== select ==
model          latents    co_bps   cycle  stateR2  rateR2
---------------------------------------------------------
latent_size=5        5    0.0680   1.000    0.863   0.377 *front
latent_size=20      20    0.1000   0.950    0.746   0.553 <- recommended
latent_size=64      64    0.1095   0.807    0.697   0.623 *front

WARNING RECONSTRUCTION_TRAP
  The best-reconstructing model latent_size=64_seed=0_max_epochs=25_batch=25 gains
  +0.0095 co_bps over latent_size=20_seed=0_max_epochs=25_batch=25 but its cycle-
  consistency is 0.81 vs 0.95. Ranking by reconstruction alone would pick a model whose
  extra latent structure the data do not constrain. Ground truth agrees: State R² 0.70
  vs 0.75.

== validation against the task-trained ground truth ==
3 bundles with ground truth

detector                    tier   TP   FP   FN   TN  precision  recall
FEATURE_INVENTION              0    1    0    2    0       1.00    0.33
INPUT_INHERITED_STRUCTURE      0    0    2    0    1       0.00     n/a
UNDERFITTING                   0    3    0    0    0       1.00    1.00

against sweep-relative truth (more than 0.05 below the best model here)
detector                    tier   TP   FP   FN   TN  precision  recall
FEATURE_INVENTION              0    1    0    1    1       1.00    0.50
UNDERFITTING                   0    2    1    0    0       0.67    1.00

disagreements
  latent_size=20_seed=0_max_epochs=25_batch=25 FEATURE_INVENTION tier 0: missed (truth: present); INPUT_INHERITED_STRUCTURE tier 0: flagged (truth: absent)
  latent_size=5_seed=0_max_epochs=25_batch=25 FEATURE_INVENTION tier 0: missed (truth: present); INPUT_INHERITED_STRUCTURE tier 0: flagged (truth: absent)
