run_id	started_at	source_commit	family	part	class	status	decision	val_bpb	total_seconds	num_steps	memory_gb	note	overrides_json	log_path
20260311_232737_00000	2026-03-12T04:27:37Z	ace33e7	muon_gate	muon_lr_gate	Gate	ok	discard	0.859849	66.6	75	3.074	muon gate N=0.14 A=0.00 F=0.04 max=1.12	{"AUTORESEARCH_MUON_SALIENCE_GATE_MAX": "1.12", "AUTORESEARCH_MUON_SALIENCE_W_A": "0.00", "AUTORESEARCH_MUON_SALIENCE_W_F": "0.04", "AUTORESEARCH_MUON_SALIENCE_W_N": "0.14"}	C:\ars2\worktrees\alphazero-proxy-lab\proxy_runs\20260311_232737_00000_muon_gate.log
20260311_232927_00001	2026-03-12T04:29:27Z	ace33e7	saliencew_gains	adamw_second_moment	Gate	ok	discard	0.863594	66.6	73	3.074	saliencew gains P=0.24 Q=0.02 Xi=0.02	{"AUTORESEARCH_SALIENCEW_AVERSIVE_GAIN": "0.02", "AUTORESEARCH_SALIENCEW_CONFLICT_GAIN": "0.02", "AUTORESEARCH_SALIENCEW_GATE_MAX": "1.15", "AUTORESEARCH_SALIENCEW_GATE_MIN": "0.98", "AUTORESEARCH_SALIENCEW_PROMOTIVE_GAIN": "0.24"}	C:\ars2\worktrees\alphazero-proxy-lab\proxy_runs\20260311_232927_00001_saliencew_gains.log
20260311_234823_00002	2026-03-12T04:48:23Z	ace33e7	muon_gate	muon_lr_gate	Gate	ok	discard	0.868138	66.7	73	3.074	Exploit the only promising direction so far: more novelty-dominant muon salience gating, lower friction, and a slightly wider gate ceiling.	{"AUTORESEARCH_MUON_SALIENCE_GATE_MAX": "1.16", "AUTORESEARCH_MUON_SALIENCE_W_A": "0.00", "AUTORESEARCH_MUON_SALIENCE_W_F": "0.02", "AUTORESEARCH_MUON_SALIENCE_W_N": "0.18"}	C:\ars2\worktrees\alphazero-proxy-lab\proxy_runs\20260311_234823_00002_muon_gate.log
20260312_000837_00003	2026-03-12T05:08:37Z	ace33e7	muon_gate	muon_lr_gate	Gate	ok	discard	0.865078	68.2	73	3.074	Bake in the only proxy-improving direction so far by reducing novelty gain and increasing fatigue penalty in Muon salience gating.	{}	C:\ars2\worktrees\alphazero-proxy-lab\proxy_runs\20260312_000837_00003_muon_gate.log
20260312_001216_00004	2026-03-12T05:12:16Z	ace33e7	muon_gate	muon_lr_gate	Gate	ok	discard	0.867364	68.5	72	3.074	Revert Muon salience defaults to the only proxy-improving direction observed so far: less novelty amplification and more fatigue penalty.	{}	C:\ars2\worktrees\alphazero-proxy-lab\proxy_runs\20260312_001216_00004_muon_gate.log
20260312_002839_00005	2026-03-12T05:28:39Z	ace33e7	saliencew_tight_bias	adamw_second_moment	Gate	ok	keep	0.854037	66.8	75	3.074	Probe the near-miss saliencew direction with a milder promotive bias and a tighter near-unity gate so second-moment modulation helps without amplifying outlier updates.	{"AUTORESEARCH_SALIENCEW_AVERSIVE_GAIN": "0.00", "AUTORESEARCH_SALIENCEW_CONFLICT_GAIN": "0.01", "AUTORESEARCH_SALIENCEW_GATE_MAX": "1.10", "AUTORESEARCH_SALIENCEW_GATE_MIN": "0.99", "AUTORESEARCH_SALIENCEW_PROMOTIVE_GAIN": "0.20"}	C:\ars2\worktrees\alphazero-proxy-lab\proxy_runs\20260312_002839_00005_saliencew_tight_bias.log
20260312_010048_00006	2026-03-12T06:00:48Z	ace33e7	saliencew_microtight	adamw_second_moment	Gate	ok	discard	0.858506	66.3	74	3.074	Test whether the gain comes from a tiny trust-region nudge by trimming promotive bias and compressing the saliencew gate even closer to unity.	{"AUTORESEARCH_SALIENCEW_AVERSIVE_GAIN": "0.00", "AUTORESEARCH_SALIENCEW_CONFLICT_GAIN": "0.01", "AUTORESEARCH_SALIENCEW_GATE_MAX": "1.08", "AUTORESEARCH_SALIENCEW_GATE_MIN": "0.995", "AUTORESEARCH_SALIENCEW_PROMOTIVE_GAIN": "0.18"}	C:\ars2\worktrees\alphazero-proxy-lab\proxy_runs\20260312_010048_00006_saliencew_microtight.log