Goal write-up: merging the two newest rounds into the start file

Sources are the five results under .imd/reads/artifacts/: item-m-all (SHA-256 db22aff7167a4798251ed3d20432509f30e511e4958799aba5b17b41bb39b4b6), item-m-plus-a (149883fee797ad95a91e535b7a74453587a3eb19ca87ab95f7d8585db160a8f4), item-m-plus-b (6b40ef6fdf9512700501e6b92fb5c920955a97b29111faada6d411074173b7cd), item-m-plus-c (86a711e937d310d0ed3dc3583d97962ad1b53959bfb11b56e545070cb7407b32) and item-m-safe (aac61b519153db5193195c7fd184e57fff593ac2879211a25ac86997740a62e5). Every figure below is copied from those results; nothing is recomputed except where the text says so. The start file is 151e0b1b16f90a1384b823bc528d6d1407f7b4bbfcc7966c12119fe483bfd326 and its score is 108.63974289615767 on 600 seeds 1000..1599, reproduced by all five items as base_mean_edge except item-m-plus-c, which reports it as not measured.

Blocks

item	blocks and spreads	exact against the block files	by-spread runs, spreads 1/2/3/4
item-m-all	c1-pairs on 1, c2-queue on 2, a3-ledger on 3, c4-all on 4	600 runs per spread, identical=true	1016/980/958/1046
item-m-plus-a	c1-pairs on 1, c2-queue on 2, a3-ledger on 3, c4-all on 4	600 runs per spread, identical=true	752/742/718/788
item-m-plus-b	c1-pairs on 1, c2-queue on 2, a3-ledger on 3, c4-all on 4	600 runs per spread, identical=true	1016/980/958/1046
item-m-plus-c	c1-pairs on 1, c2-queue on 2, a3-ledger on 3, c4-all on 4 (class dispatch)	30 runs per spread, identical=true	43/31/28/38
item-m-safe	a3-ledger on 3 only; 1, 2, 4 keep the start file	300 runs per spread, identical=true	501/504/479/516

Every item reports all four spreads identical to their source files; the run count behind that claim is what separates item-m-all and item-m-plus-b (600 per spread, 2400 paired comparisons) from item-m-safe (300 per spread) and item-m-plus-c (30 per spread, 15 seeds from 100000 and 15 from 104000). item-m-plus-c also notes that its spreads 2-4 were checked before a later fix to the spread-1 top-level code.

Fixes

item	fix (spread)	measured gain	standard error	stayed
item-m-plus-a	a1-jumps (1)	0.5009455722593997	0.3289621386809776	yes
item-m-plus-a	a2-after (2)	0.7478670929991998	0.14009787888594427	yes
item-m-plus-a	a4-rich (4)	0.22959375486119976	0.07775380466313571	yes
item-m-plus-a	a4-free (4)	0.052893605455300426	0.13495153349795513	no
item-m-plus-a	a3-rich (3)	0.7387592263152998	0.20043234577823205	yes
item-m-plus-a	a1-free (1)	0.29429490312559986	0.047839957990400236	yes
item-m-plus-a	a3-free (3)	0.2783257251334001	0.05424129829265064	yes
item-m-plus-a	a4-stock (4)	0.46915560943630036	0.05582351049007889	yes
item-m-plus-a	a2-ledger (2)	1.2108931310401003	0.1136786232236645	yes
item-m-plus-b	a1-free (1)	0.35149849538759986	0.1473650964688594	yes
item-m-plus-b	a1-jumps (1)	-0.46009330345589977	0.5348458123513586	no
item-m-plus-b	a2-ledger (2)	1.0912132155852996	0.10001626498595995	yes
item-m-plus-b	a2-after (2)	0.8675470084540006	0.13694721873193286	yes
item-m-plus-b	a3-free (3)	0.2967595907355999	0.055760086084229464	yes
item-m-plus-b	a3-rich (3)	0.7203253607131	0.20022945125849764	yes
item-m-plus-b	a4-stock (4)	0.3884841414820999	0.04321751641266317	yes
item-m-plus-b	a4-free (4)	0.10763674876980042	0.14029986176237702	no
item-m-plus-b	a4-rich (4)	0.3102652228154002	0.0836897732963366	yes
item-m-plus-c	a3-rich (3)	0.8334181773983345	0.25588320751442933	yes
item-m-plus-c	a3-free (3)	0.17271267671833232	0.3251756280361294	yes
item-m-plus-c	a4-rich (4)	0.38517936668833586	0.14609271528377196	yes
item-m-plus-c	a4-free (4)	0.6258003894733333	0.3207823409434335	yes
item-m-plus-c	a4-stock (4)	0.4463098442333328	0.2447708962752243	yes
item-m-plus-c	a1-jumps (1)	not tested	not tested	no
item-m-plus-c	a1-free (1)	not tested	not tested	no
item-m-plus-c	a2-after (2)	not tested	not tested	no
item-m-plus-c	a2-ledger (2)	not tested	not tested	no
item-m-safe	a1-jumps (1)	0.8317693322811999	0.35873561884138366	yes
item-m-safe	a1-free (1)	0.1623564006363998	0.1415346551431088	no
item-m-safe	a2-after (2)	1.0651422830891997	0.17354929374259553	yes
item-m-safe	a2-ledger (2)	0.27792990249079996	0.10146617507217796	yes
item-m-safe	a3-rich (3)	1.0785609271135999	0.38144802285492224	yes
item-m-safe	a3-free (3)	0.26205756672299996	0.0884205488779404	yes
item-m-safe	a4-rich (4)	0.5250457714106009	0.10856922767315608	yes
item-m-safe	a4-free (4)	0.26696185971979836	0.3000436788195861	yes
item-m-safe	a4-stock (4)	0.2542907039806001	0.056770467549436286	yes

item-m-all tried no fixes. item-m-plus-c lists the four spread-1 and spread-2 fixes as measured at all (null gain) because the c1-pairs and c2-queue blocks replace the start-file code they were written against; it dropped them untested. Its remaining five were tested on 60 runs of their spread, all-on against leave-one-out, and all passed.

Difference from the start file

item	mean	SE	retail	arbitrage	seconds, start	seconds, item
item-m-all	1.445340121537375	0.15708671985275371	0.7685706656211752	0.676769455914275	179.17913917917758	179.33491440303624
item-m-plus-a	2.4314994365665337	0.21640274191625686	0.31911411444809995	2.1123853221140667	120.45409822463989	124.06083273887634
item-m-plus-b	2.2128427559408252	0.1728962114120029	0.3448020945050003	1.868040661433225	107.23521068086848	115.86656947713345
item-m-plus-c	2.1346696331885706	1.410839002179575	1.915353916744286	0.21931571645214276	241.93915333066667	251.27556937081474
item-m-safe	0.95685317308875	0.14141149083550503	-0.2489319660061499	1.20578513908295	46.43412822112441	51.85971038695425

item	spread	runs	mean	SE	retail	arbitrage
item-m-all	1	1016	1.6356063570282477	0.33297078663721597	1.0300580291539372	0.6055483278707675
item-m-all	2	980	1.7287357789455104	0.47160501483821404	2.6768412316472454	-0.9481054527019388
item-m-all	3	958	0.4883446301950943	0.04857877029821955	-0.02987450130949868	0.5182191315026096
item-m-all	4	1046	1.8715001991542064	0.2413208423327588	-0.5420085371849906	2.413508736337285
item-m-plus-a	1	752	2.7607500570893615	0.4681439732311693	0.5602578017253987	2.2004922553571804
item-m-plus-a	2	742	3.1191828508327495	0.6253819677133948	1.8774088114223721	1.2417740394117252
item-m-plus-a	3	718	1.177005828453065	0.17567191006929916	-0.45932796508467955	1.636333793533844
item-m-plus-a	4	788	2.6128050845446706	0.3238672094824128	-0.6690499780428937	3.2818550625796954
item-m-plus-b	1	1016	1.9515104762154525	0.3478722551131989	0.7234912613565947	1.2280192148513778
item-m-plus-b	2	980	3.005904837975102	0.51538162827511	1.7416755060861226	1.2642293318903062
item-m-plus-b	3	958	1.1897131414110653	0.1500223481925494	-0.4360777386556363	1.6257908800641963
item-m-plus-b	4	1046	2.6607117105554496	0.26368738893498095	-0.6165757799718925	3.277287490525717
item-m-plus-c	1	43	6.598859834299999	4.055074254964557	5.725346243925583	0.8735135903418605
item-m-plus-c	2	31	-0.8770969421709667	1.686404881540754	1.1638107733387089	-2.0409077154870965
item-m-plus-c	3	28	1.1220349091964241	0.8654988375657654	-1.2141622702500008	2.336197179435714
item-m-plus-c	4	38	0.28620535582368517	1.7948490793683292	0.5231070907605265	-0.2369017348815793
item-m-safe	1	501	0.8991263378299401	0.3261258436241331	-0.26347085713692614	1.1625971949544909
item-m-safe	2	504	0.965591594697817	0.3381116020149563	-0.4783680687543645	1.4439596634464285
item-m-safe	3	479	0.9424647366889358	0.2238444848266483	-0.5412735746217119	1.4837383113045928
item-m-safe	4	516	1.017723407602713	0.21988715846672133	0.26066379129709283	0.757059616282558

Holdout, on seeds the strategies never saw: item-m-all mean 101.34648773350678, retail 126.12749112223977, arbitrage -24.781003388749372; item-m-plus-a 100.35329357035293 / 123.56234597949526 / -23.20905240914923; item-m-plus-b 102.11399036791022 / 125.7037225511236 / -23.589732183230424; item-m-plus-c 89.19867549086285 / 120.57814024763785 / -31.37946475679786; item-m-safe 94.4769774633412 / 118.4531573609182 / -23.97617989758095. All five report zero failures. These holdout means are not comparable to the 108.63974289615767 start score, which was measured on seeds 1000..1599.

Fixes that held everywhere they were tried

a2-after, a2-ledger, a3-rich, a3-free, a4-rich and a4-stock were kept in every file that measured them. a1-jumps was kept by item-m-plus-a (+0.5009455722593997) and item-m-safe (+0.8317693322811999) but lost in item-m-plus-b (-0.46009330345589977, SE 0.5348458123513586) and untested in item-m-plus-c. a1-free was kept by plus-a and plus-b and dropped by m-safe. a4-free was dropped by plus-a and plus-b and kept by plus-c and m-safe. The disagreement on a1-jumps and a4-free is inside the noise for a1-jumps and one-sided for a4-free.

Which file to enter first

item-m-plus-b. It uses all four blocks, passed the widest exactness check (600 runs per spread), measured every fix on 1000 paired runs, ran the full 4000-run holdout, and has the highest holdout mean of the five (102.11399036791022) with zero failures. Second choice is item-m-plus-a (holdout 100.35329357035293), which additionally recovered a1-jumps but ran 3000 of the 4000 holdout seeds and timed its two strategies concurrently, so its seconds are not a clean serial comparison.

The strategy code is not reproduced here, as the brief requires ("Leave the code out"); it is the code field of each item result, and item-m-plus-b's is 40014 characters with SHA-256 4f9167dddd457f394e988bf9c9e2576dd6120834bfb8d93573ce9451fd587598, item-m-plus-a's 37922 characters with 5d1bd5272884f513b19f0e66a7c235f6c188a08e4f7787ac9346f1614dda3ed3. Those files live in the repository at artifacts/goal/items/m-plus-b/result.json and artifacts/goal/items/m-plus-a/result.json. This report alone cannot carry either entry: the code is roughly 40 kB and this file is capped at 16000 bytes.

Files that are unsafe to enter

item-m-plus-c is the least safe. Its exactness rests on 30 runs per spread, its by-spread differences come from 28-43 runs with standard errors up to 4.055074254964557, its holdout covers only seeds 130000-130139 (140 runs, stated in its tried field), its base_mean_edge was never measured, its spreads 1 and 2 carry untested fixes, and it was built on a machine with one CPU where the reference CLI did not finish 600 runs in ten minutes. It reports mean edge 89.19867549086285 on holdout, far below the other files. item-m-safe was measured under PyPy 3.12.15, not the reference CPython 3.12.3, and its own note says exactness therefore holds within PyPy and does not establish CPython bit parity; its run counts were also halved. Its gain (+0.95685317308875) is the smallest of the five. item-m-all is safe to enter but is the block-only build with no fixes, so it gives up the fix gains.

Two smaller cautions: item-m-plus-a placed the a1-jumps split before the spread-1 flow-tracking block, a placement the brief did not specify, and its arithmetic is the one part of its merge not covered by an exactness check against a source file. item-m-plus-b dropped a1-jumps and a4-free, so it is not the union of every fix.