schema: cognia-cdct-levelized-preregistration/1
frozen-before-results: true
parent-result: H3 INCONCLUSIVE because all arms reached the accuracy ceiling
seeds: 11,23,37,41,53,67,79,83,97,101
hypothesis: H3b
claim: an exact declared loopback preserves calibrated binary memory beyond the training horizon better than trained GRU and LSTM controls
training: 64 sequences x 32 ticks; identical set/reset/hold observations; 400 full-batch AdamW steps; lr 0.05
test: balanced set/reset cues followed by holds; checkpoints at 1000, 10000 and 100000 ticks; no retraining or tuning
primary-metric: survival score = 1 - mean Brier error over both bits and all three checkpoints
secondary-metrics: minimum confidence margin from 0.5 and first incorrect checkpoint
decision: SUPPORTED only when lower 95% Student interval of Levelized CDCT survival score exceeds upper intervals of both controls
null-result: overlapping intervals
falsification: either control interval overlaps or exceeds the Levelized CDCT interval
ceiling-policy: binary accuracy alone is descriptive and cannot support H3b
