commit	val_bpb	memory_gb	status	description
96dfb80	1.049962	2.9	keep	baseline 5060 consumer path batch8 eval8
7a87b6d	1.122834	2.9	discard	human: hard salience gate on VE using BOS goal-fit minus mean-state redundancy
f3c6755	1.193117	2.9	discard	human: contextual VE gate inputs using BOS context minus sequence-mean redundancy
e7620ce	1.125752	2.9	discard	human: sequence-level contextual controller over attention and MLP residual branches
9a52d53	1.063856	2.9	discard	human: q-head redundancy regularizer coeff 0.01
fa2d7b6	1.062289	2.9	discard	human: q-head redundancy regularizer coeff 0.001
72c3e7d	1.125527	2.9	discard	human: inverted-U novelty curriculum over per-sequence training losses
11c07fd	1.067392	2.9	discard	full attention window pattern L on all layers
cebc334	0.970973	6.2	keep	disable activation checkpointing on RTX 5060 profile
0b1c731	0.946895	6.2	keep	no checkpointing plus WARMDOWN_RATIO 0.3
8a91ebd	0.924877	6.2	keep	no checkpointing plus WARMDOWN_RATIO 0.2
e68fceb	0.921273	6.2	keep	no checkpointing plus WARMDOWN_RATIO 0.1
f69c531	0.924117	6.2	discard	no checkpointing plus constant LR warmdown 0.0
f272afd	0.948483	6.2	discard	human: adaptive LR multiplier from AGI meta layer
f279649	1.053944	6.0	discard	human: anti-salience ablation disabling explicit value-residual pathway
30c79cf	0.932085	6.2	discard	human: bipolar promotive/aversive LR controller
29d1f4f	0.937846	6.2	discard	trunk: MATRIX_LR 0.06
87f0423	0.903525	6.2	discard	human: saliencew optimizer v1 on AdamW parameter groups
f196724	0.867978	6.2	keep	human: saliencew optimizer v2 on AdamW parameter groups
b8d903b	0.871641	6.2	discard	human: first salience gate on Muon matrix path
eff3bbf	0.861187	6.2	keep	human: tuned salience gate on Muon matrix path
adam2m01	0.876374	6.2	discard	human: salience-aware AdamW second-moment rule
valmuon1	0.983434	6.1	discard	human: move value_embeds from AdamW path to Muon
valeadam1	0.871933	6.2	discard	human: value_embeds on plain AdamW with lower LR
scaladam1	0.857426	6.2	keep	human: scalar optimizer back to plain AdamW
wteadam1	0.852087	6.2	keep	human: token embeddings back to plain AdamW
decaysw1	0.877244	6.2	discard	human: salience-controlled decay on saliencew groups
lmheps1	0.876828	6.2	discard	human: lm_head epsilon 1e-8