Player leaderboard

Every registered players/ module on every board it can play, under one paired protocol: identical held-out cases, identical memoized oracle verdicts on open (per-candidate seeds replayed to every arm — see bench/drivers/policy/policy_baseline.py), identical contraction instances scored by one cost model. ratio is the contraction board's cost/baseline figure; the notes column carries the qualitative flags (drain artifacts, premature stops, unsupported boards) that the numbers alone hide.

• date: 2026-10-20T00:00:14+00:00 · git: 2d6e5e8 (dirty) · torch 2.14.0+cu130 on cpu · seed 0
• players: all-registered
• meta_cases: 16
• meta_budget: 12
• meta_held_out: gen_cases(seed+7919)
• open_pages: 36
• open_page: 16
• open_budget: 17
• open_split: odd pages eval / even pages train (neural)
• oracle_pairing: memoized per-candidate-seeded verdicts
• neural_epochs: 300
• neural_hidden: 32
• contraction_scales: [8, 12]
• contraction_instances: 4
• contraction_samples: 32
• meta: {'cases': 16, 'budget': 12}
• open: {'pool': 576, 'useful': 7, 'random_ap': 0.0122, 'train_pairs': 576}
• contraction: {'scales': [8, 12], 'instances': 4}

meta

player	kind	reward/ep	evals/ep	useful/eval	rank AP	cost	ratio	wall s	notes
greedy	scripted	-0.6000	12.00	—	—	—	—	0.12	drains every episode's move budget
linear	linear	0.1166	1.00	—	—	2.50	—	0.05	short lines — extracts early; certified extraction on 16/16
linear-cold	linear	-0.4188	10.25	—	—	2.43	—	0.31	certified extraction on 7/16
linear-greedy	linear	0.1166	1.00	—	—	2.50	—	0.05	short lines — extracts early; certified extraction on 16/16
playbook	scripted	-0.1000	2.00	—	—	2.94	—	0.03	short lines — extracts early; certified extraction on 16/16
random	chance	-0.5324	11.12	—	—	3.00	—	0.10	certified extraction on 4/16

open

player	kind	reward/ep	evals/ep	useful/eval	rank AP	cost	ratio	wall s	notes
linear	linear	0.2852	13.72	0.0121	0.1284	—	—	1.14	—
linear-cold	linear	0.1699	9.31	0.0119	0.0123	—	—	1.14	—
linear-greedy	linear	0.3590	16.00	0.0104	0.1284	—	—	1.25	drains every page — no stopping learned
neural	neural	0.3089	16.00	0.0104	0.0412	—	—	1.33	drains every page — no stopping learned
playbook	scripted	0.3595	16.00	0.0104	0.0450	—	—	1.26	drains every page — no stopping learned
random	chance	0.1339	8.67	0.0096	0.0122	—	—	0.97	—

contraction

player	kind	reward/ep	evals/ep	useful/eval	rank AP	cost	ratio	wall s	notes
oe-greedy	external	—	—	—	—	28,164.00	1.000	0.00	0 beats / 8 ties / 0 loses vs oe-greedy
oe-rand-greedy	external	—	—	—	—	27,591.80	0.885	0.00	4 beats / 3 ties / 1 loses vs oe-greedy
distilled	distilled	—	—	—	—	27,775.50	0.930	0.03	5 beats / 1 ties / 2 loses vs oe-greedy
distilled-argmax	distilled	—	—	—	—	28,708.80	1.125	0.00	3 beats / 0 ties / 5 loses vs oe-greedy
greedy	scripted	—	—	—	—	86,419.50	5.199	0.00	0 beats / 0 ties / 8 loses vs oe-greedy
random	chance	—	—	—	—	26,847,963.80	2872.670	0.00	0 beats / 0 ties / 8 loses vs oe-greedy