

EXPERIMENTAL DESIGN:
- Spatial frequencies tested: [10.0, 20.0, 30.0, 40.0, 50.0] cycles/image
- CRF sizes tested: [3, 5, 7, 9] pixels
- eCRF diameter: 37 pixels (fixed)
- Contrast: 1.0 (100%, fixed)
- Total conditions: 5 × 4 = 20

PROTOCOL:
For each (spatial_frequency, CRF_size) pair:
1. Train decoder on CRF gratings at that SF and size
2. Generate eCRF gratings at that SF
3. Run multi-orientation optimization (6 orientations: [0, 30, 60, 90, 120, 150])
4. Rotate and average to get orientation-invariant structure

KEY FINDINGS:

Excitatory Adjustment Strength (mean absolute):
[[6.70693262e-05 4.95117711e-05 5.15595602e-05 5.19177702e-05
  4.89650265e-05]
 [5.38474487e-05 5.44530049e-05 5.06076831e-05 4.27818159e-05
  3.95175011e-05]
 [5.07107616e-05 5.13898121e-05 5.17304434e-05 4.77490830e-05
  4.54332039e-05]
 [5.99074738e-05 6.50418297e-05 5.11763501e-05 6.75721822e-05
  4.71675230e-05]]

Inhibitory Adjustment Strength (mean absolute):
[[3.49911497e-05 3.42165476e-05 3.67712855e-05 3.63108338e-05
  3.77063698e-05]
 [3.58346952e-05 3.17976665e-05 3.25544570e-05 3.46463457e-05
  2.87400653e-05]
 [3.10085197e-05 3.28388560e-05 3.65981250e-05 3.39465332e-05
  3.37933125e-05]
 [3.67000794e-05 4.47330094e-05 3.48606191e-05 4.67279388e-05
  2.96658181e-05]]

E-I Balance (mean):
[[-5.88172188e-05 -4.35165784e-05 -2.13890780e-05 -3.43000647e-05
  -1.68737679e-05]
 [-3.51686722e-05 -4.43812314e-05 -2.66553579e-05 -2.09512036e-05
  -1.87936803e-05]
 [-3.92346337e-05 -2.27063101e-05 -9.82691836e-06 -2.67494652e-05
  -2.37269960e-05]
 [-3.99877899e-05 -3.71561000e-05 -1.61734952e-05 -4.94544620e-05
  -1.61111257e-05]]

Final Loss (lower = better rescue):
[[5.49761518e-04 2.91019493e-04 2.25833640e-04 5.96473258e-03
  9.61575045e-05]
 [5.95927661e-05 1.47209427e-04 1.86049459e-04 5.51299130e-05
  3.25371734e-05]
 [6.98393718e-05 2.44837933e-05 5.98376482e-05 3.99050477e-05
  1.04168424e-04]
 [1.81748292e-05 3.77895900e-05 2.64222533e-05 2.48950096e-05
  6.59387241e-05]]

INTERPRETATION:
These results reveal how the recurrent compensation mechanism depends on:
- Scale (spatial frequency): Does recurrence prefer coarse or fine features?
- Size (CRF diameter): How does the size of the perturbed region affect rescue?

Key patterns to look for:
- SF selectivity: Does recurrence work better at certain scales?
- Size dependence: Larger perturbations harder to rescue?
- E/I balance shifts: How do E and I contributions change with scale/size?