psn-mechanism-benchmark-small

passes_gate: False (validity=True, confirmatory=False) full_passed: True | credited mechanisms: 2/4 chances: {'state': 0.16666666666666666, 'plan': 0.25, 'task': 0.20833333333333331}

Model comparison (task_acc = mean(state_acc, plan_acc))

model	group	params	carried_floats	task_acc	state_acc	plan_acc
psn	candidate	2569365	1728	0.447	0.475	0.418
mlp	fair	2657369	64	0.224	0.182	0.266
gru	fair	2486297	896	0.332	0.373	0.290
mem_tf	fair	2442329	10368	0.267	0.282	0.253
windowed_tf	fair	2431961	1024	0.238	0.223	0.253
full_tf	ceiling	2431961	3072*	0.213	0.210	0.216

PSN - best fair baseline (gru): 0.115 (margin needed 0.08) ceiling full_tf task_acc: 0.21321614583333334 (reported, NOT required to beat)

Confirmatory checks

• PASS full_learns_state
• PASS full_learns_plan
• PASS full_decisive
• PASS param_match_fair_baselines
• PASS beats_fair_baselines_by_margin
• FAIL enough_mechanisms_bite

Ablation credit (per-mechanism axis)

ablation	axis	full	ablated	delta	min_delta	credited
no_m	state_acc	0.475	0.307	0.168	0.05	yes
no_hierarchy	state_acc	0.475	0.502	-0.027	0.05	no
no_slow_level	state_acc	0.475	0.396	0.079	0.05	yes
no_rollout	plan_acc	0.418	0.376	0.042	0.1	no