# Merge 01 — batch r6-cg-joint-t (items t01–t05)

Task of each item: CMA-ES search over CG's 19-8-2 head (178 weights) plus T's 41 constants on the replica, starting from CG's frozen head and T's constants, then a real-engine check against IB. This merge reads the five item result files under `.imd/reads/artifacts/`. I checked the sha256 of each file against the id it was given; all five match. For each item that has code, I also decoded `code_zb64` (base64, then zlib) and confirmed the sha256 of the decoded code equals that item's `code_sha256`. Each `params` has 178 head weights and 41 constants.

## Summary

| key | mode | replica mk/s | gens | search min | holdout over T (n=8192) | holdout over base (CG) | real vs IB (n=1024) | afterSwap peak | OOG | valid | last curve point (min, gain over T, sigma) |
|---|---|---|---|---|---|---|---|---|---|---|---|
| t01 | real-only | 308.9 | 0 | 0 | +1.952 se 0.012 | 0 (is CG) | +0.166 se 0.020 | 235,611 | 0 | yes | none (no search) |
| t02 | search | 634.9 | 92 | 46.84 | +2.193 se 0.013 | +0.241 se 0.007 | +0.395 se 0.024 | 237,723 | 0 | yes | 45.35, +2.244, 0.351 |
| t03 | search | 1358.4 | 1598 | 55.33 | **+2.533 se 0.013** | **+0.581 se 0.008** | **+0.819 se 0.060** | 235,575 | 0 | yes | 55.3, +2.585, 0.271 |
| t04 | real-only | 256.9 | 0 | 0 | +1.952 se 0.012 | 0 (is CG) | +0.183 se 0.024 | 238,362 | 0 | yes | 0, +2.002, n/a (start point only) |
| t05 | blocked | — | — | — | — | — | — | — | — | no | — |

Every gate that ran passed. The item files report the same Gate B figure, CG−T on the replica = +1.920 se 0.044 on seeds 7232768..7233279 against the +2.00 target, so each is within 0.15. Gate A/C maximum fee differences were 2.9e-5 to 5.7e-5 bps for t01, t02 and t03, and 1.1e-3 bps for t04, which used a different trace generator with trades up to 15% of reserves. All of these are below the 0.01 / 0.1 bps limits.

**Best (valid search row with the highest holdout over_T): t03.** Replica holdout: +2.533 se 0.013 over T and +0.581 se 0.008 over CG. Real engine: +0.819 se 0.060 over IB, median +0.704, 85.8% of seeds positive. One seed, 8588481, gives +55.66. Without it t03 reports +0.765 se 0.026; the headline figure keeps every seed. In 11.08M afterSwap calls the afterSwap peak was 235,575 with 0 OOG, revert or halt.

## Where results agree and disagree

- **Replica baseline agrees.** All four runs measured the same CG−T holdout, +1.9519 se 0.0119 on seeds 40000000..40008191, and the same Gate B. t02 and t03 also start their monitor at the same value (+2.0025).
- **Real-only runs agree.** t01 and t04 each ran CG unchanged against IB on different fresh seeds and got +0.166 se 0.020 and +0.183 se 0.024. Pooled in this merge from their sums: n=2048, mean +0.1747, se 0.0155.
- **Speed differs a lot between machines.** The same replica and CG port ran at 257 to 1358 markets/s. The two slow machines (t01, t04) skipped the search under the <500 markets/s rule. The difference reflects worker count (8 for t01/t04, 8 processes for t02, 30 fork processes for t03), not the code.
- **The two searches are consistent in direction.** t03's larger gain over t02 matches its about 17× more generations (1598 vs 92). t03 ran population 16 with step scale 0.03/0.02; t02 ran population 32 with 0.05/0.03.
- **Replica-to-real transfer.** This is my inference, not something an item measured. Real candidate−CG is roughly real candidate−IB minus the pooled CG−IB (+0.175). That gives about +0.64 for t03 (replica says +0.581) and about +0.22 for t02 (replica says +0.241). The comparison uses different seed sets, so it is only indicative. It fits the pilot's finding that the replica predicts real gains within about 0.1.
- **Retail/arb split differs by method.** t01, t02 and t03 report null retail/arb edge deltas and give Y-volume deltas instead, because the wheel does not split edge. t04 reconstructed approximate edge deltas: retail −0.261 se 0.023, arb +0.444 se 0.012. Volume deltas vs IB in Y: t01 retail −39.5, arb −70.5; t02 retail +9.4, arb −161.1; t03 retail +281.9, arb −51.8; t04 retail −38.2, arb −69.3.

## Item details

- **t01** (real-only, real seeds 8580000..8581023): mean edge CG 527.27 vs IB 527.11. Worst five paired differences: −2.52, −2.39, −2.23, −1.71, −1.68. Disclosed: (a) after a JSON serialization error the first 64 real seeds were run again, which breaks strict never-reuse for that chunk; (b) a callback stress test with trades capped at 15% of X reserves hit a division-by-zero revert after extreme reserve depletion, so the reported gate uses a 0.1% cap.
- **t02** (search, CMA-ES seed 772, pop 32, training seeds 1413000000..1413047103, real seeds 8584000..8585023): mean edge 528.07 vs IB 527.68. Worst five: −3.05 (8584774), −1.84, −1.80, −1.62, −1.51. Disclosed: IB values for seeds 8584000 and 8584001 were carried over from a preliminary IB probe and not rerun.
- **t03** (search, CMA-ES seed 773, pop 16, training seeds 1416000000..1416818175, real seeds 8588000..8589023): ran in 9 segments of under 8 minutes each. The frozen point is the monitored CMA mean at 55.3 min. Mean edge 526.70 vs IB 525.89. Worst five: −2.44 (8588793), −1.43, −1.40, −1.32, −1.30. Disclosed: the IB arm ran at the same time as a replica search segment (CPU sharing only). The afterInitialize peak is 243,234; IB's code path peaks at 243,227 too.
- **t04** (real-only, real seeds 8592000..8593023): mean edge CG 533.15 vs IB 532.97. Worst five: −6.72 (8592640), −2.25, −2.19, −2.11, −2.06. Disclosed: (a) an exploratory unbounded synthetic replay reverted with Solidity panic 0x12 (division by zero) after more than 20 sequences; (b) one CG-only diagnostic run on seed 7000100 is excluded from the statistics; (c) gate C was not run separately, since the frozen code is byte-identical to CG, so gate A stands for it.
- **t05**: blocked at SETUP step `uv venv --python 3.12 .venv`: "exit 126: scratch/bin/uv: cannot execute binary file". The host reported Darwin arm64, but the prescribed binaries are Linux x86_64. Only the wheel's sha256 check passed. This item has no data, and nothing is guessed in its place.

## Notes

- **Failures seen:** t05 blocked at setup. t01 and t04 were too slow and did no search. t01 and t04 hit JSON serialization errors (numpy bool) and recovered. Both t01 and t04 hit reverts (division by zero / panic 0x12) on synthetic traces with large trades. The real engine showed none: 0 OOG/revert/halt across every real arm.
- **Curves were still rising.** In t02 every 5-minute monitor point increased (+2.002 → +2.244 over 45.35 min, sigma 0.351 at the end). In t03 every point increased (+2.002 → +2.585 over 55.3 min, sigma 0.271 at the end; the t03 file also says so in its notes). Neither had plateaued. Longer training from t03's frozen params is the obvious next step.
- **Distance to goal (inference).** The brief puts the top 10 at about +3.1 over HA, where HA is T + 1.257. That is about +4.36 over T. t03's replica holdout is +2.53 over T, so a gap of about 1.8 remains. HA itself was not measured in any item.
- **Sources that did not hold on re-read:** t01 (3), t02 (2), t04 (3). The only URLs in these files are the api.imdnotdead.fun input artifacts (named in t02's `inputs`; t01 and t04 cite artifact hashes). Each item checked those artifacts' sha256 locally, and every performance, gate and gas number in this merge was measured locally by the item. It was not quoted from those sources. So the failed re-reads affect only where the inputs came from. They do not affect the measurements here, but the provenance of t01, t02 and t04 inputs cannot be independently re-confirmed from those URLs. t03 cites no web sources.
- **Validity scope:** "valid" means zero OOG, afterSwap peak ≤ 245,000, gates passed, and 1024 paired real seeds. No item claims safety on arbitrary inputs or a leaderboard placement.

## Merged JSON

Numbers are strings. `best` holds t03's `code_zb64`, `code_sha256` and `params` copied whole from its result file.

```json
{
 "rows": [
  {
   "key": "t01",
   "mode": "real-only",
   "result_sha256": "3e7e4e15262ab016b9b7dfc19ab55b2598c17f45a3098b450416d14b27acb769",
   "markets_per_s": "308.9133342601684",
   "generations": "0",
   "search_minutes": "0",
   "holdout_over_T": {
    "mean": "1.9519172921936434",
    "se": "0.011938004553452971",
    "n": "8192"
   },
   "holdout_over_base": {
    "mean": "0.0",
    "se": "0.0",
    "note": "frozen = CG byte-identical"
   },
   "real_vs_IB": {
    "mean": "0.16625697263110506",
    "se": "0.02000975410375256",
    "n": "1024",
    "seeds": "8580000..8581023"
   },
   "afterSwap_peak": "235611",
   "OOG": "0",
   "valid": "true",
   "last_curve_point": null,
   "code_sha256": "e1ea8fcb105e00754d17061efbe4b299a634b0fcc1b5c293d65b631654b7204b"
  },
  {
   "key": "t02",
   "mode": "search",
   "result_sha256": "299a35aa4fe520e76d1d68f829c3fee76a6a35e9526aebfc3bb6982cd752cff4",
   "markets_per_s": "634.8716471599907",
   "generations": "92",
   "search_minutes": "46.83893282853145",
   "holdout_over_T": {
    "mean": "2.1931730568964474",
    "se": "0.012962698235887802",
    "n": "8192"
   },
   "holdout_over_base": {
    "mean": "0.24125576470280372",
    "se": "0.007037874154145838"
   },
   "real_vs_IB": {
    "mean": "0.3948459133695228",
    "se": "0.02374619862551985",
    "n": "1024",
    "seeds": "8584000..8585023"
   },
   "afterSwap_peak": "237723",
   "OOG": "0",
   "valid": "true",
   "last_curve_point": {
    "minutes": "45.352",
    "monitor_gain_over_T": "2.243668650066411",
    "sigma": "0.3506911484020197"
   },
   "code_sha256": "c9332fb36581acdcb4a4b94b5ae39f5127962b9a55e62ffa59eee9c207ee515c"
  },
  {
   "key": "t03",
   "mode": "search",
   "result_sha256": "2333ecf28d4356bd232b85c1e55842e94e1ada05e1cd067830b5625dfdbc5336",
   "markets_per_s": "1358.35",
   "generations": "1598",
   "search_minutes": "55.33",
   "holdout_over_T": {
    "mean": "2.533096306362695",
    "se": "0.01349951591946062",
    "n": "8192"
   },
   "holdout_over_base": {
    "mean": "0.5811790141690515",
    "se": "0.008338720514523593"
   },
   "real_vs_IB": {
    "mean": "0.8189009411467073",
    "se": "0.059561465105944955",
    "n": "1024",
    "seeds": "8588000..8589023"
   },
   "afterSwap_peak": "235575",
   "OOG": "0",
   "valid": "true",
   "last_curve_point": {
    "minutes": "55.3",
    "monitor_gain_over_T": "2.5846",
    "sigma": "0.2713"
   },
   "code_sha256": "cb6d945511e1e10af78baa64c2a6cb25806f0998403ce6440654bfc3c70092d9"
  },
  {
   "key": "t04",
   "mode": "real-only",
   "result_sha256": "85358a26f86a6f0f8f60cc29e4d7d5aaa1a975ab623ca1d4dd3e683a15207aa0",
   "markets_per_s": "256.90032706117245",
   "generations": "0",
   "search_minutes": "0",
   "holdout_over_T": {
    "mean": "1.9519172921936434",
    "se": "0.011938004553452971",
    "n": "8192"
   },
   "holdout_over_base": {
    "mean": "0.0",
    "se": "0.0",
    "note": "frozen = CG byte-identical"
   },
   "real_vs_IB": {
    "mean": "0.1831314467005002",
    "se": "0.023686881060165477",
    "n": "1024",
    "seeds": "8592000..8593023"
   },
   "afterSwap_peak": "238362",
   "OOG": "0",
   "valid": "true",
   "last_curve_point": {
    "minutes": "0",
    "monitor_gain_over_T": "2.0024990491448156",
    "sigma": null,
    "note": "initial point only, no search"
   },
   "code_sha256": "e1ea8fcb105e00754d17061efbe4b299a634b0fcc1b5c293d65b631654b7204b"
  },
  {
   "key": "t05",
   "mode": "blocked",
   "result_sha256": "b6e559e3cb9258e52ed07deaec1e43e0b76c64207cc1358ca6ca53f5175c37dc",
   "markets_per_s": null,
   "generations": null,
   "search_minutes": null,
   "holdout_over_T": null,
   "holdout_over_base": null,
   "real_vs_IB": null,
   "afterSwap_peak": null,
   "OOG": null,
   "valid": "false",
   "last_curve_point": null,
   "code_sha256": null,
   "blocked": {
    "step": "SETUP: sh scratch/setup.sh, line 6 (uv venv --python 3.12 .venv)",
    "error": "exit 126: scratch/bin/uv: cannot execute binary file",
    "cause": "The prescribed setup downloads uv-x86_64-unknown-linux-gnu and a manylinux2014_x86_64 simulator wheel; the execution host reports Darwin arm64."
   }
  }
 ],
 "best": {
  "key": "t03",
  "holdout_over_T": {
   "mean": "2.533096306362695",
   "se": "0.01349951591946062",
   "n": "8192"
  },
  "code_sha256": "cb6d945511e1e10af78baa64c2a6cb25806f0998403ce6440654bfc3c70092d9",
  "code_zb64": "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",
  "params": {
   "W1": [
    [
     -0.122003,
     0.731454,
     -0.465368,
     2.998833,
     -2.325042,
     0.583646,
     -1.465719,
     2.330367,
     -5.296259,
     -1.804534,
     -0.852343,
     -0.171263,
     0.445759,
     -2.173843,
     -0.471323,
     -0.232548,
     -1.223397,
     -0.141278,
     1.469197
    ],
    [
     0.723577,
     0.803271,
     2.012153,
     2.714905,
     -0.15748,
     -0.282939,
     -0.464975,
     0.413107,
     -0.374623,
     0.209898,
     -0.085672,
     -1.819724,
     0.155218,
     -0.150565,
     -0.090415,
     1.03029,
     0.352852,
     0.060971,
     1.075809
    ],
    [
     1.136941,
     3.756676,
     -1.203927,
     0.753486,
     0.931579,
     2.18083,
     -1.749807,
     1.029621,
     1.28779,
     -0.682605,
     0.429824,
     -0.075276,
     -0.430937,
     0.395573,
     0.304551,
     1.171626,
     -2.518045,
     -0.429065,
     -2.231623
    ],
    [
     0.38127,
     -1.388284,
     1.016535,
     -0.407249,
     -1.012472,
     -0.390365,
     0.096328,
     0.065385,
     0.788451,
     0.171605,
     0.180177,
     -1.77325,
     -0.281683,
     0.176087,
     0.248675,
     -1.12143,
     1.042201,
     -0.458772,
     -0.871662
    ],
    [
     1.327491,
     0.535888,
     4.87539,
     0.001265,
     -1.974526,
     -1.26398,
     -1.11645,
     -0.628563,
     -0.409856,
     0.563594,
     0.003287,
     -2.352466,
     -0.044409,
     -1.113051,
     0.319107,
     -0.710736,
     2.261434,
     -0.084005,
     0.495262
    ],
    [
     1.286688,
     -1.941086,
     0.444896,
     -1.373127,
     3.641454,
     1.422972,
     3.746961,
     0.525433,
     -0.534447,
     0.051663,
     0.011805,
     3.715596,
     0.577559,
     -1.171091,
     0.156784,
     0.495838,
     -1.786088,
     -0.276209,
     -1.605988
    ],
    [
     -1.091095,
     1.136029,
     -0.019796,
     -0.728152,
     2.535019,
     0.747419,
     -1.205788,
     2.259411,
     1.227282,
     -1.521974,
     1.247646,
     1.17961,
     1.415836,
     -1.102357,
     -1.135424,
     -1.465166,
     0.961808,
     -0.916514,
     2.464189
    ],
    [
     -2.543156,
     -2.722593,
     3.353545,
     3.590386,
     -0.223427,
     -0.855418,
     -1.730654,
     2.482834,
     0.435976,
     -0.90223,
     2.473317,
     -2.129974,
     1.382278,
     0.140581,
     0.699543,
     0.93941,
     -0.394647,
     -1.848609,
     1.259045
    ]
   ],
   "b1": [
    1.919914,
    -1.602846,
    -4.340146,
    -0.866279,
    -1.483061,
    -2.196219,
    -0.574283,
    0.399937
   ],
   "W2": [
    [
     0.185543,
     -2.066947,
     0.949148,
     -0.630713,
     -3.080369,
     -0.74658,
     -0.935833,
     0.951334
    ],
    [
     0.412422,
     -2.048122,
     -1.701156,
     3.503979,
     0.290637,
     -2.178203,
     0.888547,
     -0.182852
    ]
   ],
   "b2": [
    1.499011,
    2.997006
   ],
   "constants": {
    "ELAPSED_CAP": 7,
    "SIGNAL_THRESHOLD": 0.001171470033649966,
    "DIR_DECAY": 0.7335249441444244,
    "ACT_DECAY": 0.3765498595421264,
    "SIZE_DECAY": 0.48738084586012276,
    "TOX_DECAY": 0.9286329950353552,
    "SIGMA_DECAY": 0.7327584924491856,
    "LAMBDA_DECAY": 0.9918504260432562,
    "SIZE_BLEND_DECAY": 0.8397180057889105,
    "TOX_BLEND_DECAY": 0.2359008403605123,
    "ACT_BLEND_DECAY": 0.987140557453247,
    "PHAT_ALPHA": 0.24980215859066515,
    "PHAT_ALPHA_RETAIL": 0.05426450370635507,
    "PHAT_ALPHA_SAME": 0.14768044281339848,
    "DIR_IMPACT_MULT": 4,
    "GATE_SIGMA_MULT": 6.7367233899854675,
    "MIN_GATE": 0.03055236895744703,
    "TRADE_TOX_BOOST": 0.3050034835515307,
    "STALE_ATTRACT_FRAC": 0.6899010688573033,
    "RET_CAP": 0.02307934747517747,
    "TOX_CAP": 0.20799000239681092,
    "TRADE_RATIO_CAP": 0.18401740086912755,
    "LAMBDA_CAP": 1.1767903030740083,
    "STEP_COUNT_CAP": 64,
    "BASE_FEE": 0.000394715277459523,
    "SIGMA_COEF": 0.20585950793588467,
    "LAMBDA_COEF": 0.00127231278988692,
    "FLOW_SIZE_COEF": 0.5039180701096855,
    "TOX_COEF": 0.030557276512942967,
    "TOX_QUAD_COEF": 1.5845537652051427,
    "ACT_COEF": 8.530156956319056,
    "DIR_COEF": 0.001787397826556854,
    "DIR_TOX_COEF": 0.007644979769584344,
    "STALE_DIR_COEF": 0.7509802734859878,
    "SIGMA_TOX_COEF": 0.06014956771610196,
    "TAIL_KNEE": 0.05332132369252848,
    "TAIL_SLOPE_PROTECT": 0.8584369839037304,
    "TAIL_SLOPE_ATTRACT": 0.8500966921207127,
    "ANCHOR_COUNT_CAP": 715,
    "ANCHOR_LINEAR_K": 0.6779948673968177,
    "ANCHOR_QUADRATIC_K": 1.1108717241619794
   },
   "scaling": {
    "feature_order": [
     "sigmaHat",
     "lambdaHat",
     "sizeHat",
     "toxEma",
     "actEma",
     "direction",
     "(spot-staleHat)/staleHat",
     "signed_trade_ratio",
     "arb_probability",
     "first_in_step",
     "uncertainty",
     "min(silent,8)/8",
     "anchor_mean",
     "raw_size",
     "step_trade_count",
     "prev_bid_fee",
     "prev_ask_fee",
     "steps_since_arb_like_print",
     "router_implied_vanilla_spot_dev"
    ],
    "feature_multipliers": [
     1000.0,
     2.0,
     500.0,
     500.0,
     500.0,
     1.0,
     500.0,
     500.0,
     2.0,
     2.0,
     1000000.0,
     1.0,
     500.0,
     647.054817,
     1.395287,
     220.617911,
     228.148131,
     0.412565,
     274.335453
    ],
    "feature_offsets": [
     -1.0,
     -1.0,
     -1.0,
     -1.0,
     -1.0,
     -1.0,
     0.0,
     0.0,
     -1.0,
     -1.0,
     -1.0,
     0.0,
     -1.0,
     -0.618293,
     -2.037262,
     -0.803057,
     -0.808615,
     -0.823811,
     0.001772
    ],
    "weight_fixed_point_scale": 1000000,
    "input_clip": [
     -4,
     4
    ],
    "hidden_relu_clip": [
     0,
     4
    ],
    "output_unit_fee": 0.0001,
    "note": "unchanged from CG"
   }
  }
 },
 "real_only": [
  {
   "key": "t01",
   "candidate": "CG unchanged (code_sha256 e1ea8fcb…)",
   "seeds": "8580000..8581023",
   "n": "1024",
   "mean": "0.16625697263110506",
   "se": "0.02000975410375256",
   "sum": "170.24713997425158",
   "sumsq": "447.7343909937491"
  },
  {
   "key": "t04",
   "candidate": "CG unchanged (code_sha256 e1ea8fcb…)",
   "seeds": "8592000..8593023",
   "n": "1024",
   "mean": "0.1831314467005002",
   "se": "0.023686881060165477",
   "sum": "187.5266014213122",
   "sumsq": "622.0902736069509"
  },
  {
   "key": "pooled t01+t04 (computed in this merge from the sums above)",
   "seeds": "8580000..8581023 and 8592000..8593023",
   "n": "2048",
   "mean": "0.17469420966580262",
   "se": "0.015501032280757216",
   "sum": "357.7737413955638",
   "sumsq": "1069.8246646007"
  }
 ],
 "notes": [
  "t05 failed at SETUP (uv binary not executable; host reported Darwin arm64): no speed, gates, holdout, real or gas data.",
  "t01 and t04 measured replica speed below 500 markets/s (308.9 and 256.9) and skipped the search; both froze CG unchanged.",
  "Both search curves were still rising at their time limits: t02 every 5-min monitor point increased, ending +2.2437 at 45.35 min; t03 every point increased, ending +2.5846 at 55.3 min."
 ]
}
```