p4 Q matrix on Colab: q-matrix-t4

Pinned commits: {"phi": "d039f00225c5749220045e14ea4bb8f0fa32880b", "p4": "0157bbda3f6349494953605e4e995c61e942f08a", "store": "bdd69831bad870d663d0b39fae48f36c4f140385"} Trained on: Tesla T4, cpu

k	seeds	one large	Zeckendorf ensemble	beats it every seed	φ-routed	2:4-routed	route hit (φ / 2:4)	chance	pretrained on	one large's steps
2	2	0.185	0.172	no	0.161	0.152	0.77 / 0.75	0.50	cuda Tesla T4	35870, 36072
3	2	0.191	0.219	no	0.179	0.160	0.67 / 0.54	0.33	cuda Tesla T4	36120, 34976
4	2	0.196	0.186	no	0.152	0.196	0.40 / 0.48	0.25	cuda Tesla T4	35331, 36677

Tree (k = 4) held out: medium 0.158, high 0.158.

.pt source	with	without	kept
ranker	0.3751	0.3788	no
medium	0.3757	0.3788	no
high	0.3652	0.3788	no

Teacher claude-fable-5-1: 24 pools compared, agreement with the store's labels precision 0.5075, recall 0.6048, f1 0.5081.

trained on	scored against the store's labels	scored against the teacher's
the store's labels	0.165	0.1652
the teacher's labels	0.1606	0.2019

Verdicts

k = 4 is the only node count where the Zeckendorf ensemble beats one large model on every seed: does not hold
the matrix route finds the best node above chance at every k: holds
φ routing ranks better than 2:4 routing at every k: does not hold
the tree at high ranks better than at medium (held out): does not hold
the JSON ranker keeps the tree's .pt weight at high: does not hold
training on the teacher's labels predicts the store's own labels at least as well as training on them: does not hold