# Root merge 4 of 5: 25 results from batches r5-cg-continue, r5-cg-joint, r5-cg-soup, r6-confirm, r6-j5-joint, r6-j2-joint, r6-cg-joint-t and r7-t03-joint

## How this merge was made

- **Inputs.** Each of the 25 inputs under `.imd/reads/artifacts/` was hashed. All 25 SHA-256 values equal the ones the brief gives:
  - b20-merge-03 `6263d593…`, b20-merge-04 `e07ec300…`, b20-merge-05 `a425df1f…`, b20-merge-06 `0d4d7fa9…`
  - b21-merge-01 `b649dddb…`, b21-merge-02 `3094e2b0…`
  - b22-m1 `8e84cb78…`, b22-m2 `4143cd04…`, b22-m3 `f3a4ccd3…`
  - b23-j5-a `e9e5de43…`, b23-j5-b `dcae08e6…`, b23-j5-c `d9b4e012…`, b23-j2-a `2a17dab9…`, b23-j2-b `2cca2139…`
  - b24-merge-01..05 `ec371799…`, `1c5b4333…`, `f8adae4a…`, `6a51b8e4…`, `744fd70f…`
  - b25-merge-01 `43d165d7…`, b25-merge-02 `58749d22…`
  - b26-merge-01 `7336002b…`, b26-merge-02 `1907fd29…`
  - b27-merge-01 `7d0480b5…`, b27-merge-02 `12fc50cc…`
- **How the figures were copied.** Every figure below is copied from the input named in brackets after it, for example [b26-merge-01 t03]. Nothing was re-run: no setup, replica, EVM or real engine.
  - Statements marked **(merge inference)** are my own reading across inputs. They are not measurements.
  - b26-merge-01/02 and b27-merge-01/02 were read directly, and their JSON was parsed.
  - b20–b25 were read through structured digests that copy figures exactly. About twenty headline figures were then checked verbatim against the input files with grep.
- **What "real" means.** Unless noted, "real" is the amm_sim_rs engine, with each strategy run alone against the 30 bps VanillaStrategy. Results are paired differences: mean_diff, se, n. "Holdout" is the replica block, seeds 40000000–40008191, n 8192.
- **Strategy names used.**
  - **HA** is the brief's base, code `99e3a2ec…`.
  - **CG** is a 19-8-2 learned head on T paths, code `e1ea8fcb105e00754d17061efbe4b299a634b0fcc1b5c293d65b631654b7204b`.
  - **IB** has code `301941829a44bb60977bb28181a5e4bfa6544f3696763f987054789b2611d686`.
  - **J2** has code `4da9e3ae826b3b6f9170948bd2c2c0a494e78521414341656425b76e71c0f964`.
  - **J5** has code `dd53db1fbe42628e040c6c7b7a82689867ab601263c5e740506299a3a186b38c`.
  - **T3** is the frozen t03 policy, code `cb6d945511e1e10af78baa64c2a6cb25806f0998403ce6440654bfc3c70092d9` [b26-merge-01 t03]. b27's items use the same code hash as their base [b27-merge-01 rows j001, j004–j007, j009, j010; b27-merge-02 j012–j020].

## Headline findings

1. **The strongest policy in this group is the T3 line: CG's head searched jointly with T's 41 constants.**
   - **t03 search** [b26-merge-01 t03]: 1598 generations in 55.33 min at 1358.35 markets/s.
     - Replica holdout +2.533096306362695 se 0.01349951591946062 over T, and +0.5811790141690515 se 0.008338720514523593 over CG.
     - Real vs IB +0.8189009411467073 se 0.059561465105944955 (n 1024, seeds 8588000..8589023). One seed, 8588481, gives +55.66. Without it t03 reports +0.765 se 0.026; the headline keeps every seed.
     - afterSwap peak 235,575, 0 OOG.
   - **T3 measured directly against HA on fresh seeds** [b27-merge-01 real_only; b27-merge-02 real_only]: every complete block of n 1024 lies between **+1.391 and +1.480**:

     | item | T3−HA (n 1024) | se | seeds |
     |---|---|---|---|
     | j001 | +1.45048348111568 | 0.028873752433627383 | 8700000–8701023 |
     | j004 | +1.4362048955996398 | 0.027602045062884637 | 8712000–8713023 |
     | j005 | +1.4797368158721356 | 0.029505186507147115 | 8716000–8717023 |
     | j006 | +1.407090372492083 | 0.02816919616327196 | 8720000–8721023 |
     | j007 | +1.428108981774716 | 0.026646742758385772 | 8724000–8725023 |
     | j010 | +1.3912012819315906 | 0.027478565453361074 | 8736000–8737023 |
     | j012 | +1.422506343834193 | 0.027643194836678354 | 8744000–8745023 |
     | j013 | +1.4157425515465254 | 0.028080851146871755 | 8748000–8749023 |
     | j016 | +1.457630693249603 | 0.035181489120951086 | 8760000–8761023 |
     | j017 | +1.39808719639552 | 0.030798113181104068 | 8764000–8765023 |
     | j018 | +1.4290625806545059 | 0.02862930205026834 | 8768000–8769023 |
     | j019 | +1.4038528303191982 | 0.028410793125475006 | 8772000–8773023 |
     | j020 | +1.3987860088194743 | 0.02801705702712375 | 8776000–8777023 |

     - j004 and j005 are qualified because seeds were replayed.
     - Partial blocks: j009 n 376, +1.4081816527821103 se 0.04737533056303693; j014 n 110, +1.3296317787617005 se 0.07662170690252351 (invalid); j015 n 768, +1.388825437661555 se 0.03288134555686133 (invalid).
     - T3 afterSwap peaks run from 235,417 to 238,344, with 0 OOG everywhere.
   - **For comparison**, the brief puts HA at +1.257 over T, while T3's replica holdout is +2.533 over T. **(merge inference)** That fits a real T3−HA of about +1.3 to +1.4. No input measured T3 against T on the real engine.

2. **Training from T3 gave a further real gain of about +0.10 to +0.15** [b27-merge-01, b27-merge-02]. Each search was confirmed on two fresh blocks of n 1024, as candidate − T3:

   | item | gens / min / markets/s | holdout over T (se) | holdout over T3 (se) | real block 1 | real block 2 | afterSwap peak |
   |---|---|---|---|---|---|---|
   | j002 (selected by b27-merge-01) | 634 / 55.09 / 1510.38 | 2.661187927824997 (0.013621024005457819) | 0.12809162146230213 (0.005058444194093168) | 0.09977518363328142 se 0.013348523757600515 (8704000–8705023) | 0.13130668170728926 se 0.01507168112300746 (8706000–8707023) | 238228 |
   | j003 | 708 / 56.00 / 1429.90 | 2.6609045115671237 (0.013979521862778801) | 0.127808205204429 (0.005198235526106936) | **0.15194192934876266** se 0.015571034749698908 (8708000–8709023) | **0.14804815060852383** se 0.014164199335487463 (8710000–8711023) | 237620 |
   | j008 | 1581 / 48.48 / 2601.39 | 2.6552981202581853 (0.013812575327513585) | 0.12220181389549067 (0.00539708690596292) | 0.14366070182899537 se 0.016104754519997037 (8728000–8729023) | 0.1456780978790791 se 0.015510616631913937 (8730000–8731023) | 238108 |
   | j011 (best in b27-merge-02) | 479 / 57.15 / 984.47 | 2.585427599128418 (0.013773366131475423) | 0.052331292765723555 (0.0040791354368885134) | 0.05209238533409244 se 0.012078705135008127 (8740000–8741023) | 0.050786838435640025 se 0.011025002516698666 (8742000–8743023) | 237733 |

   - **Code hashes:**
     - j002 `5758f1fbbdc71c8f56469d89d9ea183f282b83456c805382f6ab7a2f63c89d8c`
     - j003 `9386b3d1b0e09fc376db314a9105a8facdf8646b5c19723702635fbb8408266d`
     - j008 `1d97c3cabc56293ad49459c3513a5739dacce604de2270952517c5aabfb365ff`
     - j011 `67f6e1a239b0896dafc9ff26415c7677b1c53de3f4d97e4d321fc7291afb63bf`
   - **Selection.** b27-merge-01 picks j002 by highest holdout. It says this "is a selection by point estimate, not evidence that j002 beats j003". **(merge inference)** On real blocks, j003 and j008 are the two highest, and j002's first block is the lowest of the three. j002, j003 and j008 have overlapping holdouts, and no candidate-vs-candidate test exists.
   - **Gas and validity.** All four have 0 OOG and are valid=true.

3. **Distance to the goal is still large.**
   - **What the inputs say.** No input claims or measures a top-10 placement. Every merge states so [all b20–b27 merges].
   - **Gap estimate (merge inference)**, adding figures from different seed sets, which b27-merge-01 explicitly declines to do: T3 ≈ +1.39 to +1.48 over HA, plus about +0.15 for j003/j008, gives about **+1.55 to +1.6 over HA**. That is still about **1.5 short** of the brief's ~+3.1 over HA for the top 10.
   - **b26-merge-01's own estimate (inference)**: "a gap of about 1.8 remains" on the replica, from T3 at +2.53 over T against a target of about +4.36 over T.

4. **Long search beats short search, and gains track generations, not wall-clock minutes:**

   | lineage | short search | long search | source |
   |---|---|---|---|
   | CG head | ~0.003–0.021 over CG at ≤ 852 gens (k11–k25, k27, k28) | +0.07735360415319258 replica / +0.09340801795559883 real over CG at 1762 gens (k26) | b20-merge-03…06 |
   | J5 joint | ≤ 0.0305 at 62–276 gens | +0.11621016505621401 replica over J5 at 1225 gens (k15) | b24-merge-01…05 |
   | CG head + T constants | +0.241 over CG at 92 gens (t02); +0.2495 at 248 gens (t07) | +0.581 at 1598 gens (t03) | b26-merge-01/02 |
   | T3 | +0.052 over T3 at 479 gens (j011, population 32, 256 markets/gen) | +0.122–0.128 at 634–1581 gens (j002, j003, j008) | b27 |

   Throughput decided whether any of this happened. Fallback items, at < 200 or < 500 markets/s depending on batch, ran no search and only re-measured the unchanged base.

5. **The replica predicted real gains well wherever both were measured.** All pairs below are (merge inference) put side by side from the cited figures. Seeds differ between replica and real. Pairs are replica holdout gain over base → real gain over base:

   | candidate | replica | real | source |
   |---|---|---|---|
   | k26 | +0.0774 | +0.0934 | b20-merge-06 |
   | q04 | +0.1581 | +0.1633 | b21-merge-01 |
   | q10 | +0.1761 | +0.1714 | b21-merge-02 |
   | m1 | +0.0084 | +0.0107 | b22-m1 |
   | j002 | +0.128 | +0.100 / +0.131 | b27 |
   | j003 | +0.128 | +0.152 / +0.148 | b27 |
   | j008 | +0.122 | +0.144 / +0.146 | b27 |
   | j011 | +0.052 | +0.052 / +0.051 | b27 |
   | t03 vs CG | +0.581 | ≈ +0.64 | b26-merge-01's own inference via IB |

   The one non-transfer is J2 p04: replica +0.02 → real +0.00285 se 0.0162 vs J2 [b25-merge-01]. It is within noise. This fits the pilot's "within about 0.1".

6. **Gas and safety.**
   - No input reports an OOG call on the real engine.
   - Every candidate afterSwap peak reported is ≤ 245,000. The highest is k35 at 239,076 [b24-merge-04]. afterInitialize peaks of about 243,234–243,237 appear in several items, and that is a separate callback.
   - **Synthetic stress failures.** Many items found a **Solidity Panic 0x12 (division by zero)** revert under unbounded or large-trade synthetic callback stress in CG-, J2-, J5- and T3-family code. A recurring case is seed 310037 at callback 1424 with reserve_x about 0.117–0.119 [b24-merge-01 k05/k10; b25-merge-02 p12/p19; b26-merge-01 t01/t04; b26-merge-02 t09/t10; b25-merge-01].
   - None was seen on the real engine, and none is an OOG. Several items then ran gates A/C on bounded generators, which differ across items.
   - This is a robustness caveat that the write-up should keep.

## Cross-batch real-engine ladder

Each figure is a paired real difference. Seeds differ across rows, so compare rows only roughly.

| comparison | values reported | source |
|---|---|---|
| CG − HA | 0.841315854907559 se 0.022079922132378915 (n 1024) | b20-merge-03 k15 |
| CG − HA | 0.84843 (k17), 0.78296 (k20) | b20-merge-04 |
| CG − HA | 0.78736 (k29, n 1024), 0.82929 (k30, n 1024), 0.8504305867396483 (k26 seeds, n 512) | b20-merge-06 |
| CG − HA | 0.8021924523188793 se 0.0430277742245398 (q01, n 336); 0.8607145007859232 se 0.022870907839139858 (q05) | b21-merge-01 |
| CG − HA | 0.8292160850501822 (q06); 0.8337773834686922 (q07, n 512); 0.8121446972390833 (q09) | b21-merge-02 |
| CG − HA | 0.8457899594142979 se 0.023503383739650566 | b22-m2 |
| CG − HA | 0.8530640014249975 se 0.022365734742669955 | b22-m3 |
| CG − IB | 0.16625697263110506 se 0.02000975410375256 (t01); 0.1831314467005002 se 0.023686881060165477 (t04); pooled n 2048 0.17469420966580262 se 0.015501032280757216 | b26-merge-01 (pool computed by that merge) |
| CG − IB | 0.15796822515594006 se 0.02322446305343043 (t09); 0.17235208388954698 se 0.022915112671975223 (t10) | b26-merge-02 |
| CG − IB | 0.20257227807566258 se 0.029922403213797133 | b20-merge-06 k26 seeds |
| IB − HA | 0.6537484851255795 (j5-a), 0.6664083815241173 (j5-b), 0.6354357019129704 (j5-c), 0.6823391730051074 (j2-a), 0.6645143221886062 (j2-b) | b23 |
| IB − HA | 0.6478583086639857 | b20-merge-06 k26 seeds |
| J2 − IB | 0.4986902269965671 se 0.02642097882399952 (j2-a); 0.5044490667916387 se 0.02043384923317343 (j2-b, n 960, invalid) | b23 |
| J2 − IB | pooled 0.511085525233829 se 0.009912380342386177 (n 7648) | b25-merge-01 |
| J2 − IB | pooled weighted mean 0.5145 (n 8192, including p17/p20) | b25-merge-02 |
| J2 − HA | 1.1810294000016746 (j2-a); 1.171380375175169 (j2-b) | b23 |
| J5 − IB | 0.5588050326637388 se 0.027926702577425504 (j5-a); 0.5512091583742078 se 0.026392635592456706 (j5-b); 0.5996675992133282 se 0.025380482616284075 (j5-c) | b23 |
| J5 − IB | pooled 0.5357 se 0.0092 (n 9152) / valid-only 0.5294 se 0.0113 | b24-merge-01 |
| J5 − IB | pooled 0.5516774482988679 se 0.009866094440459106 (n 9088) | b24-merge-03 |
| J5 − IB | pooled 0.543197 se 0.009448 (n 8032) | b24-merge-05 |
| J5 − HA | 1.2125535177893183 se 0.03069729017656727 (j5-a); 1.217617539898325 se 0.027741015794190308 (j5-b); 1.2351033011262986 se 0.02810132988960955 (j5-c) | b23 |
| J5 − HA | J5 artifact "real vs_HA +1.219 se 0.038 on 512 seeds", quoted from the J5 artifact, not re-measured | b24-merge-05 |
| J5/J2/IB/HA − T | J5 2.351974211923312, IB 1.7931691792595732, HA 1.1394206941339937 (j5-a); J5 2.3511576507443515, IB 1.7999484923701439, HA 1.1335401108460266 (j5-b); J2 2.2759821554094475, IB 1.7772919284128803, HA 1.094952755407773 (j2-a) | b23 |
| T3 − IB | 0.8189009411467073 se 0.059561465105944955 | b26-merge-01 t03 |
| T3 − HA | +1.391 to +1.480 (see finding 1) | b27 |
| best T3 descendants − T3 | +0.10 to +0.15 (see finding 2) | b27 |

**Note on HA − T.** On the real engine, the b23 items measure HA − T at about +1.09 to +1.14. The goal brief states +1.257 se 0.062 on 512 seeds. **(merge inference)** That is a difference of about 0.12–0.16. It may come from seed sets or from the "kit T" arm used in b23. The inputs do not resolve it.

## Per-batch results

### r5-cg-continue: long CMA-ES on CG's 178-weight head (k11–k30) [b20-merge-03…06]

- **Shared reference figures** [b20 digests]:
  - Gate A for CG: max 8.814065310946972e-05 bps.
  - Gate B: replica CG−T 1.920372244773669 se 0.04445924198500627 (n 512, seeds 7232768–7233279) vs a supplied reference of 2.0. The error is 0.0796, under the 0.15 tolerance. This is a single run, copied by every item.
  - Unchanged CG on the holdout: over_T 1.9519172921936434 se 0.011938004553452971.
  - Replica "over_IB" = over_T − 1.7690. That is a supplied reference, not a paired IB run.
- **Best: k26** [b20-merge-06], code `0d64a2a4b7dab21fb7f2c6d47bfa3ee616585dbd0d3c8b01782da386bcee507c`, result `cc4eec1e…b556`.
  - Search: 1762 generations, 57.85 min, 1932.1 markets/s at 16 processes. Frozen at the generation-1710 ES mean.
  - Curve "still trending up when stopped".
  - Holdout over_CG 0.07735360415319258 se 0.004116200860154031.
  - Real vs CG 0.09340801795559883 se 0.015501325635931389 (n 512, seeds 8100000–8100511).
  - Real vs HA 0.9438386046952472 se 0.03338080992274326.
  - Real vs IB 0.2707710634788416 se 0.018932727303537015 (n 1024).
  - afterSwap 235508, 0 OOG. Gate C 0.00016010090368029362.
- **Other searches gained little** (real vs CG, n 512):
  - k11 +0.014165628921265716 se 0.012998753166375516 (852 gens; flat after generation 730) [b20-merge-03]
  - k12 +0.016189; k13 −0.020488 (outlier seed 8048118 at −11.03) [b20-merge-03]
  - k16 +0.0019040595538156735 se 0.013461735525514261 [b20-merge-04]
  - k18 +0.017109 (candidate run incomplete) [b20-merge-04]
  - k19 +0.010448 [b20-merge-04]
  - k22 +0.008517 [b20-merge-05]
  - k23 −0.002257 [b20-merge-05]
  - k24 +0.04582070456441234 se 0.013295072596870941 (501 gens; b20-merge-05's pick) [b20-merge-05]
  - k25 +0.016985 [b20-merge-05]
  - k27 +0.004196; its notes say "CG looks close to a local optimum" [b20-merge-06]
  - k28 +0.006311; "not a still-rising curve" [b20-merge-06]
- **Items that kept or re-measured CG without searching:**
  - k15, k21, k29, k30: speed under 200 markets/s.
  - k17 and k20: kept CG after search.
- **Blocked:** k14 at setup line 6, `uv venv`. The error was "cannot execute binary file", status 126 [b20-merge-03].
- **Replica speed differs about 10× across machines**: 85.7 (k15) to 4260.5 (k22) markets/s [b20-merge-03, -05].
- **Protocol notes** [b20-merge-06]: k30's CG run_batch took 623.24 s, which exceeds the 8-minute command cap, and compliance is unresolved.

### r5-cg-joint: CG head plus 33 CA/T constants (q01–q10) [b21-merge-01/02]

- **Best: q10** [b21-merge-02], code `fc5472952e3eaea3d442c75d78e45bc75fb51765b2bb45e0dcbde2afbc0a1b36`.
  - Search: constants only (33 dims), CG head frozen, 374 generations / 77.2 min.
  - Holdout over_CG 0.1760548182543897 se 0.007961112369178684; over_T 2.127972110448033.
  - Real vs CG 0.17140910861079756 se 0.029715347872309228 (n 1024, seeds 8156000..8157023). The worst seed is 8156150 at −6.79.
  - afterSwap 235472.
  - Protocol deviation: first-segment scores were computed on raw genotypes and then discarded.
- **Runner-up: q04** [b21-merge-01], code `098fa26f6cc37123f1f9b05429bb8b32e6b3a622e52051c251604de18afa6b12`.
  - Search: 563 generations.
  - Holdout over_CG 0.15806236600512724 se 0.007850588212593258.
  - Real vs CG 0.16331755441657997 se 0.031047175666201313 (n 1024). afterSwap 237625.
- **Other searches:**
  - q02: holdout 0.1394, real 0.1667 (n 512) [b21-merge-01]
  - q03: holdout 0.1553, real 0.1487 se 0.0213 (n 1024) [b21-merge-01]
  - q08: holdout 0.1376, real 0.1352 [b21-merge-02]
- **Line scan from CG toward T constants** [b21-merge-01/02]: the monitor gain over CG peaks at t ≈ 0.75, at 0.1092 to 0.1093.
- **Real-only items** (no search, CG only): q01, q05, q06, q07, q09.
- **No vs-HA measurement** for any trained q candidate.

### r5-cg-soup: model soup of CG/S1/S2 heads (m1–m3) [b22-m1, b22-m2, b22-m3]

- **m1** [b22-m1]:
  - Every soup mixing point was negative. For example, the CG+S1 average at t=1 gave −0.0252, so the start stayed at CG.
  - Then a head CMA ran: 331 gens, 76.75 min, 552.2 markets/s.
  - Candidate `e874a1b3bb3ea5b5e148d6ce19ff2d38d7dd42447ceaa19b12244a5b5821e5dc`: holdout over_CG 0.008362494351767946 se 0.0029317836479103533; real vs CG 0.010691210244455407 se 0.007703009891722952 (n 1024).
  - In its own words: "does not establish an improvement over CG".
- **m2** [b22-m2]:
  - Speed 155.26 markets/s, so it fell back: no soup, no search, CG unchanged.
  - Real CG−HA 0.8457899594142979 se 0.023503383739650566 (n 1024, seeds 8164000..8165023).
  - Gas: CG 235582, HA 231572, 0 OOG.
  - **Source flag**: see the source section below.
- **m3** [b22-m3]:
  - Speed 118.3 markets/s, about 1.2 effective cores, so no search.
  - Real CG−HA 0.8530640014249975 se 0.022365734742669955 (n 1024).
- **Artifacts used**:
  - kit `0d8194e9…1786`
  - CG `ebbd90c6…f399979`
  - S1 `aaa42d3f…`
  - S2 `a97fa964…`
  - HA `2801efc8…`

### r6-confirm: fresh-seed real checks of unchanged J5 and J2 (j5-a/b/c, j2-a/b) [b23]

All five items are real-only on seeds from 7800000 to 7841023, with no replica or gate work. Gas had 0 OOG and 0 revert in every arm. Retail/arb figures are Y-volume deltas, because the engine has no edge split.

- **J5 is ahead of J2 in every block.**
  - J5−IB: +0.5588 / +0.5512 / +0.5997.
  - J2−IB: +0.4987 / +0.5044.
  - J5−HA: +1.2126 / +1.2176 / +1.2351.
  - J2−HA: +1.1810 / +1.1714.
- **Retail volume vs IB goes in opposite directions:** J5 +62.9 to +70.2, J2 −86.1 and −95.1.
- **afterSwap peaks:** J5 237600–238232, J2 237046, IB up to 237137, HA up to 233730, T up to 225591.
- **j2-b is invalid**: "J2 seeds 7840384..7840447 were interrupted by a 70-second timeout before results returned; never rerun", so n is 960. The other four are valid.

### r6-j5-joint: joint search from J5, 178 head weights plus 41 constants (k01–k50) [b24-merge-01…05]

All real figures in this batch are candidate − IB. Every item that ran it reports the same J5 holdout over T, 2.307544068661547 se 0.013528690579041706, and Gate B replica J5−IA 0.8655072079761197 vs 0.876.

- **Best: k15** [b24-merge-02], code `adca0a458ebc1764c181589dc94fc3803a4b916d3d1262ba323b5067486a1ec5`, result `0a81ce52…248d`.
  - Search: 1225 generations, 75.03 min, 1330.88 markets/s.
  - Holdout over T 2.4237542337177613 se 0.013706147517921614; over J5 0.11621016505621401 se 0.005561273717733301.
  - Real vs IB 0.6680468841626501 se 0.025156192821243612 (n 1024, seeds 8356000-8357023).
  - Edge subset (n 640): retail −0.600, arb +1.260.
  - afterSwap 237601, 0 OOG. Gates: A max 5.673911469025565e-05, C 7.246861740035915e-05.
  - Curve still rising at the end, at 2.4701.
  - Caveat: IB chunk 9 was replayed after a residual check. Not measured vs HA.
  - **(merge inference)** J5 is about +1.22 over HA and k15 adds about +0.12 over J5, which puts k15 at roughly +1.3 to +1.35 over HA. That is below T3.
- **Other searches:**

  | item | holdout over J5 (se) | real vs IB | gens | source |
  |---|---|---|---|---|
  | k02 | +0.019836 (0.004635) | 0.582983 se 0.033611 | 96 | b24-merge-01 |
  | k11 | +0.021728617 | 0.58039 | 276 | b24-merge-02 |
  | k13 | +0.030488887 | 0.61502 | 79 | b24-merge-02 |
  | k14 | froze J5 | — | — | b24-merge-02 |
  | k24 | −0.0007686244609929754 | 0.55090 | 62 | b24-merge-03 |
  | k44 | froze J5 | — | 47 | b24-merge-05 |
  | k47 | −0.008517114560603077 | 0.557461731467046 | 175 | b24-merge-05 |

  k47's search ran despite a first speed reading of 487.3, below the threshold [b24-merge-05].
- **All other items were real-only re-measures of J5.** Pooled J5−IB per merge is in the ladder table.
- **Items marked invalid:**
  - k05: IB seeds lost [b24-merge-01]
  - k06: smoke-run seed reuse [b24-merge-01]
  - k08: first 64 seeds rerun [b24-merge-01]
  - k19: launcher took 715.47 s, over the 8-minute cap [b24-merge-02]
  - k21: two chunks lost [b24-merge-03]
  - k22: rerun after exporter failure [b24-merge-03]
  - k27 and k28 also reran batches but kept valid=true. b24-merge-03 flags this as inconsistent with k22.
- **Partial items:** k48 (n 352) and k50 (n 512) [b24-merge-05].
- **Blocked:**
  - k34 [b24-merge-04]: "A path (…amm_sim_rs…whl) dependency is incompatible with the current platform … you're on macOS (`macosx_27_0_arm64`)"
  - k40 [b24-merge-04]: "scratch/bin/uv: cannot execute binary file", exit 126
- **Other notes:**
  - k16 reports that HA `99e3a2ec` returned `{"error":"unknown_artifact"}`. b24-merge-02 explains that this used the HA *code* hash as an artifact id. The artifact id is `2801efc8…`, so this does not show HA is unavailable.
  - k18's stored Gate B abs_error, 0.010741986066346687, is inconsistent with mean − 0.876 = −0.010492792023880315 [b24-merge-02].

### r6-j2-joint: joint search from J2 (p01–p20) [b25-merge-01/02]

- **Most items re-measured unchanged J2:** 8 of p01–p10 and most of p11–p20.
- **p04** [b25-merge-01], code `80a0a3a2eacdac5b6e9371a44313f6da526ca87ba161a5fa170006425e8c7876`.
  - Search: 89 gens, 50.56 min.
  - Holdout over_T 2.2952935121348492.
  - Real vs J2 directly: +0.00285 se 0.0162 (n 1024), so "not established on the real engine".
  - Curve rising, 2.2750 → 2.3293.
- **p20** [b25-merge-02], code `ec2aba8a07d4b12d0170047489ce72ed4fedd96e78f72b3d311eec6adae45557`.
  - Search: 381 gens.
  - Holdout over_T 2.2927866507020807, over J2 0.01828408942738982.
  - Real vs IB 0.5056683746643662 se 0.020930570866425847, which sits inside the unchanged-J2 range of 0.448–0.534. Curve rising.
- **p17** [b25-merge-02]: real vs IB 0.5675951900793546 se 0.024229478582731372, the highest in the batch; holdout over J2 0.01577. Curve rising.
- **p09** [b25-merge-01]: searched 308 gens, holdout over J2 0.027216, real vs IB 0.553899. It is **invalid**: gates A/C failed, with "max 2.70 bps for gate C, plus an EVM div-by-zero revert".
- **Other invalid items:** p02 (redundant IB chunk), p03 (n 960), p12 (64 seeds replayed), p14 (nothing run).
- **p16 is blocked**: "downloads from api.imd.fun were denied on its machine".
- **Gate B in this batch** [b25-merge-02]: replica 0.4948 se 0.0289 vs reference 0.578.

### r6-cg-joint-t: CG head plus T's 41 constants, the origin of T3 (t01–t10) [b26-merge-01/02]

- **t03**: the best result, detailed in finding 1. Run in 9 segments under 8 min. Population 16, CMA seed 773.
  - Curve "every point increased (+2.002 → +2.585 over 55.3 min, sigma 0.271)".
  - Gates A/C 2.9e-5 to 5.7e-5 bps.
  - Y-volume vs IB: retail +281.9, arb −51.8.
- **t02**: 92 gens. Holdout over_T 2.1931730568964474; over CG 0.24125576470280372 se 0.007037874154145838. Real vs IB 0.3948459133695228 se 0.02374619862551985. Code `c9332fb3…515c`. Curve rising.
- **t07** [b26-merge-02]: 248 gens / 70.12 min. Holdout over_T 2.201379289577117 se 0.013026305534368038; over CG 0.24950579460119215. Real vs IB 0.41779049090156767 se 0.03659417300518825, with worst seed 8604910 at −26.720525170836822. afterSwap 238885. Code `42df194bdb72d1c7484e3adaa25188e3d87fd85d6855d03a068919fdefe554c3`. Slither found 4 Medium and 84 Informational issues, none High.
- **Real-only CG:** t01, t04, t09, t10. Values are in the ladder table.
- **Unusable items:**
  - t06 is invalid: 45.49 markets/s, and the Forge creation bytecode exceeded the EVM size limit, so it had no real run.
  - t05 is blocked: `uv venv` exit 126 on a Darwin arm64 host.
  - t08 is blocked: the manylinux wheel is incompatible with the Apple Silicon host.
- **Direction of the two batch merges.** b26-merge-01 picks t03 and b26-merge-02 picks t07. They covered different items, and t03 dominates t07 on every metric.

### r7-t03-joint: continue from T3 (j001–j020) [b27-merge-01/02]

- **Searches:** j002, j003, j008 and j011, detailed in finding 2. j011 used population 32, ES seed 811, sigma0 0.5 and 256 markets/gen.
- **Real-only:** the other 16 items, all T3 − HA, listed in finding 1.
- **Retail/arb edge reconstructions** (analytic, not engine output):
  - j005, T3−HA: retail +0.8116174665395955 se 0.0305, arb +0.6672241848415338 se 0.0155, n 1016 [b27-merge-01].
  - j011, candidate−T3: retail +0.2457 / +0.2262, arb −0.1936 / −0.1754 [b27-merge-02].
- **Qualifications:**
  - j004 replayed its first 64 seeds.
  - j005 is protocol-invalid: an HA chunk was replayed. An unreplayed subset gives n 896, 1.4805939994995803 se 0.032424159389675515.
  - j009 completed only 376 of 1024 seeds.
  - j008 corrected its T baseline after segment 1 [b27-merge-01].
  - j014 is invalid: gate A ran 16 of 50 sequences and the real run covered 110 of 1024 pairs.
  - j015 is invalid: 768 of 1024 pairs, a T-proxy speed measurement and inherited gates [b27-merge-02].
- **Gate B caveat.** Gate B in every T3 item is the inherited CG−T calibration (1.920 vs 2.00), "not a new T3-minus-T real calibration" [b27-merge-01, -02]. The brief's gate "the replica reproduces its real gain over T within 0.15" has therefore **not been shown for T3 itself** in these inputs. The candidate-minus-T3 transfer figures in finding 5 are the nearest evidence.

## Where results disagree

- **Choice of best within a batch:**
  - b27-merge-01 picks j002 by holdout, but j003 and j008 have higher real blocks.
  - b26-merge-02's best, t07, is far below b26-merge-01's t03.
  - b20-merge-05 picks k24, which is below b20-merge-06's k26.
- **Replica speed on the "same" code** ranges from 15.5 (j014) to 4260.5 (k22) markets/s, depending on host load and worker count. It decided search vs fallback in most batches [b20-merge-03, b26-merge-01, b27-merge-02].
- **Validity flags are inconsistent for replays.**
  - Some items set valid=false for re-running seeds after an exporter failure: k22, p12, k08, j005.
  - Others with the same kind of re-execution kept valid=true: k27, k28, t01, j004, k15's IB chunk [b24-merge-03; b26-merge-01; b27-merge-01].
- **"Holdout se" means different things across items.** It is paired candidate−base in some and candidate−T in others [b20-merge-03, -06; b21-merge-02]. Those se values are not poolable.
- **HA − T on the real engine:** about +1.09 to +1.14 in b23, against the brief's +1.257 (see the ladder note).
- **T3 − IB:** a single block gave 0.819 with an se of 0.060, inflated by one seed at +55.66 [b26-merge-01]. **(merge inference)** T3−HA of about 1.43 minus IB−HA of about 0.66 gives about 0.77, consistent with t03's outlier-excluded +0.765.
- **Upward curves vs flat curves.**
  - Still rising at the time limit: k26, k15, t02, t03, t07, p04, p17, p20 [b20-merge-06, b24-merge-02, b26-merge-01/02, b25].
  - Flat or plateaued: k11, k27, k28, k24, k14, k44, p19 [b20-merge-03, -06; b24-merge-02, -03, -05; b25-merge-02].
  - The curve shape for j002, j003 and j008 at the end is not stated in the inputs I read. Their last points are j002 2.711710557456891 at 55.06 min and j003 2.694345615085421 at 56.00 min [b27-merge-01].

## Sources that failed recheck, and their weight

- **Named by this brief:**
  - **r5-cg-soup/m2 (2 sources)** [b22-m2]. m2 lists eight artifact URLs, all marked verified: kit `0d8194e9…`, CG `ebbd90c6…`, S1 `aaa42d3f…`, S2 `a97fa964…`, HA `2801efc8…`, t_upstream `6ac5e013…`, replica_upstream `e8a7a2af…` and port_upstream `d845a4b9…`. It also claims "source_quotes_verified: 14", and it uses the supplied references +2.00 (Gate B) and 1.7690 (IB).
    - The file does not say which two failed. They are probably among these provenance URLs or quoted references (the digest's inference).
    - What m2 measured locally, CG−HA 0.8458 and the gas figures, does not rest on those pages.
    - Its Gate B pass and its over_IB figure do rest on the quoted references, so weigh those as unconfirmed.
    - m2 produced no new policy, so no selection depends on it.
  - **r6-confirm j5-a, j5-b, j5-c, j2-a, j2-b (1 source each)** [b23]. The five files contain no URLs. They rely only on downloaded artifacts: the J5/J2/IB/HA code and weights, the kit T code, and the wheel and commit pin.
    - The failed sources most likely concern artifact provenance, meaning which upstream J5/J2/IB they are. They do not concern the locally measured paired differences.
    - Supporting evidence: the code hashes these files report (J5 `dd53db1f…`, J2 `4da9e3ae…`, IB `30194182…`, HA `99e3a2ec…`) match the base hashes used in b24, b25 and b20-merge-06. Treat the J5/J2 identities as internally consistent but not externally re-confirmed.
- **Flags recorded inside the merged inputs themselves** (the exact failed URLs were never identified):
  - b20: k11 (4), k12 (3), k13 (3), k15 (1), k16–k19 (1 each), k20 (3); every k21–k25 citation; k26 (2), k27–k30 (1 each).
  - b21: q03 (1), q05 (2), q06 (1), q09 (2).
  - b24: most k01–k50 items, 1–4 each. k34 and k40 had none.
  - b25: p02–p07, p09–p13, p17–p20.
  - b26: t01 (3), t02 (2), t04 (3), t06 (1), t07 (2), t09 (1), t10 (3).
  - b27: j001, j002, j003 (1 each), j005 (2), j006 (1), j007 (3), j011 (2), j012 (1), j013 (2), j015, j016, j018, j019, j020 (1 each).
- **The merges that discuss the flags (b20-merge-03, b25-merge-01, b26-merge-01, b26-merge-02, b27-merge-01, b27-merge-02) agree on their scope.** The flagged sources back input provenance and supplied references: Gate B's 2.0 / 0.876 / 0.578, the IB offset 1.7690, and J5's quoted +1.219 vs HA. They do not back the locally measured replica, gate, gas or real figures.
- **The flags still matter in two places:**
  - (a) Every Gate B pass is conditional on an unconfirmed reference.
  - (b) The selected bests (j002 flagged 1, k15 flagged 2, k26 flagged 2, t07 flagged 2) are provisional on worker reports. t03 cites no web sources [b26-merge-01], and j008 is unflagged [b27-merge-01].

## Failures and blocked items

| item | what happened | source |
|---|---|---|
| k14 | Setup line 6 `uv venv`: "cannot execute binary file", status 126 | b20-merge-03 |
| k40 | Setup line 6 `uv venv`: "cannot execute binary file", status 126 | b24-merge-04 |
| t05 | Setup line 6 `uv venv`: status 126, Darwin arm64 host | b26-merge-01 |
| k34 | Wheel incompatible with macOS arm64 | b24-merge-04 |
| t08 | Wheel incompatible with Apple Silicon | b26-merge-02 |
| p16 | Downloads from api.imd.fun denied | b25-merge-02 |
| t06 | Bytecode over the EVM size limit, so no real run | b26-merge-02 |
| p14 | Nothing run | b25-merge-02 |

Many more items were incomplete or replayed seeds. Each is listed in its batch section above.

## Not found in the inputs

- A real-engine T3 − T measurement, or a T3-specific Gate B calibration.
- Any real vs-HA figure for the trained q/k-J5/p/t/j candidates. The only exceptions are the k-CG items in b20-merge-03/04/06.
- Paired candidate-vs-candidate tests among j002, j003 and j008.
- Any leaderboard submission or rank.
- Which exact URLs failed recheck, for any flagged item.
- Full code payloads for the bests. Those are embedded in b26-merge-01 (t03), b26-merge-02 (t07), b27-merge-01 (j002) and b27-merge-02 (j011), and are not copied here. Retrieve them from those inputs by code_sha256.

## What the results point to (their own notes, plus labelled inference)

- **The results' own notes:**
  - Long training from saved weights is what produced every material gain: k26, k15, t03, and the j002/j003/j008 line.
  - b26-merge-01 says "Longer training from t03's frozen params is the obvious next step", and b27 did that.
  - Machines under about 500 markets/s contributed only re-measurements. b22-m3 notes that the next attempt "needs a host with more free cores".
- **(merge inference)** The next step would be continued long training from j003 or j008 (or j002), with real confirmation against T3 and a real T3−T calibration. The remaining gap to about +3.1 over HA is roughly +1.5.