[22:19:57] phase A: base models
[22:23:15] cm28 (, 4 ep, ex/cl_train0.jsonl): val 0.9342 test 0.8076
[22:25:29] u2 (p2c, 8 ep, ex/cl_train0.jsonl): val 0.8319 test 0.7698
[22:26:44] u4 (p4c, 4 ep, ex/cl_train0.jsonl): val 0.8710 test 0.7936
[22:28:18] u8 (p8c, 4 ep, ex/cl_train0.jsonl): val 0.9084 test 0.8335
[22:30:24] u16 (p16c, 4 ep, ex/cl_train0.jsonl): val 0.9403 test 0.8298
[22:30:26] phase B: coverage-tuned slate (targets D 0.950, B 0.960, C 0.968; A = plain 4L)
[22:30:26] coverages: D(2L)=75% B(8L)=70% C(16L)=70%
[22:38:19] cld2 (p2c, 16 ep, ex/cl_full75.jsonl): val 0.9829 test 0.9804