# Root merge 3: fee-policy training and confirmation

This merge covers all 25 supplied JSON results. It merges existing experiments; it does not run a new training or simulator campaign. Source labels below identify both the supplied snapshot and its original result path. Unless explicitly called a task reference, numbers are worker-reported local measurements, not certified leaderboard scores.

## Findings and limits

The strongest reported incremental real-engine J5 gain among complete confirmations is **j11: frozen minus START +0.09149265413632557, SE 0.017772594845482537, n=512**. Its independent replica holdout is +0.06539728179912, SE 0.004994016923914225. It is a reasonable continuation candidate, subject to its three flagged external sources; this is a merge assessment, not a new validation. [S22]

**j07 has the largest reported complete frozen-minus-IB mean in this group: +0.6686299709565341, SE 0.03788344799968138, n=512.** Its own starting J5 already scores +0.633416005462111 on that seed block, so its incremental real gain is only +0.03521396549442313, SE 0.020458822991768485. Its replica holdout increment is +0.07686137824073459, SE 0.0056237057367839365. One external source is flagged. j07 and j11 use different confirmation seeds; their absolute gains cannot establish a paired winner. Exact returned parameters for both are preserved below. [S18, S22]

Longer training did produce additional gains from IB itself: ib-s30 ran 314 generations and reports +0.1152033416652738 (SE 0.016026430471105104) versus IB; ib-s20 ran 1568 and reports +0.13356288225988888 (SE 0.0220021711355587). Both have n=512 real confirmations. These are separate runs from the same starting model, not successive legs whose generation counts or gains can be added. Each has two flagged sources. [S06, S07]

The slope and wide experiments provide small or absent improvements over their shared IG starting policy. FA retraining improves FA but does not establish superiority to IB. Several J5 legs improve the replica while their real increment is uncertain or absent; j13 is negative versus START on its real block. Some items return unchanged START or have incomplete confirmations. Their statuses remain explicit below. [S01–S25]

The goal brief supplies the historical context **IB = 522.00, 20th; T plus a learned 19-8-2 correction, +0.59 over HA**; sister goal cc45e1a7 reportedly gained **+0.33 in 210 CMA generations and +0.59 in 910**, still rising. These are requester-provided context, not results measured by this merge. The original IB reference is [IB artifact](https://api.imd.fun/artifacts/afd5183e9ff3235334e041c00b0cd19079599a2c6464db95f8c1586712105557); required decoded code SHA-256 is `301941829a44bb60977bb28181a5e4bfa6544f3696763f987054789b2611d686`. The requested [KIT-B artifact](https://api.imd.fun/artifacts/48cd6a10f39fc22a79c5f12398474bd65282803d59314819acd586f187015808) supports IA inputs and 1–64 hidden units; the brief describes IAD afterSwap peak as 132k versus 238k before dieting.

## Evidence policy and disagreements

The requester reports failed source rereads for the counts in the registry. The failed URL/passage identities and separate sources.json files were not supplied here for most items, so a specific passage cannot be rehabilitated or identified from the warning count alone. A flagged external quotation or inherited score is therefore provisional. Local gate/search/gas measurements remain attributed claims from the supplied result, not independently reproduced facts. No warning is silently converted into either proof or a blanket rejection of every measurement. Zero in the registry means no warning was listed, not independently verified. This merge reads the named local inputs and checks embedded payload integrity; it does not fetch the external pages again.

The goal requires search only at **>=300 markets/s**. Several workers instead say their own requester override allowed **>=100**, using M=256 below 300. That override is reported by the items but is not present in this root brief. Runs below 300 therefore do **not** satisfy the root speed rule even where their own `valid=true`; their empirical measurements are retained with that qualification. No-search branches below 100 do not demonstrate training progress. Hardware, pool startup, concurrent gate jobs and timing scopes differ, so throughput comparisons are not controlled algorithm comparisons. [S01, S05, S09, S12, S14–S16, S20, S24, S25]

Gates are not interchangeable: slope/wide Gate B targets START/IG gain over IA (historical real reference about +0.507); IB retraining targets IB over HA (real +0.5915730521292681); FA uses a +0.569 reference; J5 legs compare J5 minus IA with historical real +0.876. Thus passing Gate B does not establish the final frozen model beats IB. In sg-c10, B is copied from the supplied START real result (+0.5074247543096038), not newly reproduced replica gain. [S01–S09, S12–S25]

Real total-edge comparisons use separate markets for each strategy, each alone against the 30 bps VanillaStrategy, with `amm_sim_rs.run_batch(X, VanillaStrategy, configs, 8)` and MatchRunner configuration draws, as reported by completed items. Replica paths differ from Rust paths; a gate checks aggregate gain transfer, not path identity. Paired SE is sample SD of differences divided by sqrt(n). Shared seeds across arms are intentional; generation blocks are fresh across generations. Reused fixed monitor/holdout blocks across items are not new independent holdouts, and no pooled gain or combined certainty is claimed. [S02–S25, where measurements exist]

Retail/arb deltas require units: some workers reconstruct separate **edge** contributions from constant-product reserves, prior fees and fair prices, checking recovered arb volume against the engine; others only report **Y volumes**, because the engine does not directly expose split edge. These findings do not conflict and must not be added across units. IG block A explicitly has only volume deltas; block B reconstructs edge deltas. The numerical tables preserve original field names. [S02–S25]

`valid` is the worker’s protocol/gas decision, not statistical significance. Incomplete missing calls have unknown outcomes, even when every saved call has zero observed OOG. IG block A, j04, j05 and j14 remain incomplete; j03 has no real/gas result; FA-s20 and j12 have no measured campaign. j04 separately marks its fully evaluated candidate valid, while top-level validity is false because START confirmation is incomplete. [S08, S10, S14–S16, S23, S25]

## Source registry

Starting-artifact locators reported by the workers (provisional external provenance): [IG/START](https://api.imd.fun/artifacts/ebbd90c663c396598965b23341aea97a4501b4c971db183a8fe94a880f399979) [S02, S11]; [IA](https://api.imd.fun/artifacts/df0ddfdf122c88641513dd179980fce475f42a912a184f812d213f27db7c14df) [S02]; [J5](https://api.imd.fun/artifacts/fe2320b5c091489337b3eb06bcbaea19bf4a8af09d959d67504ba63184f1dab6) [S23].

Every data row and per-item section is sourced to its label. Each filename is under `.imd/reads/artifacts/`. Its original path is `artifacts/goal/batches/B/items/I/result.json`, where I is the filename after its first hyphen, and B is b13→slope-ig, b14→wide-ig, b15→ib-re, b16→fa-re, b17→confirm-ig, b18→j5-a, b19→j5-b, b20→j5-c. These original paths are provenance references, not files assumed present in Git. All 25 named local inputs were present and parsed. The SHA-256 column records the exact supplied snapshot bytes read here.

| Source | Supplied input / original result | SHA-256 | Flagged sources |
|---|---|---|---|
| S01 | `b13-sg-c10-p32` | `fd7e16c83aa6de3e7658a864a281a796ff929d0e5a7baf823f37e93cf45e3f5c` | 1 |
| S02 | `b13-sg-c15-p16` | `5191d37078a0fa561203d6b1065d3aca84b6f991bc72f2b73c4a5f040ed1f721` | 2 |
| S03 | `b14-w12-s30-p16` | `428ca853a6deaff555512189d6f2ce4f095c865428a69333cee6602bb546e9b9` | 1 |
| S04 | `b14-w16-s20-p32` | `6f3b876ffa0640c279d7ecd0abe2e099d18d45fce36ac22e8fb03695ca760a47` | 1 |
| S05 | `b14-w16-s30-p16` | `e58badcd15820886fcf199a39ff41377d95b0451bca81916b1d2b35914ca241c` | 1 |
| S06 | `b15-ib-s20-h60-p16` | `5f2b3956647c563440866e377258a049e495da3e45933de87fd4e201e9f34ce0` | 2 |
| S07 | `b15-ib-s30-h40-p16` | `4ec4a3eb93d3bdb3c4a299df11fed6d5bd16f7a83f5d64fda5406831260913cd` | 2 |
| S08 | `b16-fa-s20-h60-p16` | `0bd0c81b61b4efbdba3955469f93eb2cefe6de1b1e42f452fdc65939b53bbf5f` | 0 |
| S09 | `b16-fa-s30-h40-p16` | `5b07b464065e372377a612ecba7c9504793e31a5bdda28853893577816831be3` | 1 |
| S10 | `b17-ig-block-a` | `c3744ee2374260593a93a25c1433833878f1eba61ead5ca6787bc3e47f38478c` | 1 |
| S11 | `b17-ig-block-b` | `8434630e62bafc71e6090199b3900defcdea3b616c62b0cc5354dea678acedbc` | 1 |
| S12 | `b18-j01-s03-h45-p16` | `a098eb2f29c48538cdd525599333c2c9cb651475d34fbfd3df1f672bf4d4ffbe` | 1 |
| S13 | `b18-j02-s05-h30-p16` | `c7746d87085ee24fff012982bb0a162145fa5e2fcd62d9838e91268f6d0bd228` | 1 |
| S14 | `b18-j03-s05-h45-p16` | `d71de94797acf467a90894818f3f65c4467b1a4b23610eb3a70bf8a8996a5b73` | 0 |
| S15 | `b18-j04-s07-h30-p16` | `5724d30d7757ea6c2ac159061a598d3ca3b605e865418c3444418353f6ef386a` | 2 |
| S16 | `b18-j05-s03-h45-p24` | `1fdd7ddc99a379d92dd252f38d92bd084166b9cda2b6fadba177854156f8ea2e` | 2 |
| S17 | `b19-j06-s05-h30-p24` | `d05709c4420c25c02703b81ede62dcc191183d981546e37343dd84d077960788` | 2 |
| S18 | `b19-j07-s07-h45-p16` | `b6f8959118eb10e10567a727753126b3e46da51e537df0167ebc288e604daa3c` | 1 |
| S19 | `b19-j08-s03-h30-p16` | `41a13f71c696cd81bc8857f8bb9c8668df4b2789faf38be8405e0f54e29f52f8` | 1 |
| S20 | `b19-j09-s05-h45-p24` | `b35381e250d7e9663c8018035f4e88512ec03028bb6c918a606d9c86d8499b3b` | 2 |
| S21 | `b19-j10-s07-h30-p24` | `0e6c2058ee3338a7915ec537e127ef5112ddc5119a35ddcd7a30e1543198c9f0` | 2 |
| S22 | `b20-j11-s05-h30-p16` | `cd8559ed66c20cec062b6edf56ccd990c7c8638a70c457a8315bda6b61fdebce` | 3 |
| S23 | `b20-j12-s03-h45-p16` | `361be18400e7d426a0b097fd395de6de583c8bac6ed0db7130453fd7ccb148dd` | 0 |
| S24 | `b20-j13-s05-h45-p16` | `d668d825945e543c12ffdeb926a3e3077d9a0a721ed7b5787be005aad8cfd56d` | 3 |
| S25 | `b20-j14-s07-h30-p16` | `13a9ddb0a56fce1c6428a5a331695594ae71707dc59245608da0bd8e16bbc6a2` | 0 |

## Training, throughput and status

Values are from each source’s `speed`, `search`, `valid` and `blocked` fields. “Below root threshold” qualifies any positive generation count at speed <300. Confirmation-only items need no search benchmark. Minutes are reported search time, not normalized wall time.

| Source | Markets/s | Processes/CPUs as reported | Generations | Search minutes | M | Worker valid | Root speed assessment |
|---|---:|---|---:|---:|---:|---|---|
| [S01] | 37.44444384033162 | 8 | 0 | 0.0 | — | true | no search |
| [S02] | 806.9348092338565 | 6 | 446 | 71.97192095214656 | 512 | true | meets root threshold |
| [S03] | 393.15235185098953 | 8 | 194 | 77.02760719450889 | 512 | true | meets root threshold |
| [S04] | 602.8247723458653 | 4 | 166 | 75.40858339468637 | 512 | true | meets root threshold |
| [S05] | 88.04087193247261 | 8 | 0 | 0.0 | — | true | no search |
| [S06] | 2624.8473120914646 | 16 | 1568 | 82.8132987777392 | 512 | true | meets root threshold |
| [S07] | 669.9 | 4 | 314 | 75.58 | 512 | true | meets root threshold |
| [S08] | — | 4 | 0 | 0 | — | false | no search |
| [S09] | 235.4841679805308 | 6 | 418 | 78.61431800127029 | 256 | true | below root threshold |
| [S10] | — | — | — | — | — | false | no search |
| [S11] | — | — | — | — | — | true | no search |
| [S12] | 129.25649534652277 | 8 | 199 | 79.67113597393036 | 256 | true | below root threshold |
| [S13] | 699.5105212690025 | 4 | 246 | 81.05412381490072 | 512 | true | meets root threshold |
| [S14] | 198.14328392031248 | 6 | 49 | 18.88 | 256 | null | below root threshold |
| [S15] | 177.88188611697032 | 2 | 210 | 78.28937974075282 | 256 | false | below root threshold |
| [S16] | 130.76260468034607 | 8 | 169 | 65.34039610849383 | 256 | false | below root threshold |
| [S17] | 500.2158943434953 | 16 | 144 | 80.9957004706065 | 512 | true | meets root threshold |
| [S18] | 1129.5603045175874 | 32 | 628 | 76.97522556673502 | 512 | true | meets root threshold |
| [S19] | 411.4962315358452 | 8 | 331 | 80.62948489586512 | 512 | true | meets root threshold |
| [S20] | 92.80300652177785 | 32 | 0 | 0 | 256 | true | no search |
| [S21] | 189.71387850813545 | 8 | 144 | 75.16549744537718 | 256 | true | below root threshold |
| [S22] | 1100 | 8 | 492 | 62.1 | 512 | true | meets root threshold |
| [S23] | — | — | 0 | 0 | — | false | no search |
| [S24] | 141.36224805456527 | 8 | 196 | 79.69724660714468 | 256 | true | below root threshold |
| [S25] | 108.7927101481026 | 2 | 113 | 80.4637520511945 | 256 | false | below root threshold |

## Per-item findings and reproducibility

### S01 — b13-sg-c10-p32

20-8-2 expansion adds lambdaHat × sizeHat. No search at 37.4444 markets/s; the new input was scaled on seeds 1083999000..1083999255 but all eight new W1 entries and output effects remain zero. No new holdout or real ranking. Gate B is inherited START evidence, not a new replica test. Gas table is probe only; A/C cover 50 sequences split into three chunks, exact START callback identity.

### S02 — b13-sg-c15-p16

20th input uses original lambdaHat and sizeHat; T slopes use their flow-adjusted versions. Searches 186 weights and four log slopes (population 16, ES seed 20261172); 446 generations in 71.97192095214656 minutes. Best monitor edge 528.4625140632904. Independent holdout improvement is positive, real frozen-minus-START is only 0.009930904580458644 with SE 0.022296707195525955. Gas probe peak 228752 with new W1 entries 1e-6.

### S03 — b14-w12-s30-p16

Expands START 19-8-2 to 19-12-2 with new output weights zero; no step-down. All 266 head coefficients trained; feature normalization and T constants fixed. Population 16, ES seed 20261183, sigma0 .03, cap .030*0.5**(minutes/40), selected generation 160 of 194; best monitor increment 0.023551228782892286. Holdout +0.018992856880398395 (SE 0.004408275219825183), real incremental +0.02444950146297109 (SE 0.017878636638830957). Code creation/deployed bytes 15946/15918 below 24576. All-nonzero probe peak 132111, probe runtime 15250. 15 kit tests and IA restoration reported passed; custom loop replaces IA-bound train guard with measured gates/identity/gas.

### S04 — b14-w16-s20-p32

Expands to 19-16-2, 354 coefficients, population 32, ES seed 20261182; sigma0 .02 with 40-minute half-life. Selected generation 60 of 166, monitor gain 0.012634580469400997 (SE 0.0061098586990112865), no later monitor beat it. Holdout +0.01429172404260385; real incremental +0.04103670947871141. Diet shell, no step-down. Every-zero-weight replacement all-nonzero probe peak 132787; expanded/probe runtime sizes 14659/15504 (<24576). Four-worker benchmark excludes compile/startup. 15 kit tests and exact IA restoration reported.

### S05 — b14-w16-s30-p16

No proposals: speed 88.04087193247261 below worker cutoff 100 and root cutoff 300. Wide policy preserves START, zero added output weights. New real block is a confirmation of START behavior, not training gain. All-nonzero probe peak 132898; expanded/frozen runtime 14659, probe 15381 (<24576). Port identity instrumentation was repaired to clear reference state between sequences. 15 kit tests and exact IA restoration reported.

### S06 — b15-ib-s20-h60-p16

Starts from IB frozen params, trains 178 head weights only; population 16, ES seed 20261192, sigma0 .02, cap .020*0.5**(minutes/60), four START-group RMS scales. 1568 generations; selected generation 1470; best monitor gain 0.1769099057861091 (SE 0.010794469264562678), 12845056 proposal-market evaluations, final sigma 0.004621710863527907. Frozen port binds weights by default; original START port defaults to zero head unless params loaded. Repeated C after whitespace-only emitter fix gives same error. Two source warnings; new local measurements remain attributed.

### S07 — b15-ib-s30-h40-p16

IB is START. Population 16, seed 20261191, sigma0 .03, 40-minute cap half-life, 178 weights; selected ES mean generation 300 of 314. Best monitor gain 0.09761464839780229 (SE 0.009412629749005417); worker describes curve still rising slowly. Only _head constants replaced at 1e-6 precision; generator reproduces original IB byte-for-byte before freezing. Steady benchmark 669.9 markets/s, actual search generation throughput 582.6. Real deltas are Y volumes, not edge split; holdout contains split edge. IB timing probe used 8000000..8000007 before assigned confirmation.

### S08 — b16-fa-s20-h60-p16

Blocked before setup: proxy refused GitHub, api.imd.fun and PyPI with “Proxy CONNECT aborted”, curl exit 56/HTTP 000. No repository, wheel, solc or input artifact obtained; no weights, gates, speed, holdout or real run. Zero generations does not mean a measured pass-through. Worker says next leg must start from original FA artifact d3bff2aa…; full identifier is not present in this result.

### S09 — b16-fa-s30-h40-p16

FA frozen params.float used in its own feature layout; only head weights changed, rounded 1e-6. Population 16, ES seed 20261195, sigma0 .03, cap half-life 40 minutes; selected generation 350 of 418. Four fixed RMS block scales. Replica gain over HA 0.5562015907196444 (SE 0.03472238605864462), reference .569. Real improvement over FA is +0.06950588940397406, while frozen minus IB is -0.02660337068547436. This is improvement of FA, not evidence to replace IB. Worker uses lower-speed M=256 branch, contrary to this root’s >=300 rule. Real deltas are volume; holdout deltas are edge.

### S10 — b17-ig-block-a

Confirmation of IG code e1ea8fcb… without search. Requested 1024 seeds; IB all complete, candidate only 960. Missing candidate chunk 8067072..8067135 timed out after 470 seconds (exit 124), no retained edge or gas; not retried. All reported paired statistics and arm means use 960 completed pairs. Quartiles were assigned from all 1024 IB volumes, leaving unequal paired counts 240/242/243/235. Retail/arb deltas -38.70179132610792/-67.66675274150839 are Y volumes; edge splits null. Top-level invalid even with zero observed OOG on retained calls.

### S11 — b17-ig-block-b

Complete independent 1024-seed IG confirmation, code e1ea8fcb…; no training or replica speed benchmark. All 16 chunks of 64 complete. Positive difference share 0.61328125. Edge split reconstructed from arb profit, max recovered-vs-engine arb-volume error 9.31e-05 Y across 2048 runs. Quiet quartiles benefit more than busy quartiles on this block. A pre-block IB probe used 6990000..6990063.

### S12 — b18-j01-s03-h45-p16

J5 start; 219-dimensional joint weights/constants search, population 16, sigma0 .3, 45-minute cap half-life; selected generation 159 of 199. Real incremental +0.012268521997918613 (SE 0.01640176105536101) is not clear improvement, although frozen beats IB. Eight-worker benchmark median of three trials 129.25649534652277, so root speed requirement fails. Port exposes POLICY, fresh float64[40]. Two JSONs referenced by worker but only supplied result is available here.

### S13 — b18-j02-s05-h30-p16

J5 joint search, 246 generations, M512, four-worker speed 699.5105212690025. Gate B rerun on same validation seeds because first result could not serialize NumPy bool; not extra independent evidence. Initial IB holdout used wrong J5 vector, discarded; only IB recomputed with its own TH[HEAD0:] layout and EVM checked (max 0.0000664202309125228 bps). Frozen/START rows retained; no reselection. Port exposes FROZEN_VECTOR. Real incremental +0.008882924562107264 (SE 0.01871773212847921).

### S14 — b18-j03-s05-h45-p16

Only 49 generations/~18.88 minutes; no monitored ES mean beat START (5/10/15-minute gains about -.014/-.018/-.019, SE about .009), so frozen is J5 byte-for-byte. Benchmark 198.14328392031248 was measured while gate A ran on another core; below root threshold. Gate A first attempt timed out at 470 seconds, then rerun in four chunks covering all 50 sequences. Real attempt on 8070656..8071167 killed by 460-second command limit before first arm finished; no real/gas record, valid=null. Those seeds consumed without retained results. Frozen-minus-START zero is identity deduction only, not completed real confirmation.

### S15 — b18-j04-s07-h30-p16

J5 joint search selected generation 202 of 210; M256 at 177.88188611697032 fails root speed rule. Full 512 IB/frozen rows support frozen-vs-IB and candidate_valid=true, but START only 374 rows; top-level valid=false. Missing START seeds 8071744..8071753 and 8072064..8072191; slow one-market batch canceled before eight-minute limit, seeds not replayed. Frozen-minus-START uses only 374 pairs. Initial wrong-layout IB holdout raised ZeroDivisionError; corrected IB alone with its own port, retaining frozen/START once. Packed port differs: concatenate TH, W, anchors; exposes FROZEN_V.

### S16 — b18-j05-s03-h45-p24

169 generations, 65.34039610849383 search minutes; best fixed-monitor mean at final generation 169, gain 0.01732728936470443 (SE 0.007326702636469501). M256 at 130.76260468034607 fails root speed rule. First 64 frozen real rows 8072704..8072767 were simulated but lost after collector TypeError, not replayed; frozen pairs n448 versus 512 START/IB. Missing gas unknown; valid=false. Replica holdout and all gates complete. Partial real incremental +0.07707211805534733 (SE 0.015815972400617426) must not be promoted as full confirmation.

### S17 — b19-j06-s05-h30-p24

Population 24, ES seed 20261206, sigma0 .5, 30-minute cap half-life; M512, 144 generations. Best monitor at 75.18256259032836 minutes: gain .034576596235919244, SE .008433174980112292. Pickled optimizer retains CMA RandomState; monitoring only ES means. Real incremental +0.04540510577683776 (SE 0.015375485007378) and replica +0.03172520010159618 agree in direction. Edge reconstruction maximum arb-volume error 0.00012459568461054005.

### S18 — b19-j07-s07-h45-p16

Population 16, ES seed 20261207, sigma0 .7 per item designation, 45-minute schedule; 628 generations. Best monitor gain 0.0804131909666305 (SE 0.01023727439480771). Real best absolute IB delta among complete group results, but own J5 start is already strong on these seeds; incremental gain is uncertain. Real incremental retail edge -0.1114894893538062 and arb +0.14670345484822978; replica incremental retail -.0694463686533843, arb +.14630774689411896. Full params retained below.

### S19 — b19-j08-s03-h30-p16

219 dimensions, population 16, ES seed 20261208, sigma0 .3, 30-minute cap half-life. 331 generations; selected 329 at 80.1272975564003 minutes, monitor gain .02195927556426472. Initial IA-style vector for IB holdout discarded; corrected IB own TH_VALUES plus HEAD0 weights, frozen/START retained, EVM max 6.64202309125228e-05 bps. Real incremental +0.030015616678395585 (SE .015765590198014223), driven by retail edge +.12100793998513737 versus arb -.09099232330674226.

### S20 — b19-j09-s05-h45-p24

No training: final idle benchmark 92.803 markets/s (32-process pool, no concurrent validation/search) confirms below-100 cutoff; earlier timing 94.264. Returns exact J5. Real uses 1024 seeds per worker’s no-search branch; frozen/START identity zeros are deductions, not duplicate runs. Real split edge missing; Y-volume deltas only. This is a useful unchanged-J5 confirmation, not training improvement.

### S21 — b19-j10-s07-h30-p24

144 generations, 75.16549744537718 active minutes; stopped several minutes before requested cutoff, explicitly disclosed. M256 at 189.71387850813545 fails root speed rule. Checkpoint replay gives identical next proposals from independent pickle loads; explicit global RNG state added late. Gate B rerun after NumPy-bool serialization issue. Initial wrong-layout IB warm-up stopped before producing usable holdout; corrected IB only. Real incremental +.001998574428341837 (SE .017293277018584068) does not establish improvement despite +.6465089430735999 versus IB. Real split edge absent; volumes only.

### S22 — b20-j11-s05-h30-p16

219 dimensions, population 16, ES seed 20261211, sigma0 .5, 30-minute cap half-life. 492 generations, 62.1 minutes, selected ES mean generation 397 at 50.12 minutes; monitor gain .06676734765286546 (SE .009163386377570052). Curve flat within SE after ~40 minutes: +.067 at 50.1, +.062 at 55, +.066 at 60. First short segment continued, not restarted; fresh train seeds. Speed about 1100 on search pool (standalone fresh pools 883–930). Strongest complete incremental J5 real result here; volumes only in real, replica edge split retained. IB timing probe 7999000..7999015 is below fresh range, outside assigned confirmation. Full params retained below; feature scale strings resolved from numeric IA values in S18 for handoff.

### S23 — b20-j12-s03-h45-p16

Blocked by budget: worker reports only about 18 minutes left versus required 30-minute reserve, so no setup/gates/search/holdout/real. Passes through J5 params/code from fe2320b5… unchanged and checks decompressed code SHA. No seeds used. valid=false; cannot inherit J5 performance as new evidence.

### S24 — b20-j13-s05-h45-p16

196 generations, selected 179; M256 at 141.36224805456527 fails root speed rule. Best monitor selection retained independently of confirmation. Holdout increment +.0054288436524801975 (SE .005577484363092883); real increment -.015857299816971815 (SE .026140638702418593): no demonstrated new gain. Initial wrong-vector IB holdout discarded and corrected, verified EVM; frozen/START retained. Real worst frozen-minus-START event is seed 8081133, -9.56556750661889. Runtime/evaluation source maps embedded in worker result; external simulator/numerical packages remain prerequisites.

### S25 — b20-j14-s07-h30-p16

113 generations, 80.4637520511945 minutes; no monitor beat START, so exact J5 retained. M256 at 108.7927101481026 fails root speed rule. Real only 192/512 seeds per arm: completed 8081920..8082111; unrun 8082112..8082431. Observed engine ~1.8 seconds per market/arm on two CPUs. All gates/observed gas pass, observed_gas_valid=true, but valid=false because confirmation incomplete. Parameterization text says sigma0=1 while search.sigma0 and curve show .7; preserve this internal inconsistency rather than resolve by guess.

Training seed provenance from `search`: monitor range is 30000000..30002047 where reported. S02 `{"train_seed_rule":"1076000000 + M*g through +M-1, unique per generation"}`; S03 `{"training_seeds":[1092000000,1092099327]}`; S04 `{"train_seeds":[1088000000,1088084991]}`; S06 `{"train_seeds":[1100000000,1100802815]}`; S07 `{"train_seeds":[1096000000,1096160767]}`; S12 `{"train_seeds":[1112000000,1112050943]}`; S13 `{"train_seeds":[1116000000,1116125951]}`; S14 `{"train_seeds":"1120000000 + 256g .. +255, g = 0..48, each used once"}`; S15 `{"train_seeds":"1124000000 + M*g .. +M-1; g starts at 0, shared by all proposals and never reused in another generation"}`; S17 `{"training_seed_start":1132000000,"training_seed_end":1132073727}`; S19 `{"train_seeds":"1140000000 + M*g .. 1140000000 + M*(g+1)-1; g starts at zero; each block shared by 16 proposals and never reused by later generations"}`; S21 `{"training_seed_start":1148000000}`; S22 `{"train_seeds":[1152000000,1152251903]}`; S24 `{"train_seeds":[1160000000,1160050175]}`; S25 `{"train_seed_start":1164000000,"train_seed_end":1164028927}`; 

## Gates

A/C errors are maximum fee differences in bps; thresholds are <0.01/<0.1. Reported successful differential checks cover 50 sequences × 2000 swaps (100000 swap callbacks plus initialization). B varies in comparator as explained above. Nested checks are retained separately; “not reported” is not a pass.

| Source | A maximum / pass | B measurement and pass | C maximum / pass |
|---|---|---|---|
| [S01] | 8.814065310946972e-05 / True | 512; 0.5074247543096038; SE 0.024315433183016485; pass=True | 8.814065310946972e-05 / True |
| [S02] | start: 8.814065310946972e-05 / True; changed: 8.814065310946972e-05 / True | 512; 0.4330137321358012; SE 0.025468389375728855; pass=True | 0.00010949271638666103 / True |
| [S03] | 8.814065310946972e-05 / True | 512; 0.47212047649777866; SE 0.026045930005405876; pass=True | 0.00017306289269714004 / True |
| [S04] | 8.814065310946972e-05 / True | 512; 0.47212047649777866; SE 0.026045930005405876; pass=True | wide16_expanded: 8.814065310946972e-05 / True; frozen: 6.470419887383283e-05 / True |
| [S05] | 8.814065310946972e-05 / True | 512; 0.47212047649777866; SE 0.026045930005405876; pass=True | 8.814065310946972e-05 / True |
| [S06] | 6.64202309125228e-05 / True | 512; 0.5863817609921741; SE 0.03113175959364559; pass=True | 0.00011144819013833573 / True |
| [S07] | 6.64202309125228e-05 / True | 512; 0.5863817609921741; SE 0.03113175959364559; pass=True | 5.142037923927534e-05 / True |
| [S08] | not reported | `not reported` | not reported |
| [S09] | 0.0001210463733067324 / True | 512; 0.5562015907196444; SE 0.03472238605864462; pass=True | 9.059166486413694e-05 / True |
| [S10] | not reported | `not reported` | 8.814065310946972e-05 / True |
| [S11] | not reported | `not reported` | 8.814065310946972e-05 / True |
| [S12] | 0.00015195374049314925 / True | 512; 0.8655072079761197; SE 0.033774160243544064; pass=True | 8.787459513520668e-05 / True |
| [S13] | 0.00015195374049314925 / True | 512; 0.8655072079761197; SE 0.033774160243544064; pass=True | 6.316124622016833e-05 / True |
| [S14] | 0.00015195374049314925 / True | 512; 0.8655072079761197; SE 0.033774160243544064; pass=True | 0.00015195374049314925 / True |
| [S15] | 0.00015195374049314925 / True | 512; 0.8655072079761197; SE 0.033774160243544064; pass=True | 0.00024764513002761834 / True |
| [S16] | 0.00015195374049314925 / True | 512; 0.8655072079761197; SE 0.033774160243544064; pass=True | 8.259751042849484e-05 / True |
| [S17] | 0.00015195374049314925 / True | 512; 0.8655072079761197; SE 0.033774160243544064; pass=True | 0.00012808946306397395 / True |
| [S18] | 0.00015195374049314925 / True | 512; 0.8655072079761197; SE 0.033774160243544064; pass=True | 0.00010524190317828808 / True |
| [S19] | 0.00015195374049314925 / True | 512; 0.8655072079761197; SE 0.033774160243544064; pass=True | 0.00011908859325754362 / True |
| [S20] | 0.00015195374049314925 / True | 512; 0.8655072079761197; SE 0.033774160243544064; pass=True | 0.00015195374049314925 / True |
| [S21] | 0.00015195374049314925 / True | 512; 0.8655072079761197; SE 0.033774160243544064; pass=True | 0.00017685660963495664 / True |
| [S22] | 0.00015195374049314925 / True | 512; 0.8655072079761197; SE 0.033774160243544064; pass=True | 0.0006865420456442228 / True |
| [S23] | not reported | `not reported` | not reported |
| [S24] | 0.00015195374049314925 / True | 512; 0.8655072079761197; SE 0.033774160243544064; pass=True | 5.272290053755929e-05 / True |
| [S25] | 0.00015195374049314925 / True | 512; 0.8655072079761197; SE 0.033774160243544064; pass=True | 0.00015195374049314925 / True |

## Replica holdout

Each cell is `n; mean; SE`. Missing means not measured/returned. Reported holdout block is 41010000..41018191 (8192 markets). For J5 starts the repeated baseline is 0.5470720195549598, SE 0.010443134156081715; this is one common block, not multiple independent replications. Components are below.

| Source | Frozen–IB | Frozen–START |
|---|---|---|
| [S01] | — | — |
| [S02] | — | 8192; 0.027054020725567403; 0.007173224472079851 |
| [S03] | — | 8192; 0.018992856880398395; 0.004408275219825183 |
| [S04] | — | 8192; 0.01429172404260385; 0.0029765337593881357 |
| [S05] | — | 8192; 0.0; 0.0 |
| [S06] | 8192; 0.17044804098135635; 0.005398841121826615 | — |
| [S07] | 8192; 0.08233047770828306; 0.004762919106195313 | — |
| [S08] | — | — |
| [S09] | 8192; -0.01405736486138344; 0.008249843360785148 | 8192; 0.07034607081645153; 0.00416902072858915 |
| [S10] | — | — |
| [S11] | — | — |
| [S12] | 8192; 0.5503179020855475; 0.010428421049060478 | 8192; 0.003245882530587753; 0.003914954943238957 |
| [S13] | 8192; 0.5662670736390588; 0.010230602859518313 | 8192; 0.01919505408409907; 0.005635176275632321 |
| [S14] | 8192; 0.5470720195549598; 0.010443134156081715 | 8192; 0.0; 0.0 |
| [S15] | 8192; 0.565259978331962; 0.010542338927579806 | 8192; 0.018187958777002203; 0.00509454654697262 |
| [S16] | 8192; 0.5740715403606981; 0.010635226221674461 | 8192; 0.026999520805738385; 0.004212110150491084 |
| [S17] | 8192; 0.578797219656556; 0.010496293672001968 | 8192; 0.03172520010159618; 0.004732529066434342 |
| [S18] | 8192; 0.6239333977956945; 0.010529115819171243 | 8192; 0.07686137824073459; 0.0056237057367839365 |
| [S19] | 8192; 0.5802226613887239; 0.0105750099426588 | 8192; 0.03315064183376415; 0.004187444397037236 |
| [S20] | 8192; 0.5470720195549598; 0.010443134156081715 | 8192; 0.0; 0.0 |
| [S21] | 8192; 0.5782110728044033; 0.010549572271598756 | 8192; 0.031139053249443487; 0.004487051967727445 |
| [S22] | 8192; 0.6124693013540797; 0.01062049802808107 | 8192; 0.06539728179912; 0.004994016923914225 |
| [S23] | — | — |
| [S24] | 8192; 0.55250086320744; 0.010431119403812607 | 8192; 0.0054288436524801975; 0.005577484363092883 |
| [S25] | 8192; 0.5470720195549598; 0.010443134156081715 | 8192; 0.0; 0.0 |

## Real-engine confirmation

Each cell is `n; paired mean edge Y; SE`, exact source values. Each strategy runs alone versus 30 bps Vanilla, never head to head. START is the item’s own starting model; for IB retraining it is IB itself. Missing is not a measured zero. Seed intervals are inclusive and incomplete availability is disclosed per item.

| Source | Seed interval | Frozen–IB | START–IB | Frozen–START |
|---|---|---|---|---|
| [S01] | — | — | — | — |
| [S02] | 8057344..8057855 | 512; 0.22695204266577074; 0.03175727784735289 | — | 512; 0.009930904580458644; 0.022296707195525955 |
| [S03] | 8061440..8061951 | 512; 0.17779997571571715; 0.02822322016318099 | 512; 0.15335047425274606; 0.02907933726536353 | 512; 0.02444950146297109; 0.017878636638830957 |
| [S04] | 8060416..8060927 | 512; 0.19005801156038404; 0.03220712242997874 | 512; 0.14902130208167264; 0.03645193124901563 | 512; 0.04103670947871141; 0.018688512834151867 |
| [S05] | 8059392..8059903 | 512; 0.20805723535847154; 0.02825037920751212 | — | 512; 0.0; 0.0 |
| [S06] | 8063488..8063999 | 512; 0.13356288225988888; 0.0220021711355587 | — | — |
| [S07] | 8062464..8062975 | 512; 0.1152033416652738; 0.016026430471105104 | — | — |
| [S08] | — | — | — | — |
| [S09] | 8064512..8065023 | 512; -0.02660337068547436; 0.03103584682537744 | 512; -0.09610926008944842; 0.03175523823443851 | 512; 0.06950588940397406; 0.017370464207075453 |
| [S10] | 8066560..8067583 | 960; 0.20598168829020178; 0.021724313802593535 | — | — |
| [S11] | 8067584..8068607 | 1024; 0.16993832110719154; 0.021373645540853452 | — | — |
| [S12] | 8068608..8069119 | 512; 0.5457838169650835; 0.03438573508545672 | 512; 0.5335152949671649; 0.03525578842746537 | 512; 0.012268521997918613; 0.01640176105536101 |
| [S13] | 8069632..8070143 | 512; 0.5551475896520158; 0.03298331861314544 | 512; 0.5462646650899086; 0.034391972682963956 | 512; 0.008882924562107264; 0.01871773212847921 |
| [S14] | — | — | — | — |
| [S15] | 8071680..8072191 | 512; 0.5423615704766399; 0.03707830899244414 | 374; 0.5671625600522833; 0.04235546865927746 | 374; -0.016875949647023335; 0.017724454252724547 |
| [S16] | 8072704..8073215 | 448; 0.6057792143102388; 0.04704350753872425 | 512; 0.504891012317938; 0.04398183768418221 | 448; 0.07707211805534733; 0.015815972400617426 |
| [S17] | 8073728..8074239 | 512; 0.5625601320468081; 0.03673276091111249 | 512; 0.5171550262699703; 0.03833604611435274 | 512; 0.04540510577683776; 0.015375485007378 |
| [S18] | 8074752..8075263 | 512; 0.6686299709565341; 0.03788344799968138 | 512; 0.633416005462111; 0.039535897428196534 | 512; 0.03521396549442313; 0.020458822991768485 |
| [S19] | 8075776..8076287 | 512; 0.6391871199649286; 0.03954013346069903 | 512; 0.609171503286533; 0.04039486263689885 | 512; 0.030015616678395585; 0.015765590198014223 |
| [S20] | 8076800..8077823 | 1024; 0.5622450587308492; 0.029805346917661797 | 1024; 0.5622450587308492; 0.029805346917661797 | 1024; 0.0; 0.0 |
| [S21] | 8077824..8078335 | 512; 0.6465089430735999; 0.04095074170111508 | 512; 0.644510368645258; 0.042450801215130896 | 512; 0.001998574428341837; 0.017293277018584068 |
| [S22] | 8078848..8079359 | 512; 0.5953971189639615; 0.03903783775809266 | 512; 0.503904464827636; 0.038335551317079955 | 512; 0.09149265413632557; 0.017772594845482537 |
| [S23] | — | — | — | — |
| [S24] | 8080896..8081407 | 512; 0.5754337905954569; 0.040093369112042726 | 512; 0.5912910904124287; 0.037945747151973894 | 512; -0.015857299816971815; 0.026140638702418593 |
| [S25] | 8081920..8082111 | 192; 0.5737535304263682; 0.08422676297905593 | 192; 0.5737535304263682; 0.08422676297905593 | 192; 0.0; 0.0 |

### Worst five frozen-versus-IB paired outcomes

Entries are exact `[seed,difference]` pairs from `real` (or top-level `paired` in IG confirmations). Only the frozen/candidate comparison is shown here; other comparator tail lists, raw individual rows, uncentered sums/sumsq, arm means and redundant zero-identity tails are intentionally condensed, not treated as separate findings. j13’s exceptional incremental loss is disclosed above.

- [S02] n=512: `[[8057770,-1.9787981184393857],[8057786,-1.5130404061295621],[8057671,-1.4064012514821798],[8057790,-1.34673340829454],[8057535,-1.3208781860872705]]`

- [S03] n=512: `[[8061870,-2.762572878272522],[8061911,-2.2167420619541645],[8061524,-2.13263196878745],[8061460,-2.037324042197497],[8061850,-2.0030402484719616]]`

- [S04] n=512: `[[8060619,-2.31707568391937],[8060433,-2.0993801822769456],[8060734,-1.9608816579828954],[8060889,-1.894282668789515],[8060635,-1.8074326569571895]]`

- [S05] n=512: `[[8059524,-1.8300764525731665],[8059870,-1.7282762290050186],[8059887,-1.4786906370306951],[8059730,-1.460216266650832],[8059899,-1.3257681502368541]]`

- [S06] n=512: `[[8063842,-1.995629538595722],[8063938,-1.9002869295961204],[8063633,-1.8419108841010257],[8063768,-1.6963065221120814],[8063764,-1.2650151039480306]]`

- [S07] n=512: `[[8062650,-1.3304122513377479],[8062469,-1.2187797989678302],[8062471,-0.8457679379177989],[8062575,-0.822400406977863],[8062493,-0.8205279635374154]]`

- [S09] n=512: `[[8064976,-2.137901215079978],[8064630,-2.073675281443002],[8064814,-1.893027828217896],[8064822,-1.7772686741627695],[8064759,-1.6998737182996138]]`

- [S10] n=960: `[[8066910,-2.9532248823760483],[8066719,-2.822517115187111],[8067392,-2.126880274861037],[8067556,-1.6508272998768234],[8067037,-1.5738449371061733]]`

- [S11] n=1024: `[[8067714,-4.566089831615955],[8068022,-2.772031701880337],[8067942,-2.271758106701327],[8067854,-2.1125330467669983],[8068141,-1.7528778897875554]]`

- [S12] n=512: `[[8069008,-2.751334062587148],[8069056,-1.6837872916108836],[8068939,-1.4343408309442793],[8068798,-1.3097548584067908],[8068663,-1.1955212238803483]]`

- [S13] n=512: `[[8069836,-1.700432339414931],[8069653,-1.5965681700936898],[8069663,-1.4657204111764486],[8070032,-1.4278135378136767],[8070126,-1.2862005593948425]]`

- [S15] n=512: `[[8072035,-2.5662105987933046],[8072044,-2.3285005598904718],[8071917,-2.1936162648945583],[8071893,-2.089700133501424],[8072099,-1.8443633985079941]]`

- [S16] n=448: `[[8073136,-7.2928743841087],[8072974,-3.0060921848603357],[8073030,-2.0619770818043435],[8073214,-1.7488710324240628],[8073083,-1.5943869310535774]]`

- [S17] n=512: `[[8074081,-3.8746857302076023],[8073807,-1.8898587035091623],[8073946,-1.3618517317750047],[8074055,-1.3167247374620388],[8074065,-1.1583538579817514]]`

- [S18] n=512: `[[8074817,-2.6786915021930326],[8074752,-2.178169694313169],[8075184,-1.623710085591938],[8075243,-1.4770562099942026],[8074838,-1.4250533573538746]]`

- [S19] n=512: `[[8075943,-2.038004957654948],[8076048,-1.6884886031750739],[8076002,-1.3861638021481895],[8075810,-1.3652831121089548],[8075839,-1.3222712924368807]]`

- [S20] n=1024: `[[8076856,-2.685477260887751],[8077327,-2.370204198351985],[8076869,-2.0849037618762623],[8077277,-1.9215875915491551],[8076958,-1.7751697770449937]]`

- [S21] n=512: `[[8078168,-3.810006180304754],[8078296,-2.1341395410236146],[8078313,-1.5386585051289785],[8078014,-1.369760008822709],[8078030,-1.3689613430429972]]`

- [S22] n=512: `[[8079205,-1.6012559639332267],[8078997,-1.5898618343579756],[8079213,-1.5002212455750623],[8078953,-1.360066774525194],[8079138,-1.3005615645563466]]`

- [S24] n=512: `[[8081133,-7.464843112315293],[8080972,-3.2188796732821174],[8081170,-3.1061918980153678],[8081333,-1.740900225016162],[8080915,-1.7236354254270054]]`

- [S25] n=192: `[[8081988,-5.518813981572066],[8082011,-3.973478131503043],[8081992,-1.8736909860336368],[8081942,-1.4520295761565194],[8081940,-1.3864664022554507]]`

## Retail/arb decomposition and quartiles

All values in this secondary table are rounded to 9 significant digits. `h`=replica holdout, `r`=real; `FI`=frozen minus IB, `FS`=frozen minus START. Component objects preserve source field names and units; component values are means (component SE condensed). `Q` entries are [n,mean] for q1..q4 in quiet-to-busy IB retail-volume order. START component/quartile rows are condensed where the frozen and incremental rows already convey the policy finding. The recurring J5 holdout START-vs-IB edge split is retail -0.5183724507510308 (SE 0.01010538505858936), arb +1.0654444703059904 (SE 0.005673180849003477). [S13, S15, S18–S21]

| Source | Components | Real quartiles FI; FS |
|---|---|---|
| [S02] | `{"hFS":{"R":-0.440183756,"A":0.467237776},"rFI":{"R":-0.700707313,"A":0.927659355},"rFS":{"R":-0.467026404,"A":0.476957309}}` | `{"FI":[[128,0.330500953],[128,0.148959752],[128,0.169354766],[128,0.258992699]]}` |
| [S03] | `{"hFS":{"R":-0.049080576,"A":0.0680734329},"rFI":{"R":-0.344071515,"A":0.521871491},"rFS":{"R":-0.0552726055,"A":0.0797221069}}` | `{"FI":{"q1":[128,0.252712889],"q2":[128,0.195882628],"q3":[128,0.233168517],"q4":[128,0.0294358688]}}` |
| [S04] | `{"hFS":{"R":-0.158502245,"A":0.172793969},"rFI":{"R":-0.435042987,"A":0.625100999},"rFS":{"R":-0.125884214,"A":0.166920923}}` | `{"FI":{"q1":[128,0.230235471],"q2":[128,0.183252835],"q3":[128,0.270647479],"q4":[128,0.076096262]}}` |
| [S05] | `{"hFS":{"R":0.0,"A":0.0},"rFI":{"R":-0.238342362,"A":0.446399597},"rFS":{"R":0.0,"A":0.0}}` | `{"FI":{"q1":[128,0.259003498],"q2":[128,0.153214303],"q3":[128,0.148261557],"q4":[128,0.271749583]}}` |
| [S06] | `{"rFI":{"R":-0.008538117,"A":0.142100999}}` | `{"FI":{"q1":[128,0.1649474],"q2":[128,0.0984710113],"q3":[128,0.183495748],"q4":[128,0.0873373695]}}` |
| [S07] | `{"hFI":{"RE":-0.133527825,"AE":0.215858302},"rFI":{"RV":-108.42991,"AV":-90.0482123}}` | `{"FI":[[128,0.20303708],[128,0.0880644722],[128,0.077912376],[128,0.091799438]]}` |
| [S09] | `{"hFI":{"RE":-0.48385313,"AE":0.469795765},"hFS":{"RE":0.0212200078,"AE":0.049126063},"rFI":{"RV":-198.462236,"AV":-132.916524},"rFS":{"RV":46.714476,"AV":-29.7017899}}` | `{"FI":{"q1":[128,-0.00943516523],"q2":[128,-0.0405922178],"q3":[128,-0.102631388],"q4":[128,0.0462452885]}}` |
| [S10] | `{"rRetail":-38.7017913,"rArb":-67.6667527}` | `{"FI":{"q1":[240,0.241766591],"q2":[242,0.282192178],"q3":[243,0.18218066],"q4":[235,0.115566092]}}` |
| [S11] | `{"rRetail":{"mean":-0.275086425,"se":0.0221272456},"rArb":{"mean":0.445024746,"se":0.00906371994}}` | `{"FI":{"q1":[256,0.279749016],"q2":[256,0.23508074],"q3":[256,0.0695569476],"q4":[256,0.0953665808]}}` |
| [S12] | `{"rFI":{"R":-0.348827613,"A":0.89461143,"RV":152.424768},"rFS":{"R":0.15869678,"A":-0.146428258,"RV":85.5120608}}` | `{"FI":{"q1":[128,0.667448072],"q2":[128,0.471540272],"q3":[128,0.569380008],"q4":[128,0.474766916]},"FS":{"q1":[128,0.0142522327],"q2":[128,0.00138551258],"q3":[128,0.0436672421],"q4":[128,-0.0102308994]}}` |
| [S13] | `{"hFI":{"R":-0.557671082,"A":1.12393816},"hFS":{"R":-0.0392986312,"A":0.0584936853},"rFI":{"R":-0.562935499,"A":1.11808309,"RV":-16.1680507,"AV":-204.076221},"rFS":{"R":-0.0475259898,"A":0.0564089143,"RV":-82.9263436,"AV":-31.330231}}` | `{"FI":{"q1":[128,0.547416547],"q2":[128,0.558563781],"q3":[128,0.494869802],"q4":[128,0.619740228]},"FS":{"q1":[128,-0.0275849213],"q2":[128,0.0115939784],"q3":[128,0.0824234353],"q4":[128,-0.0309007941]}}` |
| [S15] | `{"hFI":{"RE":-0.474340212,"AE":1.03960019},"hFS":{"RE":0.0440322384,"AE":-0.0258442797},"rFI":{"RE":-0.516440572,"AE":1.05880214,"RV":44.3334083,"AV":-146.636654},"rFS":{"RE":0.00660977932,"AE":-0.023485729,"RV":-13.4697985,"AV":28.184115}}` | `{"FI":{"q1":[128,0.63390406],"q2":[128,0.574295328],"q3":[128,0.519122835],"q4":[128,0.442124059]},"FS":{"q1":[100,-0.0342515487],"q2":[98,-0.0681419804],"q3":[89,0.00514121615],"q4":[87,0.0383206385]}}` |
| [S16] | `{"retail_arb_volume_y_deltas":{"FI":{"R":-53.206011,"A":-285.036479},"FS":{"R":-115.786148,"A":-110.454562}}}` | `{"FI":{"q1":[116,0.604846531],"q2":[107,0.580664638],"q3":[108,0.749343717],"q4":[117,0.497150878]}}` |
| [S17] | `{"rFI":{"R":-0.576790415,"A":1.13935055,"RV":-27.1646284,"AV":-222.634997},"rFS":{"R":-0.0270685514,"A":0.0724736572,"RV":-83.8819329,"AV":-51.5574987}}` | `{"FI":{"q1":[128,0.522067574],"q2":[128,0.58595759],"q3":[128,0.544493481],"q4":[128,0.597721883]},"FS":{"q1":[128,0.0477123882],"q2":[128,0.0422725465],"q3":[128,0.0330421685],"q4":[128,0.0585933199]}}` |
| [S18] | `{"rFI":{"RE":-0.557374598,"AE":1.22600457,"RV":48.0982732,"AV":-195.767026},"rFS":{"RE":-0.111489489,"AE":0.146703455,"RV":-24.5362857,"AV":-19.2587371},"hSplit":{"FS":{"retail":{"n":8192,"mean_diff":-0.0694463687,"se":0.00573379011},"arb":{"n":8192,"mean_diff":0.146307747,"se":0.00250637105}},"FI":{"retail":{"n":8192,"mean_diff":-0.587818819,"se":0.0101505484},"arb":{"n":8192,"mean_diff":1.21175222,"se":0.0057496466}}}}` | `{"FI":{"q1":[128,0.728099901],"q2":[128,0.645731421],"q3":[128,0.575358832],"q4":[128,0.72532973]},"FS":{"q1":[128,0.0531554767],"q2":[128,0.0257320324],"q3":[128,-0.0455500096],"q4":[128,0.107518363]}}` |
| [S19] | `{"hFI":{"RE":-0.394908175,"AE":0.975130837},"hFS":{"RE":0.123464275,"AE":-0.0903136336},"rFI":{"R":-0.336908162,"A":0.976095282,"RV":115.651252,"AV":-91.1290498},"rFS":{"R":0.12100794,"A":-0.0909923233,"RV":29.4976089,"AV":76.1554043}}` | `{"FI":{"q1":[128,0.706822748],"q2":[128,0.665389949],"q3":[128,0.499122549],"q4":[128,0.685413234]},"FS":{"q1":[128,0.0539243289],"q2":[128,0.0484605025],"q3":[128,-0.0173532818],"q4":[128,0.0350309172]}}` |
| [S20] | `{"retail_arb_deltas":{"FI":{"RV":{"n":1024,"mean_diff":73.7380871,"se":8.09625299},"AV":{"n":1024,"mean_diff":-168.812316,"se":5.11518845}},"FS":{"RV":{"n":1024,"mean_diff":0.0,"se":0.0},"AV":{"n":1024,"mean_diff":0.0,"se":0.0}}}}` | `{"FI":{"q1":[256,0.584054946],"q2":[256,0.602117388],"q3":[256,0.60837157],"q4":[256,0.45443633]},"FS":{"q1":[256,0.0],"q2":[256,0.0],"q3":[256,0.0],"q4":[256,0.0]}}` |
| [S21] | `{"hFI":{"RE":-0.200166746,"AE":0.778377819},"hFS":{"RE":0.318205705,"AE":-0.287066651},"rFI":{"RV":185.690579,"AV":-23.8341663},"rFS":{"RV":110.858432,"AV":152.421102}}` | `{"FI":{"q1":[128,0.556282302],"q2":[128,0.720343429],"q3":[128,0.598403381],"q4":[128,0.711006661]},"FS":{"q1":[128,-0.01245603],"q2":[128,0.0156223864],"q3":[128,0.00135489108],"q4":[128,0.00347305018]}}` |
| [S22] | `{"rFI":{"RV":115.065659,"AV":-116.891155},"rFS":{"RV":54.8425878,"AV":47.8310003},"hretail_edge_delta_frozen_vs_start":0.129053161,"harb_edge_delta_frozen_vs_start":-0.0636558794}` | `{"FI":[[128,0.584771583],[128,0.566764826],[128,0.599421873],[128,0.630630194]],"FS":[[128,0.0786046025],[128,0.121966059],[128,0.0812680884],[128,0.084131867]]}` |
| [S24] | `{"rFI":{"R":-0.637030446,"A":1.21246424,"RV":150.608541},"rFS":{"R":-0.157161246,"A":0.141303946,"RV":70.3600996}}` | `{"FI":{"q1":[128,0.630096936],"q2":[128,0.611601293],"q3":[128,0.513707885],"q4":[128,0.546329048]},"FS":{"q1":[128,-0.0346699209],"q2":[128,0.0492698231],"q3":[128,-0.0182509021],"q4":[128,-0.0597781994]}}` |
| [S25] | `{"rFI":{"R":-0.528232874,"A":1.1019864},"rFS":{"R":0.0,"A":0.0}}` | `{"FI":{"q1":[48,0.422812332],"q2":[48,0.669721448],"q3":[48,0.384570302],"q4":[48,0.81791004]},"FS":{"q1":[48,0.0],"q2":[48,0.0],"q3":[48,0.0],"q4":[48,0.0]}}` |

Field abbreviations: RE/AE=retail/arb edge; RV/AV=retail/arb volume Y; R/A retain source plain retail/arb semantics (edge for reconstructed results, volume only for S10 and S16/S20/S22 where plain deltas occur in volume containers). Rse/Ase are their SE. S10 has no split-edge estimate. S19 real split-edge estimate is absent; its R/ A volume entries are explicitly RV/AV. S09 and S07 real deltas are explicitly volume fields. See per-item caveats for reconstruction versus directly exposed volume. Source quartile range endpoints and START quartile details are explicitly condensed; all retained candidate quartile summaries preserve their actual n.

## Gas and code integrity

Gas comes from worker `amm_sim_rs.gas_stats()` output or their aggregates. Only Submission-role callback counters are summarized, so the opponent Vanilla counters are not mistaken for policy gas. Table entries are `arm: afterSwap_peak / afterInitialize_peak / observed_OOG / other_failures / afterSwap_calls`. All retained reported OOG counters are zero; missing chunks remain unknown. A missing value is “?”. No new simulator gas measurement was run in this merge.

- [S01] probe_submission: 131829 / 176924 / 0 / 0 / 174797.

- [S02] IB: 234958 / 243227 / 0 / 0 / 5527360; START: 235611 / 243237 / 0 / 0 / 5524944; frozen: 236193 / 243240 / 0 / 0 / 5518164.

- [S03] IB: 234925 / 243227 / 0 / 0 / 5525743; START: 235582 / 243237 / 0 / 0 / 5521947; Frozen: 133634 / 176934 / 0 / 0 / 5524350.

- [S04] IB: 234868 / 243227 / 0 / 0 / 5531746; START: 235501 / 243237 / 0 / 0 / 5530199; Frozen: 135702 / 176960 / 0 / 0 / 5527831.

- [S05] IB: 237569 / 243227 / 0 / 0 / 5476015; START: 238266 / 243237 / 0 / 0 / 5472292; Frozen: 131613 / 176904 / 0 / 0 / 5472292.

- [S06] stats/IB: 237621 / 243227 / 0 / 0 / 5526609; stats/frozen: 237594 / 243227 / 0 / 0 / 5532274.

- [S07] candidate/gas_stats: 237021 / 243227 / 0 / 0 / 5528564; IB: 237021 / ? / 0 / 0 / 5533232.

- [S08] not measured / no retained gas.

- [S09] ib/gas_stats: 234864 / 243227 / 0 / 0 / 5565257; start/gas_stats: 234677 / 243227 / 0 / 0 / 5536485; frozen/gas_stats: 234665 / 243227 / 0 / 0 / 5539118.

- [S10] ib/gas_stats: 236979 / 243227 / 0 / 0 / 11005769; cand/gas_stats: 237703 / 243237 / 0 / 0 / 10313913.

- [S11] IB/gas_stats: 237045 / 243227 / 0 / 0 / 11077216; CAND/gas_stats: 237769 / 243237 / 0 / 0 / 11072110.

- [S12] IB: 234872 / 243227 / 0 / 0 / 5532847; J5: 237399 / 243234 / 0 / 0 / 5553004; frozen: 237387 / 243234 / 0 / 0 / 5560400.

- [S13] frozen/gas_stats: 235494 / 243234 / 0 / 0 / 5546225; start/gas_stats: 235385 / 243234 / 0 / 0 / 5553422; ib/gas_stats: 234837 / 243227 / 0 / 0 / 5534296.

- [S14] not measured / no retained gas.

- [S15] IB/stats: 237091 / 243227 / 0 / 0 / 5487737; START/stats: 238030 / 243234 / 0 / 0 / 3997824; frozen/stats: 238060 / 243234 / 0 / 0 / 5506500.

- [S16] frozen/stats: 237690 / 243234 / 0 / 0 / 4827573; START/stats: 237687 / 243234 / 0 / 0 / 5524387; IB/stats: 237142 / 243227 / 0 / 0 / 5506633.

- [S17] frozen/raw: 237595 / 243234 / 0 / 0 / 5521118; START/raw: 237586 / 243234 / 0 / 0 / 5529312; IB/raw: 234868 / 243227 / 0 / 0 / 5510127.

- [S18] not measured / no retained gas.

- [S19] IB/gas_stats: 234912 / 243227 / 0 / 0 / 5537736; START/gas_stats: 237586 / 243234 / 0 / 0 / 5557988; frozen/gas_stats: 237568 / 243234 / 0 / 0 / 5563611.

- [S20] START: 238447 / 243234 / 0 / 0 / 11069766; IB: 237129 / 243227 / 0 / 0 / 11032279; frozen: 238447 / 243234 / 0 / 0 / 11069766.

- [S21] not measured / no retained gas.

- [S22] frozen: 237673 / 243234 / 0 / 0 / 5537398; START: 237644 / ? / 0 / 0 / ?; IB: 234965 / ? / 0 / 0 / ?.

- [S23] not measured / no retained gas.

- [S24] IB: 232231 / 243227 / 0 / 0 / 5554065; START: 237387 / 243234 / 0 / 0 / 5574496; frozen: 237402 / 243234 / 0 / 0 / 5574080.

- [S25] ib: 237009 / 243227 / 0 / 0 / 2052708; start: 237586 / 243234 / 0 / 0 / 2059215; frozen: 237586 / 243234 / 0 / 0 / 2059215.

Embedded Solidity/port payloads were decoded here where present, and compared with their supplied decoded SHA-256 fields. This checks internal identity, not EVM correctness or external provenance. The original artifact SHA-256s in the registry can be compared with the requester’s pinned inputs. Code imports and simulator dependencies were not built in this merge.

| Source | Frozen Solidity SHA-256 | Local payload hash check |
|---|---|---|
| [S01] | `a99b46ffc41a73d00ec32aa01ae8ae657f90b30c394fc12de1fb28da8f5f3700` | code + port match |
| [S02] | `186b378d73ffdd799027e6f75e3e911f9d0a55d410edfc7aaef541befc17bec3` | code + port match |
| [S03] | `cf9becbe2bb7958513e1edc4721f1551c6697cf2d4514a3f905f803d9ba99036` | code match |
| [S04] | `7d55b389179d5ac1f1af98f3b593803e818cfa56e625ae7b83c30996dc8e0bf0` | code match |
| [S05] | `76d845f2893fa626082b580e0cc316646fd5120b920effdf1b28a2069da376eb` | code match |
| [S06] | `1cfa3d2f0b0405d3d26befc8884c34e47cf67ac25be4e9da007fa437c25d7253` | code + port match |
| [S07] | `e932fe1800cac626e7f4e6fa5168d6e42275cc32449b141205fbc42584f30028` | code + port match |
| [S08] | `None` | no code/port payload |
| [S09] | `113ad10237ea6aad79e1ec8fbd7df8f8a84d661d5d094bae29eb76d07d121797` | port match; plain code: match |
| [S10] | `IG-code` | plain code: match |
| [S11] | `IG-code` | plain code: match |
| [S12] | `484fa0dc8c6d64152277ad3dc90cbfd81431bb791e09ed13fadd3cbd802b754a` | code + port match |
| [S13] | `a5a95e360b1ec9bdefb11730ead044cece654922c248eec66dc2ac3072301fd5` | code + port match |
| [S14] | `J5-code` | code + port match |
| [S15] | `ac69ad854bcbaffef326ddc81bfd2ee623b81de2227451f0c820aaa5eaf2a66b` | code + port match |
| [S16] | `183b01709e461784adfa0709ce7ce80c54fc8c70418b8724232f0494fd2d4d1f` | code + port match |
| [S17] | `abd6b7358f8c1c312546d53cc94297f70a0fa6d635910861fb089c2dc043ad19` | code + port match |
| [S18] | `a57088f013824939a4afe40f1de3574e9447ec7d0af2a0a5d59b36095e4f0f1d` | code + port match |
| [S19] | `21642bc08c48b79448f6cc4cddeb406f0d9be53ad49a37d8fe1f9d9d41904f74` | code + port match |
| [S20] | `J5-code` | code + port match |
| [S21] | `c444b5d93f3581244e88a4447bf56aa1e5eb69ce9e0988bb6b8ad9320531fe6e` | code + port match |
| [S22] | `ea4047d3a6deedeaeeb155797b58cee3517c2004f8c5961c68485598f133ff4e` | code + port match |
| [S23] | `J5-code` | code match |
| [S24] | `3999afb6acdcd3bdb6dbee03603c4a78dcd3979d55a30f3560809ceab6671846` | code + port match; source maps match |
| [S25] | `J5-code` | code + port match; replica match |

Digest aliases: `IB-code` = `301941829a44bb60977bb28181a5e4bfa6544f3696763f987054789b2611d686`; `J5-code` = `dd53db1fbe42628e040c6c7b7a82689867ab601263c5e740506299a3a186b38c`; `IG-code` = `e1ea8fcb105e00754d17061efbe4b299a634b0fcc1b5c293d65b631654b7204b`.

Port SHA-256s are condensed from this table; matching embedded bytes are fixed by each source snapshot hash. The hash check covered both code and port where present.

## Continuation parameters retained in this root

The following lossless parameter archive copies the complete returned `params` of j11 and j07 exactly; they are frozen policies, not final CMA search coordinates. j11’s feature multipliers/offsets are strings referring to unchanged IA values; the numeric values and feature order are present in j07’s archived object. Both items report those features fixed. For a restart, keep each model’s own constants and its own weights together. Constants are `[original IA value, frozen value]` in these results; use the second member. Do not pass an IA/J5 packed vector to IB’s distinct port API. [S18, S22]

J5-family network layout: 19 clipped inputs [-4,4], 8 hidden ReLUs clipped [0,4], two output fee corrections in bps added to T. Common packed vector is `[1,feature_multipliers(19),feature_offsets(19),W1 row-major(152),b1(8),W2 row-major(16),b2(2),anchor_count_cap,anchor_linear_k,anchor_quadratic_k,TH_VALUES]`; NW=217 and TH begins at 220. Set named TH constants to the frozen second values, fresh float64 state[40] per market, then `strat_init(S,W,x0,y0)` and `strat_after(S,W,buy,amount_x,amount_y,timestamp,reserve_x,reserve_y)`. Amounts use real units. S15 has a different vector order documented above, so this common layout must not be applied to every port. Frozen head weights use 1e6 fixed point. A real continuation must reconstruct/obtain the matching port and Solidity and rerun gates; these parameter objects alone are not a newly runnable submission. [S12, S15, S17–S25]

### Exact frozen parameter archive [S22, S18]

The following is base64(zlib(UTF-8 JSON)), with keys S22 and S18 and unchanged full `params` objects. It is data, not executable code. Decoded JSON SHA-256: `99a54183d2bda4ef1bef981d0f5345634b700c3f3e985c80703c4520b5b48aa9`. Numeric IA feature scaling is in S18; S22 retains its source strings. Decode with any standard base64/zlib/JSON implementation.

```text
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
```

## Explicit condensation and missing evidence

To fit the root size bound, this merge retains every item’s substantive finding/status, the paired confirmation statistics, candidate-vs-IB worst five, holdout comparisons, component and quartile patterns, gas maxima/failures, source warnings and source/code identities. Full training curves, raw per-market rows, aggregate gas sums, repeated arm means, callback coverage histograms, duplicate identity tails and most optimizer-coordinate restart arrays are condensed. Solidity/base64 ports and frozen parameter arrays for other candidates are not duplicated; their source snapshot and decoded Solidity hashes are retained, and they must not be inferred from j11/j07. This is an explicit payload omission, not a claim that those arrays were absent from the supplied inputs. j11/j07 are retained for continuation, not declared a proven common-seed winner.

Most workers cite separate sources.json, scripts/checkpoints or additional deliverables that are not among the 25 supplied inputs; their contents could not be found here and are not vouched for. No source failure is silently repaired, no missing real/gas values are filled, no leaderboard placement is inferred from local means, and no synthetic market results are added. Embedded-payload integrity checks carry no independent authority over performance.