[atlas] Per-bucket hidden layers: every output bucket gets its own L1 and L2 instead of sharing one, at S1000 against a same-seed control.

[atlas] Decay tail 1 on the zero-blend base, testing whether the tail is still worth its old +2.8 on a new base.

[atlas] LTC decider for the zero game-result blend at prod scale, which was +6.1 at STC in test 3766.

[atlas] Zero game-result blend at prod scale S4000 no-tails, against the pre-tails prod net 23C62E38.

[atlas] Zero blend with WIN adjudication off, to size how much conversion gain the default 3-move win-adj hides against test 3759's +5.3.

LTC confirmation of the zero game-result blend, which was +5.3 at STC in test 3759.

Zero game-result blend (wdl 0.00) vs prod blend 0.20, both S1000, dev normalised to the control's eval scale at 103.

Third resumed decay tail on the pawn-pair lineage, testing whether the tail gain stacks a third time or the recipe settles at two stages.

Third resumed decay tail on the pawn-pair lineage, testing whether the tail gain stacks a third time or the recipe settles at two stages.

AdamW at a reduced learning rate versus Ranger at S2000, STC disaster check.

AdamW at the production learning rate versus Ranger at S2000, STC disaster check.

AdamW at a reduced learning rate versus Ranger at S2000, controlling for Lookahead damping Ranger's effective step size.

AdamW vs Ranger at S2000 prod LR, testing faster vs better convergence.

Same S2000 max-score pair with tablebases off.

Try --max-score 32000 (s2000)

Try --max-score 32000 (s2000)