# Root merge 3 of 5: learned AMM fee policies

All 25 supplied files are present and nonempty. They include nested w01–w20 and k01–k10 records; k10 is missing in b20-merge-02. Figures remain worker reports, not independently rerun experiments. Manifest IDs bind inputs; /field, rows[key], evidence[key] and notes[key].reported locate evidence.

The largest reported score gains here come from joint-search j2/j5, about+1.20/+1.22 over HA; no top-10 or +3.1-over-HA result is established. CG's independent confirmations are positive on average but include a roughly−10-edge tail loss. kit-c demonstrates much lower gas with identical measured IB behavior. Weight-only/union continuations mostly add small gains or retain their start policy. Exact figures and qualifications follow. [b16-j2/j5:/real;b19-cf-a/b/c;b18-kit-c;b14-merge-01–04;b15-u1–u4;b17-f1–f5;b20-merge-01/02]

## Goal and evidence limits

Requester baseline: HA=T(RS2+per-run toxicity anchor)+13-8-2 learned head; +1.257 se 0.062 over T on 512 fresh real-engine seeds, 0 OutOfGas, afterSwap peak 231k; about +1.53 over RS2 (board 520.05, 27th). Top10 needs about +3.1 over HA. Pilot282fbc03 predicted real gains within about0.1 with curves rising at limits. Supplied context. HA artifact https://api.imd.fun/artifacts/2801efc8190e86318dd922d4a7179f63684d77a226047fc2f3c78232963eb690 ; codeSHA `99e3a2eceae0e5543d27473859e10b005d0ce0aa7067f2e1a24d1e0e4a7390fe`. [requester goal]

Real protocol: each strategy alone versus30-bps Vanilla, `amm_sim_rs.run_batch(X, VanillaStrategy, configs, 8)`, paired identical configs/seeds. MatchRunner draw order: size U(19,21), arrival U(.6,1), volatility U(.000882,.001008);10000steps, initial price100,x/y100/10000,mu0,dt1,retail size sigma1.2,buy probability.5. These are worker implementation claims. [b14-merge-03:evidence[w11].fields.real.config; b13-c5:/real; b19-cf-a/b/c:/notes]

Gate A: start-port/EVM <.01bps on50×2000swaps; C:frozen-port/Solidity <.1bps. Replay normally310000–310049 plus50initializations. B:start-policy replica gain over T within.15 of historical real reference, usually without a new real T arm; it does not validate the candidate itself. Candidate OOG invalidates; afterSwap limit245000. Reported valid does not certify improvement. Null=unrun/missing; identity0 is separate. [requester gates; results:/gates,/valid]

Most replica holdouts use40000000–40008191,n8192; monitors30000000–30002047,n2048,usually every5active minutes. Shared holdouts are not independent replications; best-monitor scores are selection-biased. Seed freshness is reported, not globally certified. Search time usually excludes setup/JIT/segment gaps; totals do not imply an >8-minute command. [b13-c5;b14-merge-02/03/04;b15-u1/u3/u4;b16-j1–j5;b17-f2/f4/f5;b20-merge-01/02]

Some retail/arb deltas are Y-volume because the lightweight wheel lacks trade-type edge counters. Others reconstruct arb edge from fair price, pre-step reserves/fees, with retail=total−arb. Reconstruction is not a native counter; do not pool volume and edge or mismatched decomposition samples. [b13-c5:/notes;b15-u4:/real;b17-f2–f5:/notes;b19-cf-a/b/c:/notes]

## Provenance and source failures

All25 supplied hashes match requester pins. All31 nonempty compressed payloads with sibling SHA fields decode and match (strategy/port/kit/merged-best files). This establishes bytes only: no external page, quote, raw log, checkpoint or experiment was independently rerun. To fit, this merge explicitly omits weight/code blobs, intermediate curve rows, callback coverage/timings, redundant arm means/control tails and full per-seed arrays; it retains findings, identities, paired summaries/candidate tails, units and limits. Metrics already omitted by prior merges are unrecoverable.

Failed-source rereads flagged by requester: c5=5;u1/2/3/4=3/2/2/4;j1/2/3/4/5=2/1/3/4/1;f1–f5=2each;kit-c=2;cf-a/b/c=1each. Specific failed URL/passages/audits are absent, so failures cannot be resolved. Upstream provenance/hash-at-URL,historical scores,harness behavior and B references remain qualified. Retain local measurements as item reports: outside citation failure alone neither validates nor refutes them, even if worker checks claimed success. [requester warning; named results:/notes]

Inherited failed-source counts: w01/02/03/05=2/3/5/5(w04 unflagged);w06–09=3each;w11/12/14/15=3each;w16–20 all flagged,counts absent. k01/02/03/04/05=1/1/2/3/3;k06/07/08/09=1/3/1/1. k04 claims8 live quotes passed and cf-b claims final refetched quotes passed; later reread flags still govern. [b14-merge-01/02/03/04:notes;b20-merge-01/02:source warnings;b19-cf-b:/notes]

Common reported upstream identities are T code `ebd9de9bf6af6e55d5e308fd15dbb88edec11fd684891d02a30aec185acd19b3`, replica `1e4c44b4bffbdeed65872ee15d46dbde209bf5aa11014329d13d100507b4040c`, RS2 port `9eac3553357528b06eeae3ed8d18283d83401e1968145ceba16be356a147e271`, IB code `301941829a44bb60977bb28181a5e4bfa6544f3696763f987054789b2611d686`, IB port `f63227c5e0f63633efd132d6fb402ede665e5c9cb24481911471e8dc20a3362d`, FA code `aa2c4a5637a4060088263d71ab8b109339b101bb25d64a45e9c7de83710d8728`, FA port `243a04fe20939682a8ed127f07895fdf7dce610c30d9af0a6ea0b74a536202b9`, CG code `e1ea8fcb105e00754d17061efbe4b299a634b0fcc1b5c293d65b631654b7204b`. Setup is reported pinned to repository commit `6b6921f8d463314aa7c96234e1be25a25245f704` and wheel SHA `aedb68280aca34c110968f00148387c95885207913fb2cf19a1ab4c5cfd6887a`. These are attributed identities, not a fresh external authentication. [b15-u1:/gates.hashes; b15-u4:/gates.input_hashes; b16-j1/j5:/notes; b17-f1:/notes; b19-cf-a/b/c:/cg_code_sha256; b20-merge-01/02]

### Input manifest

IDs denote `.imd/reads/artifacts/<ID>` (not in Git). SHA-256 binds supplied bytes; family/item names identify original logical sources.

| ID | logical result(s) | SHA-256 |
|---|---|---|
| b13-c5 | r4-ib-continue/c5 | 10233ef4dc4e1f2284b897d320f4ab0595e6cc18d8715935da219ff65640a97c |
| b14-merge-01 | r4-ib-wide merge 01 | 0178ac525541f84d85870091cfa9be61a70a9670c6686e1951671e6d5deaacd8 |
| b14-merge-02 | r4-ib-wide merge 02 | a434de487b0a30acd8f6ef24ab17de93d8b72ca4b48334d302f6cf54559c3d0a |
| b14-merge-03 | r4-ib-wide merge 03 | b1505dbab6f01f1cd95558e407e28c2cc496356c814b0bc8d6e5e0f430b70271 |
| b14-merge-04 | r4-ib-wide merge 04 | e3b514885cef61b3fcfe0d8600294994f8986826eb6e16858fc36e8c08e25f70 |
| b15-u1 | r4-ib-union/u1 | d18af14a68e6c7651dff2772b354a6087190d5cb3244b73b4a4f866e2300f08e |
| b15-u2 | r4-ib-union/u2 | d585917bbd06330138375ea6bc6be42cf4a39040b82900cfe21a6f12ec8fe20c |
| b15-u3 | r4-ib-union/u3 | 4584767768c2b1ff77109448eb2710b997bf5ec42a8fd321081de1179b7e77af |
| b15-u4 | r4-ib-union/u4 | 47731657597be8a347f15f61b7ee693823a7f798a14cdf7ed6d29b4d00eadebd |
| b16-j1 | r4-ib-joint/j1 | e1e3672e192028922bbd917c8cb6f2281515a8a70ece498723c7d98cd90d8c30 |
| b16-j2 | r4-ib-joint/j2 | 4d396adc0e3056c010e59ab8c3d413eb0380a14e8cefa6cd825e809fa2494999 |
| b16-j3 | r4-ib-joint/j3 | 2746ffc11654afd24f90670243b7a0b9c7849b1882705886ef7c96a53597d582 |
| b16-j4 | r4-ib-joint/j4 | a266fdaf78689a56c354142bfc1651a9d0e814a5ebd6992c5b2b326e86f8f8b1 |
| b16-j5 | r4-ib-joint/j5 | fe2320b5c091489337b3eb06bcbaea19bf4a8af09d959d67504ba63184f1dab6 |
| b17-f1 | r4-fa-continue/f1 | 29df468c1e046ab6ccdf8b2a5a559418e44cddeceb4a8bf84b4a4549da340e61 |
| b17-f2 | r4-fa-continue/f2 | 27f399ab9380f78712a6835acba6ba5d63eaa623b3feee7bb6d5bc5ee5f65008 |
| b17-f3 | r4-fa-continue/f3 | 82afff1f3f27e7979da1652260e0044495aed85ecbfcb993deb8a9ebfdc8fb64 |
| b17-f4 | r4-fa-continue/f4 | 85a97b190c29e80d3aa2f5ef88b0a9da073ed08851676ebd6aa073a8d4b952d8 |
| b17-f5 | r4-fa-continue/f5 | 9e51be6bc457f72885541d8e9bc159dce4e5c5c1cdad655d2bf04430d3e01845 |
| b18-kit-c | r4-ib-kit/kit-c | 71695e4eb1e3c194637eafe539ddf967f9e1a27f2f7f6690fb86a6e373f1d706 |
| b19-cf-a | r5-cg-confirm/cf-a | 43fbda381d45d36627d594159b5ead7aa216881e67b74255ad6ab4a984b93680 |
| b19-cf-b | r5-cg-confirm/cf-b | 39491ce07390332f94bf7993fcd56f3c22171e174149b7236988c4d7b742d258 |
| b19-cf-c | r5-cg-confirm/cf-c | cd9c91c2b46eabe1b95d5cca006813f9d206c18ba1715a40e522ec916eda35d1 |
| b20-merge-01 | r5-cg-continue merge 01 | 762cc7a1fc7dc0083d67747680cf23d8448bab9b9c9d3e004c89e64a1f2bd3ea |
| b20-merge-02 | r5-cg-continue merge 02 | 89b5c10b32b13abaebbe5d138966aba4da29d6ccbc1324272153587c67d6ec17 |

Numerical evidence blocks below use up to nine significant digits to fit this bounded merge; narrative figures retain supplied precision unless marked approximate; tables also use up to nine significant digits. Abbreviations: m=paired mean_diff, se=paired standard error, w5=worst5 [seed,difference], r/a=retail/arb delta, rv/av=retail/arb Y-volume delta, codeSHA/portSHA=SHA-256 of decoded payload. Comparison keys IB/HA/FA/CG/base mean candidate minus that arm; base=IB for j1–j4 and IA for j5. Units are stated in the surrounding family text. Paired summaries use [n,m,se,w5] (w5 omitted if absent); component summaries use [n-or-null,m,se]; p holds this array where extra component fields exist. CF redundant arm means are explicitly omitted. Control tails are omitted; candidate tails remain. Rounding can cause tiny residuals.

## IB continuation c5

c5:CMA-ES178IB weights,pop16,M512,ES405,train740000000–740114687,sigma.08×blockRMS/30-min half-life;60.1min/224gens,nine saved segments. Speed362.55091250798057 markets/s(6processes);notes362–479 and ~540generation throughput. Initial monitor1.8095714223753105 overT remains best:dip~1.665 at5min,recovery1.7698336986581793 at55.25,1.781365419055533 at60.03. Frozen code is byte-identical IB. Suggested restart sigma~.02 is untested. [b13-c5:/speed,/search,/curve,/notes]

A/C maximum error is 6.64202309125228e-05 bps; C reuses identical frozen policy semantics. B = 1.708612109706731, SE 0.04425219850674254 versus 1.749, absolute error 0.04038789029326906, on 7132768–7133279. Holdout over T = 1.768984109173352, SE 0.01134997051758743; over HA = 0.6521520466734023, SE 0.007848748423475268; over IB = 0, SE 0 (identity). Replica retail/arb edge over T = 5.050611674119493/−3.2816275649461413, over HA = 0.9975757379632462/−0.3454236912898442. Holdout reproduces the cited IB validation exactly, as locally reported. Gas swap/init = 234852/243227, OOG/other failures 0; HA 231516/243147, OOG 0. A T timing probe used 7000000–7000007 outside confirmation. [b13-c5:/gates,/holdout,/gas,/notes]

Real evidence [b13-c5:/real]; repeated IB_vs_HA duplicates vs_HA and is omitted:

```json
{"seeds":[7556000,7556511],"IB":{"p":[512,0.0,0.0,[[7556000,0.0],[7556001,0.0],[7556002,0.0],[7556003,0.0],[7556004,0.0]]],"rv":0.0,"av":0.0,"pnl":0.0},"HA":{"p":[512,0.62418701,0.0312659858,[[7556331,-2.8813798],[7556345,-1.94207148],[7556020,-1.46063661],[7556322,-1.30946277],[7556486,-1.29135485]]],"rv":642.590547,"av":483.967519,"pnl":0.602004488},"r":{"HA":642.590547,"IB":0.0,"unit":"retail volume Y per market"},"a":{"HA":483.967519,"IB":0.0,"unit":"arb volume Y per market"}}
```

## IB-wide: w01–w20

w01–05 source=b14-merge-01,w06–10=02,w11–15=03,w16–20=04,all `/rows`. S/V=searched/reported-valid;H_IB/H_HA=replica mean(SE)/mean;real comparisons follow;pop/M=population/markets per generation;ABC=reported gate flags(null=unrun).

| item/source | S/V | speed | minutes/gens | pop/M | H_IB (SE) | H_HA | swap/OOG | ABC |
|---|---|---|---|---|---|---|---|---|
| w01/b14-merge-01 | True/True | 1838.40371 | 45.4032456/605 | 16/512 | 0.106491282 (0.00594010586) | 0.758643329 | 236944/0 | {"A":true,"B":true,"C":true} |
| w02/b14-merge-01 | False/True | 116.769634 | 0/0 | null/null | null | null | 237710/0 | {"A":true,"B":true,"C":true} |
| w03/b14-merge-01 | True/True | 328.512711 | 58.1745733/66 | 32/512 | 0.0540609506 (0.0048304321) | 0.706212997 | 237686/0 | {"A":true,"B":true,"C":true} |
| w04/b14-merge-01 | False/True | 18.8248371 | 0/0 | null/null | null | null | 234817/0 | {"A":true,"B":true,"C":true} |
| w05/b14-merge-01 | False/True | 68.1101301 | 0/0 | 16/512 | null | null | 234872/0 | {"A":true,"B":true,"C":true} |
| w06/b14-merge-02 | True/True | 275.965926 | 60.7095502/77 | 16/512 | 0.0330306945 (0.00410195241) | 0.685182741 | 237993/0 | {"A":true,"B":true,"C":true} |
| w07/b14-merge-02 | False/True | 116.578081 | 0/0 | null/null | null | null | 237016/0 | {"A":true,"B":true,"C":true} |
| w08/b14-merge-02 | True/True | 507.433409 | 75.0603884/148 | 32/512 | 0.0419425875 (0.00502999216) | 0.694094634 | 237039/0 | {"A":true,"B":true,"C":true} |
| w09/b14-merge-02 | False/True | 95.2479549 | 0/0 | null/null | null | null | 234946/0 | {"A":true,"B":true,"C":true} |
| w10/b14-merge-02 | False/False | null | null/null | null/null | null | null | null/null | {"A":null,"B":null,"C":null} |
| w11/b14-merge-03 | False/True | 98.2859139 | 0/0 | null/null | null | null | 237040/0 | {"A":true,"B":true,"C":true} |
| w12/b14-merge-03 | False/True | 119.565473 | 0/0 | null/null | null | null | 237040/0 | {"A":true,"B":true,"C":true} |
| w13/b14-merge-03 | False/False | null | null/null | null/null | null | null | null/null | {"A":null,"B":null,"C":null} |
| w14/b14-merge-03 | True/True | 554.8 | 46.17/174 | 16/512 | 0.0570679846 (0.00511087894) | 0.709220031 | 234971/0 | {"A":true,"B":true,"C":true} |
| w15/b14-merge-03 | True/True | 658.396593 | 75.6805648/367 | 16/512 | 0.0464569109 (0.00578840956) | 0.698608958 | 236922/0 | {"A":true,"B":true,"C":true} |
| w16/b14-merge-04 | True/True | 368.578489 | 65.2504203/194 | 16/512 | 0.0736512359 (0.00569917129) | 0.725803283 | 234860/0 | {"A":true,"B":true,"C":true} |
| w17/b14-merge-04 | True/True | 1086.11944 | 67.672857/299 | 16/1024 | 0.111991568 (0.0057495358) | 0.764143615 | 234954/0 | {"A":true,"B":true,"C":true} |
| w18/b14-merge-04 | False/True | [106.2,91.2,85.1,82.9] | 0/0 | null/null | null | null | 237056/0 | {"A":null,"B":true,"C":null} |
| w19/b14-merge-04 | False/True | 85.3 | 0/0 | null/null | null | null | 232015/0 | {"A":null,"B":true,"C":null} |
| w20/b14-merge-04 | True/True | 236.378541 | 60.8860768/87 | 16/512 | 0.032099437 (0.00537759933) | 0.684251484 | 238423/0 | {"A":true,"B":true,"C":true} |

Nested winner codeSHA [w01=adbf4f8c79c70cebfdecf2c0db65d869aa506068374223eddac636196b40b63a; w08=5dfd0a55d44c3d48ffd101ed650b0e0ded5ae201349c2c723ae817162ce4773a; w14=bc83647a36a46038b52798e41d304ffea1527d4c4de277a5209e6dc8a60e37b0; w17=ac6beb29a0982926a08b8db19310bac18b77cd19f474767e18c308d312c4b9a1]. [b14-merge-01/02/03/04:/best]

Only w01,w03,w06,w08,w14,w15,w16,w17,w20 trained. Submerge bests:w01,w08,w14,w17;w17 highest supplied holdout. Shared holdouts/different real blocks preclude naive pooling or a cross-policy significance claim. b14-merge-02's “two combined SE” comparison w06/w08 is heuristic because paired covariance is absent. [b14-merge-01/02/03/04:rows,best,notes]

w02,w04,w05,w07,w09,w11,w12,w18,w19 retain IB without valid training. Null vs_IB(w02,w12,w18,w19) stays null; other fallback zeros are identity. w04 completed176/512 pairs. w10 setup uv venv failed126(Linux x86_64 binary incompatible);w13 uv download/extraction failed HTTP500/exit2. Neither measured gates/gas/candidate. [b14-merge-01/02/03/04:notes]

w05 speed68.11013007875535 conflicts with note155.1368;both trigger fallback. Wrong-binding/stale-helper training was discarded;0gens means retained work;pop/M are plans. Scalar volume-delta comparator is unstated(IB/HA only contextual);HA OOG detail was undelivered scratch. [b14-merge-01:notes,rows[w05].real_evidence]

Throughput caveats: w07 initial109.871 overlapped parity,final116.57808131918684;serialization-bug7584000–7584031 dropped,not rerun. w15 initial658.3965933647448 vs later8/16/32-worker146.07128051840684/100.08795565302971/129.48771645212986 unresolved;8workers retained. w18 four repeats[106.2,91.2,85.1,82.9],not CI. w19 initial85.3,repeats82.9/86.8,serial130.3,2/4/8-worker123.8/124.2/105.7:more processes did not help. [b14-merge-02:notes;b14-merge-03:evidence[w15].fields.speed;b14-merge-04:notes[w18/w19].reported.speed]

Freeze details:w14 gen154/min40.69,monitor1.8665033988487048,total174gens;prose says3checks but lists4,saturation unproven. w15 gen367,monitor1.8445838655997129. w16 discarded missing-RNG-checkpoint runs,restarted disjoint seeds;reported sampling cap differs from later post-tell/monitor sigma logs. w17 gen288/min65.147989144444,monitor1.9135761204552957(SE .023760272075181197),total299. w20 min60.7179281442892,monitor1.85920417454552. [b14-merge-03:notes,evidence.search;b14-merge-04:notes.reported.search]

w06 monitor~1.8096→1.8228/60.7min,w08~1.8096→1.8551/75.1min,final sigma~.009–.012. Replica/real vsIB .0330/.0525 and .0419/.0629 agree within~.1,with small gains. w01 freezes best scheduled checkpoint. [b14-merge-01/02:notes]

IB-wide A mostly6.64202309125228e-05bps;w08/w09 instead5.224879067594612e-05 on nominally same replay seeds(generator differences suggested,unresolved). Both pass. Common B1.708612109706731(SE .04425219850674254)vs1.749,seeds7132768–7133279 is deterministic historical calibration. w18/w19 A/C explicitly unrun(null);identity fallbacks often reuse A for C. [b14-merge-01/02/03/04:gates,notes]

w20 originally includes7636000–7636511 despite prior nominal-config IB probe7636000,contradicting “excluded” note. Strict-fresh repair7636001–7636512 drops it/adds unseen7636512,n512:vsIB .024600454236423097(SE .01743357408105432),vsHA .7308999796478691(SE .03633675002718842). Table uses repair;below preserves original volumes,not repaired values(no repaired decomposition/means supplied). Gas includes supplement. w06/w08 both probed7590000 outside confirmation;diagnostics were not globally fresh. [b14-merge-04:notes[w20];b14-merge-02:notes]

Real evidence blocks cite supplied submerge fields; original item files are unavailable. Default components are Y-volume except w06/w07/w15 reconstructed edge(w15 also volumes). Total-edge candidate tails retained; component/control tails omitted explicitly. [b14-merge-01/02/03/04:real evidence,notes]

**w01** [b14-merge-01:rows[w01].real_evidence]

```json
{"seeds":[7560000,7560511],"r":{"IB":[512,46.2835171,4.34978573],"HA":[512,665.853678,11.3703381]},"a":{"IB":[512,10.9428661,3.44576848],"HA":[512,490.508328,5.30165251]},"IB":[512,0.0865100837,0.0202784541,[[7560356,-3.10378268],[7560416,-1.37946049],[7560399,-1.22517549],[7560081,-1.15318576],[7560137,-1.14944686]]],"HA":[512,0.686797303,0.0360548751,[[7560152,-6.49687087],[7560356,-3.16990096],[7560348,-1.83427322],[7560345,-1.71819109],[7560497,-1.59291893]]]}
```

**w02** [b14-merge-01:rows[w02].real_evidence]

```json
{"seeds":[7564000,7565599],"IB":null,"HA":[1600,0.636625142,0.017239056,[[7564122,-1.72897983],[7565430,-1.58881786],[7564656,-1.5485268],[7564426,-1.36026451],[7564324,-1.35788355]]],"requested_512":{"seeds":[7564000,7564511],"HA":[512,0.652047124,0.0313498951,[[7564122,-1.72897983],[7564426,-1.36026451],[7564324,-1.35788355],[7564457,-1.00040832],[7564440,-0.968585742]]]},"r":636.028533,"a":486.314145,"r_se":4.88021862,"a_se":2.9016679,"pnl":[1600,0.632555108,0.0217580991,[[7564871,-3.05470003],[7564932,-2.53734716],[7564556,-2.40970875],[7564839,-2.30774832],[7564624,-2.29123477]]]}
```

**w03** [b14-merge-01:rows[w03].real_evidence]

```json
{"seeds":[7568000,7568511],"IB":[512,0.0422627245,0.0151476655,[[7568008,-1.41915875],[7568331,-1.06149567],[7568384,-1.04237728],[7568188,-1.0073256],[7568253,-0.97844181]]],"HA":[512,0.657738642,0.0302440082,[[7568340,-2.75076621],[7568326,-1.92465396],[7568400,-1.21101025],[7568302,-1.07553827],[7568311,-1.0326803]]],"IB_vs_HA_same_seeds":[512,0.615475917,0.0308939449],"r":{"metric":"mean retail_volume_y of the submission (the real engine result exposes volumes, not a retail/arb edge split)","IB":45.2654692,"HA":668.016915},"a":{"metric":"mean arb_volume_y of the submission","IB":54.5129649,"HA":536.037389},"pnl":{"IB":0.0439523525,"HA":0.715967511}}
```

**w04** [b14-merge-01:rows[w04].real_evidence]

```json
{"seeds":[7572000,7572175],"IB":[176,0.0,0.0,[]],"HA":[176,0.566162736,0.0482466436,[[7572053,-1.73151979],[7572167,-1.42011151],[7572149,-0.764423536],[7572077,-0.577417763],[7572013,-0.570962931]]],"r":621.597318,"a":479.912139,"pnl":0.593090123}
```

**w05** [b14-merge-01:rows[w05].real_evidence]

```json
{"seeds":[7576000,7576511],"paired_n":512,"IB":[512,0.0,0.0,[[7576000,0.0],[7576001,0.0],[7576002,0.0],[7576003,0.0],[7576004,0.0]]],"HA":[512,0.645612471,0.030933494,[[7576188,-1.74250003],[7576106,-1.34557038],[7576103,-1.1801779],[7576047,-1.0642844],[7576502,-1.00037939]]],"r":613.504514,"a":476.394647}
```

**w11** [b14-merge-03:evidence[w11].fields.real]

```json
{"seeds":[7600000,7601823],"IB":[1824,0.0,0.0,[[7600000,0.0],[7600001,0.0],[7600002,0.0],[7600003,0.0],[7600004,0.0]]],"HA":[1824,0.66786689,0.0177394507,[[7601227,-6.78130524],[7601158,-2.9163325],[7601245,-2.71180922],[7600056,-2.60628599],[7601806,-2.27914436]]],"r":{"HA":629.748815,"IB":0.0,"unit":"mean retail_volume_y per market"},"a":{"HA":482.533978,"IB":0.0,"unit":"mean arb_volume_y per market"}}
```

**w12** [b14-merge-03:evidence[w12].fields.real]

```json
{"seeds":[7604000,7606943],"IB":null,"HA":[2944,0.649862381,0.0130495297,[[7604900,-3.27357245],[7604424,-2.60756392],[7604149,-1.96252462],[7605985,-1.91781246],[7606493,-1.78911537]]],"HA_vs_IB":[2944,-0.649862381,0.0130495297],"r":{"metric":"retail_volume_y","IB_minus_HA":635.322774},"a":{"metric":"arb_volume_y","IB_minus_HA":488.320357}}
```

**w13** [b14-merge-03:evidence[w13].fields.real]

```json
{"seeds":[],"IB":[0,null,null,null],"HA":[0,null,null,null],"r":null,"a":null}
```

**w14** [b14-merge-03:evidence[w14].fields.real]

```json
{"seeds":[7612000,7612511],"IB":{"p":[512,0.0592349211,0.0166071711,[[7612374,-2.62299017],[7612391,-1.50396424],[7612472,-1.34525021],[7612369,-1.29374145],[7612091,-1.15395989]]],"rv":-54.8685056,"av":-64.3013375},"HA":{"p":[512,0.705180475,0.0337216024,[[7612284,-4.83070023],[7612384,-1.60901476],[7612374,-1.58962315],[7612189,-1.5601525],[7612223,-1.33816517]]],"rv":587.531298,"av":419.174653},"IB_vs_HA_same_seeds":{"p":[512,0.645945554,0.0329698778],"rv":642.399804,"av":483.47599},"r":{"IB":-54.8685056,"HA":587.531298},"a":{"IB":-64.3013375,"HA":419.174653}}
```

**w15** [b14-merge-03:evidence[w15].fields.real]

```json
{"seeds":[7616000,7616511],"r":{"IB":[null,-0.583646602,0.0186901325],"HA":[null,0.40633558,0.034082001]},"a":{"IB":[null,0.609096639,0.00860313779],"HA":[null,0.266601517,0.0162251707]},"rv":{"IB":[null,-205.777256,3.04511633],"HA":[null,431.169457,8.66492496]},"av":{"IB":[null,-278.773509,3.07472204],"HA":[null,203.918216,5.11372494]},"reconciliation_max_abs":{"IB":1.13686838e-13,"HA":1.13686838e-13},"IB":[512,0.0254500363,0.0178169316,[[7616042,-1.57845552],[7616209,-1.49780713],[7616109,-1.40476036],[7616217,-1.18090538],[7616270,-1.06282296]]],"HA":[512,0.672937097,0.034563808,[[7616150,-2.60504587],[7616335,-2.51025567],[7616429,-1.56971217],[7616207,-1.37585517],[7616303,-1.22621886]]],"IB_vs_HA":[512,0.647487061,0.0327653875]}
```

**w16** [b14-merge-04:notes[w16].reported.real]

```json
{"seeds":[7620000,7620511],"IB":[512,0.0905258122,0.017990967,[[7620356,-2.47930405],[7620002,-1.26200111],[7620360,-1.14994588],[7620024,-1.04723549],[7620332,-0.978465249]]],"HA":[512,0.713253179,0.0308617571,[[7620196,-2.45289269],[7620444,-2.20709561],[7620015,-1.67564472],[7620356,-1.6605375],[7620316,-1.07810134]]],"r":{"IB":-26.6433343,"HA":609.226},"a":{"IB":-7.58309486,"HA":472.67826}}
```

**w17** [b14-merge-04:notes[w17].reported.real]

```json
{"seeds":[7624000,7624511],"IB":[512,0.10179203,0.0185076719,[[7624079,-1.76675469],[7624279,-1.32658603],[7624355,-1.30771448],[7624354,-1.00107141],[7624273,-0.939291802]]],"HA":[512,0.696698266,0.032764446,[[7624490,-3.06238404],[7624487,-2.02464217],[7624405,-1.51374012],[7624199,-1.34483005],[7624255,-1.31487076]]],"r":{"IB":163.972481,"HA":776.905489,"unit":"mean retail_volume_y (the real engine exposes no retail/arb edge split)"},"a":{"IB":66.9918618,"HA":554.416332,"unit":"mean arb_volume_y"},"pnl":{"IB":0.0915091572,"HA":0.711808225}}
```

**w18** [b14-merge-04:notes[w18].reported.real]

```json
{"seeds":[7628000,7628927],"HA":[928,0.659134071,0.0235523611,[[7628006,-3.15138507],[7628070,-2.1617023],[7628677,-1.53624277],[7628555,-1.47133729],[7628506,-1.43963302]]],"IB":null,"r":637.890003,"a":483.392317}
```

**w19** [b14-merge-04:notes[w19].reported.real]

```json
{"seeds":[7632000,7632319],"IB":null,"IB_vs_HA":[320,0.592880778,0.0369867941],"r":631.712989,"a":486.658014,"pnl":0.575661453}
```

**w20** [b14-merge-04:notes[w20].reported.real]

```json
{"seeds":[7636000,7636511],"IB":[512,0.0241217307,0.0174352066,[[7636297,-1.97077957],[7636390,-1.53076513],[7636206,-1.29591172],[7636340,-1.28073728],[7636086,-1.11235855]]],"HA":[512,0.734361235,0.0362641266,[[7636031,-2.14920136],[7636297,-1.88470004],[7636122,-1.26826308],[7636123,-1.20132711],[7636355,-1.10469443]]],"r":{"IB":-32.8567402,"HA":603.671407},"a":{"IB":60.5669378,"HA":541.331756},"strict_fresh_confirmation":{"seeds":[7636001,7636512],"IB":[512,0.0246004542,0.0174335741,[[7636297,-1.97077957],[7636390,-1.53076513],[7636206,-1.29591172],[7636340,-1.28073728],[7636086,-1.11235855]]],"HA":[512,0.73089998,0.03633675,[[7636031,-2.14920136],[7636297,-1.88470004],[7636122,-1.26826308],[7636123,-1.20132711],[7636355,-1.10469443]]],"n":512}}
```

b14-merge-02 does not retain full paired worst5, original real blocks or all component uncertainties; those cannot be recovered here. Its notes preserve these real outcomes (mean,SE,n; retail/arb units as labeled):

- w06: seeds 7580000–7580511; vs_IB (0.0525245023517521,0.015324409084365631,512), vs_HA (0.7067310415787678,0.038053443609535005,512). Reconstructed edge vs_IB retail −0.05826141007674013 (SE 0.015563453669901545), arb 0.11078591242849026 (SE 0.00670908456815188).
- w07 unchanged IB: seeds 7584032–7585087; IB−HA (0.6642513815075862,0.021755339484222994,1056), reconstructed edge decomposition noted but component values not retained in supplied merge.
- w08: seeds 7588000–7588511; vs_IB (0.06288579925944027,0.01383265485095135,512), vs_HA (0.659927865646242,0.03460751527018567,512); real deltas are volumes but their values are not retained. Replica holdout vs_IB edge retail −0.052500830088624474, arb 0.09444341757055962.
- w09 unchanged IB: seeds 7592000–7593535; IB−HA (0.6498760072868056,0.0192648148126186,1536); real component volumes not retained; IB/HA swap peaks 234946/231555, OOG 0.
- w10: setup blocked; no measured speed, gates, holdout, real or gas.

[b14-merge-02:rows,notes; unavailable details are not filled in]

## Union, joint and FA continuation experiments

Direct rows cite each input `/speed,/search,/holdout,/gas,/valid`. H_base=IB(u1–u4,j1–j4),IA(j5),FA(f1–f5);H_HA=replica mean(SE). All measured candidate OOG0;u2/f1 gas null.

| source | valid | speed | minutes/gens | H_base (SE) | H_HA (SE) | swap peak |
|---|---|---|---|---|---|---|
| b15-u1 | true | 103.213004 | 65.4811359/112 | 0.0555595752 (0.00422961673) | 0.707711622 (0.00802537633) | 236891 |
| b15-u2 | false | upstream only 4181.9 / 5980.7 | 0/0 | null | null | null |
| b15-u3 | true | 242.999756 | 50.2733134/138 | 0.0239628883 (0.00574894747) | 0.676114935 (0.00863006526) | 237108 |
| b15-u4 | true | [454.273173,522.670618] | 70.509115/127 | 0.0514276008 (0.0052494558) | 0.703579647 (0.00869933685) | 236759 |
| b16-j1 | true | 298.511426 | 46.0100485/136 | 3.1581156e-05 (2.48603197e-05) | 0.652183628 (0.00784887249) | 237023 |
| b16-j2 | true | {"T_512_markets_24_procs":[1352.6,1411],"IB_start_512_markets":1374.2,"population_4x512_markets":1565.2} | 56.4/327 | 0.505486871 (0.00775592155) | 1.1576705 (0.00979113774) | 237095 |
| b16-j3 | true | 1570.74626 | 40.5973628/452 | 0 (0) | 0.652152047 (0.00784874842) | 236950 |
| b16-j4 | true | 251.297646 | 40.9348809/42 | 3.1581156e-05 (2.48603197e-05) | 0.652183628 (0.00784887249) | 237151 |
| b16-j5 | true | 5264.22737 | 40.81/1496 | 0.82583284 (0.00934936924) | 1.19071201 (0.0101443433) | 237419 |
| b17-f1 | false | null | 0/0 | null | null | null |
| b17-f2 | true | 1373.32558 | 45.0401248/513 | 0.169810902 (0.00554062022) | 0.73407822 (0.00765001763) | 234704 |
| b17-f3 | true | 55.1141869 | 0/0 | null | null | 236683 |
| b17-f4 | true | 293.842927 | 70.2117698/68 | 0.0483217716 (0.00364434331) | 0.612589089 (0.00719575327) | 234657 |
| b17-f5 | true | 412.91761 | 60.3571977/211 | 0.0646095507 (0.00417547881) | 0.628876869 (0.0072201928) | 231849 |

### Union: 23-8-2 head

u1,u3,u4 keep IB's original 19 inputs and add four existing-state features: T bid fee×200−1, T ask fee×200−1 (after T clamps, before head), (spot−updated pHat)/pHat×500, retail flow rate×2−1 (slot15 low uint80 / 1e18). All features clip to [−4,4], hidden ReLU clips [0,4], correction is 0.0001×(W2h+b2), final fee [0,0.1]. The 32 added W1 entries start at zero, preserving IB exactly before training. No new slots or initialization writes; architecture 23-8-2, 210 learned values. u1/u3 gas probes retained all four features; u4's 16-seed probe with new weights 0.01 measured swap 233688/init 243260, no failures. [b15-u1/u3/u4:/input_formulas,/gates,/search]

IB order retained:sigmaHat,lambdaHat,sizeHat,toxEma,actEma,direction,stale-price deviation,signed trade ratio,arb probability,first-in-step,uncertainty,silence,anchor,amountY,step count,prev bid/ask,arb-recency,router deviation. Old scaling unchanged:amountY/20−1,(count−1)×.5,prevFee/.003−1,min(time−lastArb,16)×.125−1;staleHat=.7pHat+.3fastHat. [b15-u1/u4:/input_formulas]

u2 invalid/incomplete:no implementation,weights,port,gates,holdout,real or gas. IB/FA have different19-input orders;combining requires work,not shown impossible(u1/u3/u4 do it). Its4181.9/5980.7 speeds are upstream,not measured here. [b15-u2:/notes,/speed,/measured_artifacts]

u1:CMA-ES461,pop16,M512,sigma.05RMS/half-life30,train950000000–950057343. It trained despite103.21300430284467 initial speed:do not import wide-family fallback into this brief. Monitor selected65.37582298652657min,overT1.8703596548231458,sigma.01103995273325344;32new weights nonzero. [b15-u1:/search,/curve]

u3:ES463,pop16,M512,sigma.08/half-life30,train970000000–970070655. Persistent pool after40min,256-market exact parity,seconds1.2320673959329724→1.0779741941951215;active404.99684568624923 markets/s. Best monitor1.8300930431568476(SE .023276322520075183)at45.365035269603446min vs final1.828817317724705 at50.27331336083201. [b15-u3:/speed,/search,/curve]

u4:ES464,pop16,M1024,sigma.05RMS(normalized1)/half-life30,train980000000–980130047. Aftergen38 fixed-T literal specialization(no fastmath),dynamic210weights:exact100000callback/512-market parity,speed454.27317289617497→522.6706181687712. Selected1.861666539814693 at65.38762700339235min vs final1.8584171796030873 at70.45011582461109. Rounded recheck1.8617576239700628,max rounding4.985345927188689e-07 was not selection. Config256exact,RNG roundtrip verified;12no-retail/32short probes,max residual9.896083952298795e-12. [b15-u4:/search,/gates.extra_checks,/gates.replica_specialization,/curve]

A is 6.64202309125228e-05 bps for u1/u3/u4, with zero initial union-vs-IB EVM error; C respectively 4.0791483680441276e-05 / 3.0710960424035205e-05 / 4.8719526329898954e-05 bps. Their B calibration is the common IB 1.708612109706731, SE .04425219850674254 against 1.749. Real u1/u3 components are volumes; u4 components are reconstructed edge. [b15-u1/u3/u4:/gates,/real]

### Joint search: constants and head weights

Joint outcomes:j1(178weights+82constants,pop16,M512,ES471),j3(178+83,pop16,M512,ES473),j4(81constants,head fixed,pop16,M512,ES474)never beat initial IB. Successful j2/j5 use41constants+178weights;dimensionality is a research lead,not a proven cause. j1/j4 tiny freeze differences are float/decay/WAD effects; j3 replica gain0 but tiny real differences. [b16-j1–j5:/search,/holdout,/notes]

Joint knobs:start×2^x,x∈[−1,1];decay1−(1−d0)2^x;integers rounded,structural units/modes/layout fixed;zero REVERT_U cannot move. j1 weights.05RMS/constants.03; j3/j4 scalar.03. j1 train990000000–990069631;j3 retained1010047104–1010278527 after discarded6.033571602000545min/92gen infeasible-decay pilot1010000000–1010047103(firstresume reseeds473,later restores RNG);j4 train1020000000–1020021503. Late monitors overT:j1−.8023887031495726/45.2563min,j3−5.968753129583063/40.0053,j4−3.3684654082195182/40.5465;initial retained. [b16-j1/j3/j4:/search,/curve,/notes]

j2:IB start,219dims,pop32,M512,ES472,weight.05RMS/constants.03,half-life30.41knobs=38nonzero declared constants+anchor cap/2coefficients,inlineTH fixed;train1000000000–1000167423. Monitor1.8095191677040614→2.275149399992105/45.09→2.3097660837745986/55.07(selected),still rising at56.4stop. Holdout/real vsIB .5054868709453137(SE .00775592155253877)/.578109776191504(SE .05543401962784782);overHA1.157670498774741/1.1991943876660298. No new realT arm. Rounded start holdout1.769015690329377 overT differs from original1.768984109173352. [b16-j2:/search,/curve,/holdout,/real,/notes]

j5:IA start(notIB),178+41knobs,ES475,pop16,M512,1496gens;start-freeze reproduces IA byte-for-byte. Speed5264.227367326977(32processes),T7129.71244546849,generation~1.64s;train1030000000+fresh512blocks. Monitor1.539624226348283→2.3309366632762156/35.01min→2.3568343475047095/40.02(gen1467selected),stop40.81. RealIA .8762491940424664(SE .03521012376831418)vs holdout .8258328396747168(SE .00934936923764563),gap~.050. Real decomposition:arb+.9584623776328983,retail−.08221318359043195. IA codeSHA `956566e95f46dc1ed5980fae5e550de7b9da10f9b3ad8eed62f265682cfcecf2`. [b16-j5:/speed,/search,/curve,/holdout,/real,/notes]

j1/j3 Gate A report5.224879067594612e-05 bps; j2/j4 A6.64202309125228e-05. C j1/j4=6.64202309125228e-05, j2=.00013407067983584042, j3=5.224879071064059e-05. j2 B differs slightly from the common IB value:1.708724722605055(SE .04425251372493472), error .040275277394945075 vs1.749, reflecting its own parameter freezing. j5 A6.501554232485329e-05,C.00015195374049314925; its B is IA−T1.4654026984931672(SE .04614738409203076) vs1.513 on7116384–7116895, HA−T1.0989513895749443(SE .03682188500341987). All reported pass. [b16-j1–j5:/gates]

### FA continuation

FA's19-input order differs from IB;continuations preserve scaling/architecture,change head literals only. f1 setup uv venv failed “Exec format error”,exit126:reported input/wheel/port hashes checked,but no candidate,training,gates,holdout/real. [b17-f1:/blocked,/notes]

FA training:f2 ES482,pop16,M512,sigma.05RMS,train1050000000–1050262655,45.040124844483216min/513gens;monitor1.7285938423048703→1.8864885711027362/40.2826→1.896867037766382. f4 ES484,pop32,M512,sigma.05,train1070000000–1070034815,70.21176976767214min/68gens;best1.7715039748105736/65.2481 vs final1.7714539185668006. f5 ES485,pop16,M512,sigma.03,train1080018944–1080126975,60.35719768780594min/211gens;monitor1.7285938423048703→1.7820892204611938/55.2931→1.7861934817905385. Half-life30,freeze1e-6;f5 discarded36gen without RNG checkpoint,restarted fresh with optimizer+RNG saved. [b17-f2/f4/f5:/search,/curve,/notes]

f3 unchangedFA fallback:speed49.62809118931912→55.11418693042621(final other simulations paused);no training/holdout,vsFA null,768total pairs. Decomposition704seeds7684064–7684767(first64missing,never rerun). f2 total512/components448seeds7680064–7680511. f4 total/volumes512,componentsFA448/HA512. Subset component means need not sum to full-sample totals. [b17-f2/f3/f4:/real,/notes]

FA Gate B is1.7050245300047933(SE .04995197116469658) vs1.65 on7004000–7004511; all measuring FA runs report A.0001210463733067324bps. C f2=.00010656166265360367,f3=A reused (same code),f4=6.608836767771353e-05,f5=.00019982502666437973; all pass. f3 near-zero-retail diagnostics:10 selected no-fill cases out of32 short probes, max arb-edge error7.716938199564538e-12, harness SHA `b8a1e110d746241c7201c95dd7e519f1388efffe7eaa0badeda30c3474e2447b`. f5 reconstructed arb-volume tolerance relaxed1e-5→1e-3 after two second-chunk cases (initial max.00005886385770281777,relative2.743404451853797e-9); final max.00011569842899916694, edge reconciliation1.1368683772161603e-13. Recorded outputs were recovered without rerunning seeds. Probes7693000/7693001 are outside f5 confirmation. [b17-f2–f5:/gates,/real,/decomposition_validation,/notes]

### Direct-run numerical evidence

Blocks cite each input `/real,/code_sha256,/port_sha256`;total-edge worst5 retained,component/control tails omitted. Real components:Y-volume for u1/u3,j1–j4;reconstructed edge for u4,j5,FAf2–f5. u2/f1 measurements null. Params/weight matrices remain only in pinned inputs.

**b15-u1**

```json
{"codeSHA":"0f3d777375df5e233d79223ee31134eadf14f4b03e5ceb808f9796d2e98a2eb0","portSHA":"86f6163eeca2cfedc24e81599a305598f32e34586fadf9cea008ec142f9332b9","real":{"seeds":[7640000,7640511],"r":{"IB":-177.739053,"HA":464.245688},"a":{"IB":-144.408458,"HA":341.910091},"IB":[512,0.0595947981,0.0158105803,[[7640319,-1.36104726],[7640366,-1.34996646],[7640192,-1.32593065],[7640191,-1.03116793],[7640295,-0.904488404]]],"HA":[512,0.735608067,0.0324445852,[[7640463,-2.21661907],[7640146,-1.36519056],[7640143,-1.25118427],[7640221,-1.1110548],[7640061,-1.06070623]]]}}
```

**b15-u2:** no candidate/port or real measurements (all null), as detailed above.

**b15-u3**

```json
{"codeSHA":"f3dd2db9a288fbac864afcef995b736dcd1af1687bbfed5c4959bf32f1640b9a","portSHA":"7503595bf7a6e9e72b2f63f537e95be3fbe274726d12938bc2b7346d16f4bcfa","real":{"seeds":[7648000,7648511],"IB":[512,0.00505254811,0.0196856097,[[7648331,-1.96973178],[7648238,-1.60260144],[7648485,-1.53792732],[7648071,-1.31129835],[7648416,-1.30799108]]],"HA":[512,0.609597689,0.0327024404,[[7648407,-3.53014285],[7648431,-1.75505961],[7648331,-1.57927525],[7648485,-1.52666169],[7648268,-1.40644157]]],"r":{"IB":-170.38767,"HA":464.665989},"a":{"IB":-33.9786108,"HA":442.360411}}}
```

**b15-u4**

```json
{"codeSHA":"3fb407d87bb2dd2d4a9e5966cfd15066164dfbf2d2a18a42e02940e0a2ff5e47","portSHA":"cd19ded44b0d59976b13d35680107e7c47e77ec43e44953ed44733e18404dfed","real":{"seeds":[7652000,7652511],"IB":[512,0.0505716907,0.0173647144,[[7652169,-1.75802153],[7652449,-1.40405874],[7652421,-1.36994679],[7652019,-1.22665015],[7652091,-1.18186024]]],"HA":[512,0.716145352,0.0322149454,[[7652355,-1.85484294],[7652421,-1.63801693],[7652091,-1.36459205],[7652041,-1.27116868],[7652169,-1.24679095]]],"r":{"IB":[512,-0.0496875941,0.0153679222],"HA":[512,0.964286956,0.0315693215]},"a":{"IB":[512,0.100259285,0.010786804],"HA":[512,-0.248141604,0.0153061357]}}}
```

**b16-j1**

```json
{"codeSHA":"7f40448eb831e0d9aa5f969a83f74a4b584a82190edcf32add1ac904e3add2fb","real":{"seeds":[7656000,7656511],"base":[512,3.92903851e-05,7.9720661e-05,[[7656464,-0.0223154546],[7656489,-0.00235509677],[7656076,-3.688383e-07],[7656102,-3.33059006e-07],[7656097,-7.86981218e-08]]],"HA":[512,0.658536141,0.0310461412,[[7656161,-2.12798615],[7656328,-0.942441095],[7656489,-0.925584711],[7656385,-0.909789232],[7656033,-0.841464463]]],"r":{"base":0.000252487274,"HA":634.703493},"a":{"base":-0.025450286,"HA":481.928125}}}
```

**b16-j2**

```json
{"codeSHA":"4da9e3ae826b3b6f9170948bd2c2c0a494e78521414341656425b76e71c0f964","real":{"seeds":[7660000,7660511],"base":[512,0.578109776,0.0554340196,[[7660451,-1.39087229],[7660468,-1.00170794],[7660482,-0.902463444],[7660056,-0.821612657],[7660417,-0.784384778]]],"HA":[512,1.19919439,0.0619409801,[[7660468,-0.994395439],[7660332,-0.947963884],[7660451,-0.891842667],[7660482,-0.818812169],[7660426,-0.798084075]]],"r":{"base":-77.6136349,"HA":565.215531,"unit":"retail volume Y (engine exposes no retail edge split)"},"a":{"base":-273.4837,"HA":220.921634,"unit":"arb volume Y"},"pnl":{"base":0.536288009,"HA":1.17414875},"IB_vs_HA":[512,0.621084611,0.044559022]}}
```

**b16-j3**

```json
{"codeSHA":"f103439478d625169c4f776d0a22bd70bc591c9ddd3dd2edd6acba34635e8d2d","real":{"seeds":[7664000,7664511],"base":[512,-1.44093482e-05,2.45602316e-05,[[7664028,-0.00772552785],[7664216,-0.00706536503],[7664124,-0.00318090148],[7664327,-2.57421675e-07],[7664013,-4.50768312e-09]]],"HA":[512,0.705045413,0.0297932115,[[7664246,-1.64859877],[7664234,-1.11717565],[7664442,-0.891129717],[7664413,-0.835101501],[7664409,-0.780111772]]],"r":{"base":0.010713617,"HA":644.121599},"a":{"base":0.0161025057,"HA":494.460976}}}
```

**b16-j4**

```json
{"codeSHA":"69b5e66145102eb4b3824ee2f7d313f6ffc086158d56da50d6567cd2ab2c46f1","real":{"seeds":[7668000,7668511],"n":512,"base":{"p":[512,-4.9971173e-05,7.67301715e-05,[[7668493,-0.0378422353],[7668177,-7.59631575e-08],[7668153,-6.94933533e-09],[7668449,-3.24405391e-10],[7668362,-2.71370482e-10]]],"rv":[null,-2.94327552e-06,0.00573266924],"av":[null,-0.00250346269,0.00709223446]},"HA":{"p":[512,0.641225485,0.0305476584,[[7668245,-1.80319146],[7668282,-1.74276724],[7668488,-1.64077357],[7668373,-0.970603908],[7668290,-0.96804037]]],"rv":[null,628.336052,8.20060829],"av":[null,490.541073,5.01641638]},"base_vs_HA":{"p":[512,0.641275456,0.030552393],"rv":[null,628.336055,8.2008152],"av":[null,490.543577,5.01678561]}}}
```

**b16-j5**

```json
{"codeSHA":"dd53db1fbe42628e040c6c7b7a82689867ab601263c5e740506299a3a186b38c","real":{"seeds":[7672000,7672511],"base":{"p":[512,0.876249194,0.0352101238,[[7672313,-1.76407495],[7672269,-1.57242072],[7672339,-1.39998043],[7672167,-1.34184892],[7672022,-1.24485954]]],"r":-0.0822131836,"r_se":0.0369310709,"a":0.958462378,"a_se":0.0162707509},"HA":{"p":[512,1.21930866,0.0377013497,[[7672096,-1.62713019],[7672339,-1.25358786],[7672053,-0.995538022],[7672437,-0.866704195],[7672118,-0.793600311]]],"r":0.510125816,"r_se":0.040545244,"a":0.70918284,"a_se":0.0184987617},"IA_vs_HA":{"p":[512,0.343059462,0.0273086871],"r":0.592338999,"r_se":0.0280846287,"a":-0.249279537,"a_se":0.0111817984}}}
```

**b17-f1:** no candidate; real n=0, no seeds or comparisons, components null.

**b17-f2**

```json
{"codeSHA":"55db4beb8755274d3df64c5af34b6d782b12c0894349eceb21c85855b20e29af","real":{"seeds":[7680000,7680511],"FA":[512,0.135646557,0.02171598,[[7680357,-1.95419143],[7680119,-1.54294],[7680258,-1.26878669],[7680287,-1.25765761],[7680175,-1.1874233]]],"HA":[512,0.697258477,0.0324732292,[[7680083,-2.95467721],[7680299,-2.54357456],[7680101,-1.76203631],[7680135,-1.13443455],[7680157,-1.0807101]]],"r":{"FA":[448,-0.327525257,0.0239340814],"HA":[448,0.145339691,0.0372294279],"seeds":[7680064,7680511]},"a":{"FA":[448,0.453173404,0.0115909754],"HA":[448,0.527926235,0.0180377364],"seeds":[7680064,7680511]},"reconciliation_max_abs":1.13686838e-13}}
```

**b17-f3**

```json
{"codeSHA":"aa2c4a5637a4060088263d71ab8b109339b101bb25d64a45e9c7de83710d8728","real":{"seeds":[7684000,7684767],"FA":null,"HA":[768,0.568840644,0.0216491118,[[7684757,-1.58276508],[7684429,-1.25665709],[7684571,-1.24633258],[7684737,-0.947299854],[7684097,-0.879380818]]],"r":[704,0.502624174,0.0234967091,[[7684429,-1.80519599],[7684757,-1.58172109],[7684712,-1.34652125],[7684685,-1.22448651],[7684571,-1.09757505]]],"a":[704,0.0622475249,0.0106518811,[[7684343,-1.45724828],[7684344,-1.27556162],[7684094,-1.04164898],[7684165,-0.746352943],[7684236,-0.697657985]]],"component_seeds":[7684064,7684767],"reconciliation_max_abs":1.70530257e-13}}
```

**b17-f4**

```json
{"codeSHA":"f06b3cb7c141ce8239519c9c177695cc3de9c849be5428e32d2edc3d089ae78b","real":{"seeds":[7688000,7688511],"FA":[512,0.0278306756,0.012681869,[[7688358,-1.33733362],[7688032,-0.822131145],[7688349,-0.765388942],[7688202,-0.703397803],[7688375,-0.697371737]]],"HA":[512,0.594833509,0.0285511173,[[7688126,-2.22495829],[7688033,-1.95780253],[7688058,-1.21649543],[7688344,-0.910947912],[7688350,-0.868833484]]],"FA_vs_HA":[512,0.567002833,0.0275863411],"r":{"FA":[448,-0.0327830561,0.0139538485],"HA":[512,0.467272143,0.0302118448]},"rv":{"FA":[512,-1.58350413,2.28404339],"HA":[512,377.479946,11.2302878]},"a":{"FA":[448,0.0631356709,0.00571185499],"HA":[512,0.127561366,0.0146806846]},"av":{"FA":[512,-40.580959,2.25198015],"HA":[512,334.622905,6.98948715]},"reconciliation_max_abs":{"FA":1.13686838e-13,"HA":1.0658141e-13}}}
```

**b17-f5**

```json
{"codeSHA":"08107bf7a186703848e07e0c1420f040f0e9aa9233672db051c04012d3516875","real":{"seeds":[7692000,7692511],"r":{"FA":-0.123426198,"HA":0.398682883},"a":{"FA":0.183256922,"HA":0.252575228},"r_se":{"FA":0.014977198,"HA":0.0284497712},"a_se":{"FA":0.00755006367,"HA":0.0155042181},"FA":[512,0.0598307243,0.0139761314,[[7692081,-1.07628986],[7692416,-1.05104768],[7692288,-0.878508938],[7692222,-0.876539701],[7692303,-0.825769852]]],"HA":[512,0.651258111,0.0288812975,[[7692458,-2.87979047],[7692031,-1.70536912],[7692261,-1.13277713],[7692408,-1.10372393],[7692496,-0.918548336]]],"decomposition_max_abs_arb_volume_error":0.000115698429,"edge_reconciliation_max_abs":1.13686838e-13}}
```

Additional holdout/init figures cite each input `/holdout,/gas`. Replica vsIB components:u1 retail−.24277408239529646/arb.2983336576388449;u3−.12223741142114566/.146200299680232;u4split absent. j1/j4 freeze-only4.383973362444976e-05/−1.2258577599642828e-05;j3identity0/0;j2overHA .46496409999268434/.692706398782057. Other holdout component detail is explicitly omitted,not zero.

| input | holdout over_T (SE) | init peak |
|---|---|---|
| b15-u1 | 1.82454368 (0.0112252777) | 243260 |
| b15-u2 | null | null |
| b15-u3 | 1.792947 (0.0118163847) | 243264 |
| b15-u4 | 1.82041171 (0.011745136) | 243260 |
| b16-j1 | 1.76901569 (0.01135015) | 243227 |
| b16-j2 | 2.27450256 (0.0127160832) | 243227 |
| b16-j3 | 1.76898411 (0.0113499705) | 243227 |
| b16-j4 | 1.76901569 (0.01135015) | 243227 |
| b16-j5 | 2.30754407 (0.0135286906) | 243234 |
| b17-f1 | null | null |
| b17-f2 | 1.85091028 (0.0117094266) | 243227 |
| b17-f3 | null | 243227 |
| b17-f4 | 1.72942115 (0.0115630814) | 243227 |
| b17-f5 | 1.74570893 (0.0117332555) | 243227 |

## CG independent real confirmation: cf-a, cf-b, cf-c

cf-a/b/c confirm unchangedCG/IB/HA,no training/replica gates:each alone vsVanilla,8workers,64-seed chunks,reset gas perarm,1024each on7400000–7401023/7410000–7411023/7420000–7421023. Nine kit tests and seven kit-file/strategy/artifact hashes reportedly pass. [b19-cf-a/b/c:/seeds,/notes]

All3 show positive meanCG−IB,arb improvement offset by retail loss. cf-c tail7420139=−9.972649529074829 vsIB,−10.050803134776288 vsHA;second-quartile SE elevated. Wins are positive-seed counts,not rank. [b19-cf-a/b/c:/cg_vs_ib,/cg_vs_ha,/quartiles]

**b19-cf-a** [/seeds,/cg_vs_ib,/cg_vs_ha,/ib_vs_ha,/mean_edge,/gas]

```json
{"seeds":{"start":7400000,"end":7401023,"n":1024},"CG-IB":{"p":[1024,0.178680629,0.0208853585,[[7400633,-2.8761],[7400197,-2.737],[7400686,-2.477],[7400569,-2.4228],[7400680,-2.3553]]],"sum":182.968964,"sumsq":489.633304,"wins":641,"r":-0.264309883,"a":0.442990513},"CG-HA":{"p":[1024,0.837319793,0.0216113069,[[7400208,-1.6746],[7400077,-1.4526],[7400464,-1.1793],[7400795,-1.1409],[7400478,-1.0351]]],"sum":857.415468,"sumsq":1207.18862,"wins":928,"r":0.756670226,"a":0.0806495665},"IB-HA":{"p":[1024,0.658639163,0.0225810675,[[7400015,-3.0389],[7400311,-1.6898],[7400543,-1.5937],[7400577,-1.4706],[7400478,-1.3249]]],"sum":674.446503,"sumsq":978.368475,"wins":867,"r":1.02098011,"a":-0.362340946},"gas":{"HA":{"swap":233761,"init":243147,"calls":10857903,"OOG":0,"errors":0,"reverts":0},"IB":{"swap":237142,"init":243227,"calls":11011205,"OOG":0,"errors":0,"reverts":0},"CG":{"swap":237860,"init":243237,"calls":11006378,"OOG":0,"errors":0,"reverts":0}},"valid":true}
```

Quartiles [same input:/quartiles]: IB retail Y-volume order,ties by seed,n256each,edge mean(SE).

| quartile | IB volume range | CG−IB | CG−HA |
|---|---|---|---|
| 1 | [57401.4284,69027.8224] | 0.22693926 (0.0367500461) | 0.674030998 (0.032961718) |
| 2 | [69062.0912,78084.1592] | 0.19884924 (0.0385276944) | 0.765277211 (0.0390694182) |
| 3 | [78127.7259,86812.6187] | 0.181421282 (0.0423915986) | 0.940630316 (0.0472398417) |
| 4 | [86856.6136,100285.184] | 0.107512735 (0.0483433717) | 0.969340646 (0.0491440521) |

**b19-cf-b** [/seeds,/cg_vs_ib,/cg_vs_ha,/ib_vs_ha,/mean_edge,/gas]

```json
{"seeds":{"start":7410000,"end":7411023,"n":1024},"CG-IB":{"p":[1024,0.194356569,0.0197023009,[[7410950,-2.79095717],[7410996,-2.04603595],[7410746,-1.76403766],[7410350,-1.74226481],[7410431,-1.62046134]]],"sum":199.021126,"sumsq":445.320491,"wins":659,"r":-0.246243328,"a":0.440599897},"CG-HA":{"p":[1024,0.835139414,0.0209487309,[[7410950,-1.4804735],[7410431,-1.37259225],[7410864,-1.068581],[7410363,-1.00794192],[7410808,-0.979266079]]],"sum":855.18276,"sumsq":1173.91432,"wins":939,"r":0.733633078,"a":0.101506336},"IB-HA":{"p":[1024,0.640782846,0.0212470626,[[7410712,-2.24194861],[7410053,-1.55572564],[7410396,-1.48251343],[7410595,-1.42020507],[7410199,-1.34061933]]],"sum":656.161634,"sumsq":893.361551,"wins":865,"r":0.979876406,"a":-0.339093561},"gas":{"CG":{"swap":238374,"init":243237,"calls":11009841,"OOG":0,"errors":0},"IB":{"swap":237663,"init":243227,"calls":11016515,"OOG":0,"errors":0},"HA":{"swap":234255,"init":243147,"calls":10863346,"OOG":0,"errors":0}},"valid":true}
```

Quartiles [same input:/quartiles]: IB retail Y-volume order,ties by seed,n256each,edge mean(SE).

| quartile | IB volume range | CG−IB | CG−HA |
|---|---|---|---|
| 1 | [58318.0795,69412.778] | 0.235670205 (0.0334524856) | 0.642161195 (0.0356829384) |
| 2 | [69436.9703,77735.6264] | 0.196117325 (0.0369210427) | 0.809541322 (0.038951284) |
| 3 | [77738.2128,86816.3594] | 0.138791986 (0.0402024148) | 0.890091245 (0.038458389) |
| 4 | [86868.3822,100433.44] | 0.206846759 (0.045948522) | 0.998763895 (0.0500726816) |

**b19-cf-c** [/seeds,/cg_vs_ib,/cg_vs_ha,/ib_vs_ha,/mean_edge,/gas]

```json
{"seeds":{"start":7420000,"end":7421023,"n":1024},"CG-IB":{"p":[1024,0.166172876,0.0225939786,[[7420139,-9.97264953],[7420002,-4.12815727],[7420467,-1.85510426],[7420454,-1.84767767],[7420606,-1.73957419]]],"sum":170.161025,"sumsq":563.038735,"wins":626,"r":-0.272433127,"a":0.438606003},"CG-HA":{"p":[1024,0.807819617,0.0232119281,[[7420139,-10.0508031],[7420127,-1.37172696],[7420045,-1.32958877],[7420117,-1.27743871],[7420444,-1.2282002]]],"sum":827.207287,"sumsq":1232.64859,"wins":920,"r":0.711100268,"a":0.0967193482},"IB-HA":{"p":[1024,0.641646741,0.0217685578,[[7420856,-3.65222511],[7420436,-2.47722225],[7420325,-2.02940402],[7420749,-1.42017508],[7420609,-1.26632501]]],"sum":657.046262,"sumsq":917.995172,"wins":869,"r":0.983533396,"a":-0.341886655},"gas":{"CG":{"swap":237739,"init":243237,"calls":11012570,"OOG":0,"errors":0},"IB":{"swap":237021,"init":243227,"calls":11018253,"OOG":0,"errors":0},"HA":{"swap":233620,"init":243147,"calls":10864229,"OOG":0,"errors":0}},"valid":true}
```

Quartiles [same input:/quartiles]: IB retail Y-volume order,ties by seed,n256each,edge mean(SE).

| quartile | IB volume range | CG−IB | CG−HA |
|---|---|---|---|
| 1 | [56969.8126,69216.8207] | 0.203386188 (0.0323929041) | 0.642504283 (0.0321921021) |
| 2 | [69225.1833,77828.7483] | 0.150441334 (0.0568592299) | 0.71048307 (0.0581127274) |
| 3 | [77886.0979,86484.5941] | 0.17357419 (0.0449697253) | 0.888563354 (0.0428999633) |
| 4 | [86510.886,100103.182] | 0.137289792 (0.0433288916) | 0.98972776 (0.0457445491) |

CG−HA rises across volume quartiles in each block. CG−IB stays positive in all quartiles but cf-b Q4 rebounds:monotonic decline is not universal. Arb-volume max errors:cf-a .0000869079158292152,cf-b .00010741837468231097,cf-c .00011314953735563904. cf-a worst5 supplied4decimals;per-seed arrays in all three rounded4decimals,not reproduced;sum/sumsq fullprecision upstream. cf-b chunkwalls57.048–135.323s;cf-c first-arm secondsCG/IB/HA37.789293305017054/39.125698489136994/40.22464775014669. No pooling. [b19-cf-a/b/c:/quartiles,/notes;b19-cf-a:/max_abs_arb_volume_reconstruction_error;b19-cf-b:/chunk_wall_seconds]

## Kit-c: general IB port and diet emitter

kit-c extends KB https://api.imd.fun/artifacts/48cd6a10f39fc22a79c5f12398474bd65282803d59314819acd586f187015808 . Original sim.py,diff_test.py,IAD.sol,weights.json,IA.patch.json are referenced,not embedded:kit-c alone is not standalone. Hashes below. [b18-kit-c:/unchanged_from_KB,/api,/decode_instructions]

WeightsJSONv1 adds optional inputs=IA(default,omitted on output)|IB;W[0]=1/2,0=head-off;sharedfeatures0–12,15,16 unchanged. IB13=amountY(.05,−1);14=post-print count cap64(.5,−.5);17=min(time−lastArb,16)(.125,−1),lastArb updated before read on firstInStep+arbProbability≥.5;18=router deviation×500,held from later-in-step gamma>0 print,initial0. API policy/load→W,strat_init/after unchanged;S[18/19]=lastArb/routerSpot. [b18-kit-c:/api]

Emitter supports IA/IB,arbitraryn/h/IDs/scaling,1or2nets. Eight live words:lastArb slot15bits240–255,router slot13above145. IA emission unchanged(hash c20b170819e86adc7f59fd0d2a56b1a3463883ad29e471957538037d62da99a4,distinct from IAD.sol). Diet lastArb16bits(vsIB64)equivalent only time<65536;router110bits(vs112)only spot<2^110wei-scaled.10000-step tests fit time bound;universal identity unproven. [b18-kit-c:/api,/gates.IA_emit_unchanged,/notes]

A_IB/C_IBD port/EVM max6.638269694353438e-05bps,50×2000swaps,C exactIBD-vsIB mismatches0;IA A6.501554232485329e-05passes. B unrun(no search/gain claim). Regression100seeds0–99,fresh256seeds7696000–7696255:max edge difference0,pairedmean/SE0,retail/arb delta0,worst5 firstfive seeds all0. [b18-kit-c:/gates,/regression,/real]

IB/IBD regression swap237009/131206,init243227/176930;fresh swap237033/131465,init243227/176930;OOG/other failures0. Equal swapcalls regression1089194/fresh2765933. Fresh meanedge/retail/arb532.0336250161047/563.1074558736788/−31.073830857573988;regression542.8528476636853/573.0727500210379/−30.219902357352584. Regression is not fresh confirmation. [b18-kit-c:/gas,/real,/regression]

IB benchmark512seeds6010000–6010511,10000steps,8processes,warmJITexcluded,poolstartincluded:general345.2115473506388/449.52503325501044/422.7628582589294 markets/s vs own373.01721512610106/328.6838446362/264.8665316421159. No slowdown reproduced;hotpath unchanged. Historical KB original/IA/general765.4776653194682/635.4346768129121/327.9006234923561 is separate. General-vs-own max edge difference.061228731764003896,means527.5387796104765/527.5386248245479:tiny fee errors do not imply bit-identical markets;item attributes float-order/1e-6 multiplier rounding. Inspection overhead:per-callback np.empty(19),np.empty(n),dynamic headloops. [b18-kit-c:/speed]

Tests19/19(KB15/15):IB roundtrip/flag,diet emission+IAunchanged,IA/IB feature difference,randomweight EVM differential. No training/replicagain/rank claimed. [b18-kit-c:/tests,/notes]

Kit dependency identities [b18-kit-c:/unchanged_from_KB.files and sibling payload hashes]:

```json
{"unchanged_KB":{"sim.py":"aa97e7eba379b1d2b77cdbd2778dee053c603ba1c59852f680026c23301600d3","diff_test.py":"b2ec75b04999f0a8431a049f669f4cd5b2005efda4a6d1b265025e36e3749279","IAD.sol":"7f6998c8df1b738dbd9715053edae7c9c8fdad71bf7228d67b6f1b7573c7126b","weights.json":"058c2af38034b617b3d4e5c4a9f7869a739368c1b42ba2182db7445771584ed0","IA.patch.json":"2b1cd0b4278962d3e68fab51fee530b3d379adfbf3efaffd9b332ab5d6cb8d82"},"new_payload_sha256":{"portSHA":"d5329ec002ba93b70ca0caf991f1bd3ed13d37b4cf60d990388084737693d7bc","harness_sha256":"b3897a3c16287fb129077aac4136d4d36680f759ad334b37e19bef3468513ba8","tests_sha256":"7c38e27d9bb49bfe8921402bd32a4146e8c950477a892047f6869659f4e9ec0f","IBD.sol_sha256":"dd801398ab0210c978bb3431b714fa005da4dd169adb4e9a88175f60d4578f3d","weights_IB_sha256":"8110b95ebe83aa02b77528b6dd2b610781a6e7ce14c2885f165212e803085dc9","speed_bench_sha256":"521a285d7278619735d957845953d764617027b31a8c74c41762f8065a675f09"}}
```

## CG continuation: k01–k10

CG source:b20-merge-01 k01–05,b20-merge-02 k06–09 plus missingk10.178weights,19-8-2,fixedscaling,startfrozenCG(not restart). Group scalesW1 .16007274507182634,b1 .18564858000307327,W2 .12726224434596164,b2 .1185932470281719(reported.1×HA RMS);40min half-life. Monitor every10gens scores ESmean+generation-best,retainsCG absent gain;shared generation markets,weights1e-6. [b20-merge-01/02:rows.settings]

H_CG=holdout mean(SE),reported8192seeds40000000–40008191. H_IB is H_T−1.7690,never a fresh paired IB holdout;its SE excludes reference uncertainty. [b20-merge-01:notes;b20-merge-02:opening]

| item/source | valid | speed | minutes/gens | H_T | H_CG (SE) | swap/OOG |
|---|---|---|---|---|---|---|
| k01/b20-merge-01 | true | 2246 | 42.22/1276 | 2.06197897 | 0.11006168 (0.00455495273) | 237626/0 |
| k02/b20-merge-01 | true | 239.633277 | 72.6625479/70 | 1.95959549 | 0.00767819405 (0.00198655844) | 235478/0 |
| k03/b20-merge-01 | true | 412.293304 | 49.485/150 | 1.95191575 | 0 (0) | 237819/0 |
| k04/b20-merge-01 | true | 313.44137 | 76.2053345/97 | 1.95191729 | 0 (0) | 237848/0 |
| k05/b20-merge-01 | true | 86.657451 | 0/0 | 1.95191729 | 0 (0) | 235599/0 |
| k06/b20-merge-02 | null | [77.3,53.9,52.7] | 0/0 | null | null | 235599/0 |
| k07/b20-merge-02 | true | 415.270101 | 75.3352581/143 | 1.95759407 | 0.00567677512 (0.00251376129) | 238448/0 |
| k08/b20-merge-02 | true | 304.896923 | 40.7407744/40 | 1.95401839 | 0.00210109435 (0.00163402992) | 237632/0 |
| k09/b20-merge-02 | true | 338.618078 | 79.5433077/336 | 1.95191729 | 0 (0) | 237600/0 |
| k10/b20-merge-02 | null | null | null/null | null | null | null/null |

k01 bestCG continuation:42.22min/1276gens,pop32,M256,sigma.13,ES601;speed2246/8processes,4909.7/30. Freeze best@1140 monitorCG .10515063363556332;holdout .11006167954441087(SE .004554952728974956)and real .11469920783846532(SE .017890078101032988,n512)agree. Only rounded HA result retained:+.960 SE .032,n512. Candidate mean originally includes1088seeds(extra8001024–8001087 excluded from all paired stats);do not substitute arm mean for1024/512paired results. [b20-merge-01:summary,notes,rows[k01]]

k02 pop32/M512/sigma.06/ES602 finds small gain;k07 pop32/M512/.06/ES607 realCG .012292115408336524(SE .009340649957697306)is inconclusive;k08 pop16/M1024/.03/ES608 positive tiny holdout but realCG−.04220597965497963(SE .048652853788024586). Replica point estimates do not guarantee real gain. [b20-merge-01:rows[k02];b20-merge-02:rows[k07/k08]]

k03(pop16/M512/.06/ES603),k04(pop32/M256/.10/ES604),k09(pop32/M128/.13/ES609)search but retain CG:zeroH_CG/R_CG identity. k05 speed86.65745095289543 triggers no-search(plannedpop32/M128/.13/ES605),realIB976/HA512pairs. k06 speed77.3/53.9/52.7,no search,holdout/realnull,validnull,gasCG. k10 absent,allnull. [b20-merge-01:rows[k03–k05];b20-merge-02:rows[k06/k09/k10]]

Training:k01 begins1500000000/M256;k02 1510000000/M512;k03 1520000000–1520076799;k04 1530000000–1530024831;k05/06 planned1540000000/1550000000,unexecuted;k07 1560000000–1560073215;k08 1570000000–1570040959;k09 1580000000–1580043007. Real starts k01–05:8000000/8004000/8008000/8012000/8016000. k07–09 exact confirmation blocks absent from submerge. Both submerges omit real worst5/components and most HA results;unrecoverable here. [b20-merge-01/02:rows.settings,real,notes]

A8.814065310946972e-05bps for CG; B1.920372244773669(SE .04445924198500627,n512,seeds7232768–7233279) vs2.0, absolute gap .07962775522633092. b20-merge-01 reports all A/B/C pass; C k01~7.78e-05,k02~8.15e-05bps. b20-merge-02 C k07 .00012367863025386505,k08 4.658967827037095e-05,k09=A (unchanged). k06 C is null/not applicable; A segment17–49 is from printed progress after optional replay error, not a saved segment JSON, limiting auditability. k06's T head-zero port matches kit TH_T on64 replica markets (max edge diff0). k07 reports9 kit tests and exact baseline emission. [b20-merge-01:notes; b20-merge-02:rows.gates]

CG baseline differs:k03H_T1.9519157504253348(SE .011938024070038173)vs others1.9519172921936434(SE .011938004553452971),~1.5e-6,retained. k09SE0 means H_CG identity only,not H_T/H_IB. Small-step explanation for k03/k04 no-gain is untested:hardware,processes,budgets and step sizes confounded. [b20-merge-01:notes,rows.holdout;b20-merge-02:rows[k09].holdout]

Real paired moments/identities below cite `source:rows[key]`;sum/sumsq are uncentered moments,not extra observations. Unavailable tails/components not inferred.

**k01** [b20-merge-01:rows[k01]]

```json
{"codeSHA":"0ab25ecd54f9a816119546408b08906ede14f74faa99d9c10cdfec7d6d6c1be6","real":{"IB":{"p":[1024,0.315953012,0.0203650421],"sum":323.535885,"sumsq":536.678553},"CG":{"p":[512,0.114699208,0.0178900781],"sum":58.7259944,"sumsq":90.4724272}}}
```

**k02** [b20-merge-01:rows[k02]]

```json
{"codeSHA":"02b4f502af8e57de7deb7920e3a73267844d8d3a83797abd257b15336a239e96","real":{"IB":{"p":[1024,0.179854691,0.0214543303],"sum":184.171204,"sumsq":515.299974},"CG":{"p":[512,0.0176295212,0.00828185596],"sum":9.02631483,"sumsq":18.104243}}}
```

**k03** [b20-merge-01:rows[k03]]

```json
{"codeSHA":"e1ea8fcb105e00754d17061efbe4b299a634b0fcc1b5c293d65b631654b7204b","real":{"IB":{"p":[1024,0.168283484,0.0199040692],"sum":172.322288,"sumsq":444.009737},"CG":{"p":[512,0.0,0.0],"sum":0.0,"sumsq":0.0}}}
```

**k04** [b20-merge-01:rows[k04]]

```json
{"codeSHA":"e1ea8fcb105e00754d17061efbe4b299a634b0fcc1b5c293d65b631654b7204b","real":{"IB":{"p":[1024,0.170843936,0.0201101418],"sum":174.944191,"sumsq":453.536831},"CG":{"p":[512,0.0,0.0],"sum":0.0,"sumsq":0.0}}}
```

**k05** [b20-merge-01:rows[k05]]

```json
{"codeSHA":"e1ea8fcb105e00754d17061efbe4b299a634b0fcc1b5c293d65b631654b7204b","real":{"IB":{"p":[976,0.181945137,0.0224437225],"sum":177.578454,"sumsq":511.650133},"CG":{"p":[976,0.0,0.0],"sum":0.0,"sumsq":0.0}}}
```

**k06** [b20-merge-02:rows[k06]]

```json
{"codeSHA":"e1ea8fcb105e00754d17061efbe4b299a634b0fcc1b5c293d65b631654b7204b","real":{"IB":null,"CG":null}}
```

**k07** [b20-merge-02:rows[k07]]

```json
{"codeSHA":"1a08a19cb6038bac950a3a00af1c2239d96f39ac72d76a543acaa5c6c974ae61","real":{"IB":{"p":[1024,0.175353565,0.01970072],"sum":179.562051,"sumsq":438.06102},"CG":{"p":[512,0.0122921154,0.00934064996],"sum":6.29356309,"sumsq":22.9041623}}}
```

**k08** [b20-merge-02:rows[k08]]

```json
{"codeSHA":"0fd2ca4331c5aded340c06ba3a499e8c29fa477cbb24f102320b98cf45152677","real":{"IB":{"p":[512,0.114918383,0.0578461888],"sum":58.8382123,"sumsq":882.229765},"CG":{"p":[512,-0.0422059797,0.0486528538],"sum":-21.6094616,"sumsq":620.221203}}}
```

**k09** [b20-merge-02:rows[k09]]

```json
{"codeSHA":"e1ea8fcb105e00754d17061efbe4b299a634b0fcc1b5c293d65b631654b7204b","real":{"IB":{"p":[1024,0.177676431,0.023211453],"sum":181.940665,"sumsq":596.717784},"CG":{"p":[512,0.0,0.0],"sum":0.0,"sumsq":0.0}}}
```

**k10** [b20-merge-02:rows[k10]]

```json
{"codeSHA":null,"real":{"IB":null,"CG":null}}
```

## What these roots establish and leave open

Evidence supports modest transferable head-only gains(w17,u1/u4,f2,k01),larger reported narrow-constant joint gains(j2/j5),and diet gas savings without observed score change. Broadjointj1/j3/j4 fail;more time alone is not improvement. Pilot rising-curve premise needs qualification:c5 recovers without beatingIB;u3/u4/f4 freeze earlier than final;j1/j3/j4 deteriorate;k03/k04/k09retainCG. [b13-c5;b14-merge-04;b15-u1/u3/u4;b16-j1–j5;b17-f2/f4;b18-kit-c;b20-merge-01/02]

Unmeasured:no new leaderboard rank/submission,+3.1-over-HA demonstration,j2-vs-j5 real comparison,or trained-policy transplant into diet emitter. Gas headroom enabling higher score is inference,not observation. Failed citations,missing logs/checkpoints,identity fallbacks,rounding,subset decomposition and w20seed repair remain qualifications. [supplied results;requester target]