Round 3 analysis (PREREG3.md)

Root <scratch>/refrun/full3; seeds [0, 1, 2, 3, 4]; 60 checkpoints per run (400 to 24000 rows); LATE = 31 checkpoints in [12000, 24000].

LATE (primary)

seed	0	1	2	3	4	mean
reference	307.6	346.7	324.5	314.2	270.3	312.7
Ajax	310.6	321.1	198.3	232.6	264.6	265.4

• LEARNS = False (Ajax mean 265.4 >= reference min 270.3)
• WITHIN_RANGE = False (reference range [270.3, 346.7])
• Welch t = 1.795, two-sided p = 0.1216 (positive t: reference higher)

S1 (16000, 18000, 20000)

seed	0	1	2	3	4	mean
reference	201.5	298.8	358.1	273.6	278.6	282.1
Ajax	373.8	354.0	250.7	270.1	280.7	305.8

• LEARNS = True (Ajax mean 305.8 >= reference min 201.5)
• WITHIN_RANGE = True (reference range [201.5, 358.1])
• Welch t = -0.678, two-sided p = 0.5166 (positive t: reference higher)

Concordance (eval >= 400)

• at 20000: reference 2/5, Ajax 2/5
• at 24000: reference 4/5, Ajax 3/5
• rate over [16000, 24000]: reference 0.457, Ajax 0.324

Reading: H2 or noise (no reference concordance beyond its late rate, no significant LATE difference)

Curves (eval mean return; every 2000 rows shown, all points used above)

rows	ref s0	ref s1	ref s2	ref s3	ref s4	ajax s0	ajax s1	ajax s2	ajax s3	ajax s4
2000	20	23	29	17	16	22	19	16	32	28
4000	94	190	208	126	148	129	153	49	125	65
6000	163	226	325	170	93	365	259	303	369	167
8000	323	100	151	56	27	60	77	117	156	31
10000	152	156	40	37	174	132	148	82	365	348
12000	500	500	138	251	150	305	325	97	177	479
14000	144	250	202	192	142	123	236	73	57	330
16000	166	162	367	246	138	227	184	121	483	275
18000	173	372	208	203	220	500	378	432	131	67
20000	265	362	500	371	478	394	500	199	196	500
22000	407	488	500	451	212	362	500	160	330	500
24000	500	479	500	500	212	500	500	500	191	128