robbiethompson2018 commited on
Commit
d41fad7
·
verified ·
1 Parent(s): 1038628

Add 590_control-untied-198_s1

Browse files
runs/590_control-untied-198_s1/eval_dev.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
runs/590_control-untied-198_s1/final.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eb19f603f8ddf9fb021a2e296f7e0ff9be3d336afc3ae3f19eb41b559d985be5
3
+ size 38268575
runs/590_control-untied-198_s1/found_targets.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
runs/590_control-untied-198_s1/lean_gate.disagree.jsonl ADDED
@@ -0,0 +1 @@
 
 
1
+ {"prompt": "THM R , ( R & R ) SEQ ( ( S & S ) > ( ( ( P v S ) v R ) & ( ( R & R ) v ( P v ( ~ Q ) ) ) ) ) PRF", "lean_text": "have n59 : R := h1 ; have n60 : ( R ∧ R ) := ⟨ n59 , n59 ⟩ ; have n61 : ( ( P ∨ S ) ∨ R ) := Or.inr n59 ; have n62 : ( ( R ∧ R ) ∨ ( P ∨ ( ¬ Q ) ) ) := Or.inl n60 ; have n63 : ( ( S ∧ S ) → ( ( ( P ∨ S ) ∨ R ) ∧ ( ( R ∧ R ) ∨ ( P ∨ ( ¬ Q ) ) ) ) ) := ( fun ( n64 : ( S ∧ S ) ) => by have n61 : ( ( ( P ∨ S ) ∨ R ) ∧ ( ( R ∧ R ) ∨ ( P ∨ ( ¬ Q ) ) ) ) := ⟨ n61 , n62 ⟩ ; exact n61 ) ; exact n63", "nd": "N1 R : PR ; N2 ( R & R ) : ANDI N1 N1 ; N3 ( ( P v S ) v R ) : ORI2 N1 ; N4 ( ( R & R ) v ( P v ( ~ Q ) ) ) : ORI1 N2 ; N5 | ( S & S ) : AS ; N6 | ( ( ( P v S ) v R ) & ( ( R & R ) v ( P v ( ~ Q ) ) ) ) : ANDI N3 N4 ; N7 ( ( S & S ) > ( ( ( P v S ) v R ) & ( ( R & R ) v ( P v ( ~ Q ) ) ) ) ) : IMPI N5 N6 ; QED", "nd_ok": false, "lean_ok": true}
runs/590_control-untied-198_s1/lean_gate.jsonl ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"utc": "2026-10-07T22:11:41Z", "samples": 2000, "parse_fail": 9, "distinct_checked": 1991, "both_ok": 1949, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 42, "parse_reasons": {"unbound": 6, "formula": 1, "no-eos": 1, "expected )": 1}, "lean_wall_s": 3.7846169471740723, "lean_proc_s": 18.49671769142151, "nd_verify_s": 0.09245419502258301, "workers": 96, "chunk": 400}
2
+ {"utc": "2026-10-07T22:12:38Z", "samples": 12000, "parse_fail": 6386, "distinct_checked": 5463, "both_ok": 1188, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 4275, "parse_reasons": {"unbound": 4002, "expected )": 1677, "exact not last": 170, "formula": 153, "expected (": 127, "name": 92, "expected ;": 55, "expected :=": 49, "no-eos": 49, "ORE box": 8, "PR position": 3, "expected None": 1}, "lean_wall_s": 22.64402461051941, "lean_proc_s": 310.9076271057129, "nd_verify_s": 0.27167320251464844, "workers": 96, "chunk": 400}
3
+ {"utc": "2026-10-07T22:13:22Z", "samples": 9480, "parse_fail": 5546, "distinct_checked": 3579, "both_ok": 74, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3505, "parse_reasons": {"unbound": 3555, "expected )": 1296, "formula": 170, "exact not last": 146, "expected (": 90, "no-eos": 85, "name": 76, "expected ;": 67, "expected :=": 54, "PR position": 4, "expected exact": 1, "expected =>": 1, "ORE box": 1}, "lean_wall_s": 14.96888279914856, "lean_proc_s": 133.16824507713318, "nd_verify_s": 0.1959824562072754, "workers": 96, "chunk": 400}
4
+ {"utc": "2026-10-07T22:14:07Z", "samples": 9456, "parse_fail": 5399, "distinct_checked": 3743, "both_ok": 82, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3661, "parse_reasons": {"unbound": 3576, "expected )": 1275, "formula": 141, "exact not last": 133, "expected (": 107, "name": 70, "expected ;": 33, "expected :=": 29, "no-eos": 19, "PR position": 9, "ORE box": 5, "expected :": 1, "expected None": 1}, "lean_wall_s": 16.008084535598755, "lean_proc_s": 147.26593375205994, "nd_verify_s": 0.18413877487182617, "workers": 96, "chunk": 400}
5
+ {"utc": "2026-10-07T22:14:55Z", "samples": 9456, "parse_fail": 5413, "distinct_checked": 3673, "both_ok": 85, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3588, "parse_reasons": {"unbound": 3582, "expected )": 1210, "expected (": 144, "formula": 113, "exact not last": 109, "name": 86, "expected :=": 76, "expected ;": 50, "no-eos": 35, "ORE box": 5, "PR position": 2, "expected =>": 1}, "lean_wall_s": 15.089941263198853, "lean_proc_s": 138.87708401679993, "nd_verify_s": 0.1943974494934082, "workers": 96, "chunk": 400}
6
+ {"utc": "2026-10-07T22:16:40Z", "samples": 12000, "parse_fail": 1685, "distinct_checked": 9916, "both_ok": 5673, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 1, "both_rej": 4242, "parse_reasons": {"unbound": 1217, "formula": 111, "name": 83, "expected ;": 76, "expected )": 62, "exact not last": 40, "no-eos": 37, "expected (": 29, "ORE box": 24, "expected :=": 3, "expected =>": 2, "PR position": 1}, "lean_wall_s": 34.293299436569214, "lean_proc_s": 841.31969165802, "nd_verify_s": 0.5830521583557129, "workers": 96, "chunk": 400}
7
+ {"utc": "2026-10-07T22:17:13Z", "samples": 5016, "parse_fail": 1282, "distinct_checked": 3382, "both_ok": 43, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3339, "parse_reasons": {"unbound": 972, "formula": 90, "name": 81, "expected ;": 34, "expected )": 32, "no-eos": 29, "expected (": 24, "exact not last": 15, "ORE box": 5}, "lean_wall_s": 14.278998136520386, "lean_proc_s": 120.56367683410645, "nd_verify_s": 0.20432186126708984, "workers": 96, "chunk": 400}
8
+ {"utc": "2026-10-07T22:17:45Z", "samples": 5016, "parse_fail": 1202, "distinct_checked": 3464, "both_ok": 53, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3411, "parse_reasons": {"unbound": 894, "formula": 71, "expected ;": 62, "name": 45, "expected )": 43, "no-eos": 37, "expected (": 36, "exact not last": 10, "expected =>": 2, "ORE box": 1, "PR position": 1}, "lean_wall_s": 16.411970853805542, "lean_proc_s": 143.9047713279724, "nd_verify_s": 0.17322373390197754, "workers": 96, "chunk": 400}
9
+ {"utc": "2026-10-07T22:18:16Z", "samples": 5016, "parse_fail": 1427, "distinct_checked": 3263, "both_ok": 34, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3229, "parse_reasons": {"unbound": 1019, "expected ;": 89, "name": 84, "formula": 78, "no-eos": 49, "expected )": 41, "ORE box": 34, "exact not last": 25, "expected (": 5, "PR position": 2, "expected :=": 1}, "lean_wall_s": 14.564558982849121, "lean_proc_s": 118.92972898483276, "nd_verify_s": 0.16830110549926758, "workers": 96, "chunk": 400}
10
+ {"utc": "2026-10-07T22:19:57Z", "samples": 12000, "parse_fail": 980, "distinct_checked": 10563, "both_ok": 7256, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3307, "parse_reasons": {"unbound": 719, "expected ;": 51, "expected (": 51, "exact not last": 34, "name": 26, "formula": 25, "ORE box": 22, "expected )": 19, "no-eos": 16, "PR position": 10, "expected :=": 6, "expected hh": 1}, "lean_wall_s": 34.841187953948975, "lean_proc_s": 908.3169329166412, "nd_verify_s": 0.632037878036499, "workers": 96, "chunk": 400}
11
+ {"utc": "2026-10-07T22:20:22Z", "samples": 3672, "parse_fail": 880, "distinct_checked": 2537, "both_ok": 29, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2508, "parse_reasons": {"unbound": 603, "formula": 67, "expected ;": 62, "expected (": 57, "no-eos": 35, "name": 16, "expected )": 14, "exact not last": 14, "ORE box": 8, "expected :=": 3, "PR position": 1}, "lean_wall_s": 11.211755514144897, "lean_proc_s": 71.61623406410217, "nd_verify_s": 0.13564229011535645, "workers": 96, "chunk": 400}
12
+ {"utc": "2026-10-07T22:20:44Z", "samples": 3672, "parse_fail": 704, "distinct_checked": 2693, "both_ok": 10, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2683, "parse_reasons": {"unbound": 510, "expected (": 42, "expected ;": 40, "PR position": 26, "ORE box": 26, "expected )": 17, "exact not last": 17, "name": 13, "formula": 10, "no-eos": 2, "expected exact": 1}, "lean_wall_s": 12.421874046325684, "lean_proc_s": 84.37383723258972, "nd_verify_s": 0.14615559577941895, "workers": 96, "chunk": 400}
13
+ {"utc": "2026-10-07T22:21:10Z", "samples": 3648, "parse_fail": 722, "distinct_checked": 2648, "both_ok": 10, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2638, "parse_reasons": {"unbound": 557, "expected ;": 44, "expected (": 33, "name": 25, "expected )": 18, "ORE box": 17, "no-eos": 9, "exact not last": 9, "formula": 6, "PR position": 3, "expected exact": 1}, "lean_wall_s": 12.177467823028564, "lean_proc_s": 81.78430318832397, "nd_verify_s": 0.1413555145263672, "workers": 96, "chunk": 400}
14
+ {"utc": "2026-10-07T22:22:49Z", "samples": 12000, "parse_fail": 1102, "distinct_checked": 10440, "both_ok": 7919, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2521, "parse_reasons": {"unbound": 767, "expected ;": 79, "name": 77, "formula": 71, "expected )": 35, "expected (": 29, "exact not last": 18, "ORE box": 17, "no-eos": 7, "trailing tokens": 2}, "lean_wall_s": 33.06440043449402, "lean_proc_s": 854.4419257640839, "nd_verify_s": 0.6352345943450928, "workers": 96, "chunk": 400}
15
+ {"utc": "2026-10-07T22:23:08Z", "samples": 3144, "parse_fail": 1046, "distinct_checked": 1940, "both_ok": 9, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1931, "parse_reasons": {"unbound": 766, "expected ;": 89, "name": 58, "formula": 42, "expected )": 30, "expected (": 27, "ORE box": 12, "no-eos": 12, "exact not last": 8, "expected fun": 1, "PR position": 1}, "lean_wall_s": 7.214571475982666, "lean_proc_s": 35.48117995262146, "nd_verify_s": 0.09534955024719238, "workers": 96, "chunk": 400}
16
+ {"utc": "2026-10-07T22:23:27Z", "samples": 3144, "parse_fail": 1028, "distinct_checked": 1937, "both_ok": 15, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1922, "parse_reasons": {"unbound": 689, "name": 119, "formula": 88, "expected ;": 58, "expected )": 22, "expected (": 22, "exact not last": 14, "ORE box": 10, "PR position": 3, "no-eos": 3}, "lean_wall_s": 6.698266983032227, "lean_proc_s": 32.51301193237305, "nd_verify_s": 0.09705352783203125, "workers": 96, "chunk": 400}
17
+ {"utc": "2026-10-07T22:23:43Z", "samples": 3120, "parse_fail": 901, "distinct_checked": 2042, "both_ok": 22, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2020, "parse_reasons": {"unbound": 678, "expected ;": 60, "formula": 54, "ORE box": 26, "expected (": 25, "expected )": 22, "name": 21, "exact not last": 12, "PR position": 1, "expected None": 1, "expected :=": 1}, "lean_wall_s": 8.159974813461304, "lean_proc_s": 40.952595710754395, "nd_verify_s": 0.10168576240539551, "workers": 96, "chunk": 400}
18
+ {"utc": "2026-10-07T22:25:30Z", "samples": 12000, "parse_fail": 975, "distinct_checked": 10553, "both_ok": 8329, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2224, "parse_reasons": {"unbound": 752, "expected (": 54, "name": 49, "formula": 28, "expected )": 21, "expected ;": 16, "no-eos": 16, "exact not last": 15, "expected :=": 10, "ORE box": 7, "PR position": 6, "expected None": 1}, "lean_wall_s": 33.070029497146606, "lean_proc_s": 865.68732380867, "nd_verify_s": 0.6911237239837646, "workers": 96, "chunk": 400}
19
+ {"utc": "2026-10-07T22:25:50Z", "samples": 2784, "parse_fail": 859, "distinct_checked": 1791, "both_ok": 6, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1785, "parse_reasons": {"unbound": 598, "expected (": 59, "expected ;": 56, "name": 54, "expected )": 24, "formula": 23, "expected :=": 14, "exact not last": 12, "no-eos": 9, "PR position": 4, "ORE box": 4, "expected exact": 1, "trailing tokens": 1}, "lean_wall_s": 6.966033458709717, "lean_proc_s": 32.895395278930664, "nd_verify_s": 0.12120461463928223, "workers": 96, "chunk": 400}
20
+ {"utc": "2026-10-07T22:26:08Z", "samples": 2784, "parse_fail": 891, "distinct_checked": 1741, "both_ok": 7, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1734, "parse_reasons": {"unbound": 668, "expected (": 75, "name": 41, "formula": 34, "exact not last": 17, "ORE box": 15, "expected )": 13, "no-eos": 11, "expected ;": 11, "PR position": 6}, "lean_wall_s": 7.67464017868042, "lean_proc_s": 33.02237796783447, "nd_verify_s": 0.10727047920227051, "workers": 96, "chunk": 400}
21
+ {"utc": "2026-10-07T22:26:25Z", "samples": 2760, "parse_fail": 791, "distinct_checked": 1827, "both_ok": 4, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1823, "parse_reasons": {"unbound": 661, "expected (": 33, "name": 29, "formula": 22, "expected )": 13, "exact not last": 9, "expected :=": 8, "expected ;": 6, "PR position": 5, "ORE box": 2, "no-eos": 2, "eof": 1}, "lean_wall_s": 7.6069958209991455, "lean_proc_s": 35.87954020500183, "nd_verify_s": 0.09193825721740723, "workers": 96, "chunk": 400}
22
+ {"utc": "2026-10-07T22:28:01Z", "samples": 12000, "parse_fail": 816, "distinct_checked": 10707, "both_ok": 8624, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2083, "parse_reasons": {"unbound": 602, "name": 43, "expected (": 36, "expected ;": 25, "formula": 24, "expected )": 22, "exact not last": 22, "ORE box": 16, "no-eos": 15, "expected :=": 6, "trailing tokens": 3, "expected False": 1, "PR position": 1}, "lean_wall_s": 32.36218070983887, "lean_proc_s": 851.4265160560608, "nd_verify_s": 0.6679849624633789, "workers": 96, "chunk": 400}
23
+ {"utc": "2026-10-07T22:28:18Z", "samples": 2544, "parse_fail": 654, "distinct_checked": 1753, "both_ok": 2, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1751, "parse_reasons": {"unbound": 485, "name": 53, "expected (": 26, "no-eos": 24, "expected ;": 22, "formula": 19, "exact not last": 14, "expected )": 10, "PR position": 1}, "lean_wall_s": 7.045920133590698, "lean_proc_s": 31.59938144683838, "nd_verify_s": 0.08098673820495605, "workers": 96, "chunk": 400}
24
+ {"utc": "2026-10-07T22:28:29Z", "samples": 2544, "parse_fail": 702, "distinct_checked": 1662, "both_ok": 3, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1659, "parse_reasons": {"unbound": 512, "expected (": 66, "exact not last": 31, "name": 27, "formula": 16, "expected )": 15, "ORE box": 15, "expected ;": 14, "trailing tokens": 5, "expected False": 1}, "lean_wall_s": 5.116694688796997, "lean_proc_s": 21.532093286514282, "nd_verify_s": 0.09893298149108887, "workers": 96, "chunk": 400}
25
+ {"utc": "2026-10-07T22:28:46Z", "samples": 2520, "parse_fail": 701, "distinct_checked": 1683, "both_ok": 5, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1678, "parse_reasons": {"unbound": 467, "expected (": 47, "expected ;": 44, "name": 43, "formula": 32, "ORE box": 20, "no-eos": 18, "expected )": 17, "exact not last": 5, "trailing tokens": 4, "PR position": 3, "expected :=": 1}, "lean_wall_s": 7.386868953704834, "lean_proc_s": 31.80172634124756, "nd_verify_s": 0.12150788307189941, "workers": 96, "chunk": 400}
26
+ {"utc": "2026-10-07T22:30:28Z", "samples": 12000, "parse_fail": 956, "distinct_checked": 10578, "both_ok": 8862, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1716, "parse_reasons": {"unbound": 725, "name": 69, "expected )": 44, "expected (": 32, "exact not last": 28, "formula": 22, "no-eos": 15, "expected :=": 7, "ORE box": 6, "expected ;": 5, "PR position": 3}, "lean_wall_s": 34.63088774681091, "lean_proc_s": 908.6563501358032, "nd_verify_s": 0.6407496929168701, "workers": 96, "chunk": 400}
27
+ {"utc": "2026-10-07T22:30:43Z", "samples": 2280, "parse_fail": 837, "distinct_checked": 1327, "both_ok": 1, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1326, "parse_reasons": {"unbound": 617, "name": 65, "expected )": 53, "expected (": 34, "formula": 21, "expected :=": 15, "expected ;": 13, "exact not last": 9, "no-eos": 8, "ORE box": 2}, "lean_wall_s": 5.398378133773804, "lean_proc_s": 17.884498357772827, "nd_verify_s": 0.057172536849975586, "workers": 96, "chunk": 400}
28
+ {"utc": "2026-10-07T22:30:57Z", "samples": 2280, "parse_fail": 778, "distinct_checked": 1393, "both_ok": 2, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1391, "parse_reasons": {"unbound": 569, "name": 92, "expected )": 28, "formula": 28, "no-eos": 28, "expected (": 20, "expected ;": 8, "ORE box": 4, "expected :=": 1}, "lean_wall_s": 5.7022294998168945, "lean_proc_s": 21.550528526306152, "nd_verify_s": 0.07017993927001953, "workers": 96, "chunk": 400}
29
+ {"utc": "2026-10-07T22:31:07Z", "samples": 2256, "parse_fail": 814, "distinct_checked": 1350, "both_ok": 3, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1347, "parse_reasons": {"unbound": 691, "name": 34, "expected (": 32, "expected )": 21, "ORE box": 19, "exact not last": 11, "expected ;": 6}, "lean_wall_s": 5.2906482219696045, "lean_proc_s": 18.06407403945923, "nd_verify_s": 0.06550383567810059, "workers": 96, "chunk": 400}
30
+ {"utc": "2026-10-07T22:32:51Z", "samples": 12000, "parse_fail": 970, "distinct_checked": 10586, "both_ok": 9104, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1482, "parse_reasons": {"unbound": 675, "name": 100, "expected (": 67, "formula": 35, "ORE box": 23, "expected )": 23, "expected ;": 18, "exact not last": 11, "expected :=": 7, "PR position": 6, "no-eos": 5}, "lean_wall_s": 34.89832592010498, "lean_proc_s": 910.1571400165558, "nd_verify_s": 0.6665403842926025, "workers": 96, "chunk": 400}
31
+ {"utc": "2026-10-07T22:33:06Z", "samples": 2136, "parse_fail": 763, "distinct_checked": 1267, "both_ok": 7, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1260, "parse_reasons": {"unbound": 529, "name": 67, "expected (": 60, "formula": 28, "ORE box": 27, "expected ;": 25, "expected )": 16, "no-eos": 5, "PR position": 3, "exact not last": 2, "expected exact": 1}, "lean_wall_s": 5.313301086425781, "lean_proc_s": 16.961365222930908, "nd_verify_s": 0.06611013412475586, "workers": 96, "chunk": 400}
32
+ {"utc": "2026-10-07T22:33:21Z", "samples": 2136, "parse_fail": 830, "distinct_checked": 1215, "both_ok": 0, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1215, "parse_reasons": {"unbound": 597, "name": 83, "formula": 38, "expected (": 34, "expected )": 34, "ORE box": 19, "expected ;": 17, "exact not last": 5, "no-eos": 2, "PR position": 1}, "lean_wall_s": 5.087433099746704, "lean_proc_s": 15.550572872161865, "nd_verify_s": 0.06650733947753906, "workers": 96, "chunk": 400}
33
+ {"utc": "2026-10-07T22:33:34Z", "samples": 2136, "parse_fail": 931, "distinct_checked": 1143, "both_ok": 2, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1141, "parse_reasons": {"unbound": 609, "name": 132, "expected (": 73, "expected )": 35, "ORE box": 23, "formula": 20, "expected :=": 20, "expected ;": 16, "no-eos": 1, "PR position": 1, "exact not last": 1}, "lean_wall_s": 4.862018346786499, "lean_proc_s": 14.243491172790527, "nd_verify_s": 0.05267977714538574, "workers": 96, "chunk": 400}
34
+ {"utc": "2026-10-07T22:35:17Z", "samples": 12000, "parse_fail": 887, "distinct_checked": 10713, "both_ok": 9227, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1486, "parse_reasons": {"unbound": 660, "name": 76, "expected (": 56, "formula": 34, "exact not last": 13, "ORE box": 11, "expected )": 10, "expected ;": 8, "no-eos": 8, "expected :=": 7, "PR position": 2, "trailing tokens": 1, "expected False": 1}, "lean_wall_s": 32.83972358703613, "lean_proc_s": 871.3997395038605, "nd_verify_s": 0.6967160701751709, "workers": 96, "chunk": 400}
35
+ {"utc": "2026-10-07T22:35:32Z", "samples": 2064, "parse_fail": 728, "distinct_checked": 1226, "both_ok": 7, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1219, "parse_reasons": {"unbound": 535, "name": 61, "expected (": 41, "formula": 34, "expected )": 22, "ORE box": 15, "expected ;": 7, "no-eos": 7, "exact not last": 3, "PR position": 2, "expected None": 1}, "lean_wall_s": 4.966235637664795, "lean_proc_s": 15.183523893356323, "nd_verify_s": 0.07897186279296875, "workers": 96, "chunk": 400}
36
+ {"utc": "2026-10-07T22:35:46Z", "samples": 2040, "parse_fail": 748, "distinct_checked": 1177, "both_ok": 1, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1176, "parse_reasons": {"unbound": 557, "expected (": 59, "name": 55, "formula": 24, "no-eos": 17, "expected ;": 15, "expected )": 11, "ORE box": 4, "PR position": 3, "expected :=": 2, "exact not last": 1}, "lean_wall_s": 4.7945780754089355, "lean_proc_s": 13.842448234558105, "nd_verify_s": 0.05059075355529785, "workers": 96, "chunk": 400}
37
+ {"utc": "2026-10-07T22:35:57Z", "samples": 2040, "parse_fail": 726, "distinct_checked": 1218, "both_ok": 9, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1209, "parse_reasons": {"unbound": 562, "name": 81, "expected (": 43, "formula": 12, "expected ;": 7, "expected )": 6, "no-eos": 5, "expected :=": 4, "PR position": 2, "exact not last": 2, "ORE box": 2}, "lean_wall_s": 4.85127067565918, "lean_proc_s": 14.763352632522583, "nd_verify_s": 0.07502245903015137, "workers": 96, "chunk": 400}
38
+ {"utc": "2026-10-07T22:37:43Z", "samples": 12000, "parse_fail": 890, "distinct_checked": 10647, "both_ok": 9402, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1245, "parse_reasons": {"unbound": 727, "name": 52, "expected (": 38, "ORE box": 20, "formula": 19, "expected ;": 14, "expected )": 10, "no-eos": 3, "exact not last": 3, "PR position": 2, "expected None": 1, "expected :=": 1}, "lean_wall_s": 36.66239547729492, "lean_proc_s": 968.2424907684326, "nd_verify_s": 0.6724472045898438, "workers": 96, "chunk": 400}
39
+ {"utc": "2026-10-07T22:37:55Z", "samples": 1872, "parse_fail": 798, "distinct_checked": 969, "both_ok": 2, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 967, "parse_reasons": {"unbound": 602, "name": 63, "formula": 42, "expected (": 40, "expected )": 32, "ORE box": 11, "no-eos": 4, "expected ;": 3, "exact not last": 1}, "lean_wall_s": 3.6425814628601074, "lean_proc_s": 9.59370493888855, "nd_verify_s": 0.05515122413635254, "workers": 96, "chunk": 400}
40
+ {"utc": "2026-10-07T22:38:05Z", "samples": 1872, "parse_fail": 795, "distinct_checked": 992, "both_ok": 3, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 989, "parse_reasons": {"unbound": 651, "name": 61, "expected (": 38, "expected ;": 25, "ORE box": 13, "formula": 2, "expected )": 1, "expected None": 1, "exact not last": 1, "trailing tokens": 1, "no-eos": 1}, "lean_wall_s": 4.058251857757568, "lean_proc_s": 11.181692600250244, "nd_verify_s": 0.04317212104797363, "workers": 96, "chunk": 400}
41
+ {"utc": "2026-10-07T22:38:16Z", "samples": 1848, "parse_fail": 717, "distinct_checked": 1050, "both_ok": 4, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1046, "parse_reasons": {"unbound": 624, "name": 33, "expected (": 21, "expected ;": 20, "formula": 9, "expected )": 5, "ORE box": 4, "exact not last": 1}, "lean_wall_s": 4.254538536071777, "lean_proc_s": 12.424631118774414, "nd_verify_s": 0.06738495826721191, "workers": 96, "chunk": 400}
42
+ {"utc": "2026-10-07T22:40:01Z", "samples": 12000, "parse_fail": 824, "distinct_checked": 10793, "both_ok": 9562, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1231, "parse_reasons": {"unbound": 646, "name": 47, "expected (": 43, "formula": 36, "expected ;": 20, "ORE box": 11, "expected )": 11, "exact not last": 4, "no-eos": 3, "expected False": 1, "PR position": 1, "expected :": 1}, "lean_wall_s": 36.03303837776184, "lean_proc_s": 962.2108108997345, "nd_verify_s": 0.6672346591949463, "workers": 96, "chunk": 400}
43
+ {"utc": "2026-10-07T22:40:12Z", "samples": 1800, "parse_fail": 565, "distinct_checked": 1109, "both_ok": 4, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1105, "parse_reasons": {"unbound": 437, "expected (": 77, "name": 23, "expected ;": 14, "formula": 6, "ORE box": 4, "expected )": 2, "expected False": 1, "no-eos": 1}, "lean_wall_s": 4.521115779876709, "lean_proc_s": 12.517682075500488, "nd_verify_s": 0.0600130558013916, "workers": 96, "chunk": 400}
44
+ {"utc": "2026-10-07T22:40:21Z", "samples": 1800, "parse_fail": 827, "distinct_checked": 896, "both_ok": 2, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 894, "parse_reasons": {"unbound": 610, "expected (": 63, "name": 44, "formula": 35, "ORE box": 34, "expected ;": 24, "expected )": 12, "exact not last": 2, "no-eos": 2, "expected False": 1}, "lean_wall_s": 3.5676140785217285, "lean_proc_s": 8.320209980010986, "nd_verify_s": 0.05598735809326172, "workers": 96, "chunk": 400}
45
+ {"utc": "2026-10-07T22:40:31Z", "samples": 1800, "parse_fail": 704, "distinct_checked": 997, "both_ok": 2, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 995, "parse_reasons": {"unbound": 526, "name": 75, "formula": 39, "expected (": 26, "expected )": 25, "expected ;": 7, "exact not last": 3, "ORE box": 2, "PR position": 1}, "lean_wall_s": 3.671114683151245, "lean_proc_s": 9.598406076431274, "nd_verify_s": 0.07662010192871094, "workers": 96, "chunk": 400}
46
+ {"utc": "2026-10-07T22:42:16Z", "samples": 12000, "parse_fail": 895, "distinct_checked": 10683, "both_ok": 9512, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1171, "parse_reasons": {"unbound": 635, "name": 67, "expected (": 65, "formula": 39, "expected )": 26, "exact not last": 24, "expected ;": 16, "ORE box": 15, "no-eos": 5, "PR position": 2, "expected exact": 1}, "lean_wall_s": 33.97800803184509, "lean_proc_s": 897.8315439224243, "nd_verify_s": 0.691852331161499, "workers": 96, "chunk": 400}
47
+ {"utc": "2026-10-07T22:42:26Z", "samples": 1752, "parse_fail": 779, "distinct_checked": 892, "both_ok": 1, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 891, "parse_reasons": {"unbound": 530, "expected (": 73, "formula": 56, "name": 45, "expected )": 27, "expected ;": 26, "no-eos": 10, "ORE box": 6, "exact not last": 2, "expected fun": 2, "PR position": 1, "expected :=": 1}, "lean_wall_s": 3.3323307037353516, "lean_proc_s": 7.667317628860474, "nd_verify_s": 0.04273104667663574, "workers": 96, "chunk": 400}
48
+ {"utc": "2026-10-07T22:42:36Z", "samples": 1728, "parse_fail": 654, "distinct_checked": 987, "both_ok": 2, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 985, "parse_reasons": {"unbound": 437, "name": 93, "expected (": 68, "ORE box": 19, "expected )": 15, "expected ;": 12, "formula": 9, "exact not last": 1}, "lean_wall_s": 4.053038835525513, "lean_proc_s": 10.570454120635986, "nd_verify_s": 0.04715394973754883, "workers": 96, "chunk": 400}
49
+ {"utc": "2026-10-07T22:42:47Z", "samples": 1728, "parse_fail": 747, "distinct_checked": 920, "both_ok": 2, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 918, "parse_reasons": {"unbound": 478, "expected (": 74, "name": 53, "expected )": 53, "formula": 43, "expected ;": 24, "ORE box": 13, "no-eos": 3, "exact not last": 2, "trailing tokens": 2, "expected None": 1, "PR position": 1}, "lean_wall_s": 3.8213605880737305, "lean_proc_s": 9.066309928894043, "nd_verify_s": 0.03421187400817871, "workers": 96, "chunk": 400}
50
+ {"utc": "2026-10-07T22:44:38Z", "samples": 12000, "parse_fail": 954, "distinct_checked": 10612, "both_ok": 9575, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1037, "parse_reasons": {"unbound": 615, "name": 138, "expected ;": 67, "expected (": 60, "formula": 32, "expected )": 13, "no-eos": 13, "exact not last": 9, "ORE box": 6, "trailing tokens": 1}, "lean_wall_s": 35.72919321060181, "lean_proc_s": 939.3376688957214, "nd_verify_s": 0.6853852272033691, "workers": 96, "chunk": 400}
51
+ {"utc": "2026-10-07T22:44:49Z", "samples": 1704, "parse_fail": 813, "distinct_checked": 824, "both_ok": 1, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 823, "parse_reasons": {"unbound": 515, "name": 128, "expected (": 63, "expected ;": 57, "formula": 33, "expected )": 11, "ORE box": 3, "no-eos": 2, "exact not last": 1}, "lean_wall_s": 3.5260653495788574, "lean_proc_s": 7.358459711074829, "nd_verify_s": 0.04607248306274414, "workers": 96, "chunk": 400}
52
+ {"utc": "2026-10-07T22:44:59Z", "samples": 1680, "parse_fail": 911, "distinct_checked": 699, "both_ok": 0, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 699, "parse_reasons": {"unbound": 671, "name": 82, "expected ;": 77, "expected (": 45, "expected )": 22, "ORE box": 5, "no-eos": 4, "formula": 4, "trailing tokens": 1}, "lean_wall_s": 3.0931832790374756, "lean_proc_s": 5.860424995422363, "nd_verify_s": 0.026239395141601562, "workers": 96, "chunk": 400}
53
+ {"utc": "2026-10-07T22:45:10Z", "samples": 1680, "parse_fail": 740, "distinct_checked": 858, "both_ok": 1, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 857, "parse_reasons": {"unbound": 454, "name": 102, "expected ;": 57, "expected (": 56, "formula": 39, "expected )": 10, "exact not last": 6, "ORE box": 5, "no-eos": 3, "expected None": 3, "trailing tokens": 3, "expected =>": 1, "expected :=": 1}, "lean_wall_s": 3.0981385707855225, "lean_proc_s": 6.704622507095337, "nd_verify_s": 0.054720401763916016, "workers": 96, "chunk": 400}
54
+ {"utc": "2026-10-07T22:46:56Z", "samples": 12000, "parse_fail": 862, "distinct_checked": 10721, "both_ok": 9650, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1071, "parse_reasons": {"unbound": 581, "name": 106, "expected (": 55, "formula": 43, "expected )": 28, "expected ;": 26, "exact not last": 9, "no-eos": 6, "ORE box": 5, "expected None": 1, "expected hh": 1, "expected fun": 1}, "lean_wall_s": 33.875598192214966, "lean_proc_s": 895.7266855239868, "nd_verify_s": 0.6469309329986572, "workers": 96, "chunk": 400}
55
+ {"utc": "2026-10-07T22:47:07Z", "samples": 1680, "parse_fail": 693, "distinct_checked": 910, "both_ok": 0, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 910, "parse_reasons": {"unbound": 450, "name": 131, "expected (": 55, "expected )": 28, "expected ;": 7, "formula": 7, "exact not last": 4, "ORE box": 4, "no-eos": 4, "expected fun": 2, "expected exact": 1}, "lean_wall_s": 3.9257493019104004, "lean_proc_s": 9.439648628234863, "nd_verify_s": 0.04067659378051758, "workers": 96, "chunk": 400}
56
+ {"utc": "2026-10-07T22:47:16Z", "samples": 1680, "parse_fail": 777, "distinct_checked": 826, "both_ok": 0, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 826, "parse_reasons": {"unbound": 586, "expected ;": 53, "name": 48, "formula": 41, "expected )": 23, "expected (": 19, "exact not last": 4, "ORE box": 3}, "lean_wall_s": 3.2405476570129395, "lean_proc_s": 7.005871772766113, "nd_verify_s": 0.03750872611999512, "workers": 96, "chunk": 400}
57
+ {"utc": "2026-10-07T22:47:27Z", "samples": 1656, "parse_fail": 740, "distinct_checked": 842, "both_ok": 0, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 842, "parse_reasons": {"unbound": 505, "name": 97, "formula": 55, "expected (": 46, "expected )": 20, "expected ;": 9, "no-eos": 6, "expected None": 2}, "lean_wall_s": 3.338042974472046, "lean_proc_s": 7.039812803268433, "nd_verify_s": 0.03567910194396973, "workers": 96, "chunk": 400}
58
+ {"utc": "2026-10-07T22:49:15Z", "samples": 12000, "parse_fail": 801, "distinct_checked": 10773, "both_ok": 9629, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1144, "parse_reasons": {"unbound": 582, "name": 57, "expected (": 39, "expected ;": 32, "expected )": 25, "formula": 17, "expected :=": 17, "no-eos": 16, "exact not last": 8, "ORE box": 6, "expected None": 1, "expected :": 1}, "lean_wall_s": 33.56905555725098, "lean_proc_s": 895.1724514961243, "nd_verify_s": 0.6677272319793701, "workers": 96, "chunk": 400}
59
+ {"utc": "2026-10-07T22:49:26Z", "samples": 1656, "parse_fail": 689, "distinct_checked": 889, "both_ok": 1, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 888, "parse_reasons": {"unbound": 466, "expected (": 58, "name": 43, "formula": 34, "expected ;": 25, "expected )": 22, "no-eos": 21, "expected :=": 7, "ORE box": 5, "expected None": 3, "exact not last": 2, "expected =>": 2, "expected exact": 1}, "lean_wall_s": 3.3635947704315186, "lean_proc_s": 7.921581506729126, "nd_verify_s": 0.04089689254760742, "workers": 96, "chunk": 400}
60
+ {"utc": "2026-10-07T22:49:37Z", "samples": 1656, "parse_fail": 726, "distinct_checked": 864, "both_ok": 3, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 861, "parse_reasons": {"unbound": 548, "expected ;": 52, "name": 44, "expected (": 31, "expected )": 26, "no-eos": 8, "expected :=": 6, "formula": 6, "exact not last": 2, "expected None": 1, "PR position": 1, "trailing tokens": 1}, "lean_wall_s": 3.7127902507781982, "lean_proc_s": 8.23496413230896, "nd_verify_s": 0.04343986511230469, "workers": 96, "chunk": 400}
61
+ {"utc": "2026-10-07T22:49:47Z", "samples": 1656, "parse_fail": 676, "distinct_checked": 904, "both_ok": 0, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 904, "parse_reasons": {"unbound": 514, "expected (": 44, "name": 34, "formula": 27, "expected ;": 25, "expected )": 23, "no-eos": 3, "exact not last": 2, "ORE box": 2, "expected :=": 1, "expected =>": 1}, "lean_wall_s": 3.741650104522705, "lean_proc_s": 9.144235610961914, "nd_verify_s": 0.04424762725830078, "workers": 96, "chunk": 400}
62
+ {"utc": "2026-10-07T22:51:40Z", "samples": 12000, "parse_fail": 874, "distinct_checked": 10713, "both_ok": 9677, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1036, "parse_reasons": {"unbound": 658, "name": 57, "expected (": 42, "formula": 25, "expected ;": 21, "no-eos": 21, "expected )": 20, "exact not last": 14, "ORE box": 9, "expected None": 5, "trailing tokens": 1, "eof": 1}, "lean_wall_s": 35.413288593292236, "lean_proc_s": 938.9061851501465, "nd_verify_s": 0.6585297584533691, "workers": 96, "chunk": 400}
63
+ {"utc": "2026-10-07T22:51:51Z", "samples": 1632, "parse_fail": 759, "distinct_checked": 793, "both_ok": 0, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 793, "parse_reasons": {"unbound": 572, "name": 61, "expected (": 47, "expected )": 38, "expected ;": 27, "no-eos": 5, "formula": 2, "ORE box": 2, "PR position": 2, "expected None": 2, "expected exact": 1}, "lean_wall_s": 3.1027181148529053, "lean_proc_s": 6.097707271575928, "nd_verify_s": 0.04024338722229004, "workers": 96, "chunk": 400}
64
+ {"utc": "2026-10-07T22:52:01Z", "samples": 1608, "parse_fail": 728, "distinct_checked": 815, "both_ok": 0, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 815, "parse_reasons": {"unbound": 529, "expected (": 56, "name": 53, "no-eos": 30, "expected )": 25, "expected ;": 16, "ORE box": 6, "formula": 5, "expected None": 5, "expected :=": 1, "expected .elim": 1, "exact not last": 1}, "lean_wall_s": 3.4211013317108154, "lean_proc_s": 7.123792409896851, "nd_verify_s": 0.03533649444580078, "workers": 96, "chunk": 400}
65
+ {"utc": "2026-10-07T22:52:10Z", "samples": 1608, "parse_fail": 637, "distinct_checked": 893, "both_ok": 10, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 883, "parse_reasons": {"unbound": 485, "name": 59, "formula": 28, "ORE box": 16, "no-eos": 15, "expected ;": 14, "expected )": 5, "exact not last": 5, "expected (": 5, "expected :=": 2, "trailing tokens": 2, "expected None": 1}, "lean_wall_s": 3.505915641784668, "lean_proc_s": 8.41605520248413, "nd_verify_s": 0.0424189567565918, "workers": 96, "chunk": 400}
66
+ {"utc": "2026-10-07T22:53:56Z", "samples": 12000, "parse_fail": 878, "distinct_checked": 10728, "both_ok": 9773, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 955, "parse_reasons": {"unbound": 653, "name": 101, "expected ;": 44, "formula": 26, "expected (": 22, "expected )": 15, "no-eos": 9, "ORE box": 4, "exact not last": 3, "expected None": 1}, "lean_wall_s": 35.101577281951904, "lean_proc_s": 932.3634111881256, "nd_verify_s": 0.6851181983947754, "workers": 96, "chunk": 400}
67
+ {"utc": "2026-10-07T22:54:05Z", "samples": 1584, "parse_fail": 820, "distinct_checked": 718, "both_ok": 0, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 718, "parse_reasons": {"unbound": 626, "name": 76, "expected (": 30, "expected ;": 27, "formula": 24, "ORE box": 18, "expected )": 12, "expected None": 2, "no-eos": 1, "expected hh": 1, "PR position": 1, "exact not last": 1, "expected =>": 1}, "lean_wall_s": 3.089698076248169, "lean_proc_s": 5.863642930984497, "nd_verify_s": 0.025526762008666992, "workers": 96, "chunk": 400}
68
+ {"utc": "2026-10-07T22:54:15Z", "samples": 1584, "parse_fail": 682, "distinct_checked": 831, "both_ok": 1, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 830, "parse_reasons": {"unbound": 513, "name": 67, "expected ;": 40, "expected )": 29, "ORE box": 10, "expected (": 9, "no-eos": 9, "formula": 2, "expected None": 2, "exact not last": 1}, "lean_wall_s": 3.2444589138031006, "lean_proc_s": 7.012999534606934, "nd_verify_s": 0.03846597671508789, "workers": 96, "chunk": 400}
69
+ {"utc": "2026-10-07T22:54:23Z", "samples": 1584, "parse_fail": 816, "distinct_checked": 721, "both_ok": 1, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 720, "parse_reasons": {"unbound": 613, "name": 88, "expected ;": 42, "formula": 33, "expected (": 19, "expected )": 18, "exact not last": 1, "trailing tokens": 1, "expected None": 1}, "lean_wall_s": 3.056405782699585, "lean_proc_s": 5.948251485824585, "nd_verify_s": 0.03201699256896973, "workers": 96, "chunk": 400}
70
+ {"utc": "2026-10-07T22:56:15Z", "samples": 12000, "parse_fail": 866, "distinct_checked": 10745, "both_ok": 9765, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 980, "parse_reasons": {"unbound": 623, "name": 110, "no-eos": 26, "formula": 26, "expected (": 25, "expected ;": 24, "expected )": 16, "exact not last": 9, "trailing tokens": 5, "expected None": 1, "expected hh": 1}, "lean_wall_s": 32.941861629486084, "lean_proc_s": 870.0195224285126, "nd_verify_s": 0.6572699546813965, "workers": 96, "chunk": 400}
71
+ {"utc": "2026-10-07T22:56:26Z", "samples": 1560, "parse_fail": 662, "distinct_checked": 837, "both_ok": 6, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 831, "parse_reasons": {"unbound": 503, "name": 76, "formula": 36, "expected (": 13, "expected ;": 12, "trailing tokens": 7, "expected )": 6, "exact not last": 4, "no-eos": 3, "expected None": 2}, "lean_wall_s": 3.473053216934204, "lean_proc_s": 7.447371006011963, "nd_verify_s": 0.042382001876831055, "workers": 96, "chunk": 400}
72
+ {"utc": "2026-10-07T22:56:36Z", "samples": 1560, "parse_fail": 677, "distinct_checked": 828, "both_ok": 0, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 828, "parse_reasons": {"unbound": 433, "name": 83, "expected ;": 57, "expected (": 42, "formula": 29, "no-eos": 23, "expected )": 4, "exact not last": 3, "ORE box": 1, "expected None": 1, "trailing tokens": 1}, "lean_wall_s": 3.5155117511749268, "lean_proc_s": 7.4980528354644775, "nd_verify_s": 0.038084983825683594, "workers": 96, "chunk": 400}
73
+ {"utc": "2026-10-07T22:56:46Z", "samples": 1560, "parse_fail": 801, "distinct_checked": 699, "both_ok": 2, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 697, "parse_reasons": {"unbound": 529, "name": 181, "no-eos": 32, "expected )": 25, "expected (": 15, "expected ;": 12, "formula": 3, "exact not last": 2, "expected :=": 1, "ORE box": 1}, "lean_wall_s": 2.923445224761963, "lean_proc_s": 5.54842472076416, "nd_verify_s": 0.036624908447265625, "workers": 96, "chunk": 400}
74
+ {"utc": "2026-10-07T23:02:43Z", "samples": 70912, "parse_fail": 11883, "distinct_checked": 45649, "both_ok": 29025, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 16624, "parse_reasons": {"unbound": 9085, "name": 910, "expected (": 574, "no-eos": 307, "expected ;": 275, "expected )": 246, "ORE box": 199, "formula": 131, "expected :=": 70, "exact not last": 46, "expected None": 27, "trailing tokens": 8, "PR position": 3, "expected fun": 1, "expected by": 1}, "lean_wall_s": 119.12897419929504, "lean_proc_s": 9430.691316843033, "nd_verify_s": 2.7739779949188232, "workers": 96, "chunk": 400}
75
+ {"utc": "2026-10-07T23:02:52Z", "samples": 2000, "parse_fail": 2, "distinct_checked": 1998, "both_ok": 1973, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 25, "parse_reasons": {"unbound": 2}, "lean_wall_s": 2.9326798915863037, "lean_proc_s": 13.388767719268799, "nd_verify_s": 0.0920114517211914, "workers": 96, "chunk": 400}
76
+ {"utc": "2026-10-07T23:08:06Z", "samples": 70912, "parse_fail": 11883, "distinct_checked": 45649, "both_ok": 29025, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 16624, "parse_reasons": {"unbound": 9085, "name": 910, "expected (": 574, "no-eos": 307, "expected ;": 275, "expected )": 246, "ORE box": 199, "formula": 131, "expected :=": 70, "exact not last": 46, "expected None": 27, "trailing tokens": 8, "PR position": 3, "expected fun": 1, "expected by": 1}, "lean_wall_s": 117.96959280967712, "lean_proc_s": 9292.331460475922, "nd_verify_s": 2.7619974613189697, "workers": 96, "chunk": 400}
runs/590_control-untied-198_s1/pretrain.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0b5ba557ead858e79ca7df87647b058303c00c029ade773d96a7f0784993bb7e
3
+ size 38268899
runs/590_control-untied-198_s1/source/harness.py ADDED
@@ -0,0 +1,529 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ # imports the fork (model, tokenizer, train, ...), which is not resolvable here
3
+ # pyright: reportMissingImports=false
4
+ """FROZEN harness for the test-time-loops autoresearch loop. The research agent does not edit it.
5
+
6
+ Forked from ../autoresearch/harness.py (sha 2aa52f23e7ec) on 2026-10-07. What changed: the RL
7
+ stage belongs to the agent, behind a wall-clock budget and a sampling oracle, and the metric is
8
+ measured with the model's test-time compute knob (KNOB) set to EVAL_LEVEL.
9
+
10
+ One timed run = harness owns the clocks, the oracle and the eval; `pretrain.py` (agent-editable)
11
+ owns pretraining and RL:
12
+
13
+ 1. load frozen data (Dan's cap-6 Stage-1 file, ladder RL-target pool, ladder transfer pool)
14
+ 2. call pretrain.pretrain(train_recs, PRETRAIN_SECS, ...) -> (model, tok, stats); charged for
15
+ everything including model construction and torch.compile. Overrun past the grace -> timeout.
16
+ 3. check n_params <= PARAM_CAP, tok.mode == TOK_MODE and that the model has the KNOB attribute
17
+ 4. call pretrain.rl(model, tok, targets, train_recs, oracle, RL_SECS, ...): expert iteration in
18
+ whatever form the agent likes. The targets are statements only, and `oracle.sample` is the
19
+ only way to turn them into training data: it samples the model's own attempts and returns
20
+ the ones Lean AND nd_verify accept. RL_SECS is wall-clock inside rl() minus the time the
21
+ oracle spent in the Lean gate (pod CPUs differ; GPU work is what is rationed). Overrun past
22
+ the grace, or an oracle call after the budget is spent -> timeout.
23
+ 5. eval on the dev half of the transfer pool with the harness setting the model's KNOB:
24
+ METRIC = dev theorems solved at L_true >= 7 with EVAL_K attempts at EVAL_LEVEL.
25
+ The same eval at DIAG_LEVEL is recorded as a diagnostic, not scored.
26
+ 6. SIGALRM at TOTAL_SECS from process start kills everything -> status "timeout"
27
+
28
+ The holdout half of the transfer pool is never scored here. Everything is written to --out.
29
+
30
+ python3 harness.py --out artifacts/x # on the pod (see pod.py / remote.sh)
31
+ ND_FORK=... python3 harness.py --out /tmp/smoke --smoke # CPU, ~1 min, plumbing check
32
+ """
33
+
34
+ import argparse
35
+ import collections
36
+ import hashlib
37
+ import importlib
38
+ import json
39
+ import os
40
+ import random
41
+ import shutil
42
+ import signal
43
+ import sys
44
+ import time
45
+ import traceback
46
+ from pathlib import Path
47
+
48
+ HERE = os.path.dirname(os.path.abspath(__file__))
49
+ FORK = os.environ.get("ND_FORK", "/workspace/nd-takehome")
50
+ sys.path.insert(0, HERE)
51
+ sys.path.insert(
52
+ 0, FORK
53
+ ) # fork first: its train.py provides batch/loss_on; the agent's file is pretrain.py
54
+ import torch # noqa: E402
55
+ from eval_set import judge, wilson # noqa: E402
56
+ from sample import generate # noqa: E402
57
+
58
+ # ---------------------------------------------------------------- the rules
59
+ # (hash-checked into every result)
60
+ TOTAL_SECS = int(
61
+ os.environ.get("ND_TOTAL_SECS", 5400)
62
+ ) # process start -> summary written; SIGALRM past this. env: benchmarks only
63
+ PRETRAIN_SECS = 300 # agent's pretraining budget, compile included
64
+ PRETRAIN_GRACE = 15 # seconds the agent may overrun before the run is a "timeout"
65
+ RL_SECS = 1800 # agent's RL budget: wall-clock inside rl() minus the oracle's Lean-gate time
66
+ RL_GRACE = 60
67
+ PARAM_CAP = 10_000_000 # unique trainable params, embeddings included; checked after RL too
68
+ N_TARGETS, TEMP, MAX_NEW, SAMPLE_BATCH = 1500, 0.8, 512, 1024
69
+ EVAL_K = 64
70
+ # The model's test-time compute knob: an integer attribute the harness sets before each eval and
71
+ # rl() sets before oracle calls. A sibling loop for another architecture changes these three
72
+ # constants and nothing else in this file.
73
+ KNOB = "n_loops" # loops of the shared core
74
+ EVAL_LEVEL = 8 # the metric: the knob's value during the scored eval
75
+ DIAG_LEVEL = 2 # the same eval again at this value, recorded but not scored
76
+ TOK_MODE = "lean_seq" # required tokenizer mode (fork lean_tok.py)
77
+ DEV = "cuda" if torch.cuda.is_available() else "cpu"
78
+
79
+
80
+ def read(fn):
81
+ return [json.loads(line) for line in open(fn) if line.strip()]
82
+
83
+
84
+ def log(msg):
85
+ print(f"[{time.strftime('%H:%M:%S')}] {msg}", flush=True)
86
+
87
+
88
+ def sha(fn):
89
+ return hashlib.sha256(open(fn, "rb").read()).hexdigest()[:12]
90
+
91
+
92
+ class Timeout(Exception):
93
+ pass
94
+
95
+
96
+ class BudgetSpent(Timeout):
97
+ """rl() asked the oracle for samples after RL_SECS was used up."""
98
+
99
+
100
+ # ---------------------------------------------------------------- bookkeeping
101
+
102
+
103
+ def save_checkpoint(a, model, tok, stage):
104
+ """Publish complete weight snapshots atomically; these are not optimizer-resume files."""
105
+ path = Path(a.out) / f"{stage}.pt"
106
+ pending = path.with_suffix(".pt.tmp")
107
+ try:
108
+ torch.save(
109
+ {
110
+ "state": model.state_dict(),
111
+ "tok_mode": getattr(tok, "mode", None),
112
+ "stage": stage,
113
+ "seed": a.seed,
114
+ "harness_sha": sha(f"{HERE}/harness.py"),
115
+ "pretrain_sha": sha(f"{HERE}/pretrain.py"),
116
+ },
117
+ pending,
118
+ )
119
+ os.replace(pending, path)
120
+ finally:
121
+ pending.unlink(missing_ok=True)
122
+
123
+
124
+ def lstar(solved_names, recs, need=5):
125
+ ls_solved = [r["n_lines"] for r in recs if r["name"] in solved_names]
126
+ ge = {L: sum(1 for x in ls_solved if x >= L) for L in range(7, 15)}
127
+ return max([L for L, c in ge.items() if c >= need], default=0), ge
128
+
129
+
130
+ def by_bin(solved_names, recs):
131
+ by = collections.defaultdict(lambda: [0, 0])
132
+ for r in recs:
133
+ by[r["n_lines"]][0] += r["name"] in solved_names
134
+ by[r["n_lines"]][1] += 1
135
+ return {
136
+ str(L): {"solved": k, "n": n, "rate": k / n, "ci": list(wilson(k, n))}
137
+ for L, (k, n) in sorted(by.items())
138
+ }
139
+
140
+
141
+ def gate_seconds():
142
+ """Seconds the fork's lean_gate has logged so far (Lean wall-clock plus nd_verify)."""
143
+ fn = os.environ.get("LEAN_GATE_LOG")
144
+ if not fn or not os.path.exists(fn):
145
+ return 0.0
146
+ return sum(rec["lean_wall_s"] + rec["nd_verify_s"] for rec in read(fn))
147
+
148
+
149
+ def sample_and_judge(model, tok, recs, k, temperature, seed, batch, max_new):
150
+ prompts = [r["prompt"] for r in recs for _ in range(k)]
151
+ t0, g0 = time.time(), gate_seconds()
152
+ flat = generate(
153
+ model,
154
+ tok,
155
+ prompts,
156
+ greedy=(k == 1 and temperature == 0),
157
+ temperature=temperature,
158
+ batch=batch,
159
+ seed=seed,
160
+ max_new=max_new,
161
+ )
162
+ secs, gate = time.time() - t0, gate_seconds() - g0
163
+ rows = judge(recs, [flat[i * k : (i + 1) * k] for i in range(len(recs))], "n_lines")
164
+ return rows, {
165
+ "samples": len(prompts),
166
+ "secs": secs,
167
+ "gate_secs": gate,
168
+ "samples_per_sec": len(prompts) / secs,
169
+ }
170
+
171
+
172
+ def greedy_rate(model, tok, recs):
173
+ rows, _ = sample_and_judge(model, tok, recs, 1, 0, 0, SAMPLE_BATCH, MAX_NEW)
174
+ return sum(r["solved"] for r in rows) / max(len(rows), 1)
175
+
176
+
177
+ def dev_split(transfer):
178
+ return [r for r in transfer if int(hashlib.sha1(r["key"].encode()).hexdigest(), 16) % 2 == 0]
179
+
180
+
181
+ def holdout_split(transfer):
182
+ """Never scored by the loop. Score final.pt on this offline when checking in."""
183
+ return [r for r in transfer if int(hashlib.sha1(r["key"].encode()).hexdigest(), 16) % 2 == 1]
184
+
185
+
186
+ # ---------------------------------------------------------------- the RL oracle
187
+
188
+
189
+ class Oracle:
190
+ """What rl() gets instead of data: the model's own attempts at the RL targets, filtered to
191
+ the ones Lean AND nd_verify accept. It also keeps the RL clock and the record of every call.
192
+ """
193
+
194
+ def __init__(self, a, tok, targets):
195
+ self.a, self.tok = a, tok
196
+ self.targets = {t["name"]: t for t in targets}
197
+ self.found = collections.defaultdict(dict) # name -> {proof text: record}
198
+ self.calls = []
199
+ self.t0 = self.last = time.time()
200
+ self.uncharged = 0.0
201
+
202
+ def used(self):
203
+ """RL seconds charged so far: wall-clock since rl() started, minus Lean-gate time."""
204
+ return time.time() - self.t0 - self.uncharged
205
+
206
+ def left(self):
207
+ return self.a.rl_secs - self.used()
208
+
209
+ def sample(self, model, targets, k, temperature=TEMP, batch=SAMPLE_BATCH):
210
+ """k attempts at each of `targets` (records rl() was given, any subset, in any order) from
211
+ `model` as it is now, in eval mode, at whatever its KNOB is set to. Returns one
212
+ dict per target, in order: name, prompt, n_lines, proofs (the distinct accepted proof
213
+ texts), n_ok (accepted attempts) and n_tried. Rejected attempts are not returned.
214
+ Raises BudgetSpent, which ends the run as a timeout, once the budget is used up."""
215
+ if self.left() <= 0:
216
+ raise BudgetSpent(f"oracle called with {self.used():.0f}s of {self.a.rl_secs}s used")
217
+ recs = [self.targets[t["name"]] for t in targets]
218
+ call = len(self.calls) + 1
219
+ model.eval()
220
+ rows, samp = sample_and_judge(
221
+ model, self.tok, recs, k, temperature, self.a.seed * 1000 + call, batch, self.a.max_new
222
+ )
223
+ self.uncharged += samp["gate_secs"]
224
+ new = 0
225
+ for t, row in zip(recs, rows, strict=True):
226
+ for p, wl, pl in zip(
227
+ row["proofs"], row["written_lens"], row["pruned_lens"], strict=True
228
+ ):
229
+ if p not in self.found[t["name"]]:
230
+ self.found[t["name"]][p] = {
231
+ "proof": p,
232
+ "written": wl,
233
+ "pruned": pl,
234
+ "round": call,
235
+ }
236
+ new += 1
237
+ solved = {n for n, d in self.found.items() if d}
238
+ ls, ge = lstar(solved, list(self.targets.values()))
239
+ now = time.time()
240
+ self.calls.append(
241
+ {
242
+ "round": call,
243
+ "level": getattr(model, KNOB, None),
244
+ "k": k,
245
+ "n_targets": len(recs),
246
+ "sample_acc": sum(x["n_ok"] for x in rows) / max(samp["samples"], 1),
247
+ "new_proofs": new,
248
+ "targets_solved": len(solved),
249
+ "lstar": ls,
250
+ "ge": ge,
251
+ "mix_rl_records": sum(len(x["proofs"]) for x in rows), # proofs handed back
252
+ "sampling": samp,
253
+ "secs": now - self.last, # since the previous call returned: includes rl()'s work
254
+ "rl_secs_used": self.used(),
255
+ }
256
+ )
257
+ self.last = now
258
+ c = self.calls[-1]
259
+ log(
260
+ f"ei r{call}: acc {c['sample_acc']:.4f}, +{new} proofs, "
261
+ f"solved {len(solved)}/{len(self.targets)}, L*={ls}, {KNOB} {c['level']}, "
262
+ f"{len(recs)} targets x {k}, {samp['samples_per_sec']:.0f} samples/s, "
263
+ f"RL clock {c['rl_secs_used']:.0f}/{self.a.rl_secs}s"
264
+ )
265
+ return [
266
+ {
267
+ "name": t["name"],
268
+ "prompt": t["prompt"],
269
+ "n_lines": t["n_lines"],
270
+ "proofs": list(row["proofs"]),
271
+ "n_ok": row["n_ok"],
272
+ "n_tried": row["n_tried"],
273
+ }
274
+ for t, row in zip(recs, rows, strict=True)
275
+ ]
276
+
277
+
278
+ # ---------------------------------------------------------------- stages
279
+
280
+
281
+ def stage_rl(a, agent, tok, model, targets, train_recs):
282
+ oracle = Oracle(a, tok, targets)
283
+ # statements only: what the agent's RL may know about a target
284
+ visible = [{"name": t["name"], "prompt": t["prompt"], "n_lines": t["n_lines"]} for t in targets]
285
+ stats = agent.rl(model, tok, visible, train_recs, oracle, a.rl_secs, DEV, a.seed, log)
286
+ if DEV == "cuda":
287
+ torch.cuda.synchronize()
288
+ used = oracle.used()
289
+ model.eval()
290
+ with open(f"{a.out}/found_targets.jsonl", "w") as f:
291
+ for t in targets:
292
+ for x in oracle.found[t["name"]].values():
293
+ f.write(
294
+ json.dumps({"name": t["name"], "thm": t["thm"], "L_true": t["n_lines"], **x})
295
+ + "\n"
296
+ )
297
+ log(f"rl returned: {used:.0f}s of {a.rl_secs}s charged, {oracle.uncharged:.0f}s Lean gate")
298
+ return {
299
+ "rounds": oracle.calls,
300
+ "secs": time.time() - oracle.t0,
301
+ "rl_secs_used": used,
302
+ "gate_secs": oracle.uncharged,
303
+ "targets": len(targets),
304
+ "stats": dict(stats or {}),
305
+ }
306
+
307
+
308
+ def eval_at(a, tok, model, dev_pool, level):
309
+ setattr(model, KNOB, level)
310
+ rows, samp = sample_and_judge(
311
+ model, tok, dev_pool, a.eval_k, TEMP, a.seed * 1000 + 777, SAMPLE_BATCH, a.max_new
312
+ )
313
+ solved = {r["name"] for r in rows if r["solved"]}
314
+ ls, ge = lstar(solved, dev_pool)
315
+ out = {
316
+ "level": level,
317
+ "attempts_per_theorem": a.eval_k,
318
+ "solved": len(solved),
319
+ "n": len(dev_pool),
320
+ "solved_ge7": ge[7],
321
+ "lstar": ls,
322
+ "ge": ge,
323
+ "by_bin": by_bin(solved, dev_pool),
324
+ "sample_acc": sum(r["n_ok"] for r in rows) / max(samp["samples"], 1),
325
+ "max_pruned": max((max(r["pruned_lens"]) for r in rows if r["solved"]), default=0),
326
+ "sampling": samp,
327
+ }
328
+ log(
329
+ f"eval at {KNOB} {level}: solved {len(solved)}/{len(dev_pool)}, ge7 {ge[7]}, L*={ls}, "
330
+ f"per-sample {out['sample_acc']:.4f}, {samp['samples_per_sec']:.0f} samples/s"
331
+ )
332
+ return out, rows
333
+
334
+
335
+ def stage_eval(a, tok, model, dev_pool, heldout_recs):
336
+ out, rows = eval_at(a, tok, model, dev_pool, a.eval_level)
337
+ out["heldout_greedy"] = greedy_rate(model, tok, heldout_recs)
338
+ with open(f"{a.out}/eval_dev.jsonl", "w") as f:
339
+ for r in rows:
340
+ keys = ("name", "n_lines", "solved", "n_ok", "proofs", "written_lens", "pruned_lens")
341
+ f.write(json.dumps({k: r[k] for k in keys}) + "\n")
342
+ diag, _ = eval_at(a, tok, model, dev_pool, a.diag_level)
343
+ out["diag"] = diag
344
+ setattr(model, KNOB, a.eval_level)
345
+ log(
346
+ f"eval: solved {out['solved']}/{len(dev_pool)}, ge7 {out['solved_ge7']} at "
347
+ f"{KNOB} {a.eval_level} ({diag['solved_ge7']} at {a.diag_level}), L*={out['lstar']}, "
348
+ f"max pruned {out['max_pruned']}, held-out greedy {out['heldout_greedy']:.3f}"
349
+ )
350
+ return out
351
+
352
+
353
+ def lean_gate_totals():
354
+ """Sum the fork lean_gate's per-generate() log lines (remote.sh sets LEAN_GATE_LOG)."""
355
+ fn = os.environ.get("LEAN_GATE_LOG")
356
+ if not fn or not os.path.exists(fn):
357
+ return None
358
+ keys = ("samples", "parse_fail", "distinct_checked", "both_ok", "nd_ok_lean_rej")
359
+ tot = dict.fromkeys(keys + ("nd_rej_lean_ok", "lean_wall_s"), 0)
360
+ for rec in read(fn):
361
+ for k in tot:
362
+ tot[k] += rec[k]
363
+ return tot
364
+
365
+
366
+ RULES = (
367
+ "TOTAL_SECS",
368
+ "PRETRAIN_SECS",
369
+ "RL_SECS",
370
+ "PARAM_CAP",
371
+ "N_TARGETS",
372
+ "EVAL_K",
373
+ "KNOB",
374
+ "EVAL_LEVEL",
375
+ "DIAG_LEVEL",
376
+ "TOK_MODE",
377
+ )
378
+
379
+
380
+ def write_summary(a, status, **kw):
381
+ s = {
382
+ "status": status,
383
+ "metric": kw.get("eval", {}).get("solved_ge7"),
384
+ "metric_diag": kw.get("eval", {}).get("diag", {}).get("solved_ge7"),
385
+ "metric_rl": (kw.get("rl", {}).get("rounds") or [{}])[-1].get(
386
+ "targets_solved"
387
+ ), # secondary: cumulative RL targets solved
388
+ "harness_sha": sha(f"{HERE}/harness.py"),
389
+ "pretrain_sha": sha(f"{HERE}/pretrain.py"),
390
+ "device": DEV,
391
+ "gpu": torch.cuda.get_device_name(0) if DEV == "cuda" else "cpu",
392
+ "torch": torch.__version__,
393
+ "lean_gate": lean_gate_totals(),
394
+ "total_secs": time.time() - a.T0,
395
+ "rules": {k: globals()[k] for k in RULES},
396
+ **kw,
397
+ }
398
+ with open(f"{a.out}/summary.json", "w") as f:
399
+ json.dump(s, f, indent=1)
400
+ log(
401
+ f"SUMMARY status={status} metric={s['metric']} metric_diag={s['metric_diag']} "
402
+ f"metric_rl={s['metric_rl']} total {s['total_secs']:.0f}s"
403
+ )
404
+ return s
405
+
406
+
407
+ def main():
408
+ ap = argparse.ArgumentParser()
409
+ ap.add_argument("--out", required=True)
410
+ ap.add_argument("--seed", type=int, default=0)
411
+ ap.add_argument(
412
+ "--pretrain-only", action="store_true", help="save pretrain.pt and stop before eval/RL"
413
+ )
414
+ ap.add_argument(
415
+ "--smoke", action="store_true", help="tiny everything, for a CPU plumbing check"
416
+ )
417
+ a = ap.parse_args()
418
+ a.pretrain_secs, a.rl_secs, a.eval_k, a.max_new, a.n_targets = (
419
+ PRETRAIN_SECS,
420
+ RL_SECS,
421
+ EVAL_K,
422
+ MAX_NEW,
423
+ N_TARGETS,
424
+ )
425
+ a.eval_level, a.diag_level = EVAL_LEVEL, DIAG_LEVEL
426
+ if a.smoke:
427
+ a.pretrain_secs, a.rl_secs, a.eval_k, a.max_new, a.n_targets = 10, 30, 2, 120, 8
428
+ os.makedirs(a.out, exist_ok=True)
429
+ # Preserve discarded variants too: weights alone do not specify the model architecture.
430
+ source_dir = Path(a.out) / "source"
431
+ source_dir.mkdir(exist_ok=True)
432
+ for filename in ("pretrain.py", "harness.py", "remote.sh"):
433
+ shutil.copy2(Path(HERE) / filename, source_dir / filename)
434
+ a.T0 = time.time()
435
+ torch.manual_seed(a.seed)
436
+
437
+ def on_alarm(signum, frame):
438
+ raise Timeout()
439
+
440
+ signal.signal(signal.SIGALRM, on_alarm)
441
+ signal.alarm(TOTAL_SECS if not a.smoke else 900)
442
+
443
+ train_recs = read(f"{FORK}/data/train.jsonl")
444
+ assert all(r["n_lines"] <= 6 for r in train_recs), "pretrain file violates cap 6"
445
+ heldout_all = read(f"{FORK}/data/heldout.jsonl")
446
+ # Length-ordered file: take 2000 evenly spaced records, so every length 2-6 is covered.
447
+ heldout_all = [heldout_all[i * len(heldout_all) // 2000] for i in range(2000)]
448
+ heldout_recs = heldout_all if not a.smoke else heldout_all[:16]
449
+ targets_all = read(f"{FORK}/data/ladder/rl_targets.jsonl")
450
+ transfer = read(f"{FORK}/data/ladder/transfer.jsonl")
451
+ dev_pool = dev_split(transfer) if not a.smoke else dev_split(transfer)[:8]
452
+ targets = random.Random(a.seed).sample(targets_all, min(a.n_targets, len(targets_all)))
453
+ tk, dk, hk = (
454
+ {t["key"] for t in targets},
455
+ {r["key"] for r in dev_pool},
456
+ {r["key"] for r in train_recs},
457
+ )
458
+ assert not (tk & dk) and not (tk & hk) and not (dk & hk), "renaming-class leak between pools"
459
+ log(
460
+ f"pools: train {len(train_recs):,}, targets {len(targets)}, dev-transfer {len(dev_pool)}, "
461
+ f"held-out {len(heldout_recs)}; "
462
+ f"device {DEV}; harness {sha(f'{HERE}/harness.py')} pretrain {sha(f'{HERE}/pretrain.py')}"
463
+ )
464
+
465
+ pre = {}
466
+ try:
467
+ agent = importlib.import_module("pretrain")
468
+ t0 = time.time()
469
+ model, tok, pre = agent.pretrain(train_recs, a.pretrain_secs, DEV, a.seed, log)
470
+ if DEV == "cuda":
471
+ torch.cuda.synchronize()
472
+ pre = dict(pre or {})
473
+ pre["secs"] = time.time() - t0
474
+ model.eval()
475
+ n_params = sum(p.numel() for p in model.parameters())
476
+ pre["n_params"] = n_params
477
+ log(
478
+ f"pretrain returned after {pre['secs']:.0f}s (budget {a.pretrain_secs}s), "
479
+ f"{n_params:,} params"
480
+ )
481
+ if pre["secs"] > a.pretrain_secs + PRETRAIN_GRACE and not a.smoke:
482
+ over = pre["secs"] - a.pretrain_secs
483
+ write_summary(
484
+ a, "timeout", pretrain=pre, reason=f"pretrain overran budget by {over:.0f}s"
485
+ )
486
+ return
487
+ if getattr(tok, "mode", None) != TOK_MODE:
488
+ write_summary(a, "crash", pretrain=pre, reason=f"tok.mode {tok.mode!r} != {TOK_MODE!r}")
489
+ return
490
+ if n_params > PARAM_CAP:
491
+ write_summary(a, "crash", pretrain=pre, reason=f"{n_params} params > cap {PARAM_CAP}")
492
+ return
493
+ if not isinstance(getattr(model, KNOB, None), int):
494
+ reason = f"model has no integer {KNOB} for the harness to set at eval"
495
+ write_summary(a, "crash", pretrain=pre, reason=reason)
496
+ return
497
+ save_checkpoint(a, model, tok, "pretrain")
498
+ if a.pretrain_only:
499
+ write_summary(a, "ok", stage="pretrain_only", pretrain=pre)
500
+ return
501
+ pre["heldout_greedy"] = greedy_rate(model, tok, heldout_recs)
502
+ log(f"pretrain held-out greedy {pre['heldout_greedy']:.3f}")
503
+ rl = stage_rl(a, agent, tok, model, targets, train_recs)
504
+ if rl["rl_secs_used"] > a.rl_secs + RL_GRACE and not a.smoke:
505
+ over = rl["rl_secs_used"] - a.rl_secs
506
+ reason = f"rl overran budget by {over:.0f}s"
507
+ write_summary(a, "timeout", pretrain=pre, rl=rl, reason=reason)
508
+ return
509
+ n_after = sum(p.numel() for p in model.parameters())
510
+ if n_after > PARAM_CAP:
511
+ reason = f"{n_after} params > cap {PARAM_CAP} after rl"
512
+ write_summary(a, "crash", pretrain=pre, rl=rl, reason=reason)
513
+ return
514
+ save_checkpoint(a, model, tok, "final")
515
+ ev = stage_eval(a, tok, model, dev_pool, heldout_recs)
516
+ write_summary(a, "ok", pretrain=pre, rl=rl, eval=ev)
517
+ except BudgetSpent as e:
518
+ write_summary(a, "timeout", pretrain=pre, reason=f"rl: {e}")
519
+ except Timeout:
520
+ write_summary(a, "timeout", pretrain=pre, reason=f"exceeded TOTAL_SECS={TOTAL_SECS}")
521
+ except Exception:
522
+ traceback.print_exc()
523
+ write_summary(a, "crash", pretrain=pre, reason=traceback.format_exc()[-2000:])
524
+ finally:
525
+ signal.alarm(0)
526
+
527
+
528
+ if __name__ == "__main__":
529
+ main()
runs/590_control-untied-198_s1/source/pretrain.py ADDED
@@ -0,0 +1,594 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # imports the fork (model, tokenizer, train, ...), which is not resolvable here
2
+ # pyright: reportMissingImports=false
3
+ """AGENT-EDITABLE: pretraining and RL. This is the only file the autoresearch agent changes.
4
+
5
+ Contract with harness.py (frozen):
6
+
7
+ pretrain(train_recs, budget_secs, device, seed, log) -> (model, tok, stats)
8
+
9
+ train_recs list of dicts {prompt, proof, n_lines<=6, key, ...}; the frozen cap-6 Stage-1 data
10
+ budget_secs wall-clock you are charged for, from call to return, INCLUDING model construction
11
+ and torch.compile. Overrun by more than the harness grace -> the run is a timeout.
12
+ model nn.Module; <= PARAM_CAP unique params (checked again after rl). The harness samples
13
+ from it with the fork's sample.generate, so it must provide:
14
+ - forward(idx) -> logits (B,T,V) and forward(idx, pos=, mask=, caches=) with a
15
+ per-block KV cache dict, exactly like the fork's GPT (model.blocks must exist)
16
+ - an integer `n_loops` (the harness's KNOB): the loop count applied in eval
17
+ mode. The harness sets it to EVAL_LEVEL for the scored eval and to DIAG_LEVEL
18
+ for the diagnostic one.
19
+ tok must provide encode_prompt, encode_proof, decode, shift_abs, pad, eos, mode, and
20
+ mode must be "lean_seq" (harness TOK_MODE): Lean text in and out, Lean-gated samples.
21
+ Decoder rule: decode() is deterministic and single-pass; it may not call the
22
+ verifier or try alternatives.
23
+ stats dict recorded in summary.json and copied into results.json. MUST contain `curve`:
24
+ a list of {secs, step, train_loss, val_loss} taken every ~CURVE_EVERY seconds
25
+ (see log_point below), so a run's trajectory survives in git. Add anything else.
26
+ val_loss is on VAL_N records held out from train_recs by this file; keep that.
27
+
28
+ rl(model, tok, targets, train_recs, oracle, budget_secs, device, seed, log) -> stats | None
29
+
30
+ targets list of dicts {name, prompt, n_lines}: the 1,500 RL target statements. No proofs.
31
+ oracle oracle.sample(model, targets_subset, k, temperature=0.8, batch=1024) samples k
32
+ attempts per target from the model in eval mode at its current n_loops, and returns
33
+ per target {name, prompt, n_lines, proofs, n_ok, n_tried}; `proofs` are the distinct
34
+ attempts Lean AND nd_verify accepted. This is the only source of new training data.
35
+ oracle.used() / oracle.left(): seconds of the RL budget charged / remaining.
36
+ budget_secs RL_SECS: wall-clock inside rl() minus the oracle's Lean-gate time. Returning more
37
+ than the grace late, or calling the oracle with nothing left, is a timeout.
38
+ The model is trained in place. Everything else about RL is yours: rounds, samples per target,
39
+ loop counts when sampling, which proofs to keep, weighting, replay, optimizer, loss.
40
+
41
+ Anything imported from the fork (ND_FORK) is fair game to copy in here and modify.
42
+ """
43
+
44
+ import math
45
+ import random
46
+ import time
47
+
48
+ import torch
49
+ import torch._inductor.config # noqa: F401 (explicit, so the attribute resolves for pyright)
50
+ from lean_tok import MAXN # fork
51
+ from model import GPT, Block # fork
52
+ from normalize import norm # fork
53
+ from tokenizer import make_tokenizer # fork
54
+ from train import loss_on # fork
55
+
56
+ # ---- knobs
57
+ # Unique blocks: one untied entry block, N_CORE shared core blocks, one untied exit block.
58
+ N_CORE = 4
59
+ # Control: run 198's trunk. The six blocks are independent and applied once, with no re-injection;
60
+ # n_loops is accepted and ignored, so the harness's two evals and rl()'s retry depths all sample
61
+ # the same model.
62
+ UNTIED = True
63
+ N_LAYER, D, N_HEAD = 2 + N_CORE, 384, 8
64
+ # Core applications (loops). Training draws one count per batch from TRAIN_LOOPS, in pretraining
65
+ # and in rl()'s fine-tune; eval and sampling apply model.n_loops, which the harness sets for its
66
+ # evals and rl() sets before each oracle call. DEFAULT_LOOPS is used for the validation curve and
67
+ # the harness's post-pretraining greedy check.
68
+ TRAIN_LOOPS = (1, 2, 3, 4, 5, 6, 7, 8)
69
+ DEFAULT_LOOPS = 2
70
+ VAL_LOOPS = (1, 2, 4, 8) # val loss at each count, once, when pretraining ends
71
+ # ---- RL knobs: the old harness's expert iteration (600 AdamW steps per round on up to 4 proofs
72
+ # per solved target, weighted x4, mixed with 20,000 replayed pretraining records), except for
73
+ # where the samples go. Each round every target gets RL_K attempts at RL_FIRST_LOOPS; a target
74
+ # all of them fail on gets RL_RETRY_K more (32 in all, as before) at ONE deeper loop count from
75
+ # RL_RETRY_LOOPS, assigned at random, so proofs that only deeper sampling finds enter the
76
+ # training set and no deep sample is spent on a target one loop already solves. The random
77
+ # assignment makes the retry calls a controlled comparison of loop counts on resistant targets.
78
+ # Rounds repeat until another one would not fit in the budget.
79
+ RL_FIRST_LOOPS, RL_K = 1, 8
80
+ RL_RETRY_LOOPS, RL_RETRY_K = (2, 4, 8), 24
81
+ # The fine-tune sees FT_RECORDS records per round (the old 600 steps x 128 records) in the same
82
+ # token-budget batches as pretraining (TOK_BUDGET below) instead of 128-record batches padded to
83
+ # their longest proof, which with 7-14 line RL proofs in nearly every batch was ~3x the content.
84
+ FT_RECORDS, FT_LR, FT_WARMUP = 600 * 128, 3e-4, 50
85
+ RETAIN, MAX_PER_THM, RL_WEIGHT = 20000, 4, 4
86
+ # MLP width held at 151's 1280: the fork's Block hardcodes 4 * D (1536 at D 384 would pass the
87
+ # 10M cap), so PeriLNBlock swaps fc1/fc2. 1280 != 3 * D keeps muon_step's QKV-by-shape test valid.
88
+ D_FF = 1280
89
+ BS, LR, MIN_LR_FRAC, WD = 128, 1e-3, 0.3, 0.1
90
+ MUON_LR, MUON_MOMENTUM = 0.01, 0.95
91
+ # Polar Express (Amsel et al. 2025) per-iteration quintic coefficients, as in modded-nanogpt
92
+ # (num_iters=5, safety factor 2e-2): singular values land in ~[0.86, 1.14] instead of the
93
+ # fixed quintic's ~[0.68, 1.13], at the same five iterations.
94
+ NS_COEFFS = (
95
+ (8.156554524902461, -22.48329292557795, 15.878769915207462),
96
+ (4.042929935166739, -2.808917465908714, 0.5000178451051316),
97
+ (3.8916678022926607, -2.772484153217685, 0.5060648178503393),
98
+ (3.285753657755655, -2.3681294933425376, 0.46449024233003106),
99
+ (2.3465413258596377, -1.7097828382687081, 0.42323551169305323),
100
+ )
101
+ # Ten times the AdamW group's per-step multiplicative decay (lr * wd) on the Muon matrices.
102
+ MUON_WD = 10 * WD * LR / MUON_LR
103
+ TOK_MODE = "lean_seq" # required by the harness (Lean protocol)
104
+ # Hybrid heads: half the heads get ALiBi (Press et al. 2021) with the standard slope set for
105
+ # H/2 heads, 2^(-8 i / (H/2)) = 1/4 .. 1/256 with 8 heads; the other half get slope 0, i.e. no
106
+ # positional signal at all (NoPE), so they can look citations up by content at any distance.
107
+ N_ALIBI = N_HEAD // 2
108
+ ALIBI_SLOPES = tuple(2.0 ** (-8.0 * (i + 1) / N_ALIBI) for i in range(N_ALIBI)) + (0.0,) * (
109
+ N_HEAD - N_ALIBI
110
+ )
111
+ VAL_N, CURVE_EVERY = 1024, 15 # records held out for val_loss; seconds between curve points
112
+ # Token-budget batches (fairseq max-tokens): each step takes a run of length-sorted records with
113
+ # count * padded_length <= TOK_BUDGET, so a batch of short proofs has many records and a batch of
114
+ # long ones few, and the padded token count per step is ~constant. 128 x 144 matches the real
115
+ # (unpadded) tokens of the incumbent's 128-record batches (mean record ~135 tokens), whose
116
+ # padding to the batch max (~400) made every step ~3x larger than its content.
117
+ TOK_BUDGET = 128 * 144
118
+ # Multi-token prediction (DeepSeek-V3 2024, Gloeckle et al. 2024), pretraining only: one extra
119
+ # Peri-LN block reads [trunk state at t; embedding of token t+1] and predicts token t+2 through
120
+ # the model's own ln_f/head. Its loss is added at this weight (DeepSeek-V3's early-training value).
121
+ MTP_WEIGHT = 0.3
122
+
123
+
124
+ # Muon update adapted from KellerJordan/Muon, commit
125
+ # f98f1cacc0263b04290753e32be8d498c1efc806 (muon.py).
126
+ # MIT License
127
+ # Copyright (c) 2024 Keller Jordan
128
+ #
129
+ # Permission is hereby granted, free of charge, to any person obtaining a copy
130
+ # of this software and associated documentation files (the "Software"), to deal
131
+ # in the Software without restriction, including without limitation the rights
132
+ # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
133
+ # copies of the Software, and to permit persons to whom the Software is
134
+ # furnished to do so, subject to the following conditions:
135
+ # The above copyright notice and this permission notice shall be included in all
136
+ # copies or substantial portions of the Software.
137
+ # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
138
+ # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
139
+ # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
140
+ # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
141
+ # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
142
+ # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
143
+ # SOFTWARE.
144
+ @torch.no_grad()
145
+ def muon_step(params, buffers, lr):
146
+ """Nesterov momentum plus five bf16 Polar Express steps, on hidden matrices only."""
147
+ for p, momentum in zip(params, buffers, strict=True):
148
+ if p.grad is None:
149
+ continue
150
+ momentum.lerp_(p.grad, 1 - MUON_MOMENTUM)
151
+ update = p.grad.lerp(momentum, MUON_MOMENTUM).bfloat16()
152
+ # Only QKV has shape (3D, D) in this GPT; orthogonalize its roles separately.
153
+ if update.size(0) == 3 * update.size(1):
154
+ update = update.reshape(3, update.size(1), update.size(1))
155
+ transpose = update.size(-2) > update.size(-1)
156
+ if transpose:
157
+ update = update.mT
158
+ update = update / (update.norm(dim=(-2, -1), keepdim=True) * (1 + 2e-2) + 1e-6)
159
+ for a, b, c in NS_COEFFS:
160
+ gram = update @ update.mT
161
+ polynomial = b * gram + c * gram @ gram
162
+ update = a * update + polynomial @ update
163
+ if transpose:
164
+ update = update.mT
165
+ p.mul_(1 - lr * MUON_WD)
166
+ # lr may be a 0-d tensor (CUDA-graphed step); scale in p's dtype like add_(alpha=-lr)
167
+ p.sub_(lr * update.reshape_as(p).to(p.dtype))
168
+
169
+
170
+ class PeriLNBlock(Block):
171
+ """P22 Peri-LN (Kim et al. 2025): the fork's pre-LN block plus a LayerNorm on each branch
172
+ output, so every residual write has unit scale times a learned gain. Attention has no RoPE:
173
+ `mask` is the additive ALiBi bias (B or 1, H, T, S) that LoopedGPT builds once per forward
174
+ and cos/sin are unused."""
175
+
176
+ def __init__(self, d, h):
177
+ super().__init__(d, h)
178
+ self.fc1, self.fc2 = torch.nn.Linear(d, D_FF), torch.nn.Linear(D_FF, d)
179
+ self.ln_attn_out, self.ln_mlp_out = torch.nn.LayerNorm(d), torch.nn.LayerNorm(d)
180
+
181
+ def attn(self, x, cos, sin, mask, cache):
182
+ B, T, D = x.shape
183
+ q, k, v = self.qkv(x).view(B, T, 3, self.h, D // self.h).permute(2, 0, 3, 1, 4)
184
+ if cache is not None: # the fork's preallocated KV cache, unchanged
185
+ if "k" not in cache:
186
+ cache["k"] = k.new_empty(B, self.h, cache["max"], D // self.h)
187
+ cache["v"] = v.new_empty(B, self.h, cache["max"], D // self.h)
188
+ cache["n"] = 0
189
+ n0, n1 = cache["n"], cache["n"] + T
190
+ cache["k"][:, :, n0:n1] = k
191
+ cache["v"][:, :, n0:n1] = v
192
+ cache["n"] = n1
193
+ k, v = cache["k"][:, :, :n1], cache["v"][:, :, :n1]
194
+ y = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask.to(q.dtype))
195
+ return self.proj(y.transpose(1, 2).reshape(B, T, D))
196
+
197
+ def forward(self, x, cos, sin, mask=None, cache=None):
198
+ x = x + self.ln_attn_out(self.attn(self.ln1(x), cos, sin, mask, cache))
199
+ return x + self.ln_mlp_out(self.fc2(torch.nn.functional.gelu(self.fc1(self.ln2(x)))))
200
+
201
+
202
+ class LoopedGPT(GPT):
203
+ """Looped trunk (Universal Transformer, Dehghani et al. 2018; sandwich layout and input
204
+ injection as in Geiping et al. 2025): blocks[0] is the entry block, blocks[1:-1] the shared
205
+ core applied n times, blocks[-1] the exit block. Every core pass starts from
206
+ ln_loop(state) + e, with e the entry block's output, so each pass sees an input of the same
207
+ scale and the prompt is re-injected instead of being buried under accumulated writes.
208
+
209
+ P34: the fork's GPT with ALiBi (Press et al. 2021) instead of RoPE. Head h's attention
210
+ logits get -slope_h * (query position - key position), slopes 2^-1 .. 2^-8 for 8 heads, so
211
+ no rotation of q/k anywhere. Positions come from `pos` (sample.generate's left-padded
212
+ batches) or arange; key positions are kept in caches[0] so cached decoding sees the same
213
+ bias as the full forward. Causality/padding come from `mask` when given, else from rel < 0."""
214
+
215
+ def __init__(self, vocab, n_layer, d, n_head, seed, max_len=1024):
216
+ super().__init__(vocab, n_layer, d, n_head, max_len)
217
+ slopes = torch.tensor(ALIBI_SLOPES).view(1, n_head, 1, 1)
218
+ self.register_buffer("slopes", slopes, persistent=False)
219
+ self.ln_loop = torch.nn.LayerNorm(d)
220
+ self.n_loops, self.loop_rng = DEFAULT_LOOPS, random.Random(seed)
221
+
222
+ def forward(self, idx, pos=None, mask=None, caches=None):
223
+ return self.head(self.ln_f(self.features(idx, pos, mask, caches)[0]))
224
+
225
+ def features(self, idx, pos=None, mask=None, caches=None):
226
+ """The trunk: (final hidden state before ln_f, the ALiBi bias used)."""
227
+ B, T = idx.shape
228
+ if pos is None: # training/val: one (1, H, T, T) bias broadcast over the batch
229
+ pos = torch.arange(T, device=idx.device)[None]
230
+ key_pos = pos
231
+ if caches is not None: # incremental decoding: keys are every cached token
232
+ c = caches[0]
233
+ if "pos" not in c:
234
+ c["pos"], c["n_pos"] = pos.new_empty(pos.size(0), c["max"]), 0
235
+ n0, n1 = c["n_pos"], c["n_pos"] + T
236
+ c["pos"][:, n0:n1] = pos
237
+ c["n_pos"] = n1
238
+ key_pos = c["pos"][:, :n1]
239
+ rel = pos[:, None, :, None] - key_pos[:, None, None, :] # (B or 1, 1, T, S)
240
+ bias = -self.slopes * rel.to(self.slopes.dtype)
241
+ bias = bias.masked_fill(rel < 0 if mask is None else ~mask, float("-inf"))
242
+ if UNTIED:
243
+ x = self.emb(idx)
244
+ for j, b in enumerate(self.blocks):
245
+ x = b(x, None, None, bias, None if caches is None else caches[j])
246
+ return x, bias
247
+ n = self.loop_rng.choice(TRAIN_LOOPS) if self.training else self.n_loops
248
+ entry, *core, exit_block = self.blocks
249
+ x = e = entry(self.emb(idx), None, None, bias, None if caches is None else caches[0])
250
+ for loop in range(n):
251
+ x = self.ln_loop(x) + e
252
+ for j, b in enumerate(core, 1):
253
+ # sample.generate makes one cache dict per unique block; a shared block's keys
254
+ # and values differ per pass, so each pass gets its own sub-cache inside it.
255
+ c = caches and caches[j].setdefault(loop, {"max": caches[j]["max"]})
256
+ x = b(x, None, None, bias, c)
257
+ x = exit_block(x, None, None, bias, None if caches is None else caches[-1])
258
+ return x, bias
259
+
260
+
261
+ class MTPHead(torch.nn.Module):
262
+ """The MTP module. A separate nn.Module, never registered on the model, so the harness's
263
+ parameter count, RL AdamW and sampler never see it; it is dropped when pretrain() returns."""
264
+
265
+ def __init__(self, d, h):
266
+ super().__init__()
267
+ self.ln_h, self.ln_e = torch.nn.LayerNorm(d), torch.nn.LayerNorm(d)
268
+ self.proj = torch.nn.Linear(2 * d, d, bias=False)
269
+ self.block = PeriLNBlock(d, h)
270
+
271
+
272
+ def mtp_loss(model, mtp, x, m):
273
+ """The fork's loss_on (next-token CE on proof tokens) plus MTP_WEIGHT times the CE of the
274
+ MTP module's token-(t+2) prediction at the same proof positions. Returns (total, main)."""
275
+ h, bias = model.features(x[:, :-1])
276
+ logits = model.head(model.ln_f(h))
277
+ V = logits.size(-1)
278
+ ce = torch.nn.functional.cross_entropy(
279
+ logits.reshape(-1, V).float(), x[:, 1:].reshape(-1), reduction="none"
280
+ )
281
+ main = (ce * m[:, 1:].reshape(-1)).sum() / m[:, 1:].sum()
282
+ # Position t (0..T-3) gets h_t and emb(x_{t+1}); causal attention over these keeps every
283
+ # prediction of x_{t+2} a function of x_{<=t+1}.
284
+ z = mtp.proj(torch.cat((mtp.ln_h(h[:, :-1]), mtp.ln_e(model.emb(x[:, 1:-1]))), -1))
285
+ z = mtp.block(z, None, None, bias[..., :-1, :-1], None)
286
+ ce2 = torch.nn.functional.cross_entropy(
287
+ model.head(model.ln_f(z)).reshape(-1, V).float(), x[:, 2:].reshape(-1), reduction="none"
288
+ )
289
+ aux = (ce2 * m[:, 2:].reshape(-1)).sum() / m[:, 2:].sum().clamp(min=1)
290
+ return main + MTP_WEIGHT * aux, main, aux
291
+
292
+
293
+ def val_loss(model, ids, mask, ref_max, lens, tok, gen, autocast):
294
+ """Mean loss over the first VAL_N records (never trained on), no shift augmentation."""
295
+ model.eval()
296
+ tot = 0.0
297
+ with torch.no_grad(), autocast:
298
+ for i in range(0, VAL_N, BS):
299
+ idxs = torch.arange(i, min(i + BS, VAL_N), device=ids.device)
300
+ x, m = ids[idxs], mask[idxs]
301
+ T = int(lens[idxs].max())
302
+ tot += loss_on(model, x[:, :T], m[:, :T]).item() * len(idxs)
303
+ model.train()
304
+ return tot / VAL_N
305
+
306
+
307
+ def to_gpu(data, tok, device):
308
+ """Pad all (prompt, proof) records once into (N, T) ids + proof mask on the device."""
309
+ N, T = len(data), max(len(p) + len(q) for p, q in data)
310
+ ids = torch.full((N, T), tok.pad, dtype=torch.long)
311
+ mask = torch.zeros((N, T), dtype=torch.bool)
312
+ ref_max = torch.zeros(
313
+ N, dtype=torch.long
314
+ ) # largest N<k> index used in the proof, for the shift augmentation
315
+ for i, (p, q) in enumerate(data):
316
+ ids[i, : len(p) + len(q)] = torch.tensor(p + q)
317
+ mask[i, len(p) : len(p) + len(q)] = True
318
+ ref_max[i] = max((x - tok.ref0 + 1 for x in q if x >= tok.ref0), default=0)
319
+ return ids.to(device), mask.to(device), ref_max.to(device), (ids != tok.pad).sum(1).to(device)
320
+
321
+
322
+ def token_batches(lens, n_train, gen, rng, first=VAL_N):
323
+ """One epoch of token-budget batches: a random permutation of the n_train records from
324
+ `first` on, stably sorted by length (equal lengths stay shuffled), cut greedily into runs of
325
+ consecutive records with count * max_len <= TOK_BUDGET, and the runs' order shuffled.
326
+ Returns (order, cuts)."""
327
+ perm = torch.randperm(n_train, device=lens.device, generator=gen) + first
328
+ order = perm[torch.argsort(lens[perm], stable=True)]
329
+ sorted_lens = lens[order].cpu()
330
+ cuts, s = [], 0
331
+ while s < len(sorted_lens):
332
+ window = sorted_lens[s : s + TOK_BUDGET // int(sorted_lens[s])]
333
+ # count * max_len is non-decreasing along the window, so the fits form a prefix.
334
+ e = s + int((torch.arange(1, len(window) + 1) * window <= TOK_BUDGET).sum())
335
+ cuts.append((s, e, int(sorted_lens[e - 1]))) # sorted: last is the batch's max length
336
+ s = e
337
+ rng.shuffle(cuts)
338
+ return order, cuts
339
+
340
+
341
+ def gpu_batch(ids, mask, ref_max, lens, idxs, tok, gen, max_len=None):
342
+ """The fork's batch() + tok.shift_abs, vectorised: a random per-row offset is added to
343
+ every ref token."""
344
+ x, m = ids[idxs], mask[idxs]
345
+ # Known on the host for training batches: passing it avoids a GPU sync every step.
346
+ T = int(lens[idxs].max()) if max_len is None else max_len
347
+ x, m = x[:, :T].clone(), m[:, :T]
348
+ if tok.mode in ("abs", "lean_seq"): # lean_seq: names n1.. get the same start offset
349
+ room = MAXN - ref_max[idxs]
350
+ shift = (
351
+ (torch.rand(len(idxs), device=x.device, generator=gen) * (room + 1))
352
+ .long()
353
+ .clamp(max=room)
354
+ )
355
+ is_ref = (x >= tok.ref0) & m
356
+ x = torch.where(is_ref, x + shift[:, None], x)
357
+ return x, m
358
+
359
+
360
+ def pretrain(train_recs, budget_secs, device, seed, log):
361
+ t0 = time.time()
362
+ tok = make_tokenizer(TOK_MODE)
363
+ data = [(tok.encode_prompt(r["prompt"]), tok.encode_proof(r["proof"])) for r in train_recs]
364
+ ids, mask, ref_max, lens = to_gpu(data, tok, device)
365
+ gen = torch.Generator(device=device)
366
+ gen.manual_seed(seed)
367
+ model = LoopedGPT(tok.vocab_size, N_LAYER, D, N_HEAD, seed)
368
+ model.blocks = torch.nn.ModuleList(PeriLNBlock(D, N_HEAD) for _ in range(N_LAYER))
369
+ model.blocks.apply(model._init) # fork init: zero biases
370
+ model = model.to(device)
371
+ mtp = MTPHead(D, N_HEAD)
372
+ mtp.apply(model._init)
373
+ mtp = mtp.to(device)
374
+ # Compile the repeated blocks once each (torch >= 2.9; dynamic shapes, ~14 s): ~14% faster
375
+ # steps. The coalescing tiling analysis crashes on dynamic shapes, so it is off. Undone at
376
+ # handoff so the harness's RL fine-tune and KV-cache sampling run eager.
377
+ torch._inductor.config.triton.coalesce_tiling_analysis = False
378
+ muon_lr, muon_graph = torch.zeros((), device=device), None
379
+ compiled = list(model.blocks) + [mtp.block]
380
+ for b in compiled if device == "cuda" else []: # CPU smoke runs stay eager
381
+ b.compile(dynamic=True)
382
+ for p in list(model.blocks.parameters()) + list(mtp.parameters()):
383
+ if p.ndim == 2:
384
+ # Orthogonal matrices have element second moment gain^2 / max(rows, cols).
385
+ torch.nn.init.orthogonal_(p, gain=0.02 * math.sqrt(max(p.shape)))
386
+ log(
387
+ f"pretrain: {model.n_params():,} params, {len(data):,} records on {device} "
388
+ f"after {time.time() - t0:.0f}s, budget {budget_secs}s"
389
+ )
390
+ # The output readout joins the hidden matrices under Muon; embeddings stay on AdamW.
391
+ trainable = list(model.parameters()) + list(mtp.parameters())
392
+ matrices = [p for p in model.blocks.parameters() if p.ndim == 2] + [model.head.weight]
393
+ matrices += [p for p in mtp.parameters() if p.ndim == 2]
394
+ matrix_ids = {id(p) for p in matrices}
395
+ auxiliary = [p for p in trainable if id(p) not in matrix_ids]
396
+ momentum_buffers = [torch.zeros_like(p) for p in matrices]
397
+ opt = torch.optim.AdamW(
398
+ auxiliary, lr=LR, weight_decay=WD, betas=(0.9, 0.95), fused=device == "cuda"
399
+ )
400
+ autocast = torch.autocast("cuda", dtype=torch.bfloat16, enabled=(device == "cuda"))
401
+ model.train()
402
+ step, last_loss, records_seen = 0, float("nan"), 0
403
+ curve, recent, recent_aux, next_point = [], [], [], CURVE_EVERY
404
+ n_train = len(data) - VAL_N # records [0, VAL_N) are the val slice
405
+ rng, cuts, batches_per_epoch = random.Random(seed), [], None
406
+ order = torch.empty(0, dtype=torch.long, device=device)
407
+ while True:
408
+ frac = (
409
+ time.time() - t0
410
+ ) / budget_secs # cosine over wall-clock, so the cap is the schedule
411
+ if frac >= 1.0:
412
+ break
413
+ step += 1
414
+ # No warmup (as in modded-nanogpt): Muon's step is normalized and Adam's is ~lr per
415
+ # element from step 1, and Peri-LN normalizes every branch write.
416
+ lr = LR * MIN_LR_FRAC + 0.5 * LR * (1 - MIN_LR_FRAC) * (1 + math.cos(math.pi * frac))
417
+ for g in opt.param_groups:
418
+ g["lr"] = lr
419
+ if not cuts: # new epoch
420
+ order, cuts = token_batches(lens, n_train, gen, rng)
421
+ batches_per_epoch = len(cuts)
422
+ s, e, T = cuts.pop()
423
+ idxs = order[s:e]
424
+ records_seen += e - s
425
+ x, m = gpu_batch(ids, mask, ref_max, lens, idxs, tok, gen, T)
426
+ with autocast:
427
+ loss, main, aux = mtp_loss(model, mtp, x, m)
428
+ model.zero_grad(set_to_none=False) # grads keep fixed addresses for the Muon CUDA graph
429
+ mtp.zero_grad(set_to_none=False)
430
+ loss.backward()
431
+ torch.nn.utils.clip_grad_norm_(trainable, 1.0)
432
+ opt.step()
433
+ muon_lr.fill_(MUON_LR * lr / LR)
434
+ # Capture once; later steps replay one graph instead of ~400 kernel launches. CPU smoke
435
+ # runs call the step directly.
436
+ if device != "cuda":
437
+ muon_step(matrices, momentum_buffers, muon_lr)
438
+ elif muon_graph is None:
439
+ state = matrices + momentum_buffers
440
+ saved = [t.clone() for t in state]
441
+ side = torch.cuda.Stream()
442
+ side.wait_stream(torch.cuda.current_stream())
443
+ with torch.cuda.stream(side):
444
+ muon_step(matrices, momentum_buffers, muon_lr) # warm-up (lazy cuBLAS init)
445
+ torch.cuda.current_stream().wait_stream(side)
446
+ muon_graph = torch.cuda.CUDAGraph()
447
+ with torch.cuda.graph(muon_graph):
448
+ muon_step(matrices, momentum_buffers, muon_lr)
449
+ with torch.no_grad():
450
+ for t, v in zip(state, saved, strict=True):
451
+ t.copy_(v) # undo the warm-up and capture updates
452
+ muon_graph.replay()
453
+ else:
454
+ muon_graph.replay()
455
+ if step % 20 == 0:
456
+ last_loss = main.item() # the next-token CE, comparable with earlier runs
457
+ recent.append(last_loss)
458
+ recent_aux.append(aux.item())
459
+ if (
460
+ time.time() - t0 >= next_point
461
+ ): # curve point: mean recent train loss + val loss on the held-out slice
462
+ curve.append(
463
+ {
464
+ "secs": round(time.time() - t0),
465
+ "step": step,
466
+ "train_loss": sum(recent) / len(recent) if recent else last_loss,
467
+ "val_loss": val_loss(model, ids, mask, ref_max, lens, tok, gen, autocast),
468
+ "mtp_loss": sum(recent_aux) / len(recent_aux) if recent_aux else None,
469
+ }
470
+ )
471
+ log(
472
+ f"pretrain step {step} train {curve[-1]['train_loss']:.4f} "
473
+ f"val {curve[-1]['val_loss']:.4f} lr {lr:.2e} {curve[-1]['secs']}s"
474
+ )
475
+ recent, recent_aux, next_point = [], [], next_point + CURVE_EVERY
476
+ val_by_loops = {}
477
+ for n in VAL_LOOPS if step else (): # no steps: a checkpoint rebuild, not a training run
478
+ model.n_loops = n
479
+ val_by_loops[n] = val_loss(model, ids, mask, ref_max, lens, tok, gen, autocast)
480
+ model.n_loops = DEFAULT_LOOPS
481
+ log(f"pretrain val loss by loops: {val_by_loops}")
482
+ for b in compiled:
483
+ b._compiled_call_impl = None
484
+ model.eval()
485
+ return (
486
+ model,
487
+ tok,
488
+ {
489
+ "steps": step,
490
+ "records_seen": records_seen,
491
+ "epochs": records_seen / n_train,
492
+ "batches_per_epoch": batches_per_epoch,
493
+ "tok_budget": TOK_BUDGET,
494
+ "last_loss": last_loss,
495
+ "train_loops": TRAIN_LOOPS,
496
+ "default_loops": DEFAULT_LOOPS,
497
+ "val_by_loops": val_by_loops,
498
+ "curve": curve,
499
+ },
500
+ )
501
+
502
+
503
+ def finetune(model, tok, data, rng, device):
504
+ """The old harness's RL fine-tune: AdamW, linear warmup then cosine to lr/10, over FT_RECORDS
505
+ records in token-budget batches (to_gpu / token_batches / gpu_batch, as pretraining). The
506
+ model is in train mode, so every batch draws its own loop count from TRAIN_LOOPS. Returns
507
+ (last loss, steps)."""
508
+ ids, mask, ref_max, lens = to_gpu(data, tok, device)
509
+ gen = torch.Generator(device=device)
510
+ gen.manual_seed(rng.getrandbits(31))
511
+ plan, seen = [], 0 # (order, start, end, padded length) per step, epoch after epoch
512
+ while seen < FT_RECORDS:
513
+ order, cuts = token_batches(lens, len(data), gen, rng, first=0)
514
+ for s, e, T in cuts:
515
+ if seen >= FT_RECORDS:
516
+ break
517
+ plan.append((order, s, e, T))
518
+ seen += e - s
519
+ opt = torch.optim.AdamW(model.parameters(), lr=FT_LR, weight_decay=0.1, betas=(0.9, 0.95))
520
+ autocast = torch.autocast("cuda", dtype=torch.bfloat16, enabled=(device == "cuda"))
521
+ model.train()
522
+ for step, (order, s, e, T) in enumerate(plan, 1):
523
+ lr = (
524
+ FT_LR * step / FT_WARMUP
525
+ if step < FT_WARMUP
526
+ else FT_LR / 10 + 0.45 * FT_LR * (1 + math.cos(math.pi * step / len(plan)))
527
+ )
528
+ for g in opt.param_groups:
529
+ g["lr"] = lr
530
+ x, m = gpu_batch(ids, mask, ref_max, lens, order[s:e], tok, gen, T)
531
+ with autocast:
532
+ loss = loss_on(model, x, m)
533
+ opt.zero_grad(set_to_none=True)
534
+ loss.backward()
535
+ torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
536
+ opt.step()
537
+ model.eval()
538
+ return loss.item(), len(plan)
539
+
540
+
541
+ def rl(model, tok, targets, train_recs, oracle, budget_secs, device, seed, log):
542
+ rng = random.Random(seed)
543
+ found = {t["name"]: {} for t in targets} # name -> {normalized proof: proof text}
544
+ rounds, round_secs, retries, ft_steps = 0, 0.0, [], []
545
+ # Start a round only if one as long as the last (plus 15%) still fits in the budget.
546
+ while oracle.left() > 1.15 * round_secs:
547
+ t0 = oracle.used()
548
+ model.n_loops = RL_FIRST_LOOPS
549
+ rows = oracle.sample(model, targets, RL_K)
550
+ failed = [t for t, row in zip(targets, rows, strict=True) if not row["proofs"]]
551
+ rng.shuffle(failed)
552
+ for i, n in enumerate(RL_RETRY_LOOPS):
553
+ part = failed[i :: len(RL_RETRY_LOOPS)]
554
+ if not part:
555
+ continue
556
+ model.n_loops = n
557
+ part_rows = oracle.sample(model, part, RL_RETRY_K)
558
+ cracked = sum(bool(row["proofs"]) for row in part_rows)
559
+ retries.append(
560
+ {"round": rounds + 1, "loops": n, "targets": len(part), "cracked": cracked}
561
+ )
562
+ rows += part_rows
563
+ for row in rows:
564
+ for proof in row["proofs"]:
565
+ found[row["name"]].setdefault(norm(proof), proof)
566
+ mix = []
567
+ for t in targets:
568
+ proofs = list(found[t["name"]].values())
569
+ rng.shuffle(proofs)
570
+ mix += [(t["prompt"], proof) for proof in proofs[:MAX_PER_THM]] * RL_WEIGHT
571
+ n_rl = len(mix)
572
+ if not n_rl:
573
+ break
574
+ mix += [(r["prompt"], r["proof"]) for r in rng.sample(train_recs, RETAIN)]
575
+ data = [(tok.encode_prompt(p), tok.encode_proof(q)) for p, q in mix]
576
+ loss, steps = finetune(model, tok, data, rng, device)
577
+ ft_steps.append(steps)
578
+ rounds += 1
579
+ round_secs = oracle.used() - t0
580
+ log(
581
+ f"rl round {rounds}: {n_rl} RL records, fine-tune loss {loss:.4f} in {steps} steps, "
582
+ f"round {round_secs:.0f}s, {oracle.left():.0f}s left"
583
+ )
584
+ model.n_loops = DEFAULT_LOOPS
585
+ return {
586
+ "rounds": rounds,
587
+ "first_loops": RL_FIRST_LOOPS,
588
+ "k_first": RL_K,
589
+ "retry_loops": RL_RETRY_LOOPS,
590
+ "k_retry": RL_RETRY_K,
591
+ "retries": retries,
592
+ "ft_records": FT_RECORDS,
593
+ "ft_steps": ft_steps,
594
+ }
runs/590_control-untied-198_s1/source/remote.sh ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env bash
2
+ # Runs on the pod. Checks out Dan's fork at the pinned commit (frozen data + verifier + Lean
3
+ # tokenizer/gate), installs Lean, runs the harness.
4
+ # Usage: bash remote.sh <out-dir> [harness.py args...]
5
+ set -euo pipefail
6
+ FORK_COMMIT=51604b3feaa5e66821d1faba22238fe6d96d46d0 # dan-pandori/nd-takehome, branch dan_lean_seed2 (lean_seq with the BOTE fix)
7
+ export ND_FORK=/workspace/nd-takehome
8
+ OUT=$1; shift
9
+
10
+ if [ ! -d "$ND_FORK/.git" ]; then
11
+ git clone -q https://github.com/dan-pandori/nd-takehome.git "$ND_FORK"
12
+ fi
13
+ git -C "$ND_FORK" cat-file -e "$FORK_COMMIT" 2>/dev/null || git -C "$ND_FORK" fetch -q origin
14
+ git -C "$ND_FORK" checkout -q "$FORK_COMMIT"
15
+ [ -f "$ND_FORK/data/train.jsonl" ] || gunzip -k "$ND_FORK/data/train.jsonl.gz"
16
+ # Lean 4 core checks every sampled proof (fork lean_gate.py, called from sample.generate): the
17
+ # reward is Lean AND nd_verify. Installed before harness.py starts, so it is not charged to TOTAL_SECS.
18
+ [ -x ~/.elan/bin/lean ] || curl -sSf https://raw.githubusercontent.com/leanprover/elan/master/elan-init.sh \
19
+ | sh -s -- -y --default-toolchain leanprover/lean4:v4.34.0 > /dev/null
20
+ # One Lean process per CPU: on these pods (nproc 96) 12 workers took 146 s per RL-round gate
21
+ # pass (099) against 67-84 s with 96 workers (098).
22
+ export LEAN_GATE_WORKERS=$(nproc) LEAN_GATE_LOG="$OUT/lean_gate.jsonl"
23
+ nvidia-smi --query-gpu=name,memory.total --format=csv,noheader
24
+ echo "cpus $(nproc), $(~/.elan/bin/lean --version)"
25
+ # Newer RunPod images put torch under `python` and leave `python3` as bare system Python.
26
+ PY=$(for p in python python3; do "$p" -c "import torch" 2>/dev/null && echo "$p" && break; done)
27
+ echo "python $($PY --version 2>&1), torch $($PY -c 'import torch; print(torch.__version__)')"
28
+ cd /workspace/autoresearch
29
+ $PY harness.py --out "$OUT" "$@" 2>&1 | tee "$OUT.log"
runs/590_control-untied-198_s1/summary.json ADDED
@@ -0,0 +1,2858 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "ok",
3
+ "metric": 785,
4
+ "metric_diag": 785,
5
+ "metric_rl": 1307,
6
+ "harness_sha": "2eec1b61666a",
7
+ "pretrain_sha": "b27dd2cb03ba",
8
+ "device": "cuda",
9
+ "gpu": "NVIDIA A40",
10
+ "torch": "2.9.1+cu129",
11
+ "lean_gate": {
12
+ "samples": 506496,
13
+ "parse_fail": 103614,
14
+ "distinct_checked": 361458,
15
+ "both_ok": 215272,
16
+ "nd_ok_lean_rej": 0,
17
+ "nd_rej_lean_ok": 1,
18
+ "lean_wall_s": 1182.3275501728058
19
+ },
20
+ "total_secs": 3700.6044154167175,
21
+ "rules": {
22
+ "TOTAL_SECS": 5400,
23
+ "PRETRAIN_SECS": 300,
24
+ "RL_SECS": 1800,
25
+ "PARAM_CAP": 10000000,
26
+ "N_TARGETS": 1500,
27
+ "EVAL_K": 64,
28
+ "KNOB": "n_loops",
29
+ "EVAL_LEVEL": 8,
30
+ "DIAG_LEVEL": 2,
31
+ "TOK_MODE": "lean_seq"
32
+ },
33
+ "pretrain": {
34
+ "steps": 5391,
35
+ "records_seen": 732958,
36
+ "epochs": 4.760518556044841,
37
+ "batches_per_epoch": 1132,
38
+ "tok_budget": 18432,
39
+ "last_loss": 0.01816931925714016,
40
+ "train_loops": [
41
+ 1,
42
+ 2,
43
+ 3,
44
+ 4,
45
+ 5,
46
+ 6,
47
+ 7,
48
+ 8
49
+ ],
50
+ "default_loops": 2,
51
+ "val_by_loops": {
52
+ "1": 0.07875961344689131,
53
+ "2": 0.07875961344689131,
54
+ "4": 0.07875961344689131,
55
+ "8": 0.07875961344689131
56
+ },
57
+ "curve": [
58
+ {
59
+ "secs": 32,
60
+ "step": 1,
61
+ "train_loss": NaN,
62
+ "val_loss": 3.8513625860214233,
63
+ "mtp_loss": null
64
+ },
65
+ {
66
+ "secs": 34,
67
+ "step": 2,
68
+ "train_loss": NaN,
69
+ "val_loss": 3.437624543905258,
70
+ "mtp_loss": null
71
+ },
72
+ {
73
+ "secs": 45,
74
+ "step": 223,
75
+ "train_loss": 0.43221770836548373,
76
+ "val_loss": 0.12727239541709423,
77
+ "mtp_loss": 0.4301349811933257
78
+ },
79
+ {
80
+ "secs": 60,
81
+ "step": 530,
82
+ "train_loss": 0.10062761182586352,
83
+ "val_loss": 0.09747785702347755,
84
+ "mtp_loss": 0.09841699600219726
85
+ },
86
+ {
87
+ "secs": 75,
88
+ "step": 833,
89
+ "train_loss": 0.0774573711057504,
90
+ "val_loss": 0.09270505420863628,
91
+ "mtp_loss": 0.08087193816900254
92
+ },
93
+ {
94
+ "secs": 90,
95
+ "step": 1137,
96
+ "train_loss": 0.0704626532892386,
97
+ "val_loss": 0.08240908943116665,
98
+ "mtp_loss": 0.07202240005135536
99
+ },
100
+ {
101
+ "secs": 105,
102
+ "step": 1438,
103
+ "train_loss": 0.07201575761040052,
104
+ "val_loss": 0.09006544388830662,
105
+ "mtp_loss": 0.0725412093102932
106
+ },
107
+ {
108
+ "secs": 120,
109
+ "step": 1744,
110
+ "train_loss": 0.06276722974143922,
111
+ "val_loss": 0.09227755293250084,
112
+ "mtp_loss": 0.0634800880216062
113
+ },
114
+ {
115
+ "secs": 135,
116
+ "step": 2050,
117
+ "train_loss": 0.07109759251276652,
118
+ "val_loss": 0.08788908645510674,
119
+ "mtp_loss": 0.07061384568611781
120
+ },
121
+ {
122
+ "secs": 150,
123
+ "step": 2353,
124
+ "train_loss": 0.059425455828507744,
125
+ "val_loss": 0.09058903809636831,
126
+ "mtp_loss": 0.06017695019642512
127
+ },
128
+ {
129
+ "secs": 165,
130
+ "step": 2657,
131
+ "train_loss": 0.06261006941397985,
132
+ "val_loss": 0.08938852604478598,
133
+ "mtp_loss": 0.06269428183635076
134
+ },
135
+ {
136
+ "secs": 180,
137
+ "step": 2960,
138
+ "train_loss": 0.05555315804667771,
139
+ "val_loss": 0.08441854082047939,
140
+ "mtp_loss": 0.05599284381605685
141
+ },
142
+ {
143
+ "secs": 195,
144
+ "step": 3265,
145
+ "train_loss": 0.056754926592111586,
146
+ "val_loss": 0.07835845742374659,
147
+ "mtp_loss": 0.05724384561181069
148
+ },
149
+ {
150
+ "secs": 210,
151
+ "step": 3568,
152
+ "train_loss": 0.06073875253399213,
153
+ "val_loss": 0.081726199015975,
154
+ "mtp_loss": 0.06188008263707161
155
+ },
156
+ {
157
+ "secs": 225,
158
+ "step": 3872,
159
+ "train_loss": 0.060468320300181705,
160
+ "val_loss": 0.08574883732944727,
161
+ "mtp_loss": 0.05994002968072891
162
+ },
163
+ {
164
+ "secs": 240,
165
+ "step": 4176,
166
+ "train_loss": 0.04860136831800143,
167
+ "val_loss": 0.07705918606370687,
168
+ "mtp_loss": 0.048752292742331824
169
+ },
170
+ {
171
+ "secs": 255,
172
+ "step": 4480,
173
+ "train_loss": 0.05323238379787654,
174
+ "val_loss": 0.09137827903032303,
175
+ "mtp_loss": 0.05253439326770604
176
+ },
177
+ {
178
+ "secs": 270,
179
+ "step": 4784,
180
+ "train_loss": 0.058249530692895256,
181
+ "val_loss": 0.08194581512361765,
182
+ "mtp_loss": 0.05817735555271308
183
+ },
184
+ {
185
+ "secs": 285,
186
+ "step": 5087,
187
+ "train_loss": 0.05181963207821051,
188
+ "val_loss": 0.07721857819706202,
189
+ "mtp_loss": 0.051862321545680365
190
+ },
191
+ {
192
+ "secs": 300,
193
+ "step": 5391,
194
+ "train_loss": 0.046261887749036154,
195
+ "val_loss": 0.07875961344689131,
196
+ "mtp_loss": 0.046748237187663716
197
+ }
198
+ ],
199
+ "secs": 300.6973946094513,
200
+ "n_params": 9558528,
201
+ "heldout_greedy": 0.9745
202
+ },
203
+ "rl": {
204
+ "rounds": [
205
+ {
206
+ "round": 1,
207
+ "level": 1,
208
+ "k": 8,
209
+ "n_targets": 1500,
210
+ "sample_acc": 0.10275,
211
+ "new_proofs": 378,
212
+ "targets_solved": 317,
213
+ "lstar": 9,
214
+ "ge": {
215
+ "7": 317,
216
+ "8": 63,
217
+ "9": 15,
218
+ "10": 1,
219
+ "11": 0,
220
+ "12": 0,
221
+ "13": 0,
222
+ "14": 0
223
+ },
224
+ "mix_rl_records": 378,
225
+ "sampling": {
226
+ "samples": 12000,
227
+ "secs": 56.55860733985901,
228
+ "gate_secs": 22.915697813034058,
229
+ "samples_per_sec": 212.16929773203842
230
+ },
231
+ "secs": 56.894145488739014,
232
+ "rl_secs_used": 33.9788544178009
233
+ },
234
+ {
235
+ "round": 2,
236
+ "level": 2,
237
+ "k": 24,
238
+ "n_targets": 395,
239
+ "sample_acc": 0.008227848101265823,
240
+ "new_proofs": 40,
241
+ "targets_solved": 354,
242
+ "lstar": 9,
243
+ "ge": {
244
+ "7": 354,
245
+ "8": 76,
246
+ "9": 18,
247
+ "10": 1,
248
+ "11": 0,
249
+ "12": 0,
250
+ "13": 0,
251
+ "14": 0
252
+ },
253
+ "mix_rl_records": 40,
254
+ "sampling": {
255
+ "samples": 9480,
256
+ "secs": 43.05401968955994,
257
+ "gate_secs": 15.164865255355835,
258
+ "samples_per_sec": 220.18849966519576
259
+ },
260
+ "secs": 43.28769612312317,
261
+ "rl_secs_used": 62.10140872001648
262
+ },
263
+ {
264
+ "round": 3,
265
+ "level": 4,
266
+ "k": 24,
267
+ "n_targets": 394,
268
+ "sample_acc": 0.008989001692047378,
269
+ "new_proofs": 50,
270
+ "targets_solved": 400,
271
+ "lstar": 9,
272
+ "ge": {
273
+ "7": 400,
274
+ "8": 89,
275
+ "9": 21,
276
+ "10": 1,
277
+ "11": 0,
278
+ "12": 0,
279
+ "13": 0,
280
+ "14": 0
281
+ },
282
+ "mix_rl_records": 50,
283
+ "sampling": {
284
+ "samples": 9456,
285
+ "secs": 45.54316973686218,
286
+ "gate_secs": 16.19222331047058,
287
+ "samples_per_sec": 207.6271821797772
288
+ },
289
+ "secs": 45.77345609664917,
290
+ "rl_secs_used": 91.68258929252625
291
+ },
292
+ {
293
+ "round": 4,
294
+ "level": 8,
295
+ "k": 24,
296
+ "n_targets": 394,
297
+ "sample_acc": 0.009412013536379018,
298
+ "new_proofs": 53,
299
+ "targets_solved": 444,
300
+ "lstar": 9,
301
+ "ge": {
302
+ "7": 444,
303
+ "8": 101,
304
+ "9": 25,
305
+ "10": 1,
306
+ "11": 0,
307
+ "12": 0,
308
+ "13": 0,
309
+ "14": 0
310
+ },
311
+ "mix_rl_records": 53,
312
+ "sampling": {
313
+ "samples": 9456,
314
+ "secs": 47.78762173652649,
315
+ "gate_secs": 15.28433871269226,
316
+ "samples_per_sec": 197.8755095228416
317
+ },
318
+ "secs": 48.01021099090576,
319
+ "rl_secs_used": 124.40845036506653
320
+ },
321
+ {
322
+ "round": 5,
323
+ "level": 1,
324
+ "k": 8,
325
+ "n_targets": 1500,
326
+ "sample_acc": 0.49383333333333335,
327
+ "new_proofs": 631,
328
+ "targets_solved": 874,
329
+ "lstar": 10,
330
+ "ge": {
331
+ "7": 874,
332
+ "8": 390,
333
+ "9": 163,
334
+ "10": 26,
335
+ "11": 3,
336
+ "12": 0,
337
+ "13": 0,
338
+ "14": 0
339
+ },
340
+ "mix_rl_records": 1130,
341
+ "sampling": {
342
+ "samples": 12000,
343
+ "secs": 74.854248046875,
344
+ "gate_secs": 34.87635159492493,
345
+ "samples_per_sec": 160.31154293989295
346
+ },
347
+ "secs": 105.02026152610779,
348
+ "rl_secs_used": 194.55279231071472
349
+ },
350
+ {
351
+ "round": 6,
352
+ "level": 2,
353
+ "k": 24,
354
+ "n_targets": 209,
355
+ "sample_acc": 0.008971291866028708,
356
+ "new_proofs": 24,
357
+ "targets_solved": 891,
358
+ "lstar": 10,
359
+ "ge": {
360
+ "7": 891,
361
+ "8": 403,
362
+ "9": 171,
363
+ "10": 30,
364
+ "11": 3,
365
+ "12": 0,
366
+ "13": 0,
367
+ "14": 0
368
+ },
369
+ "mix_rl_records": 24,
370
+ "sampling": {
371
+ "samples": 5016,
372
+ "secs": 32.617483377456665,
373
+ "gate_secs": 14.483319997787476,
374
+ "samples_per_sec": 153.78255710147067
375
+ },
376
+ "secs": 32.80847525596619,
377
+ "rl_secs_used": 212.87750339508057
378
+ },
379
+ {
380
+ "round": 7,
381
+ "level": 4,
382
+ "k": 24,
383
+ "n_targets": 209,
384
+ "sample_acc": 0.01076555023923445,
385
+ "new_proofs": 27,
386
+ "targets_solved": 911,
387
+ "lstar": 10,
388
+ "ge": {
389
+ "7": 911,
390
+ "8": 417,
391
+ "9": 182,
392
+ "10": 33,
393
+ "11": 3,
394
+ "12": 0,
395
+ "13": 0,
396
+ "14": 0
397
+ },
398
+ "mix_rl_records": 27,
399
+ "sampling": {
400
+ "samples": 5016,
401
+ "secs": 31.73079490661621,
402
+ "gate_secs": 16.58519458770752,
403
+ "samples_per_sec": 158.0798720852124
404
+ },
405
+ "secs": 31.929019927978516,
406
+ "rl_secs_used": 228.22135138511658
407
+ },
408
+ {
409
+ "round": 8,
410
+ "level": 8,
411
+ "k": 24,
412
+ "n_targets": 209,
413
+ "sample_acc": 0.006977671451355662,
414
+ "new_proofs": 18,
415
+ "targets_solved": 925,
416
+ "lstar": 10,
417
+ "ge": {
418
+ "7": 925,
419
+ "8": 427,
420
+ "9": 191,
421
+ "10": 33,
422
+ "11": 3,
423
+ "12": 0,
424
+ "13": 0,
425
+ "14": 0
426
+ },
427
+ "mix_rl_records": 18,
428
+ "sampling": {
429
+ "samples": 5016,
430
+ "secs": 30.51624369621277,
431
+ "gate_secs": 14.732860088348389,
432
+ "samples_per_sec": 164.37147539959227
433
+ },
434
+ "secs": 30.70336890220642,
435
+ "rl_secs_used": 244.19187879562378
436
+ },
437
+ {
438
+ "round": 9,
439
+ "level": 1,
440
+ "k": 8,
441
+ "n_targets": 1500,
442
+ "sample_acc": 0.635,
443
+ "new_proofs": 289,
444
+ "targets_solved": 1043,
445
+ "lstar": 11,
446
+ "ge": {
447
+ "7": 1043,
448
+ "8": 513,
449
+ "9": 255,
450
+ "10": 56,
451
+ "11": 6,
452
+ "12": 1,
453
+ "13": 0,
454
+ "14": 0
455
+ },
456
+ "mix_rl_records": 1440,
457
+ "sampling": {
458
+ "samples": 12000,
459
+ "secs": 68.91086292266846,
460
+ "gate_secs": 35.473225831985474,
461
+ "samples_per_sec": 174.13800221115153
462
+ },
463
+ "secs": 101.98323249816895,
464
+ "rl_secs_used": 310.7021028995514
465
+ },
466
+ {
467
+ "round": 10,
468
+ "level": 2,
469
+ "k": 24,
470
+ "n_targets": 153,
471
+ "sample_acc": 0.008714596949891068,
472
+ "new_proofs": 13,
473
+ "targets_solved": 1053,
474
+ "lstar": 11,
475
+ "ge": {
476
+ "7": 1053,
477
+ "8": 520,
478
+ "9": 260,
479
+ "10": 59,
480
+ "11": 8,
481
+ "12": 1,
482
+ "13": 0,
483
+ "14": 0
484
+ },
485
+ "mix_rl_records": 13,
486
+ "sampling": {
487
+ "samples": 3672,
488
+ "secs": 23.61838722229004,
489
+ "gate_secs": 11.347397804260254,
490
+ "samples_per_sec": 155.47208898897725
491
+ },
492
+ "secs": 23.760610103607178,
493
+ "rl_secs_used": 323.1150324344635
494
+ },
495
+ {
496
+ "round": 11,
497
+ "level": 4,
498
+ "k": 24,
499
+ "n_targets": 153,
500
+ "sample_acc": 0.0027233115468409588,
501
+ "new_proofs": 7,
502
+ "targets_solved": 1060,
503
+ "lstar": 11,
504
+ "ge": {
505
+ "7": 1060,
506
+ "8": 526,
507
+ "9": 265,
508
+ "10": 59,
509
+ "11": 8,
510
+ "12": 1,
511
+ "13": 0,
512
+ "14": 0
513
+ },
514
+ "mix_rl_records": 7,
515
+ "sampling": {
516
+ "samples": 3672,
517
+ "secs": 22.52391219139099,
518
+ "gate_secs": 12.568029642105103,
519
+ "samples_per_sec": 163.02674103850833
520
+ },
521
+ "secs": 22.666586875915527,
522
+ "rl_secs_used": 333.2136092185974
523
+ },
524
+ {
525
+ "round": 12,
526
+ "level": 8,
527
+ "k": 24,
528
+ "n_targets": 152,
529
+ "sample_acc": 0.0027412280701754384,
530
+ "new_proofs": 8,
531
+ "targets_solved": 1067,
532
+ "lstar": 11,
533
+ "ge": {
534
+ "7": 1067,
535
+ "8": 533,
536
+ "9": 271,
537
+ "10": 62,
538
+ "11": 10,
539
+ "12": 1,
540
+ "13": 0,
541
+ "14": 0
542
+ },
543
+ "mix_rl_records": 8,
544
+ "sampling": {
545
+ "samples": 3648,
546
+ "secs": 25.3083758354187,
547
+ "gate_secs": 12.318823337554932,
548
+ "samples_per_sec": 144.14200356921668
549
+ },
550
+ "secs": 25.451199531555176,
551
+ "rl_secs_used": 346.3459806442261
552
+ },
553
+ {
554
+ "round": 13,
555
+ "level": 1,
556
+ "k": 8,
557
+ "n_targets": 1500,
558
+ "sample_acc": 0.6931666666666667,
559
+ "new_proofs": 203,
560
+ "targets_solved": 1109,
561
+ "lstar": 11,
562
+ "ge": {
563
+ "7": 1109,
564
+ "8": 562,
565
+ "9": 292,
566
+ "10": 69,
567
+ "11": 12,
568
+ "12": 1,
569
+ "13": 0,
570
+ "14": 0
571
+ },
572
+ "mix_rl_records": 1610,
573
+ "sampling": {
574
+ "samples": 12000,
575
+ "secs": 64.8794801235199,
576
+ "gate_secs": 33.69963502883911,
577
+ "samples_per_sec": 184.95832545442667
578
+ },
579
+ "secs": 99.8878059387207,
580
+ "rl_secs_used": 412.53439831733704
581
+ },
582
+ {
583
+ "round": 14,
584
+ "level": 2,
585
+ "k": 24,
586
+ "n_targets": 131,
587
+ "sample_acc": 0.0028625954198473282,
588
+ "new_proofs": 6,
589
+ "targets_solved": 1115,
590
+ "lstar": 11,
591
+ "ge": {
592
+ "7": 1115,
593
+ "8": 567,
594
+ "9": 296,
595
+ "10": 70,
596
+ "11": 12,
597
+ "12": 1,
598
+ "13": 0,
599
+ "14": 0
600
+ },
601
+ "mix_rl_records": 6,
602
+ "sampling": {
603
+ "samples": 3144,
604
+ "secs": 18.254712343215942,
605
+ "gate_secs": 7.309921026229858,
606
+ "samples_per_sec": 172.22950112212612
607
+ },
608
+ "secs": 18.36786985397339,
609
+ "rl_secs_used": 423.592081785202
610
+ },
611
+ {
612
+ "round": 15,
613
+ "level": 4,
614
+ "k": 24,
615
+ "n_targets": 131,
616
+ "sample_acc": 0.004770992366412214,
617
+ "new_proofs": 9,
618
+ "targets_solved": 1123,
619
+ "lstar": 11,
620
+ "ge": {
621
+ "7": 1123,
622
+ "8": 574,
623
+ "9": 302,
624
+ "10": 72,
625
+ "11": 13,
626
+ "12": 1,
627
+ "13": 0,
628
+ "14": 0
629
+ },
630
+ "mix_rl_records": 9,
631
+ "sampling": {
632
+ "samples": 3144,
633
+ "secs": 18.455451011657715,
634
+ "gate_secs": 6.795320510864258,
635
+ "samples_per_sec": 170.35617271092624
636
+ },
637
+ "secs": 18.560636520385742,
638
+ "rl_secs_used": 435.3574059009552
639
+ },
640
+ {
641
+ "round": 16,
642
+ "level": 8,
643
+ "k": 24,
644
+ "n_targets": 130,
645
+ "sample_acc": 0.007371794871794872,
646
+ "new_proofs": 9,
647
+ "targets_solved": 1131,
648
+ "lstar": 11,
649
+ "ge": {
650
+ "7": 1131,
651
+ "8": 580,
652
+ "9": 308,
653
+ "10": 75,
654
+ "11": 14,
655
+ "12": 1,
656
+ "13": 0,
657
+ "14": 0
658
+ },
659
+ "mix_rl_records": 9,
660
+ "sampling": {
661
+ "samples": 3120,
662
+ "secs": 15.886699914932251,
663
+ "gate_secs": 8.2616605758667,
664
+ "samples_per_sec": 196.39069263638856
665
+ },
666
+ "secs": 16.00020694732666,
667
+ "rl_secs_used": 443.095965385437
668
+ },
669
+ {
670
+ "round": 17,
671
+ "level": 1,
672
+ "k": 8,
673
+ "n_targets": 1500,
674
+ "sample_acc": 0.7286666666666667,
675
+ "new_proofs": 151,
676
+ "targets_solved": 1154,
677
+ "lstar": 11,
678
+ "ge": {
679
+ "7": 1154,
680
+ "8": 600,
681
+ "9": 326,
682
+ "10": 81,
683
+ "11": 16,
684
+ "12": 2,
685
+ "13": 0,
686
+ "14": 0
687
+ },
688
+ "mix_rl_records": 1733,
689
+ "sampling": {
690
+ "samples": 12000,
691
+ "secs": 73.38039708137512,
692
+ "gate_secs": 33.76115322113037,
693
+ "samples_per_sec": 163.53141271084445
694
+ },
695
+ "secs": 108.6016788482666,
696
+ "rl_secs_used": 517.9371497631073
697
+ },
698
+ {
699
+ "round": 18,
700
+ "level": 2,
701
+ "k": 24,
702
+ "n_targets": 116,
703
+ "sample_acc": 0.0021551724137931034,
704
+ "new_proofs": 5,
705
+ "targets_solved": 1159,
706
+ "lstar": 11,
707
+ "ge": {
708
+ "7": 1159,
709
+ "8": 604,
710
+ "9": 328,
711
+ "10": 82,
712
+ "11": 16,
713
+ "12": 2,
714
+ "13": 0,
715
+ "14": 0
716
+ },
717
+ "mix_rl_records": 5,
718
+ "sampling": {
719
+ "samples": 2784,
720
+ "secs": 18.592900037765503,
721
+ "gate_secs": 7.087238073348999,
722
+ "samples_per_sec": 149.7345757974925
723
+ },
724
+ "secs": 18.700435638427734,
725
+ "rl_secs_used": 529.54967212677
726
+ },
727
+ {
728
+ "round": 19,
729
+ "level": 4,
730
+ "k": 24,
731
+ "n_targets": 116,
732
+ "sample_acc": 0.002514367816091954,
733
+ "new_proofs": 6,
734
+ "targets_solved": 1165,
735
+ "lstar": 11,
736
+ "ge": {
737
+ "7": 1165,
738
+ "8": 608,
739
+ "9": 331,
740
+ "10": 84,
741
+ "11": 17,
742
+ "12": 3,
743
+ "13": 0,
744
+ "14": 0
745
+ },
746
+ "mix_rl_records": 6,
747
+ "sampling": {
748
+ "samples": 2784,
749
+ "secs": 18.430371284484863,
750
+ "gate_secs": 7.78191065788269,
751
+ "samples_per_sec": 151.05501441219684
752
+ },
753
+ "secs": 18.53823494911194,
754
+ "rl_secs_used": 540.3060154914856
755
+ },
756
+ {
757
+ "round": 20,
758
+ "level": 8,
759
+ "k": 24,
760
+ "n_targets": 115,
761
+ "sample_acc": 0.0014492753623188406,
762
+ "new_proofs": 2,
763
+ "targets_solved": 1167,
764
+ "lstar": 11,
765
+ "ge": {
766
+ "7": 1167,
767
+ "8": 609,
768
+ "9": 332,
769
+ "10": 84,
770
+ "11": 17,
771
+ "12": 3,
772
+ "13": 0,
773
+ "14": 0
774
+ },
775
+ "mix_rl_records": 2,
776
+ "sampling": {
777
+ "samples": 2760,
778
+ "secs": 16.5108003616333,
779
+ "gate_secs": 7.698934078216553,
780
+ "samples_per_sec": 167.16330762580742
781
+ },
782
+ "secs": 16.612196683883667,
783
+ "rl_secs_used": 549.2192680835724
784
+ },
785
+ {
786
+ "round": 21,
787
+ "level": 1,
788
+ "k": 8,
789
+ "n_targets": 1500,
790
+ "sample_acc": 0.7539166666666667,
791
+ "new_proofs": 182,
792
+ "targets_solved": 1185,
793
+ "lstar": 11,
794
+ "ge": {
795
+ "7": 1185,
796
+ "8": 624,
797
+ "9": 344,
798
+ "10": 91,
799
+ "11": 19,
800
+ "12": 3,
801
+ "13": 0,
802
+ "14": 0
803
+ },
804
+ "mix_rl_records": 1893,
805
+ "sampling": {
806
+ "samples": 12000,
807
+ "secs": 60.40621519088745,
808
+ "gate_secs": 33.030165672302246,
809
+ "samples_per_sec": 198.65505498199553
810
+ },
811
+ "secs": 96.91976475715637,
812
+ "rl_secs_used": 613.1092293262482
813
+ },
814
+ {
815
+ "round": 22,
816
+ "level": 2,
817
+ "k": 24,
818
+ "n_targets": 106,
819
+ "sample_acc": 0.0007861635220125787,
820
+ "new_proofs": 2,
821
+ "targets_solved": 1187,
822
+ "lstar": 11,
823
+ "ge": {
824
+ "7": 1187,
825
+ "8": 625,
826
+ "9": 344,
827
+ "10": 91,
828
+ "11": 19,
829
+ "12": 3,
830
+ "13": 0,
831
+ "14": 0
832
+ },
833
+ "mix_rl_records": 2,
834
+ "sampling": {
835
+ "samples": 2544,
836
+ "secs": 15.502235412597656,
837
+ "gate_secs": 7.126906871795654,
838
+ "samples_per_sec": 164.1053649548282
839
+ },
840
+ "secs": 15.597738265991211,
841
+ "rl_secs_used": 621.5796868801117
842
+ },
843
+ {
844
+ "round": 23,
845
+ "level": 4,
846
+ "k": 24,
847
+ "n_targets": 106,
848
+ "sample_acc": 0.0011792452830188679,
849
+ "new_proofs": 2,
850
+ "targets_solved": 1189,
851
+ "lstar": 11,
852
+ "ge": {
853
+ "7": 1189,
854
+ "8": 627,
855
+ "9": 346,
856
+ "10": 91,
857
+ "11": 19,
858
+ "12": 3,
859
+ "13": 0,
860
+ "14": 0
861
+ },
862
+ "mix_rl_records": 2,
863
+ "sampling": {
864
+ "samples": 2544,
865
+ "secs": 11.628038167953491,
866
+ "gate_secs": 5.215627670288086,
867
+ "samples_per_sec": 218.78153160962134
868
+ },
869
+ "secs": 11.723161697387695,
870
+ "rl_secs_used": 628.087265253067
871
+ },
872
+ {
873
+ "round": 24,
874
+ "level": 8,
875
+ "k": 24,
876
+ "n_targets": 105,
877
+ "sample_acc": 0.001984126984126984,
878
+ "new_proofs": 4,
879
+ "targets_solved": 1193,
880
+ "lstar": 11,
881
+ "ge": {
882
+ "7": 1193,
883
+ "8": 629,
884
+ "9": 348,
885
+ "10": 93,
886
+ "11": 20,
887
+ "12": 3,
888
+ "13": 0,
889
+ "14": 0
890
+ },
891
+ "mix_rl_records": 5,
892
+ "sampling": {
893
+ "samples": 2520,
894
+ "secs": 16.824281215667725,
895
+ "gate_secs": 7.508376836776733,
896
+ "samples_per_sec": 149.7835163176679
897
+ },
898
+ "secs": 16.916361570358276,
899
+ "rl_secs_used": 637.4952027797699
900
+ },
901
+ {
902
+ "round": 25,
903
+ "level": 1,
904
+ "k": 8,
905
+ "n_targets": 1500,
906
+ "sample_acc": 0.7739166666666667,
907
+ "new_proofs": 211,
908
+ "targets_solved": 1216,
909
+ "lstar": 11,
910
+ "ge": {
911
+ "7": 1216,
912
+ "8": 647,
913
+ "9": 358,
914
+ "10": 98,
915
+ "11": 22,
916
+ "12": 3,
917
+ "13": 0,
918
+ "14": 0
919
+ },
920
+ "mix_rl_records": 2055,
921
+ "sampling": {
922
+ "samples": 12000,
923
+ "secs": 65.605473279953,
924
+ "gate_secs": 35.27163743972778,
925
+ "samples_per_sec": 182.91156819787517
926
+ },
927
+ "secs": 102.08219480514526,
928
+ "rl_secs_used": 704.3061244487762
929
+ },
930
+ {
931
+ "round": 26,
932
+ "level": 2,
933
+ "k": 24,
934
+ "n_targets": 95,
935
+ "sample_acc": 0.0004385964912280702,
936
+ "new_proofs": 1,
937
+ "targets_solved": 1217,
938
+ "lstar": 11,
939
+ "ge": {
940
+ "7": 1217,
941
+ "8": 648,
942
+ "9": 359,
943
+ "10": 98,
944
+ "11": 22,
945
+ "12": 3,
946
+ "13": 0,
947
+ "14": 0
948
+ },
949
+ "mix_rl_records": 1,
950
+ "sampling": {
951
+ "samples": 2280,
952
+ "secs": 14.946874380111694,
953
+ "gate_secs": 5.455550670623779,
954
+ "samples_per_sec": 152.5402530333544
955
+ },
956
+ "secs": 15.024343013763428,
957
+ "rl_secs_used": 713.8745698928833
958
+ },
959
+ {
960
+ "round": 27,
961
+ "level": 4,
962
+ "k": 24,
963
+ "n_targets": 95,
964
+ "sample_acc": 0.0008771929824561404,
965
+ "new_proofs": 2,
966
+ "targets_solved": 1218,
967
+ "lstar": 11,
968
+ "ge": {
969
+ "7": 1218,
970
+ "8": 649,
971
+ "9": 360,
972
+ "10": 99,
973
+ "11": 22,
974
+ "12": 3,
975
+ "13": 0,
976
+ "14": 0
977
+ },
978
+ "mix_rl_records": 2,
979
+ "sampling": {
980
+ "samples": 2280,
981
+ "secs": 13.35990571975708,
982
+ "gate_secs": 5.772409439086914,
983
+ "samples_per_sec": 170.6598869652395
984
+ },
985
+ "secs": 13.439311981201172,
986
+ "rl_secs_used": 721.5414609909058
987
+ },
988
+ {
989
+ "round": 28,
990
+ "level": 8,
991
+ "k": 24,
992
+ "n_targets": 94,
993
+ "sample_acc": 0.0013297872340425532,
994
+ "new_proofs": 2,
995
+ "targets_solved": 1220,
996
+ "lstar": 11,
997
+ "ge": {
998
+ "7": 1220,
999
+ "8": 651,
1000
+ "9": 362,
1001
+ "10": 99,
1002
+ "11": 22,
1003
+ "12": 3,
1004
+ "13": 0,
1005
+ "14": 0
1006
+ },
1007
+ "mix_rl_records": 2,
1008
+ "sampling": {
1009
+ "samples": 2256,
1010
+ "secs": 10.319400548934937,
1011
+ "gate_secs": 5.356152057647705,
1012
+ "samples_per_sec": 218.61734984527192
1013
+ },
1014
+ "secs": 10.393909692764282,
1015
+ "rl_secs_used": 726.5792157649994
1016
+ },
1017
+ {
1018
+ "round": 29,
1019
+ "level": 1,
1020
+ "k": 8,
1021
+ "n_targets": 1500,
1022
+ "sample_acc": 0.7936666666666666,
1023
+ "new_proofs": 152,
1024
+ "targets_solved": 1234,
1025
+ "lstar": 11,
1026
+ "ge": {
1027
+ "7": 1234,
1028
+ "8": 661,
1029
+ "9": 371,
1030
+ "10": 106,
1031
+ "11": 22,
1032
+ "12": 3,
1033
+ "13": 0,
1034
+ "14": 0
1035
+ },
1036
+ "mix_rl_records": 2164,
1037
+ "sampling": {
1038
+ "samples": 12000,
1039
+ "secs": 66.94237971305847,
1040
+ "gate_secs": 35.56486630439758,
1041
+ "samples_per_sec": 179.2586408107502
1042
+ },
1043
+ "secs": 104.58717608451843,
1044
+ "rl_secs_used": 795.6019032001495
1045
+ },
1046
+ {
1047
+ "round": 30,
1048
+ "level": 2,
1049
+ "k": 24,
1050
+ "n_targets": 89,
1051
+ "sample_acc": 0.0032771535580524347,
1052
+ "new_proofs": 4,
1053
+ "targets_solved": 1237,
1054
+ "lstar": 11,
1055
+ "ge": {
1056
+ "7": 1237,
1057
+ "8": 663,
1058
+ "9": 373,
1059
+ "10": 108,
1060
+ "11": 23,
1061
+ "12": 4,
1062
+ "13": 0,
1063
+ "14": 0
1064
+ },
1065
+ "mix_rl_records": 4,
1066
+ "sampling": {
1067
+ "samples": 2136,
1068
+ "secs": 14.669150829315186,
1069
+ "gate_secs": 5.379411220550537,
1070
+ "samples_per_sec": 145.61170069445097
1071
+ },
1072
+ "secs": 14.745643615722656,
1073
+ "rl_secs_used": 804.9677517414093
1074
+ },
1075
+ {
1076
+ "round": 31,
1077
+ "level": 4,
1078
+ "k": 24,
1079
+ "n_targets": 89,
1080
+ "sample_acc": 0.0,
1081
+ "new_proofs": 0,
1082
+ "targets_solved": 1237,
1083
+ "lstar": 11,
1084
+ "ge": {
1085
+ "7": 1237,
1086
+ "8": 663,
1087
+ "9": 373,
1088
+ "10": 108,
1089
+ "11": 23,
1090
+ "12": 4,
1091
+ "13": 0,
1092
+ "14": 0
1093
+ },
1094
+ "mix_rl_records": 0,
1095
+ "sampling": {
1096
+ "samples": 2136,
1097
+ "secs": 14.398267269134521,
1098
+ "gate_secs": 5.153940439224243,
1099
+ "samples_per_sec": 148.35118421359843
1100
+ },
1101
+ "secs": 14.46718692779541,
1102
+ "rl_secs_used": 814.2809979915619
1103
+ },
1104
+ {
1105
+ "round": 32,
1106
+ "level": 8,
1107
+ "k": 24,
1108
+ "n_targets": 89,
1109
+ "sample_acc": 0.0009363295880149813,
1110
+ "new_proofs": 2,
1111
+ "targets_solved": 1239,
1112
+ "lstar": 11,
1113
+ "ge": {
1114
+ "7": 1239,
1115
+ "8": 665,
1116
+ "9": 374,
1117
+ "10": 109,
1118
+ "11": 23,
1119
+ "12": 4,
1120
+ "13": 0,
1121
+ "14": 0
1122
+ },
1123
+ "mix_rl_records": 2,
1124
+ "sampling": {
1125
+ "samples": 2136,
1126
+ "secs": 13.248353004455566,
1127
+ "gate_secs": 4.914698123931885,
1128
+ "samples_per_sec": 161.22758800898796
1129
+ },
1130
+ "secs": 13.312713384628296,
1131
+ "rl_secs_used": 822.6790173053741
1132
+ },
1133
+ {
1134
+ "round": 33,
1135
+ "level": 1,
1136
+ "k": 8,
1137
+ "n_targets": 1500,
1138
+ "sample_acc": 0.7993333333333333,
1139
+ "new_proofs": 154,
1140
+ "targets_solved": 1245,
1141
+ "lstar": 12,
1142
+ "ge": {
1143
+ "7": 1245,
1144
+ "8": 670,
1145
+ "9": 378,
1146
+ "10": 112,
1147
+ "11": 24,
1148
+ "12": 5,
1149
+ "13": 0,
1150
+ "14": 0
1151
+ },
1152
+ "mix_rl_records": 2252,
1153
+ "sampling": {
1154
+ "samples": 12000,
1155
+ "secs": 66.02069735527039,
1156
+ "gate_secs": 33.536439657211304,
1157
+ "samples_per_sec": 181.76118218542337
1158
+ },
1159
+ "secs": 103.71173524856567,
1160
+ "rl_secs_used": 892.8546257019043
1161
+ },
1162
+ {
1163
+ "round": 34,
1164
+ "level": 2,
1165
+ "k": 24,
1166
+ "n_targets": 86,
1167
+ "sample_acc": 0.003391472868217054,
1168
+ "new_proofs": 6,
1169
+ "targets_solved": 1250,
1170
+ "lstar": 12,
1171
+ "ge": {
1172
+ "7": 1250,
1173
+ "8": 673,
1174
+ "9": 379,
1175
+ "10": 112,
1176
+ "11": 24,
1177
+ "12": 5,
1178
+ "13": 0,
1179
+ "14": 0
1180
+ },
1181
+ "mix_rl_records": 6,
1182
+ "sampling": {
1183
+ "samples": 2064,
1184
+ "secs": 14.294419288635254,
1185
+ "gate_secs": 5.045207500457764,
1186
+ "samples_per_sec": 144.39201469632127
1187
+ },
1188
+ "secs": 14.370021104812622,
1189
+ "rl_secs_used": 902.1791269779205
1190
+ },
1191
+ {
1192
+ "round": 35,
1193
+ "level": 4,
1194
+ "k": 24,
1195
+ "n_targets": 85,
1196
+ "sample_acc": 0.0004901960784313725,
1197
+ "new_proofs": 1,
1198
+ "targets_solved": 1251,
1199
+ "lstar": 12,
1200
+ "ge": {
1201
+ "7": 1251,
1202
+ "8": 674,
1203
+ "9": 379,
1204
+ "10": 112,
1205
+ "11": 24,
1206
+ "12": 5,
1207
+ "13": 0,
1208
+ "14": 0
1209
+ },
1210
+ "mix_rl_records": 1,
1211
+ "sampling": {
1212
+ "samples": 2040,
1213
+ "secs": 13.203391313552856,
1214
+ "gate_secs": 4.845168828964233,
1215
+ "samples_per_sec": 154.50575928216304
1216
+ },
1217
+ "secs": 13.268937349319458,
1218
+ "rl_secs_used": 910.6028928756714
1219
+ },
1220
+ {
1221
+ "round": 36,
1222
+ "level": 8,
1223
+ "k": 24,
1224
+ "n_targets": 85,
1225
+ "sample_acc": 0.004411764705882353,
1226
+ "new_proofs": 6,
1227
+ "targets_solved": 1257,
1228
+ "lstar": 12,
1229
+ "ge": {
1230
+ "7": 1257,
1231
+ "8": 678,
1232
+ "9": 381,
1233
+ "10": 112,
1234
+ "11": 24,
1235
+ "12": 5,
1236
+ "13": 0,
1237
+ "14": 0
1238
+ },
1239
+ "mix_rl_records": 6,
1240
+ "sampling": {
1241
+ "samples": 2040,
1242
+ "secs": 11.405163049697876,
1243
+ "gate_secs": 4.926293134689331,
1244
+ "samples_per_sec": 178.86636000824555
1245
+ },
1246
+ "secs": 11.475750923156738,
1247
+ "rl_secs_used": 917.1523671150208
1248
+ },
1249
+ {
1250
+ "round": 37,
1251
+ "level": 1,
1252
+ "k": 8,
1253
+ "n_targets": 1500,
1254
+ "sample_acc": 0.8188333333333333,
1255
+ "new_proofs": 150,
1256
+ "targets_solved": 1268,
1257
+ "lstar": 12,
1258
+ "ge": {
1259
+ "7": 1268,
1260
+ "8": 687,
1261
+ "9": 383,
1262
+ "10": 113,
1263
+ "11": 24,
1264
+ "12": 5,
1265
+ "13": 0,
1266
+ "14": 0
1267
+ },
1268
+ "mix_rl_records": 2382,
1269
+ "sampling": {
1270
+ "samples": 12000,
1271
+ "secs": 68.13904762268066,
1272
+ "gate_secs": 37.334842681884766,
1273
+ "samples_per_sec": 176.1104743707292
1274
+ },
1275
+ "secs": 106.75181317329407,
1276
+ "rl_secs_used": 986.5698001384735
1277
+ },
1278
+ {
1279
+ "round": 38,
1280
+ "level": 2,
1281
+ "k": 24,
1282
+ "n_targets": 78,
1283
+ "sample_acc": 0.0010683760683760685,
1284
+ "new_proofs": 2,
1285
+ "targets_solved": 1270,
1286
+ "lstar": 12,
1287
+ "ge": {
1288
+ "7": 1270,
1289
+ "8": 687,
1290
+ "9": 383,
1291
+ "10": 113,
1292
+ "11": 24,
1293
+ "12": 5,
1294
+ "13": 0,
1295
+ "14": 0
1296
+ },
1297
+ "mix_rl_records": 2,
1298
+ "sampling": {
1299
+ "samples": 1872,
1300
+ "secs": 11.373922348022461,
1301
+ "gate_secs": 3.69773268699646,
1302
+ "samples_per_sec": 164.58702132123113
1303
+ },
1304
+ "secs": 11.446675539016724,
1305
+ "rl_secs_used": 994.3182647228241
1306
+ },
1307
+ {
1308
+ "round": 39,
1309
+ "level": 4,
1310
+ "k": 24,
1311
+ "n_targets": 78,
1312
+ "sample_acc": 0.0016025641025641025,
1313
+ "new_proofs": 1,
1314
+ "targets_solved": 1271,
1315
+ "lstar": 12,
1316
+ "ge": {
1317
+ "7": 1271,
1318
+ "8": 688,
1319
+ "9": 384,
1320
+ "10": 113,
1321
+ "11": 24,
1322
+ "12": 5,
1323
+ "13": 0,
1324
+ "14": 0
1325
+ },
1326
+ "mix_rl_records": 2,
1327
+ "sampling": {
1328
+ "samples": 1872,
1329
+ "secs": 9.760677576065063,
1330
+ "gate_secs": 4.101423978805542,
1331
+ "samples_per_sec": 191.7899639048093
1332
+ },
1333
+ "secs": 9.816674709320068,
1334
+ "rl_secs_used": 1000.0335118770599
1335
+ },
1336
+ {
1337
+ "round": 40,
1338
+ "level": 8,
1339
+ "k": 24,
1340
+ "n_targets": 77,
1341
+ "sample_acc": 0.0021645021645021645,
1342
+ "new_proofs": 4,
1343
+ "targets_solved": 1273,
1344
+ "lstar": 12,
1345
+ "ge": {
1346
+ "7": 1273,
1347
+ "8": 690,
1348
+ "9": 385,
1349
+ "10": 113,
1350
+ "11": 24,
1351
+ "12": 5,
1352
+ "13": 0,
1353
+ "14": 0
1354
+ },
1355
+ "mix_rl_records": 4,
1356
+ "sampling": {
1357
+ "samples": 1848,
1358
+ "secs": 10.8306725025177,
1359
+ "gate_secs": 4.321923494338989,
1360
+ "samples_per_sec": 170.62652384424084
1361
+ },
1362
+ "secs": 10.904166221618652,
1363
+ "rl_secs_used": 1006.6157555580139
1364
+ },
1365
+ {
1366
+ "round": 41,
1367
+ "level": 1,
1368
+ "k": 8,
1369
+ "n_targets": 1500,
1370
+ "sample_acc": 0.8264166666666667,
1371
+ "new_proofs": 107,
1372
+ "targets_solved": 1276,
1373
+ "lstar": 12,
1374
+ "ge": {
1375
+ "7": 1276,
1376
+ "8": 692,
1377
+ "9": 386,
1378
+ "10": 113,
1379
+ "11": 24,
1380
+ "12": 5,
1381
+ "13": 0,
1382
+ "14": 0
1383
+ },
1384
+ "mix_rl_records": 2440,
1385
+ "sampling": {
1386
+ "samples": 12000,
1387
+ "secs": 67.64672660827637,
1388
+ "gate_secs": 36.70027303695679,
1389
+ "samples_per_sec": 177.3921755222349
1390
+ },
1391
+ "secs": 106.05201172828674,
1392
+ "rl_secs_used": 1075.9678246974945
1393
+ },
1394
+ {
1395
+ "round": 42,
1396
+ "level": 2,
1397
+ "k": 24,
1398
+ "n_targets": 75,
1399
+ "sample_acc": 0.0022222222222222222,
1400
+ "new_proofs": 4,
1401
+ "targets_solved": 1280,
1402
+ "lstar": 12,
1403
+ "ge": {
1404
+ "7": 1280,
1405
+ "8": 696,
1406
+ "9": 390,
1407
+ "10": 115,
1408
+ "11": 25,
1409
+ "12": 5,
1410
+ "13": 0,
1411
+ "14": 0
1412
+ },
1413
+ "mix_rl_records": 4,
1414
+ "sampling": {
1415
+ "samples": 1800,
1416
+ "secs": 9.615079164505005,
1417
+ "gate_secs": 4.581128835678101,
1418
+ "samples_per_sec": 187.2059469510011
1419
+ },
1420
+ "secs": 9.680747270584106,
1421
+ "rl_secs_used": 1081.067135810852
1422
+ },
1423
+ {
1424
+ "round": 43,
1425
+ "level": 4,
1426
+ "k": 24,
1427
+ "n_targets": 75,
1428
+ "sample_acc": 0.0011111111111111111,
1429
+ "new_proofs": 1,
1430
+ "targets_solved": 1281,
1431
+ "lstar": 12,
1432
+ "ge": {
1433
+ "7": 1281,
1434
+ "8": 696,
1435
+ "9": 390,
1436
+ "10": 115,
1437
+ "11": 25,
1438
+ "12": 5,
1439
+ "13": 0,
1440
+ "14": 0
1441
+ },
1442
+ "mix_rl_records": 1,
1443
+ "sampling": {
1444
+ "samples": 1800,
1445
+ "secs": 9.289007186889648,
1446
+ "gate_secs": 3.6236014366149902,
1447
+ "samples_per_sec": 193.777436467106
1448
+ },
1449
+ "secs": 9.342305421829224,
1450
+ "rl_secs_used": 1086.7858140468597
1451
+ },
1452
+ {
1453
+ "round": 44,
1454
+ "level": 8,
1455
+ "k": 24,
1456
+ "n_targets": 75,
1457
+ "sample_acc": 0.0011111111111111111,
1458
+ "new_proofs": 2,
1459
+ "targets_solved": 1283,
1460
+ "lstar": 12,
1461
+ "ge": {
1462
+ "7": 1283,
1463
+ "8": 697,
1464
+ "9": 391,
1465
+ "10": 116,
1466
+ "11": 25,
1467
+ "12": 5,
1468
+ "13": 0,
1469
+ "14": 0
1470
+ },
1471
+ "mix_rl_records": 2,
1472
+ "sampling": {
1473
+ "samples": 1800,
1474
+ "secs": 9.666643142700195,
1475
+ "gate_secs": 3.747734785079956,
1476
+ "samples_per_sec": 186.20734968987423
1477
+ },
1478
+ "secs": 9.753841400146484,
1479
+ "rl_secs_used": 1092.7919239997864
1480
+ },
1481
+ {
1482
+ "round": 45,
1483
+ "level": 1,
1484
+ "k": 8,
1485
+ "n_targets": 1500,
1486
+ "sample_acc": 0.82525,
1487
+ "new_proofs": 121,
1488
+ "targets_solved": 1285,
1489
+ "lstar": 12,
1490
+ "ge": {
1491
+ "7": 1285,
1492
+ "8": 698,
1493
+ "9": 392,
1494
+ "10": 116,
1495
+ "11": 25,
1496
+ "12": 5,
1497
+ "13": 0,
1498
+ "14": 0
1499
+ },
1500
+ "mix_rl_records": 2542,
1501
+ "sampling": {
1502
+ "samples": 12000,
1503
+ "secs": 67.25629663467407,
1504
+ "gate_secs": 34.66986036300659,
1505
+ "samples_per_sec": 178.42195601673055
1506
+ },
1507
+ "secs": 105.89979076385498,
1508
+ "rl_secs_used": 1164.0221619606018
1509
+ },
1510
+ {
1511
+ "round": 46,
1512
+ "level": 2,
1513
+ "k": 24,
1514
+ "n_targets": 73,
1515
+ "sample_acc": 0.0005707762557077625,
1516
+ "new_proofs": 1,
1517
+ "targets_solved": 1286,
1518
+ "lstar": 12,
1519
+ "ge": {
1520
+ "7": 1286,
1521
+ "8": 698,
1522
+ "9": 392,
1523
+ "10": 116,
1524
+ "11": 25,
1525
+ "12": 5,
1526
+ "13": 0,
1527
+ "14": 0
1528
+ },
1529
+ "mix_rl_records": 1,
1530
+ "sampling": {
1531
+ "samples": 1752,
1532
+ "secs": 9.13127851486206,
1533
+ "gate_secs": 3.3750617504119873,
1534
+ "samples_per_sec": 191.86798400119397
1535
+ },
1536
+ "secs": 9.18688702583313,
1537
+ "rl_secs_used": 1169.8336803913116
1538
+ },
1539
+ {
1540
+ "round": 47,
1541
+ "level": 4,
1542
+ "k": 24,
1543
+ "n_targets": 72,
1544
+ "sample_acc": 0.0011574074074074073,
1545
+ "new_proofs": 0,
1546
+ "targets_solved": 1286,
1547
+ "lstar": 12,
1548
+ "ge": {
1549
+ "7": 1286,
1550
+ "8": 698,
1551
+ "9": 392,
1552
+ "10": 116,
1553
+ "11": 25,
1554
+ "12": 5,
1555
+ "13": 0,
1556
+ "14": 0
1557
+ },
1558
+ "mix_rl_records": 1,
1559
+ "sampling": {
1560
+ "samples": 1728,
1561
+ "secs": 9.589057445526123,
1562
+ "gate_secs": 4.1001927852630615,
1563
+ "samples_per_sec": 180.20540702946948
1564
+ },
1565
+ "secs": 9.646642446517944,
1566
+ "rl_secs_used": 1175.380129814148
1567
+ },
1568
+ {
1569
+ "round": 48,
1570
+ "level": 8,
1571
+ "k": 24,
1572
+ "n_targets": 72,
1573
+ "sample_acc": 0.0011574074074074073,
1574
+ "new_proofs": 2,
1575
+ "targets_solved": 1288,
1576
+ "lstar": 12,
1577
+ "ge": {
1578
+ "7": 1288,
1579
+ "8": 700,
1580
+ "9": 394,
1581
+ "10": 118,
1582
+ "11": 25,
1583
+ "12": 5,
1584
+ "13": 0,
1585
+ "14": 0
1586
+ },
1587
+ "mix_rl_records": 2,
1588
+ "sampling": {
1589
+ "samples": 1728,
1590
+ "secs": 11.016762495040894,
1591
+ "gate_secs": 3.855572462081909,
1592
+ "samples_per_sec": 156.85188827278841
1593
+ },
1594
+ "secs": 11.061370134353638,
1595
+ "rl_secs_used": 1182.5859212875366
1596
+ },
1597
+ {
1598
+ "round": 49,
1599
+ "level": 1,
1600
+ "k": 8,
1601
+ "n_targets": 1500,
1602
+ "sample_acc": 0.8318333333333333,
1603
+ "new_proofs": 103,
1604
+ "targets_solved": 1290,
1605
+ "lstar": 12,
1606
+ "ge": {
1607
+ "7": 1290,
1608
+ "8": 702,
1609
+ "9": 395,
1610
+ "10": 119,
1611
+ "11": 26,
1612
+ "12": 6,
1613
+ "13": 0,
1614
+ "14": 0
1615
+ },
1616
+ "mix_rl_records": 2589,
1617
+ "sampling": {
1618
+ "samples": 12000,
1619
+ "secs": 73.39733624458313,
1620
+ "gate_secs": 36.414578437805176,
1621
+ "samples_per_sec": 163.4936717595882
1622
+ },
1623
+ "secs": 111.99669742584229,
1624
+ "rl_secs_used": 1258.1684112548828
1625
+ },
1626
+ {
1627
+ "round": 50,
1628
+ "level": 2,
1629
+ "k": 24,
1630
+ "n_targets": 71,
1631
+ "sample_acc": 0.0005868544600938967,
1632
+ "new_proofs": 1,
1633
+ "targets_solved": 1291,
1634
+ "lstar": 12,
1635
+ "ge": {
1636
+ "7": 1291,
1637
+ "8": 703,
1638
+ "9": 395,
1639
+ "10": 119,
1640
+ "11": 26,
1641
+ "12": 6,
1642
+ "13": 0,
1643
+ "14": 0
1644
+ },
1645
+ "mix_rl_records": 1,
1646
+ "sampling": {
1647
+ "samples": 1704,
1648
+ "secs": 10.50482964515686,
1649
+ "gate_secs": 3.5721378326416016,
1650
+ "samples_per_sec": 162.21110266034736
1651
+ },
1652
+ "secs": 10.581371307373047,
1653
+ "rl_secs_used": 1265.1772801876068
1654
+ },
1655
+ {
1656
+ "round": 51,
1657
+ "level": 4,
1658
+ "k": 24,
1659
+ "n_targets": 70,
1660
+ "sample_acc": 0.0,
1661
+ "new_proofs": 0,
1662
+ "targets_solved": 1291,
1663
+ "lstar": 12,
1664
+ "ge": {
1665
+ "7": 1291,
1666
+ "8": 703,
1667
+ "9": 395,
1668
+ "10": 119,
1669
+ "11": 26,
1670
+ "12": 6,
1671
+ "13": 0,
1672
+ "14": 0
1673
+ },
1674
+ "mix_rl_records": 0,
1675
+ "sampling": {
1676
+ "samples": 1680,
1677
+ "secs": 9.885926723480225,
1678
+ "gate_secs": 3.119422674179077,
1679
+ "samples_per_sec": 169.93854465963264
1680
+ },
1681
+ "secs": 9.922623634338379,
1682
+ "rl_secs_used": 1271.9804763793945
1683
+ },
1684
+ {
1685
+ "round": 52,
1686
+ "level": 8,
1687
+ "k": 24,
1688
+ "n_targets": 70,
1689
+ "sample_acc": 0.0005952380952380953,
1690
+ "new_proofs": 1,
1691
+ "targets_solved": 1292,
1692
+ "lstar": 12,
1693
+ "ge": {
1694
+ "7": 1292,
1695
+ "8": 704,
1696
+ "9": 396,
1697
+ "10": 120,
1698
+ "11": 26,
1699
+ "12": 6,
1700
+ "13": 0,
1701
+ "14": 0
1702
+ },
1703
+ "mix_rl_records": 1,
1704
+ "sampling": {
1705
+ "samples": 1680,
1706
+ "secs": 10.366864204406738,
1707
+ "gate_secs": 3.1528589725494385,
1708
+ "samples_per_sec": 162.05478984530993
1709
+ },
1710
+ "secs": 10.416150331497192,
1711
+ "rl_secs_used": 1279.2437672615051
1712
+ },
1713
+ {
1714
+ "round": 53,
1715
+ "level": 1,
1716
+ "k": 8,
1717
+ "n_targets": 1500,
1718
+ "sample_acc": 0.8368333333333333,
1719
+ "new_proofs": 97,
1720
+ "targets_solved": 1293,
1721
+ "lstar": 12,
1722
+ "ge": {
1723
+ "7": 1293,
1724
+ "8": 704,
1725
+ "9": 396,
1726
+ "10": 120,
1727
+ "11": 26,
1728
+ "12": 6,
1729
+ "13": 0,
1730
+ "14": 0
1731
+ },
1732
+ "mix_rl_records": 2655,
1733
+ "sampling": {
1734
+ "samples": 12000,
1735
+ "secs": 68.24231123924255,
1736
+ "gate_secs": 34.52252912521362,
1737
+ "samples_per_sec": 175.84398567525997
1738
+ },
1739
+ "secs": 107.52387070655823,
1740
+ "rl_secs_used": 1352.2454328536987
1741
+ },
1742
+ {
1743
+ "round": 54,
1744
+ "level": 2,
1745
+ "k": 24,
1746
+ "n_targets": 70,
1747
+ "sample_acc": 0.0,
1748
+ "new_proofs": 0,
1749
+ "targets_solved": 1293,
1750
+ "lstar": 12,
1751
+ "ge": {
1752
+ "7": 1293,
1753
+ "8": 704,
1754
+ "9": 396,
1755
+ "10": 120,
1756
+ "11": 26,
1757
+ "12": 6,
1758
+ "13": 0,
1759
+ "14": 0
1760
+ },
1761
+ "mix_rl_records": 0,
1762
+ "sampling": {
1763
+ "samples": 1680,
1764
+ "secs": 10.038402080535889,
1765
+ "gate_secs": 3.966425895690918,
1766
+ "samples_per_sec": 167.35731309841248
1767
+ },
1768
+ "secs": 10.09642744064331,
1769
+ "rl_secs_used": 1358.3751254081726
1770
+ },
1771
+ {
1772
+ "round": 55,
1773
+ "level": 4,
1774
+ "k": 24,
1775
+ "n_targets": 70,
1776
+ "sample_acc": 0.0,
1777
+ "new_proofs": 0,
1778
+ "targets_solved": 1293,
1779
+ "lstar": 12,
1780
+ "ge": {
1781
+ "7": 1293,
1782
+ "8": 704,
1783
+ "9": 396,
1784
+ "10": 120,
1785
+ "11": 26,
1786
+ "12": 6,
1787
+ "13": 0,
1788
+ "14": 0
1789
+ },
1790
+ "mix_rl_records": 0,
1791
+ "sampling": {
1792
+ "samples": 1680,
1793
+ "secs": 9.01930284500122,
1794
+ "gate_secs": 3.2780563831329346,
1795
+ "samples_per_sec": 186.26716819150923
1796
+ },
1797
+ "secs": 9.070981740951538,
1798
+ "rl_secs_used": 1364.1680388450623
1799
+ },
1800
+ {
1801
+ "round": 56,
1802
+ "level": 8,
1803
+ "k": 24,
1804
+ "n_targets": 69,
1805
+ "sample_acc": 0.0,
1806
+ "new_proofs": 0,
1807
+ "targets_solved": 1293,
1808
+ "lstar": 12,
1809
+ "ge": {
1810
+ "7": 1293,
1811
+ "8": 704,
1812
+ "9": 396,
1813
+ "10": 120,
1814
+ "11": 26,
1815
+ "12": 6,
1816
+ "13": 0,
1817
+ "14": 0
1818
+ },
1819
+ "mix_rl_records": 0,
1820
+ "sampling": {
1821
+ "samples": 1656,
1822
+ "secs": 10.241567373275757,
1823
+ "gate_secs": 3.3737220764160156,
1824
+ "samples_per_sec": 161.6940005024182
1825
+ },
1826
+ "secs": 10.289306163787842,
1827
+ "rl_secs_used": 1371.083619594574
1828
+ },
1829
+ {
1830
+ "round": 57,
1831
+ "level": 1,
1832
+ "k": 8,
1833
+ "n_targets": 1500,
1834
+ "sample_acc": 0.8360833333333333,
1835
+ "new_proofs": 120,
1836
+ "targets_solved": 1294,
1837
+ "lstar": 12,
1838
+ "ge": {
1839
+ "7": 1294,
1840
+ "8": 705,
1841
+ "9": 397,
1842
+ "10": 120,
1843
+ "11": 26,
1844
+ "12": 6,
1845
+ "13": 0,
1846
+ "14": 0
1847
+ },
1848
+ "mix_rl_records": 2694,
1849
+ "sampling": {
1850
+ "samples": 12000,
1851
+ "secs": 70.28077173233032,
1852
+ "gate_secs": 34.23678278923035,
1853
+ "samples_per_sec": 170.7437141655603
1854
+ },
1855
+ "secs": 109.41314959526062,
1856
+ "rl_secs_used": 1446.2602512836456
1857
+ },
1858
+ {
1859
+ "round": 58,
1860
+ "level": 2,
1861
+ "k": 24,
1862
+ "n_targets": 69,
1863
+ "sample_acc": 0.0006038647342995169,
1864
+ "new_proofs": 1,
1865
+ "targets_solved": 1295,
1866
+ "lstar": 12,
1867
+ "ge": {
1868
+ "7": 1295,
1869
+ "8": 706,
1870
+ "9": 398,
1871
+ "10": 121,
1872
+ "11": 26,
1873
+ "12": 6,
1874
+ "13": 0,
1875
+ "14": 0
1876
+ },
1877
+ "mix_rl_records": 1,
1878
+ "sampling": {
1879
+ "samples": 1656,
1880
+ "secs": 10.235370874404907,
1881
+ "gate_secs": 3.404491662979126,
1882
+ "samples_per_sec": 161.79189013473643
1883
+ },
1884
+ "secs": 10.28577709197998,
1885
+ "rl_secs_used": 1453.1412699222565
1886
+ },
1887
+ {
1888
+ "round": 59,
1889
+ "level": 4,
1890
+ "k": 24,
1891
+ "n_targets": 69,
1892
+ "sample_acc": 0.0018115942028985507,
1893
+ "new_proofs": 1,
1894
+ "targets_solved": 1296,
1895
+ "lstar": 12,
1896
+ "ge": {
1897
+ "7": 1296,
1898
+ "8": 706,
1899
+ "9": 398,
1900
+ "10": 121,
1901
+ "11": 26,
1902
+ "12": 6,
1903
+ "13": 0,
1904
+ "14": 0
1905
+ },
1906
+ "mix_rl_records": 1,
1907
+ "sampling": {
1908
+ "samples": 1656,
1909
+ "secs": 10.632396697998047,
1910
+ "gate_secs": 3.756230115890503,
1911
+ "samples_per_sec": 155.75039636282617
1912
+ },
1913
+ "secs": 10.686289072036743,
1914
+ "rl_secs_used": 1460.0713419914246
1915
+ },
1916
+ {
1917
+ "round": 60,
1918
+ "level": 8,
1919
+ "k": 24,
1920
+ "n_targets": 69,
1921
+ "sample_acc": 0.0,
1922
+ "new_proofs": 0,
1923
+ "targets_solved": 1296,
1924
+ "lstar": 12,
1925
+ "ge": {
1926
+ "7": 1296,
1927
+ "8": 706,
1928
+ "9": 398,
1929
+ "10": 121,
1930
+ "11": 26,
1931
+ "12": 6,
1932
+ "13": 0,
1933
+ "14": 0
1934
+ },
1935
+ "mix_rl_records": 0,
1936
+ "sampling": {
1937
+ "samples": 1656,
1938
+ "secs": 10.189783334732056,
1939
+ "gate_secs": 3.785897731781006,
1940
+ "samples_per_sec": 162.5157224251761
1941
+ },
1942
+ "secs": 10.244639873504639,
1943
+ "rl_secs_used": 1466.5300872325897
1944
+ },
1945
+ {
1946
+ "round": 61,
1947
+ "level": 1,
1948
+ "k": 8,
1949
+ "n_targets": 1500,
1950
+ "sample_acc": 0.839,
1951
+ "new_proofs": 129,
1952
+ "targets_solved": 1300,
1953
+ "lstar": 12,
1954
+ "ge": {
1955
+ "7": 1300,
1956
+ "8": 709,
1957
+ "9": 401,
1958
+ "10": 121,
1959
+ "11": 26,
1960
+ "12": 6,
1961
+ "13": 0,
1962
+ "14": 0
1963
+ },
1964
+ "mix_rl_records": 2782,
1965
+ "sampling": {
1966
+ "samples": 12000,
1967
+ "secs": 73.00217175483704,
1968
+ "gate_secs": 36.071818351745605,
1969
+ "samples_per_sec": 164.3786713674706
1970
+ },
1971
+ "secs": 113.6829993724823,
1972
+ "rl_secs_used": 1544.1415979862213
1973
+ },
1974
+ {
1975
+ "round": 62,
1976
+ "level": 2,
1977
+ "k": 24,
1978
+ "n_targets": 68,
1979
+ "sample_acc": 0.0,
1980
+ "new_proofs": 0,
1981
+ "targets_solved": 1300,
1982
+ "lstar": 12,
1983
+ "ge": {
1984
+ "7": 1300,
1985
+ "8": 709,
1986
+ "9": 401,
1987
+ "10": 121,
1988
+ "11": 26,
1989
+ "12": 6,
1990
+ "13": 0,
1991
+ "14": 0
1992
+ },
1993
+ "mix_rl_records": 0,
1994
+ "sampling": {
1995
+ "samples": 1632,
1996
+ "secs": 9.96399736404419,
1997
+ "gate_secs": 3.1429615020751953,
1998
+ "samples_per_sec": 163.78968604399586
1999
+ },
2000
+ "secs": 10.014247417449951,
2001
+ "rl_secs_used": 1551.0125451087952
2002
+ },
2003
+ {
2004
+ "round": 63,
2005
+ "level": 4,
2006
+ "k": 24,
2007
+ "n_targets": 67,
2008
+ "sample_acc": 0.0,
2009
+ "new_proofs": 0,
2010
+ "targets_solved": 1300,
2011
+ "lstar": 12,
2012
+ "ge": {
2013
+ "7": 1300,
2014
+ "8": 709,
2015
+ "9": 401,
2016
+ "10": 121,
2017
+ "11": 26,
2018
+ "12": 6,
2019
+ "13": 0,
2020
+ "14": 0
2021
+ },
2022
+ "mix_rl_records": 0,
2023
+ "sampling": {
2024
+ "samples": 1608,
2025
+ "secs": 10.451392889022827,
2026
+ "gate_secs": 3.456437826156616,
2027
+ "samples_per_sec": 153.85509061561487
2028
+ },
2029
+ "secs": 10.498980522155762,
2030
+ "rl_secs_used": 1558.0550870895386
2031
+ },
2032
+ {
2033
+ "round": 64,
2034
+ "level": 8,
2035
+ "k": 24,
2036
+ "n_targets": 67,
2037
+ "sample_acc": 0.006218905472636816,
2038
+ "new_proofs": 4,
2039
+ "targets_solved": 1301,
2040
+ "lstar": 12,
2041
+ "ge": {
2042
+ "7": 1301,
2043
+ "8": 710,
2044
+ "9": 402,
2045
+ "10": 122,
2046
+ "11": 27,
2047
+ "12": 6,
2048
+ "13": 0,
2049
+ "14": 0
2050
+ },
2051
+ "mix_rl_records": 8,
2052
+ "sampling": {
2053
+ "samples": 1608,
2054
+ "secs": 8.437610149383545,
2055
+ "gate_secs": 3.5483345985412598,
2056
+ "samples_per_sec": 190.57528986658397
2057
+ },
2058
+ "secs": 8.490069150924683,
2059
+ "rl_secs_used": 1562.9968223571777
2060
+ },
2061
+ {
2062
+ "round": 65,
2063
+ "level": 1,
2064
+ "k": 8,
2065
+ "n_targets": 1500,
2066
+ "sample_acc": 0.8453333333333334,
2067
+ "new_proofs": 133,
2068
+ "targets_solved": 1302,
2069
+ "lstar": 12,
2070
+ "ge": {
2071
+ "7": 1302,
2072
+ "8": 711,
2073
+ "9": 403,
2074
+ "10": 123,
2075
+ "11": 28,
2076
+ "12": 6,
2077
+ "13": 0,
2078
+ "14": 0
2079
+ },
2080
+ "mix_rl_records": 2855,
2081
+ "sampling": {
2082
+ "samples": 12000,
2083
+ "secs": 67.82750010490417,
2084
+ "gate_secs": 35.78669548034668,
2085
+ "samples_per_sec": 176.91939082879978
2086
+ },
2087
+ "secs": 107.22025418281555,
2088
+ "rl_secs_used": 1634.4307126998901
2089
+ },
2090
+ {
2091
+ "round": 66,
2092
+ "level": 2,
2093
+ "k": 24,
2094
+ "n_targets": 66,
2095
+ "sample_acc": 0.0,
2096
+ "new_proofs": 0,
2097
+ "targets_solved": 1302,
2098
+ "lstar": 12,
2099
+ "ge": {
2100
+ "7": 1302,
2101
+ "8": 711,
2102
+ "9": 403,
2103
+ "10": 123,
2104
+ "11": 28,
2105
+ "12": 6,
2106
+ "13": 0,
2107
+ "14": 0
2108
+ },
2109
+ "mix_rl_records": 0,
2110
+ "sampling": {
2111
+ "samples": 1584,
2112
+ "secs": 8.182636022567749,
2113
+ "gate_secs": 3.115224838256836,
2114
+ "samples_per_sec": 193.58065000463426
2115
+ },
2116
+ "secs": 8.223321199417114,
2117
+ "rl_secs_used": 1639.538475036621
2118
+ },
2119
+ {
2120
+ "round": 67,
2121
+ "level": 4,
2122
+ "k": 24,
2123
+ "n_targets": 66,
2124
+ "sample_acc": 0.0006313131313131314,
2125
+ "new_proofs": 1,
2126
+ "targets_solved": 1303,
2127
+ "lstar": 12,
2128
+ "ge": {
2129
+ "7": 1303,
2130
+ "8": 711,
2131
+ "9": 403,
2132
+ "10": 123,
2133
+ "11": 28,
2134
+ "12": 6,
2135
+ "13": 0,
2136
+ "14": 0
2137
+ },
2138
+ "mix_rl_records": 1,
2139
+ "sampling": {
2140
+ "samples": 1584,
2141
+ "secs": 9.833085298538208,
2142
+ "gate_secs": 3.2829248905181885,
2143
+ "samples_per_sec": 161.08880904709312
2144
+ },
2145
+ "secs": 9.8818519115448,
2146
+ "rl_secs_used": 1646.137404680252
2147
+ },
2148
+ {
2149
+ "round": 68,
2150
+ "level": 8,
2151
+ "k": 24,
2152
+ "n_targets": 66,
2153
+ "sample_acc": 0.0006313131313131314,
2154
+ "new_proofs": 1,
2155
+ "targets_solved": 1304,
2156
+ "lstar": 12,
2157
+ "ge": {
2158
+ "7": 1304,
2159
+ "8": 712,
2160
+ "9": 403,
2161
+ "10": 123,
2162
+ "11": 28,
2163
+ "12": 6,
2164
+ "13": 0,
2165
+ "14": 0
2166
+ },
2167
+ "mix_rl_records": 1,
2168
+ "sampling": {
2169
+ "samples": 1584,
2170
+ "secs": 7.744364023208618,
2171
+ "gate_secs": 3.0884227752685547,
2172
+ "samples_per_sec": 204.53583990279975
2173
+ },
2174
+ "secs": 7.788423299789429,
2175
+ "rl_secs_used": 1650.8374028205872
2176
+ },
2177
+ {
2178
+ "round": 69,
2179
+ "level": 1,
2180
+ "k": 8,
2181
+ "n_targets": 1500,
2182
+ "sample_acc": 0.8446666666666667,
2183
+ "new_proofs": 151,
2184
+ "targets_solved": 1306,
2185
+ "lstar": 12,
2186
+ "ge": {
2187
+ "7": 1306,
2188
+ "8": 714,
2189
+ "9": 405,
2190
+ "10": 125,
2191
+ "11": 28,
2192
+ "12": 6,
2193
+ "13": 0,
2194
+ "14": 0
2195
+ },
2196
+ "mix_rl_records": 2935,
2197
+ "sampling": {
2198
+ "samples": 12000,
2199
+ "secs": 73.09983134269714,
2200
+ "gate_secs": 33.59913158416748,
2201
+ "samples_per_sec": 164.1590654805092
2202
+ },
2203
+ "secs": 113.17964172363281,
2204
+ "rl_secs_used": 1730.4182505607605
2205
+ },
2206
+ {
2207
+ "round": 70,
2208
+ "level": 2,
2209
+ "k": 24,
2210
+ "n_targets": 65,
2211
+ "sample_acc": 0.0038461538461538464,
2212
+ "new_proofs": 4,
2213
+ "targets_solved": 1306,
2214
+ "lstar": 12,
2215
+ "ge": {
2216
+ "7": 1306,
2217
+ "8": 714,
2218
+ "9": 405,
2219
+ "10": 125,
2220
+ "11": 28,
2221
+ "12": 6,
2222
+ "13": 0,
2223
+ "14": 0
2224
+ },
2225
+ "mix_rl_records": 4,
2226
+ "sampling": {
2227
+ "samples": 1560,
2228
+ "secs": 10.024504661560059,
2229
+ "gate_secs": 3.515435218811035,
2230
+ "samples_per_sec": 155.61866173617258
2231
+ },
2232
+ "secs": 10.076973915100098,
2233
+ "rl_secs_used": 1736.9794569015503
2234
+ },
2235
+ {
2236
+ "round": 71,
2237
+ "level": 4,
2238
+ "k": 24,
2239
+ "n_targets": 65,
2240
+ "sample_acc": 0.0,
2241
+ "new_proofs": 0,
2242
+ "targets_solved": 1306,
2243
+ "lstar": 12,
2244
+ "ge": {
2245
+ "7": 1306,
2246
+ "8": 714,
2247
+ "9": 405,
2248
+ "10": 125,
2249
+ "11": 28,
2250
+ "12": 6,
2251
+ "13": 0,
2252
+ "14": 0
2253
+ },
2254
+ "mix_rl_records": 0,
2255
+ "sampling": {
2256
+ "samples": 1560,
2257
+ "secs": 10.004522323608398,
2258
+ "gate_secs": 3.5535967350006104,
2259
+ "samples_per_sec": 155.92948364148828
2260
+ },
2261
+ "secs": 10.085642099380493,
2262
+ "rl_secs_used": 1743.5115225315094
2263
+ },
2264
+ {
2265
+ "round": 72,
2266
+ "level": 8,
2267
+ "k": 24,
2268
+ "n_targets": 65,
2269
+ "sample_acc": 0.001282051282051282,
2270
+ "new_proofs": 1,
2271
+ "targets_solved": 1307,
2272
+ "lstar": 12,
2273
+ "ge": {
2274
+ "7": 1307,
2275
+ "8": 715,
2276
+ "9": 405,
2277
+ "10": 125,
2278
+ "11": 28,
2279
+ "12": 6,
2280
+ "13": 0,
2281
+ "14": 0
2282
+ },
2283
+ "mix_rl_records": 1,
2284
+ "sampling": {
2285
+ "samples": 1560,
2286
+ "secs": 9.533372163772583,
2287
+ "gate_secs": 2.9600701332092285,
2288
+ "samples_per_sec": 163.63569712804232
2289
+ },
2290
+ "secs": 9.580638885498047,
2291
+ "rl_secs_used": 1750.1320621967316
2292
+ }
2293
+ ],
2294
+ "secs": 2743.5326623916626,
2295
+ "rl_secs_used": 1789.2528948783875,
2296
+ "gate_secs": 954.2544889450073,
2297
+ "targets": 1500,
2298
+ "stats": {
2299
+ "rounds": 18,
2300
+ "first_loops": 1,
2301
+ "k_first": 8,
2302
+ "retry_loops": [
2303
+ 2,
2304
+ 4,
2305
+ 8
2306
+ ],
2307
+ "k_retry": 24,
2308
+ "retries": [
2309
+ {
2310
+ "round": 1,
2311
+ "loops": 2,
2312
+ "targets": 395,
2313
+ "cracked": 37
2314
+ },
2315
+ {
2316
+ "round": 1,
2317
+ "loops": 4,
2318
+ "targets": 394,
2319
+ "cracked": 46
2320
+ },
2321
+ {
2322
+ "round": 1,
2323
+ "loops": 8,
2324
+ "targets": 394,
2325
+ "cracked": 44
2326
+ },
2327
+ {
2328
+ "round": 2,
2329
+ "loops": 2,
2330
+ "targets": 209,
2331
+ "cracked": 17
2332
+ },
2333
+ {
2334
+ "round": 2,
2335
+ "loops": 4,
2336
+ "targets": 209,
2337
+ "cracked": 20
2338
+ },
2339
+ {
2340
+ "round": 2,
2341
+ "loops": 8,
2342
+ "targets": 209,
2343
+ "cracked": 14
2344
+ },
2345
+ {
2346
+ "round": 3,
2347
+ "loops": 2,
2348
+ "targets": 153,
2349
+ "cracked": 11
2350
+ },
2351
+ {
2352
+ "round": 3,
2353
+ "loops": 4,
2354
+ "targets": 153,
2355
+ "cracked": 7
2356
+ },
2357
+ {
2358
+ "round": 3,
2359
+ "loops": 8,
2360
+ "targets": 152,
2361
+ "cracked": 7
2362
+ },
2363
+ {
2364
+ "round": 4,
2365
+ "loops": 2,
2366
+ "targets": 131,
2367
+ "cracked": 6
2368
+ },
2369
+ {
2370
+ "round": 4,
2371
+ "loops": 4,
2372
+ "targets": 131,
2373
+ "cracked": 8
2374
+ },
2375
+ {
2376
+ "round": 4,
2377
+ "loops": 8,
2378
+ "targets": 130,
2379
+ "cracked": 8
2380
+ },
2381
+ {
2382
+ "round": 5,
2383
+ "loops": 2,
2384
+ "targets": 116,
2385
+ "cracked": 5
2386
+ },
2387
+ {
2388
+ "round": 5,
2389
+ "loops": 4,
2390
+ "targets": 116,
2391
+ "cracked": 6
2392
+ },
2393
+ {
2394
+ "round": 5,
2395
+ "loops": 8,
2396
+ "targets": 115,
2397
+ "cracked": 2
2398
+ },
2399
+ {
2400
+ "round": 6,
2401
+ "loops": 2,
2402
+ "targets": 106,
2403
+ "cracked": 2
2404
+ },
2405
+ {
2406
+ "round": 6,
2407
+ "loops": 4,
2408
+ "targets": 106,
2409
+ "cracked": 2
2410
+ },
2411
+ {
2412
+ "round": 6,
2413
+ "loops": 8,
2414
+ "targets": 105,
2415
+ "cracked": 5
2416
+ },
2417
+ {
2418
+ "round": 7,
2419
+ "loops": 2,
2420
+ "targets": 95,
2421
+ "cracked": 1
2422
+ },
2423
+ {
2424
+ "round": 7,
2425
+ "loops": 4,
2426
+ "targets": 95,
2427
+ "cracked": 1
2428
+ },
2429
+ {
2430
+ "round": 7,
2431
+ "loops": 8,
2432
+ "targets": 94,
2433
+ "cracked": 2
2434
+ },
2435
+ {
2436
+ "round": 8,
2437
+ "loops": 2,
2438
+ "targets": 89,
2439
+ "cracked": 3
2440
+ },
2441
+ {
2442
+ "round": 8,
2443
+ "loops": 4,
2444
+ "targets": 89,
2445
+ "cracked": 0
2446
+ },
2447
+ {
2448
+ "round": 8,
2449
+ "loops": 8,
2450
+ "targets": 89,
2451
+ "cracked": 2
2452
+ },
2453
+ {
2454
+ "round": 9,
2455
+ "loops": 2,
2456
+ "targets": 86,
2457
+ "cracked": 5
2458
+ },
2459
+ {
2460
+ "round": 9,
2461
+ "loops": 4,
2462
+ "targets": 85,
2463
+ "cracked": 1
2464
+ },
2465
+ {
2466
+ "round": 9,
2467
+ "loops": 8,
2468
+ "targets": 85,
2469
+ "cracked": 6
2470
+ },
2471
+ {
2472
+ "round": 10,
2473
+ "loops": 2,
2474
+ "targets": 78,
2475
+ "cracked": 2
2476
+ },
2477
+ {
2478
+ "round": 10,
2479
+ "loops": 4,
2480
+ "targets": 78,
2481
+ "cracked": 2
2482
+ },
2483
+ {
2484
+ "round": 10,
2485
+ "loops": 8,
2486
+ "targets": 77,
2487
+ "cracked": 2
2488
+ },
2489
+ {
2490
+ "round": 11,
2491
+ "loops": 2,
2492
+ "targets": 75,
2493
+ "cracked": 4
2494
+ },
2495
+ {
2496
+ "round": 11,
2497
+ "loops": 4,
2498
+ "targets": 75,
2499
+ "cracked": 1
2500
+ },
2501
+ {
2502
+ "round": 11,
2503
+ "loops": 8,
2504
+ "targets": 75,
2505
+ "cracked": 2
2506
+ },
2507
+ {
2508
+ "round": 12,
2509
+ "loops": 2,
2510
+ "targets": 73,
2511
+ "cracked": 1
2512
+ },
2513
+ {
2514
+ "round": 12,
2515
+ "loops": 4,
2516
+ "targets": 72,
2517
+ "cracked": 1
2518
+ },
2519
+ {
2520
+ "round": 12,
2521
+ "loops": 8,
2522
+ "targets": 72,
2523
+ "cracked": 2
2524
+ },
2525
+ {
2526
+ "round": 13,
2527
+ "loops": 2,
2528
+ "targets": 71,
2529
+ "cracked": 1
2530
+ },
2531
+ {
2532
+ "round": 13,
2533
+ "loops": 4,
2534
+ "targets": 70,
2535
+ "cracked": 0
2536
+ },
2537
+ {
2538
+ "round": 13,
2539
+ "loops": 8,
2540
+ "targets": 70,
2541
+ "cracked": 1
2542
+ },
2543
+ {
2544
+ "round": 14,
2545
+ "loops": 2,
2546
+ "targets": 70,
2547
+ "cracked": 0
2548
+ },
2549
+ {
2550
+ "round": 14,
2551
+ "loops": 4,
2552
+ "targets": 70,
2553
+ "cracked": 0
2554
+ },
2555
+ {
2556
+ "round": 14,
2557
+ "loops": 8,
2558
+ "targets": 69,
2559
+ "cracked": 0
2560
+ },
2561
+ {
2562
+ "round": 15,
2563
+ "loops": 2,
2564
+ "targets": 69,
2565
+ "cracked": 1
2566
+ },
2567
+ {
2568
+ "round": 15,
2569
+ "loops": 4,
2570
+ "targets": 69,
2571
+ "cracked": 1
2572
+ },
2573
+ {
2574
+ "round": 15,
2575
+ "loops": 8,
2576
+ "targets": 69,
2577
+ "cracked": 0
2578
+ },
2579
+ {
2580
+ "round": 16,
2581
+ "loops": 2,
2582
+ "targets": 68,
2583
+ "cracked": 0
2584
+ },
2585
+ {
2586
+ "round": 16,
2587
+ "loops": 4,
2588
+ "targets": 67,
2589
+ "cracked": 0
2590
+ },
2591
+ {
2592
+ "round": 16,
2593
+ "loops": 8,
2594
+ "targets": 67,
2595
+ "cracked": 2
2596
+ },
2597
+ {
2598
+ "round": 17,
2599
+ "loops": 2,
2600
+ "targets": 66,
2601
+ "cracked": 0
2602
+ },
2603
+ {
2604
+ "round": 17,
2605
+ "loops": 4,
2606
+ "targets": 66,
2607
+ "cracked": 1
2608
+ },
2609
+ {
2610
+ "round": 17,
2611
+ "loops": 8,
2612
+ "targets": 66,
2613
+ "cracked": 1
2614
+ },
2615
+ {
2616
+ "round": 18,
2617
+ "loops": 2,
2618
+ "targets": 65,
2619
+ "cracked": 1
2620
+ },
2621
+ {
2622
+ "round": 18,
2623
+ "loops": 4,
2624
+ "targets": 65,
2625
+ "cracked": 0
2626
+ },
2627
+ {
2628
+ "round": 18,
2629
+ "loops": 8,
2630
+ "targets": 65,
2631
+ "cracked": 1
2632
+ }
2633
+ ],
2634
+ "ft_records": 76800,
2635
+ "ft_steps": [
2636
+ 596,
2637
+ 643,
2638
+ 664,
2639
+ 672,
2640
+ 680,
2641
+ 688,
2642
+ 702,
2643
+ 707,
2644
+ 710,
2645
+ 718,
2646
+ 721,
2647
+ 722,
2648
+ 726,
2649
+ 729,
2650
+ 738,
2651
+ 732,
2652
+ 735,
2653
+ 736
2654
+ ]
2655
+ }
2656
+ },
2657
+ "eval": {
2658
+ "level": 8,
2659
+ "attempts_per_theorem": 64,
2660
+ "solved": 785,
2661
+ "n": 1108,
2662
+ "solved_ge7": 785,
2663
+ "lstar": 12,
2664
+ "ge": {
2665
+ "7": 785,
2666
+ "8": 669,
2667
+ "9": 538,
2668
+ "10": 145,
2669
+ "11": 30,
2670
+ "12": 7,
2671
+ "13": 2,
2672
+ "14": 1
2673
+ },
2674
+ "by_bin": {
2675
+ "7": {
2676
+ "solved": 116,
2677
+ "n": 150,
2678
+ "rate": 0.7733333333333333,
2679
+ "ci": [
2680
+ 0.6999966119167443,
2681
+ 0.8330191653502627
2682
+ ]
2683
+ },
2684
+ "8": {
2685
+ "solved": 131,
2686
+ "n": 152,
2687
+ "rate": 0.8618421052631579,
2688
+ "ci": [
2689
+ 0.7980159638451529,
2690
+ 0.907828945344691
2691
+ ]
2692
+ },
2693
+ "9": {
2694
+ "solved": 393,
2695
+ "n": 477,
2696
+ "rate": 0.8238993710691824,
2697
+ "ci": [
2698
+ 0.7871669240132779,
2699
+ 0.8554563431915563
2700
+ ]
2701
+ },
2702
+ "10": {
2703
+ "solved": 115,
2704
+ "n": 219,
2705
+ "rate": 0.5251141552511416,
2706
+ "ci": [
2707
+ 0.4591136699864923,
2708
+ 0.590248747084647
2709
+ ]
2710
+ },
2711
+ "11": {
2712
+ "solved": 23,
2713
+ "n": 49,
2714
+ "rate": 0.46938775510204084,
2715
+ "ci": [
2716
+ 0.33703281897894777,
2717
+ 0.6061937298008014
2718
+ ]
2719
+ },
2720
+ "12": {
2721
+ "solved": 5,
2722
+ "n": 54,
2723
+ "rate": 0.09259259259259259,
2724
+ "ci": [
2725
+ 0.040200542070567336,
2726
+ 0.1991012753065453
2727
+ ]
2728
+ },
2729
+ "13": {
2730
+ "solved": 1,
2731
+ "n": 5,
2732
+ "rate": 0.2,
2733
+ "ci": [
2734
+ 0.03622316096978745,
2735
+ 0.6244717358814613
2736
+ ]
2737
+ },
2738
+ "14": {
2739
+ "solved": 1,
2740
+ "n": 2,
2741
+ "rate": 0.5,
2742
+ "ci": [
2743
+ 0.09452865480086614,
2744
+ 0.9054713451991339
2745
+ ]
2746
+ }
2747
+ },
2748
+ "sample_acc": 0.5537003610108303,
2749
+ "max_pruned": 15,
2750
+ "sampling": {
2751
+ "samples": 70912,
2752
+ "secs": 317.55645656585693,
2753
+ "gate_secs": 121.90295219421387,
2754
+ "samples_per_sec": 223.3051746667724
2755
+ },
2756
+ "heldout_greedy": 0.9865,
2757
+ "diag": {
2758
+ "level": 2,
2759
+ "attempts_per_theorem": 64,
2760
+ "solved": 785,
2761
+ "n": 1108,
2762
+ "solved_ge7": 785,
2763
+ "lstar": 12,
2764
+ "ge": {
2765
+ "7": 785,
2766
+ "8": 669,
2767
+ "9": 538,
2768
+ "10": 145,
2769
+ "11": 30,
2770
+ "12": 7,
2771
+ "13": 2,
2772
+ "14": 1
2773
+ },
2774
+ "by_bin": {
2775
+ "7": {
2776
+ "solved": 116,
2777
+ "n": 150,
2778
+ "rate": 0.7733333333333333,
2779
+ "ci": [
2780
+ 0.6999966119167443,
2781
+ 0.8330191653502627
2782
+ ]
2783
+ },
2784
+ "8": {
2785
+ "solved": 131,
2786
+ "n": 152,
2787
+ "rate": 0.8618421052631579,
2788
+ "ci": [
2789
+ 0.7980159638451529,
2790
+ 0.907828945344691
2791
+ ]
2792
+ },
2793
+ "9": {
2794
+ "solved": 393,
2795
+ "n": 477,
2796
+ "rate": 0.8238993710691824,
2797
+ "ci": [
2798
+ 0.7871669240132779,
2799
+ 0.8554563431915563
2800
+ ]
2801
+ },
2802
+ "10": {
2803
+ "solved": 115,
2804
+ "n": 219,
2805
+ "rate": 0.5251141552511416,
2806
+ "ci": [
2807
+ 0.4591136699864923,
2808
+ 0.590248747084647
2809
+ ]
2810
+ },
2811
+ "11": {
2812
+ "solved": 23,
2813
+ "n": 49,
2814
+ "rate": 0.46938775510204084,
2815
+ "ci": [
2816
+ 0.33703281897894777,
2817
+ 0.6061937298008014
2818
+ ]
2819
+ },
2820
+ "12": {
2821
+ "solved": 5,
2822
+ "n": 54,
2823
+ "rate": 0.09259259259259259,
2824
+ "ci": [
2825
+ 0.040200542070567336,
2826
+ 0.1991012753065453
2827
+ ]
2828
+ },
2829
+ "13": {
2830
+ "solved": 1,
2831
+ "n": 5,
2832
+ "rate": 0.2,
2833
+ "ci": [
2834
+ 0.03622316096978745,
2835
+ 0.6244717358814613
2836
+ ]
2837
+ },
2838
+ "14": {
2839
+ "solved": 1,
2840
+ "n": 2,
2841
+ "rate": 0.5,
2842
+ "ci": [
2843
+ 0.09452865480086614,
2844
+ 0.9054713451991339
2845
+ ]
2846
+ }
2847
+ },
2848
+ "sample_acc": 0.5537003610108303,
2849
+ "max_pruned": 15,
2850
+ "sampling": {
2851
+ "samples": 70912,
2852
+ "secs": 314.0251543521881,
2853
+ "gate_secs": 120.7315902709961,
2854
+ "samples_per_sec": 225.81630489534027
2855
+ }
2856
+ }
2857
+ }
2858
+ }