=== dup.c — does DIMENSION-ADDITION cross where tunnelling cannot? ===
M=16 units, 4000 sweeps, window 10000, 5 seeds, T=2e-05
F fixed: N=3 throughout D grow: N=1 ->(sw 200) 2 ->(sw 400) 3
optimal log-loss 0.7960 nats
--- guard: did the test actually run? ---
[1] both arms end at N=3 (equal capacity)? yes
[2] scoring-call budget F 640080 D 640080 (ratio 1.000; D handicapped if <1)
[3] positive control: TRUE-seeded ends at gap +0.0001 (must stay <+0.01)
guard PASSED.
PRIMARY — distinct belief states @ G=3000, ALL 5 seeds:
seed 0 1 2 3 4 mean
F fixed 7083 9159 7851 8132 9643 8374
D grow 4766 2296 5118 5378 5351 4582
(true Mess3 @ G=3000: 54821 ; learn.c collapse: 36)
full ladder, seed 0: G=30 100 300 1000 3000
F fixed 128 531 1524 3845 7083
D grow 94 369 994 2494 4766
TRUE 396 2399 8924 26394 54821
SECONDARY — log-loss gap: F +0.0052 +- 0.0022 | D +0.0192 +- 0.0068
SEPARATION F - D = -0.0140 +- 0.0074 (>0 means GROWING beat starting there)
NOTE: log-loss SECONDARY by pre-registration. Judge on the ladder.
EXIT=0