[sn1] fifty-move damping applied after the TT (raw evals in TT), qsearch TT cutoffs off at halfmove >= 90; same curve (non-regression)

[sn1] DFRC datagen d5 from A617C0B4 (kb4, gen 12): 600k games N=15000 (genfens dfrc)

[sn1] gen c16 from 43F3DD60 (gen 13): 1.5M games N=15000; adjudication A/B control (current settings)

[sn1] gen c16t from 43F3DD60 (gen 13): N=15000; adjudication A/B: TB adjudication 5-man required, no win adjudication

[atlas] Per-bucket hidden layers at prod scale S4000, four times the data per stack, after the S1000 version missed by 9.5.

[sn1] FIXED NODES: training length: hidden layer + warm-up 400 SB (43C07679) vs 260 SB (8A964D65), same data; hidden1 both sides

[sn1] STC no-worse: 8-neuron hidden layer + warm-up on gen-14 data (91057715; reordered) vs main gen 14 AE339E48

[sn1] draw fixes: insufficient material, mate at halfmove 100 (non-regression)

[sn1] fifty-move damping: none until halfmove 50, then linear to 10% at 100 (vs hmdamp: linear from 0 to 50%)

[sn1] FIXED NODES: h14w with hidden-layer inputs at shift 10 (u8 max 63; would allow a faster AVX2 kernel) vs shift 9 (max 127); same weights, hidden1 both sides

[sn1] FIXED NODES: training length: hidden layer + warm-up 400 SB (43C07679) vs 260 SB (8A964D65), same data; hidden1 both sides

[sn1] STC no-worse: hidden layer + warm-up on gen-14 data (B8809FFD; reordered FT, fused kernels) vs main gen 14 AE339E48

[sn1] FIXED NODES: hidden layer + warm-up on gen-14 data (B8809FFD, reordered) vs main gen 14 AE339E48 (same data, no hidden layer)

[sn1] gen 14 net AE339E48 (kb4, c6a..c15 + DFRC d1-d4 ~10%, 330 SB) vs 43F3DD60 (gen 13)

[sn1] Syzygy WDL result cache (512 KB) vs main; tables on both, TB and win adjudication off (non-regression)

[sn1] FIXED NODES: hidden layer + warm-up, no factoriser (FB00018C) vs with factoriser (8A964D65); hidden1 engine both sides

[sn1] FIXED NODES: per-bucket hidden layer + LR warm-up (8A964D65, 0 dead) vs per-bucket hidden layer without warm-up (E1C896AE, 17 dead); hidden1 engine both sides

[sn1] FIXED NODES: shared hidden layer + LR warm-up (03A804E8, 0 dead) vs per-bucket hidden layer E1C896AE (no warm-up, 17 dead); hidden1 engine both sides

[sn1] gen c15 from A617C0B4 (kb4, gen 12): 1.5M games N=15000

[sn1] FIXED NODES: shared hidden layer + LR warm-up (03A804E8, 0 dead) vs per-bucket hidden layer E1C896AE (no warm-up, 17 dead); same hidden1 engine

[sn1] FIXED NODES: shared hidden layer + LR warm-up (03A804E8, 0 dead) vs per-bucket hidden layer E1C896AE (no warm-up, 17 dead); same hidden1 engine

[sn1] Syzygy 5-man probing (search WDL + root DTZ) vs main without tables; TB adjudication and win adjudication off

[sn1] gen 13 net 43F3DD60 (kb4, c6a..c14 + DFRC d1-d3, 300 SB) vs A617C0B4 (gen 12)

[sn1] DFRC datagen d4 from 85B5CDD6 (steady-state ~12% DFRC share): 600k games N=15000 (genfens dfrc)

[sn1] SPSA #3854 values vs main (gen 12)

[sn1] gen c14 from A617C0B4 (kb4, gen 12): 1.5M games N=15000

[sn1] full SPSA on 85B5CDD6 (kb4 gen 11) after #3826 values: 40k iters x 2 pairs

[sn1] SHARED hidden layer (16, 8 alive) vs main, same data/recipe, FIXED NODES

[sn1] hidden layer (16, per bucket) at 10+0.1 vs main: is it no worse? (fixed nodes +25.4 #3859)

[sn1] hidden layer probe: kb4 FT 1024 -> 16 (per bucket) vs main (no hidden layer), same data/recipe, FIXED NODES