# Root merge 5 of 5: batch r7-t03-joint (merges 03–15, items j021–j150) and batch r7-t03-confirm (c1–c4) This merge combines 17 inputs: 13 earlier merges of batch **r7-t03-joint** (these continue fee policy T3 by CMA-ES on the replica, or run T3 against HA when the replica is too slow) and the 4 item results of batch **r7-t03-confirm** (T3, IB and HA, each run alone on 1,024 fresh real-engine seeds). Every input was read from `.imd/reads/artifacts/`, and its SHA-256 matched the id in the brief (checked here with `sha256sum`; listed under Sources). Labels like [m03] or [c2] point to that list. Every figure has its source next to it. Figures marked **(computed here)** were derived in this merge only from the inputs' own `n`/`sum`/`sumsq`. No engine, replica or training was run for this merge. I did not check any leaderboard placement. Terms. **T** = RS2 plus the toxicity anchor. **HA** = artifact `2801efc8…`, code `99e3a2ec…`. **T3** = artifact `2333ecf28d4356bd232b85c1e55842e94e1ada05e1cd067830b5625dfdbc5336`, code `cb6d945511e1e10af78baa64c2a6cb25806f0998403ce6440654bfc3c70092d9`. It is the policy all r7-t03 items start from. **IB** = artifact `afd5183e9ff3235334e041c00b0cd19079599a2c6464db95f8c1586712105557`, code `301941829a44bb60977bb28181a5e4bfa6544f3696763f987054789b2611d686` [c1–c4]. The inputs never say what IB is or how it was built, so I don't guess. `over_base` means over T3. Real-engine differences are paired by seed, with each strategy run alone against the 30 bps VanillaStrategy through `amm_sim_rs.run_batch(X, VanillaStrategy, configs, 8)`. ## 1. Headline 1. **No result reaches the top-10 target.** The goal needs about +3.1 over HA (goal statement). The best measured candidate, **j024**, is +0.1910 se 0.0106 over T3 on 2,048 fresh real seeds [m03; pooled in m03 and here]. T3 is about +1.42 over HA (section 3). That puts j024 at roughly **+1.61 over HA**, still about 1.5 short. This is an **inference across different seed sets, not a measured pair**: no item ran any candidate directly against HA [m03, m05, m07, m12, m14]. 2. **T3 − HA is well established on the real engine.** - Across j021–j150: 88 primary blocks from 85 real-only items, 85,288 paired seeds, all in disjoint seed ranges. Pooled: **+1.4198 se 0.0032 (computed here)**. Excluding items that flagged themselves invalid: +1.4190 se 0.0034 (77,832 pairs). - Means of blocks with ≥448 seeds run from +1.3536 to +1.5004. Cochran's Q is 94.5 on 87 df, so the blocks agree with each other (computed here). - The four confirm items find +1.4477 se 0.0156 (3,776 pairs, computed here). That is about 1.8 combined se above the j-item pool: a mild difference, not a contradiction. 3. **The confirm items (c1–c4) measure T3 against IB and HA on the same seeds.** - T3 − IB is **+0.8054 se 0.0144** (3,776 pairs). - IB − HA is **+0.6428 se 0.0119** (3,968 pairs). All pools in this item are computed here. - The two steps add up to about T3 − HA. Three of the four confirm items mark themselves `valid:false`. Their data are lost or incomplete, not bad (section 2). 4. **Continuing T3 by CMA-ES mostly gives small gains.** - 37 items were in search mode, and 33 of them report `valid:true`. Leaving out j050, whose engine was not the pinned wheel, 32 produced a frozen candidate that passed its own gates and was checked on the real engine. - Most gained only +0.003 to +0.06 over T3. - Four gained clearly more: j024 +0.191, j116 +0.149, j030 +0.123 (one block) and j021 +0.093 (computed here from block sums; sources m03 and m12). These are also the runs with the most generations (889, 1,585, 895 and 507). The fast pools behind j024 (3,780 markets/s) and j116 (4,790 markets/s) made that possible. This is an association across items, not a controlled comparison: m12 notes it cannot separate population size from generation count. 5. **The replica tracks real gains.** For every valid search with real blocks, holdout over T3 and pooled real gain over T3 differ by less than 0.04 (section 5). That is well inside the 0.15 tolerance. The one exception is j050, whose "real engine" was not the pinned wheel [m05]. 6. **The inputs disagree on whether training curves were still rising.** - m05, m06, m07 and m14 report flat curves for runs at 512–1,024 markets per generation and under about 1,500 markets/s. - m12 reports that j116 was "still rising" at 45 min: +2.585 → +2.730 on the monitor over T. - j024's last curve point is +2.7615 over T at 45.28 min [m03]. Its curve shape is not given in m03. - So the pilot's "still rising" premise holds for the long, fast runs and fails for the short or slow ones. 7. **No OOG anywhere.** Every candidate's and T3's afterSwap peak was under 245,000. The highest are T3 238,523 (j074) [m08] and candidate 238,473 (j076) [m08]. T3's afterInitialize peak is 243,234 [c1–c4; m03 j026]. That is under 245,000 but close to it. The afterSwap limit is the only one the brief states. ## 2. Confirm batch r7-t03-confirm (c1–c4): T3, IB and HA on fresh seeds Each item reports doing the following: - ran the goal SETUP at the pinned checkout `6b6921f8…`; - verified the wheel `aedb6828…` and the hashes of T3, IB, HA and KIT (`0d8194e9…`); - changed no code or weights; - ran each arm alone against the 30 bps Vanilla in 64-seed chunks, with configs drawn in MatchRunner `_build_configs` order (seed = market seed, 10,000 steps). Differences are first arm minus second arm. The se is the sample sd divided by √n. | item | seeds | T3−IB n / mean / se / wins | T3−HA n / mean / se / wins | IB−HA n / mean / se / wins | T3 afterSwap peak / OOG (coverage) | valid | |---|---|---|---|---|---|---| | c1 | 9500000–9501023 | 960 / +0.7844 / 0.0280 / 815 | 960 / +1.4023 / 0.0291 / 924 | 1024 / +0.6230 / 0.0250 / 858 | 237,689 / 0 (960 seeds) | false | | c2 | 9510000–9511023 | 960 / +0.8207 / 0.0289 / 833 | 960 / +1.4734 / 0.0298 / 926 | 1024 / +0.6588 / 0.0227 / 872 | 237,653 / 0 (all 1024; `T3_gas_pass` true) | false | | c3 | 9520000–9521023 | 960 / +0.8196 / 0.0308 / 848 | 960 / +1.4465 / 0.0358 / 925 | 1024 / +0.6194 / 0.0221 / 856 | 237,600 / 0 (960 seeds) | false | | c4 | 9530000–9530895 | 896 / +0.7962 / 0.0273 / 788 | 896 / +1.4701 / 0.0296 / 872 | 896 / +0.6739 / 0.0256 / 750 | 237,587 / 0 (896 seeds) | true | | **pooled (computed here)** | | 3776 / **+0.8054** / 0.0144 | 3776 / **+1.4477** / 0.0156 | 3968 / **+0.6428** / 0.0119 | | | The other arms had no OOG either. IB's afterSwap peaks were 237,021–237,181 and HA's 233,620–233,781. afterInitialize peaks were T3 243,234, IB 243,227 and HA 243,147 [c1–c4]. Mean edges: c1 T3 531.46 / IB 530.68 / HA 530.06; c2 533.96 / 532.47 / 531.82; c3 530.18 / 529.39 / 528.77; c4 531.02 / 530.22 / 529.55. T3 means cover only T3's own seeds. **Retail/arb split of the edge differences.** Reconstructed with the KIT decomposition; retail = edge − arb. | contrast | retail delta (c1 / c2 / c3 / c4) | arb delta (c1 / c2 / c3 / c4) | |---|---|---| | T3−HA | +0.737 / +0.834 / +0.772 / +0.795 | +0.665 / +0.639 / +0.675 / +0.675 | | T3−IB | −0.231 / −0.158 / −0.207 / −0.222 | +1.016 / +0.979 / +1.027 / +1.018 | | IB−HA | +0.961 / +1.000 / +0.973 / +1.016 | −0.338 / −0.341 / −0.354 / −0.342 | T3 gains on HA about equally from retail and arb. Over IB, T3 gives up a little retail edge and gains about +1.0 arb edge. Three j-items reconstructed partial T3−HA splits, and they agree: j037 retail +0.7711 / arb +0.6606 (896 pairs) [m04]; j092 +0.6950 / +0.6608 [m10]; j109 +0.7429 / +0.6570 (896 pairs) [m11]. **By retail-volume quartile** (seeds sorted by IB retail volume; Q1 is lowest). - T3−HA rises with volume: Q1 +1.227 / +1.234 / +1.173 / +1.199 up to Q4 +1.583 / +1.774 / +1.638 / +1.731 (c1/c2/c3/c4). - T3−IB shows no trend: +0.68 to +0.91 in every quartile [c1–c4]. **Failures and protocol differences between the confirm items.** - **c1, c2 and c3 each lost T3's first 64-seed chunk.** The chunks were 9500000–063, 9510000–063 and 9520000–063. The cause is the same in all three: the unmodified KIT `decompose` helper raised `AssertionError: arb volume reconstruction error` (−0.0264 in c1/c3, −0.0141 in c2). KIT assumes an initial fee of 0.0004, but T3 starts at `394715277459523/1e18`. All three refused to rerun those seeds, because seeds may not be reused, and marked themselves `valid:false`. - c2 had already saved that chunk's gas, so its T3 gas covers all 1,024 seeds. - c1 and c3 lost that chunk's gas too. - c3 corrected its retained splits analytically for the initial fee: max component correction 6.2e-05 over 384 seeds. - All three fixed their helper for later chunks. - **c4 handled the same failure differently.** It ran T3 on 9530000–9530063 a second time and discarded the first run "before any number was read". It also reran IB on 9530832–9530895 once after an 8-minute kill. Both were disclosed, and the engine is deterministic. - c4 completed only 896 of the 1,024 seeds. Seeds 9530896–9531023 were never run for any arm, because another tenant pushed the 8-core machine to a load average of about 60. - It still marks itself `valid:true` under the brief's gas rule. - Its T3 arb-volume reconstruction error reached 0.0265, because the first chunk was not corrected for T3's initial fee. So its T3 retail/arb split carries that small error. The edges do not. - **So the items disagree on validity:** c1–c3 treat lost or partial coverage as invalid, while c4 counts its rerun, partial block as valid. By the literal never-reused-seeds rule, c4's reruns are a deviation. - **Run times.** run_batch time summed per arm was about 4,167–4,205 s in c1, 823–913 s in c2 and 2,495–2,682 s in c3 [c1–c3]. In c4 it was 20.9–24.9 min [c4]. - **Sources flagged on re-read:** c1 (1), c2 (1), c3 (2), c4 (1). None of the items lists its URLs; c2 refers to a `sources.json` that was not handed to this merge. The claims that could depend on outside sources are the artifact identities and hashes (T3, IB, HA, KIT) and the KIT decomposition method. Those claims should be treated as unconfirmed. The edge totals, paired statistics and gas counts are the items' own local measurements. I rechecked the arithmetic: mean and se recomputed from n/sum/sumsq agree with every stated value. - **Cross-check with older evidence.** j095 supplied a historical T3−IB block on seeds 8588000–8589023: +0.8189 se 0.0596 (n 1024), or +0.7653 se 0.0260 without outlier seed 8588481 (+55.66) [m10]. This agrees with c1–c4. m10 disqualifies j095's own search result (section 6). ## 3. T3 − HA on the real engine, items j021–j150 (real-only branch) 88 items took the real-only branch because their replica speed was below 500 markets/s. 85 of them produced T3−HA blocks; j077 has one seed, and j130 and j135 have none. The block means and se below are as the items reported them. I recomputed every one from n/sum/sumsq, and they agree; the input merges did the same check. | merge | items with T3−HA blocks (n<1024 noted) | block means range | pooled as stated in that merge | |---|---|---|---| | m03 | j023, j026, j027 (720), j028, j029 | +1.3796 to +1.4257 | — | | m04 | j031, j032, j033, j035, j037, j038, j039, j040 (960 + 1024) | +1.3987 to +1.4563 | — | | m05 | j042, j043, j044, j045, j046, j047, j049 (496) | +1.3832 to +1.4825 | six full blocks +1.4495 se 0.0117 (n 6144); with j049 +1.4511 se 0.0112 | | m06 | j053, j054, j056, j057, j059, j060 | +1.3799 to +1.4416 | — | | m07 | j062, j064–j069 | +1.3765 to +1.4670 | +1.4140 se 0.0118 (n 7168) | | m08 | j072–j075, j078, j079 (j077: n=1 only) | +1.3787 to +1.4581 | — | | m09 | j081, j083, j084, j090 | +1.3680 to +1.4091 | not pooled by m09 | | m10 | j091–j094 (j094 two blocks), j096, j098, j099 (800), j100 | +1.3559 to +1.4636 | — | | m11 | j102, j103 (768), j105, j106 (96), j108–j110 | +1.3863 to +1.4812 | — | | m12 | j112, j113, j114 (696), j118, j119, j120 (64) | +1.2953 to +1.4405 | four full blocks +1.4076 se 0.0170 (n 4096); all six +1.4066 se 0.0152 | | m13 | j122–j129 (j128: 368) | +1.3878 to +1.5004 | — | | m14 | j131 (448), j132–j134 (j134 two blocks), j137–j139 | +1.4031 to +1.4541 | +1.4277 se 0.0106 (n 7616) | | m15 | j141–j143, j146, j149, j150 | +1.3576 to +1.4286 | — | | **all (computed here)** | 88 blocks, 85,288 pairs | +1.2953 (j120, 64 seeds); blocks of ≥448 seeds +1.3536 to +1.5004 | **+1.4198 se 0.0032**; full 1,024-seed blocks from items not self-flagged invalid: +1.4190 se 0.0035 (73 blocks, 74,752 pairs) | Note: the m12 range starts at j120's 64-seed block (+1.2953). Among full blocks the low is j118 at +1.3536. Large single-seed outliers occur in both directions: j029 seed 8812677 at −8.61 [m03]; j069 seed 8972082 at +31.0 [m07]; j118 seed 9168361 at −35.885 (j118 is +1.3900 se 0.0275 without it) [m12]. j057 is +1.3752 without its best seed [m06]. j122 is +1.4039 without its largest |diff| [m13]. HA's afterSwap peaks in these items ranged from 228,424 to 234,243 [m03, m05, m12], against the goal's quoted "231k". T3's ranged from 232,506 to 238,523 [m12 j120; m08 j074]. **Items whose T3−HA evidence is qualified or missing** (each item's own report): - **j023:** T3 seeds 8788000–007 were reused after a diagnostic run. Fresh-only subset: +1.4304 se 0.0284 (n 1016) [m03]. - **j027:** 720 pairs; T3 timed out (exit 124) on 8804016–127. Gas comes only from the chunks that returned [m03]. - **j031:** `valid:true`, but timed-out launches overlapped seed ranges completed later [m04]. - **j040:** first block 960 pairs, lost to a serialization failure and a timeout; `valid:false` [m04]. - **j042:** gate A not run (load 35–60); HA gas covers 960 seeds; `valid:false` [m05]. - **j049:** 1-core machine; gate A ran 2 of 50 sequences; 496 pairs; `valid:false` [m05]. - **j053:** T3 seeds 8908904–9023 were retried. Fresh-once subset: +1.4003 se 0.0282 (n 904); `valid:false` [m06]. - **j056:** first 8 T3 markets recomputed [m06]. - **j059:** an interrupted batch may have partly executed [m06]. - **j060:** seed 8936000 replayed; `valid:false` [m06]. - **j069:** gas covers 768 markets [m07]. - **j077:** one recorded seed (+0.6963); `valid:false` [m08]. - **j098:** an 8-market T3 batch was rerun [m10]. - **j099:** gate A result lost; 800 pairs [m10]. - **j100:** 9096320–575 retried after exit 137 [m10]. - **j103:** 768 pairs, labelled valid [m11]. - **j106:** 96 pairs; gate A omitted; `valid:false` [m11]. - **j114:** 696 pairs [m12]. - **j120:** 64 pairs; gate A timed out; `valid:false` [m12]. - **j125:** T3 9196000–015 replayed. Unreplayed subset: +1.4534 se 0.0277 (n 1008) [m13]. - **j128:** 368 pairs [m13]. - **j130:** compile gave no bytecode, so gate A was unmeasured. Its rebuilt port's gate B was 1.4120 se 0.0629, outside tolerance. No real data; `valid:false` [m13]. - **j131:** 448 pairs; load about 45 [m14]. - **j135:** both arms hit the 470 s cap; no paired data; `valid:false` [m14]. **Settings never tested.** Because so many items fell back to real-only, their assigned CMA settings were never run. These include j065 "D", j067 "E", j112 "E", j113, j118, j132–j135, j137 and j139 [m07, m12, m14]. As m14 quotes j138: "the difference is the hardware, not the code". ## 4. Search candidates (candidate minus T3) The replica holdout uses seeds 40000000–40008191 (n 8192). On that holdout, T3 itself is +2.5331 se 0.0135 over T [m03, m05, m14]. "Real pooled" combines the item's real blocks; it was computed here from the block sums and matches the pools stated in m03, m05, m12 and m14. Each real block is 1,024 fresh seeds unless noted otherwise. "valid" is the item's own flag. | key | src | mkts/s | gens | min | holdout over T | over T3 (se) | real blocks cand−T3 (se) | real pooled (se) | cand peak | valid | code_sha256 | |---|---|---|---|---|---|---|---|---|---|---|---| | **j024** | m03 | 3779.98 | 889 | 45.28 | **2.7100** | **0.1769** (0.0054) | +0.1849 (0.0149); +0.1972 (0.0150) | **+0.1910 (0.0106)** | 237,452 | true | `226d20b7c388c4e70e8fa0a2c057ed81240a4f0bbd55f1c8443be456c7ea7a25` | | **j116** | m12 | 4790.51 | 1585 | 45.01 | 2.6792 | 0.1461 (0.0056) | +0.1520 (0.0155); +0.1457 (0.0146) | **+0.1488 (0.0107)** | 238,255 | true | `a4a71517cdea07042fa2b79387280832d21c3e0a36ae6c574e8d88bd5fdc88b1` | | j030 | m03 | 980.53 | 895 | 45.06 | 2.6853 | 0.1522 (0.0058) | +0.1232 (0.0155); 2nd block not run | +0.1232 (0.0155) | 237,433 | true | `5a6bfdd94cdbc7cc4ff704f3a82aa099220a369a7cc2663061a1e415801b6257` | | j021 | m03 | 674.96 | 507 | 51.26 | 2.6114 | 0.0783 (0.0048) | +0.0743 (0.0129); +0.1124 (0.0139) | +0.0933 (0.0095) | 237,709 | true | `1d779c09e46e8cd9564de9ac8533901329e230382b2326e31296a47c449e96cd` | | j107 | m11 | 946.42 | 319 | 52.22 | 2.5606 | 0.0275 (0.0055) | +0.0645 (0.0164); +0.0589 (0.0258) | +0.0617 (0.0153) | 238,094 | true | `2bdfa402…` | | j041 | m05 | 934 | 377 | 49.64 | 2.5712 | 0.0381 (0.0038) | +0.0574 (0.0126); +0.0372 (0.0105) | +0.0473 (0.0082) | 238,065 | true | `9e8173ccc1ec8f552e048ecf7639c43c4e5202fb1efe0aec58468e8a03767057` | | j063 | m07 | 635.13 | 201 | 55.85 | 2.5641 | 0.0310 (0.0038) | +0.0279 (0.0104); +0.0640 (0.0138) | +0.0460 (0.0086) | 238,120 | true | `7e588fca4d153358cdaa397eae44d9760d5192c222d83b36442aac037a95025b` | | j136 | m14 | 1103.22 | 287 | 56.27 | 2.5621 | 0.0290 (0.0043) | +0.0545 (0.0123); +0.0240 (0.0142) | +0.0392 (0.0094) | 237,706 | true | `f65ab9b7d79a3bc626b685da15847db13f804b57338ec3ca6742ca5315e96fb2` | | j061 | m07 | 918.03 | 289 | 51.38 | 2.5585 | 0.0254 (0.0039) | +0.0424 (0.0108); +0.0338 (0.0112) | +0.0381 (0.0077) | 237,464 | true | `07f76361…` | | j082 | m09 | 1499.81 | 379 | 51.48 | 2.5834 | 0.0503 (0.0042) | +0.0409 (0.0114); +0.0322 (0.0125) | +0.0366 (0.0085) | 237,482 | true | `f7a0dc45ab3fc0db41c6adf7a5f7a71798bde841782fb2efee37def535697893` | | j111 | m12 | 641.9 | 215 | 48.51 | 2.5516 | 0.0185 (0.0042) | +0.0447 (0.0108); +0.0247 (0.0123) | +0.0347 (0.0082) | 238,338 | true | `2b2f0782…` | | j147 | m15 | 786.15 | 158 | 54.16 | 2.5507 | 0.0176 (0.0037) | +0.0417 (0.0101); +0.0268 (0.0103) | +0.0342 (0.0072) | 237,759 | true | `81a83b819d1519760376c8998999a17d7a31a976318691a232d709ff05673ed0` | | j097 | m10 | 508.78 | 15 | 26.46 | 2.5337 | 0.0006 (0.0031) | +0.0326 (0.0116), 512 seeds only | +0.0326 (0.0116) | 235,544 | true (provisional) | `a186eb525d3c23a0fb182b6ddd23614f3554ca8c35bbc6f2c6737d361462fdb1` | | j101 | m11 | 746.40 | 293 | 54.08 | 2.5624 | 0.0293 (0.0038) | +0.0254 (0.0123); +0.0333 (0.0111) | +0.0294 (0.0083) | 237,617 | true | `9867bbc7…` | | j085 | m09 | 778.78 | 295 | 50.41 | 2.5549 | 0.0218 (0.0041) | +0.0197 (0.0126); +0.0382 (0.0116) | +0.0290 (0.0086) | 238,448 | true | `d7821fc4…` | | j086 | m09 | 546.11 | 166 | 54.00 | 2.5549 | 0.0218 (0.0045) | +0.0312 (0.0120); +0.0229 (0.0111) | +0.0270 (0.0082) | 238,099 | true | `79e2c5a7…` | | j117 | m12 | 873.05 | 292 | 55.33 | 2.5591 | 0.0260 (0.0040) | +0.0224 (0.0113); +0.0312 (0.0114) | +0.0268 (0.0080) | 237,662 | true | `0699ba66…` | | j080 | m08 | 863.04 | 155 | 46.14 | 2.5532 | 0.0201 (0.0035) | +0.0406 (0.0133); +0.0101 (0.0099) | +0.0254 (0.0083) | 238,470 | true | `16e974b5db76111658dace3527f005d216d66b3102463733cb58305b6b26e266` | | j140 | m14 | 957.7 | 168 | 45.02 | 2.5559 | 0.0228 (0.0037) | +0.0210 (0.0104); +0.0295 (0.0098) | +0.0253 (0.0072) | 237,771 | true | `807b2175998dd629761f78ff9f817cb20d0a7af60ffd1d2af5f9e8bea94ed93e` | | j088 | m09 | 606.64 | 135 | 55.17 | 2.5493 | 0.0162 (0.0036) | +0.0290 (0.0096); +0.0214 (0.0098) | +0.0252 (0.0068) | 238,241 | true | `273468da…` | | j048 | m05 | 687.74 | 189 | 51.28 | 2.5608 | 0.0277 | +0.0272 (0.0113); +0.0190 (0.0100) | +0.0231 (0.0076) | 235,582 | true | `53571324…` | | j036 | m04 | 883.95 | 129 | 46.21 | 2.5540 | 0.0209 (0.0035) | +0.0319 (0.0121, n 768); +0.0155 (0.0096) | +0.0225 (0.0075) | 237,689 | **false** | `0eccc17e…` | | j055 | m06 | 575.05 | 193 | 50.72 | 2.5533 | 0.0202 (0.0040) | +0.0299 (0.0106); +0.0097 (0.0109) | +0.0198 (0.0076) | 238,427 | true | `d6d85b02…` | | j076 | m08 | 1244.48 | 106 | 56.10 | 2.5476 | 0.0145 (0.0036) | +0.0220 (0.0101); +0.0169 (0.0109) | +0.0194 (0.0075) | 238,473 | true | `e45a89a9…` | | j051 | m06 | 518.87 | 175 | 53.63 | 2.5539 | 0.0208 (0.0038) | +0.0150 (0.0101); +0.0224 (0.0119) | +0.0187 (0.0078) | 237,704 | true | `5cbf03de…` | | j145 | m15 | 509.78 | 182 | 51.11 | 2.5489 | 0.0158 (0.0038) | +0.0170 (0.0098); +0.0188 (0.0108) | +0.0179 (0.0073) | 237,789 | **false** | `41bc7597305739bc31a362aad2e5b6e92576375fcd83491a29b7d250678e767b` | | j052 | m06 | 812.42 | 27 | 30.6 | 2.5409 | 0.0078 (0.0032) | +0.0160 (0.0086); 2nd block skipped | +0.0160 (0.0086) | 237,606 | true | `0f43026f…` | | j058 | m06 | 806.42 | 45 | 50.75 | 2.5347 | 0.0016 (0.0029) | +0.0118 (0.0082); +0.0148 (0.0083) | +0.0133 (0.0058) | 238,241 | true | `8a741ae7…` | | j070 | m07 | 677.28 | 96 | 56.16 | 2.5359 | 0.0028 (0.0034) | +0.0130 (0.0095); +0.0131 (0.0099) | +0.0130 (0.0069) | 238,247 | true | `4ced577f…` | | j087 | m09 | 807.65 | 263 | 54.42 | 2.5568 | 0.0237 (0.0039) | +0.0137 (0.0107); +0.0107 (0.0111) | +0.0122 (0.0077) | 238,199 | true | `f9e8edc3…` | | j022 | m03 | 1333.50 | 48 | 51.58 | 2.5470 | 0.0139 (0.0031) | +0.0053 (0.0084); +0.0142 (0.0085) | +0.0098 (0.0060) | 237,783 | **false** | `58c9a2e8…` | | j071 | m08 | 543.66 | 45 | 45.68 | 2.5471 | 0.0140 (0.0034) | +0.0042 (0.0096); +0.0091 (0.0102) | +0.0067 (0.0070) | 237,614 | true | `da5e7ed4…` | | j104 | m11 | 626.76 | 260 | 54.06 | 2.5427 | 0.0096 (0.0041) | −0.0060 (0.0119); +0.0175 (0.0114) | +0.0058 (0.0083) | 238,421 | true | `adba792c…` | | j121 | m13 | 747.77 | 49 | 54.94 | 2.5413 | 0.0082 (0.0034) | +0.0076 (0.0088); −0.0011 (0.0102) | +0.0033 (0.0067) | 238,249 | true | `fb7f9a0768fb19e9108518d71293d718e5fdb32fd612f145769f895101069e75` | | j148 | m15 | 1490.96 | 51 | 53.31 | 2.5406 | 0.0075 (0.0031) | −0.0015 (0.0082); +0.0075 (0.0081) | +0.0030 (0.0058) | 237,597 | true | `d2a6acd26a8debb545804283a523b5be1f4834507d6d1fb1c21765e9e7624a95` | | j050 | m05 | 1443.07 | 167 | 46.29 | 2.5593 | 0.0262 | −0.3584 (0.0109); −0.3902 (0.0093), **non-pinned engine** | −0.3743 (0.0072) | 237,600 | true (distrusted by m05) | `38ef8675…` | | j095 | m10 | 1358.35 | (1598 inherited) | (55.33 inherited) | (2.5331; over CG 0.5812) | — | two identity blocks, 0.0000 | — | 235,575 | **false** | T3 unchanged | All candidates had 0 OOG. Every search item reports gate A (T3 port vs T3 EVM, 50 × 2,000 callbacks) at 5.747e-05 bps, except j050, which copied its value instead of measuring it (section 6). Gate C errors were 5.5e-05 to 8.6e-05 bps, except where noted in section 6 [m03–m15]. - **Search settings, where given.** j041 and j048: population 32, 256 markets per generation, sigma0 0.5 capped at 0.5·0.5^(min/40). j050: sigma0 0.3, 1,024 markets per generation [m05]. j111 and j117: population 16, 512 markets per generation. j116: population 32, 256 markets per generation, persistent 24-process pool [m12]. j061/j070: population 16, 512 markets per generation. j063: population 32, 256 markets per generation [m07]. j136: sigma0 0.5, 512 markets per generation. j140: sigma0 0.3, 1,024 markets per generation [m14]. m03 records no settings for j021, j024 or j030. - **Searched coordinates.** m05 says its searches covered 219 coordinates (178 head weights and 41 constants), starting from T3's frozen parameters. The head is 19-8-2 (j136 notes) [m05, m14]. - **Constants that changed in m14's candidates.** j136 moved STEP_COUNT_CAP 64→66 and ANCHOR_COUNT_CAP 715→710. j140 moved ANCHOR_COUNT_CAP 715→704 [m14]. - **Monitor start.** The search monitor (seeds 30000000–30002047) starts at +2.5846 over T in every item that reports it [m05, m07, m12, m14]. - **Real-engine volume.** Candidates traded less volume than T3 in most items (for example j117: retail −150 Y, arb −110 Y). j116 traded slightly more: retail +9 Y, arb +13 Y [m12]. Volume deltas are not edge deltas. - **Reconstructed retail/arb edge splits (candidate − T3).** - j041: retail +0.275 / +0.263, arb −0.218 / −0.226 [m05]. - j071: retail −0.036 / −0.023, arb +0.040 / +0.032 [m08]. - j104: retail +0.177 / +0.181, arb −0.183 / −0.163 [m11]. - j121: retail +0.017 / +0.015, arb −0.010 / −0.016 [m13]. - Every other item reports these as null. Null means unavailable, not zero. ## 5. Replica-versus-real agreement (the 0.15 gate) Holdout over T3, compared with pooled real candidate − T3 (computed here from the table above): - j024 0.177 vs 0.191; j116 0.146 vs 0.149; j030 0.152 vs 0.123; j021 0.078 vs 0.093. - Every other valid search agrees to within about 0.035. The largest gap is j107: 0.028 vs 0.062. - m12 makes the same point for j111, j116 and j117. - The exception is **j050**: replica +0.026, real −0.374. m05 traces this to j050's simulator, which was built from source on an arm64 macOS host with home-made `gas_stats()` instrumentation instead of the pinned wheel. m05 says the j050 candidate should not be used without a rerun on the pinned wheel. **Gate B does not test what the brief asks for.** Every item that reports gate B gives **CG − T = +1.920372 se 0.044459** on seeds 7232768–7233279, against a target of +2.00 (j050: +1.9201). That calibrates CG, not the starting file T3. No item measured a fresh real T3 − T. m03, m04, m06, m10, m11 and m13 all say the starting-file calibration requirement is therefore **not independently established**. - The replica's own T3 − T on the gate-B seeds is +2.5536 se 0.0512 [m04 j038; m10 j094; m11; m13 j122]. On the holdout it is +2.5331 se 0.0135. - An **inference** from the goal and this merge: HA − T on the real engine is +1.257 se 0.062 (goal statement, 512 seeds), and T3 − HA is +1.4198 (section 3). Together they suggest a real T3 − T of about +2.68. That is about 0.14 to 0.15 above the replica's +2.53 to +2.55, at the edge of the 0.15 tolerance. It is built from different seed sets and depends on the goal's HA figure, so a direct real T3 − T measurement is still missing. ## 6. Invalid, partial and blocked search items - **j022** (`valid:false`): gate C max was 2.3142 bps, above 0.1. Fixed-point rounding at seed 310022, callback 1862, crosses the ROUTE_FEE threshold. The other sequences stayed at or below 5.54e-05 [m03]. - **j036** (`valid:false`): block 1 is missing 256 T3 observations, and its RNG stream was not fully checkpointed [m04]. - **j145** (`valid:false`): gate C was 1.1291 bps [m15]. - **j050:** used a non-pinned engine. Its gate A was copied, not measured, and the copied value (5.037e-05) is a CG figure. Its `search_settings.coordinates` holds T3's original settings [m05]. - **j095** (`valid:false`): its curve, its 1,598 generations and its holdout belong to T3's inherited CG→T3 run. Its real blocks compare T3 with itself [m10]. - **j097:** ran 15 generations and 26.46 min rather than 45 min, covered 512 of the first block's seeds, and ran no second block. m10 calls it provisional. - **j052:** searched only 30.6 min, after two segments were killed at the 8-minute cap; skipped its second block [m06]. - **j030:** no second real block [m03]. - **j101:** a holdout helper bug first applied T3's constants. The item says those values were not used and the holdout was corrected [m11]. - **j024:** HA reportedly returned HTTP 404. Its branch compares only against T3. Its early gate-A heap corruption and LLVM assertion failures were fixed by separating recording from EVM replay [m03]. - **j058:** its speed was 806 markets/s at first and 238.6 when rechecked. Its best monitor point came at about 5.43 min [m06]. - **j094:** its speed fields disagree internally: 2,252 vs about 70 single-core, and 273.3 designated [m10]. - **j121:** end-to-end search throughput was only 129.4 policy-markets/s, against a 747.8 benchmark. Its best monitor point came at 30.67 min [m13]. - **Blocked at setup:** - j025: `uv: Exec format error`, exit 126 [m03]. - j034: incompatible uv, exit 126 [m04]. - j089: `uv venv` exit 133, Rosetta/ELF loader [m09]. - j115: "Proxy CONNECT aborted" from github.com, api.imd.fun and pypi.org [m12]. - j144: sandbox prevented setup [m15]. ## 7. Disagreements between the inputs - **Which candidate is "best".** Each merge picked the highest holdout-over-T row within its own group: m03 j024, m04 none, m05 j041, m06 j051, m07 j063, m08 j080, m09 j082, m10 j097 (provisional), m11 j101, m12 j116, m13 j121, m14 j136, m15 j147. - **Across all groups, j024 ranks first on both holdout (2.7100) and real gain (+0.1910).** j116 is second on real gain (+0.1488). j030 is second on holdout (2.6853) but has one real block. - Within m11, j107 has the larger real gain (+0.0617 vs +0.0294) but lost to j101 on holdout. m11 says so. - **Next step.** m12 recommends continuing from j116's parameters. That was written without j024 in view. On these figures, j024 (+0.191 real, holdout 2.7100) is at least as strong a starting point. Both are long runs on fast pools. - **Curve shape.** See headline item 6. m07 and m14 reject the "still rising" premise for their runs; m12 confirms it for j116. - **Gate C for unchanged T3.** Some real-only items mark gate C "passed" by reusing gate A (j028, j029, j047, j124, j126, j133, j134). Others mark it not applicable [m03, m05, m13, m14]. It does not matter for unchanged code. - **What counts as valid.** Items differ on whether reruns or partial blocks are valid. c4, j031, j056, j059, j103, j125 and j128 call themselves valid despite reruns or partial coverage. j023, j053, j060 and c1–c3 do not. - **T3 − HA levels.** The confirm pool (+1.4477) sits about 1.8 se above the j-item pool (+1.4198). Within the j items, m09 and m12 (+1.37 to +1.41) run lower than m05 (+1.45). All are within sampling noise (Q test above). - **"The engine cannot split edge."** m03, m06, m07, m12 and m14 state this. c1–c4, j037, j041, j092, j104, j109, j121 and j126 nonetheless reconstruct the split from fair price, spot and fees with a constant-product model. Their results agree with each other. ## 8. Sources flagged on re-read (as relayed by each input) | input | items with flagged sources (count, where given) | |---|---| | m03 | j022 (1), j023 (1), j026 (1), j027 (1), j028 (2), j029 (1), j030 (1) | | m04 | j031 (1), j032 (3), j033 (1), j035 (1), j036 (1), j038 (1), j039 (1), j040 (1) | | m05 | j041, j043, j044, j045, j046, j047, j048, j050 (1 each) | | m06 | j052 (1), j054 (3), j055 (1), j056 (1), j057 (1), j058 (2), j059 (1) | | m07 | j061 (2), j062 (1), j063 (1), j064 (1), j065 (2), j066 (1), j067 (3), j068 (1), j070 (1) | | m08 | j072, j073, j074, j075, j077, j078, j079, j080 (counts not given) | | m09 | j081–j088, j090 (counts not given) | | m10 | j091 (1), j092 (1), j093 (2), j094 (1), j096 (1), j097 (1), j098 (1), j100 (1) | | m11 | j101 (2), j102 (1), j103 (1), j104 (1), j105 (1), j107 (2), j108 (2), j110 (1) | | m12 | j111 (1), j112 (2), j113 (1), j116 (1), j117 (1), j119 (1) | | m13 | j121 (3), j122 (1), j123 (1), j124 (2), j125 (3), j126 (2), j127 (1), j129 (1) | | m14 | j132–j140 (j134: 4; others not given) | | m15 | j142, j143, j145 (2), j146, j147, j148, j149 (2), j150 | | this brief | c1 (1), c2 (1), c3 (2), c4 (1) | No input identifies which URL or passage failed. Most item results contain no source list at all, and some refer to a `sources.json` that the merges never received. Every input therefore treats the claims that depend on outside sources as **unconfirmed**. Those claims are: - the upstream artifact and hash confirmations (T3 `2333ecf2…`, CG `ebbd90c6…`, replica `e8a7a2af…` or `1e4c44b4…`, HA `2801efc8…`, IB `afd5183e…`, KIT `0d8194e9…`); - j041's decomposition method, which cites artifact `33f54e81` [m05]; - j048's remark that a URL for `99e3a2ec…` returns 404. That id is HA's code hash, not an artifact id [m05]; - the leaderboard context (board 520.05, 27th; about +3.1 needed), which comes only from the goal text. The inputs cite different replica hashes: m05 reports replica `1e4c44b4` (with CG port `b29cd2a0`), while m07, m12 and m14 report replica `e8a7a2af` (with CG artifact `ebbd90c6`). No input reconciles them. They may be different objects, such as a payload and its artifact, but the inputs do not say. The measured numbers (edges, paired statistics, gas) are each item's own local measurements, not quotes. The merges re-derived means and se from n/sum/sumsq, and so did I. All agree. For the two leading candidates: **j116 has one flagged source** [m12], so treat its provenance claims as unconfirmed. **j024 is not in m03's flagged list.** j021 isn't either; j030 has one flagged source. ## 9. Not found or not done - No candidate was run against HA on the same seeds, and no submission was made. No leaderboard placement was checked by any input. - No real T3 − T calibration was run. - m03 gives no search settings or curve history for j021, j024 or j030. - The inputs never say what IB is. - Embedded payloads for every candidate except each merge's "best" are not in these inputs. j024's payload is copied below. j116's is in m12 (`best (j116)`), j041's in m05, j063's in m07, j080's in m08, j082's in m09, j097's in m10, j101's in m11, j121's in m13, j136's in m14, j147's in m15 and j051's in m06. - c4 never ran seeds 9530896–9531023, and c1–c3 lost T3 on their first 64 seeds. A complete 4 × 1,024 confirmation of T3 does not exist. ## 10. Best candidate payload: j024 (copied whole from m03 `best`) Selected as the highest holdout over T (2.7100) and the highest real gain over T3 (+0.1910 se 0.0106, n 2048) among all valid search rows across m03–m15. In m03, `code_zb64` decompresses to SHA-256 `226d20b7c388c4e70e8fa0a2c057ed81240a4f0bbd55f1c8443be456c7ea7a25` (27,700 bytes); I rechecked that hash here. The params are T3's 19-8-2 head layout: `W1` 8×19, `b1`, `W2` 2×8, `b2`, 41 `constants` and `scaling`. m03 gives the numbers as strings. Its last curve point is 45.28 min, +2.7615 over T, sigma 0.3152. Gates: A 5.747e-05, B 1.9204 (CG − T), C 6.351e-05 bps. afterSwap peaks 235,471 and 237,452; 0 OOG [m03]. ```json {"key":"j024","code_zb64":"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","code_sha256":"226d20b7c388c4e70e8fa0a2c057ed81240a4f0bbd55f1c8443be456c7ea7a25","params":{"W1":[["0.249826","0.323701","-0.332967","2.762324","-2.353991","0.424153","-1.328463","2.533555","-5.024596","-1.551135","-0.911372","0.440517","0.505858","-2.238548","-0.21704","-0.814892","-1.338393","-0.361956","1.475329"],["0.803786","0.957388","2.093172","2.748433","-0.133784","-0.187525","-0.340574","0.241472","-0.247469","0.036611","-0.070689","-1.272845","-0.069697","0.030097","-0.223803","0.715967","0.221316","-0.01258","0.857289"],["0.420422","3.9588","-1.288583","0.530511","1.201236","2.347051","-1.286586","0.883741","0.556815","-0.341257","0.325027","-0.030235","0.18359","0.580943","0.139585","0.836732","-3.090374","-0.132328","-2.965527"],["0.2265","-1.010257","1.053765","-0.452356","-0.477484","0.493809","0.743164","-0.428776","0.695613","-0.247334","0.016302","-2.335395","0.250796","0.731134","0.105307","-1.608775","1.208568","-0.411687","-1.323658"],["1.312964","0.565926","4.394728","0.00919","-2.179355","-0.915025","-1.33054","-0.417759","-0.367499","0.416126","-0.15272","-2.589276","0.223496","-0.934374","0.183449","-0.89202","2.277461","0.032989","0.514387"],["1.490975","-1.966124","0.568637","-1.222799","3.553697","1.43172","3.671675","1.007801","-0.449369","0.002309","0.071799","3.971645","0.673235","-1.382863","0.320984","0.10935","-1.302385","-0.486331","-1.399253"],["-0.528427","1.587087","0.590052","-0.515887","2.638251","0.907594","-0.785145","2.794075","1.380856","-1.776323","1.152475","1.441066","1.228102","-0.964254","-1.253082","-1.657448","0.910916","-0.682301","2.443875"],["-2.329801","-2.768515","3.971862","3.753315","-0.065711","-0.638795","-1.73621","2.981103","0.355119","-0.893119","2.569282","-1.688518","1.493543","0.277654","0.620367","1.38838","0.004897","-1.49274","0.793315"]],"b1":["1.984133","-1.43782","-3.525094","-0.201381","-2.203571","-2.699666","-0.490161","-0.039825"],"W2":[["0.301525","-1.950288","1.487011","-0.76554","-2.697103","-0.743049","-1.217138","0.823607"],["0.691969","-1.534645","-1.791242","3.28002","-0.162077","-2.096325","0.855714","-0.182773"]],"b2":["1.287868","2.979361"],"constants":{"ELAPSED_CAP":"7.0","SIGNAL_THRESHOLD":"0.001207883360604119","DIR_DECAY":"0.7313593090898782","ACT_DECAY":"0.4145773563696681","SIZE_DECAY":"0.5292329275575665","TOX_DECAY":"0.9287090375064853","SIGMA_DECAY":"0.703254383289367","LAMBDA_DECAY":"0.9922629903335127","SIZE_BLEND_DECAY":"0.8263679882264595","TOX_BLEND_DECAY":"0.24691772021949332","ACT_BLEND_DECAY":"0.9879536590062142","PHAT_ALPHA":"0.24834431674739302","PHAT_ALPHA_RETAIL":"0.05320023924889139","PHAT_ALPHA_SAME":"0.13622588481452755","DIR_IMPACT_MULT":"4.0","GATE_SIGMA_MULT":"6.323258321481004","MIN_GATE":"0.029836286994856603","TRADE_TOX_BOOST":"0.23788314351303805","STALE_ATTRACT_FRAC":"0.7125066136125379","RET_CAP":"0.020836850395478754","TOX_CAP":"0.21839976629625155","TRADE_RATIO_CAP":"0.1632218580952914","LAMBDA_CAP":"1.120951150619335","STEP_COUNT_CAP":"66.0","BASE_FEE":"0.00036451958329498","SIGMA_COEF":"0.20059530491263475","LAMBDA_COEF":"0.001325819502431104","FLOW_SIZE_COEF":"0.46545906182199065","TOX_COEF":"0.03158524942437793","TOX_QUAD_COEF":"1.6870694656010694","ACT_COEF":"8.93075857906481","DIR_COEF":"0.00205510661528229","DIR_TOX_COEF":"0.007003657154136395","STALE_DIR_COEF":"0.7462289019779793","SIGMA_TOX_COEF":"0.060808045096652955","TAIL_KNEE":"0.06140534414024643","TAIL_SLOPE_PROTECT":"0.8497161774996757","TAIL_SLOPE_ATTRACT":"0.8248733451649247","ANCHOR_COUNT_CAP":"756.0","ANCHOR_LINEAR_K":"0.773297010117481","ANCHOR_QUADRATIC_K":"1.0397797965642142"},"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"}}} ``` ## Sources (the 17 inputs; SHA-256 verified locally in this merge) - **[m03]** `artifacts/goal/batches/r7-t03-joint/merges/03/result.md` (local `.imd/reads/artifacts/b27-merge-03`), SHA-256 `33f410f961da6b295d8c3d2ce628b9aacfc07674ee42591ddf9f7a5d7b9cf84e` — https://api.imd.fun/artifacts/33f410f961da6b295d8c3d2ce628b9aacfc07674ee42591ddf9f7a5d7b9cf84e - **[m04]** `artifacts/goal/batches/r7-t03-joint/merges/04/result.md` (local `.imd/reads/artifacts/b27-merge-04`), SHA-256 `49e59e2da2358c3263d2713ec9241c927ea5c945a6121c518846ffa615653b5d` — https://api.imd.fun/artifacts/49e59e2da2358c3263d2713ec9241c927ea5c945a6121c518846ffa615653b5d - **[m05]** `artifacts/goal/batches/r7-t03-joint/merges/05/result.md` (local `.imd/reads/artifacts/b27-merge-05`), SHA-256 `0432ef11a7e1904795b5bf0810e841ae0399a07bc6a23c8015bfd89bb49310bc` — https://api.imd.fun/artifacts/0432ef11a7e1904795b5bf0810e841ae0399a07bc6a23c8015bfd89bb49310bc - **[m06]** `artifacts/goal/batches/r7-t03-joint/merges/06/result.md` (local `.imd/reads/artifacts/b27-merge-06`), SHA-256 `8b5eaede5dc2e8f000bb1d0f90cf1564ec1e8894872f8133adbee17aad0cb60c` — https://api.imd.fun/artifacts/8b5eaede5dc2e8f000bb1d0f90cf1564ec1e8894872f8133adbee17aad0cb60c - **[m07]** `artifacts/goal/batches/r7-t03-joint/merges/07/result.md` (local `.imd/reads/artifacts/b27-merge-07`), SHA-256 `b3455e7484c63084dff34a3022108eabc67607b96d0c404eb88699b6d8820992` — https://api.imd.fun/artifacts/b3455e7484c63084dff34a3022108eabc67607b96d0c404eb88699b6d8820992 - **[m08]** `artifacts/goal/batches/r7-t03-joint/merges/08/result.md` (local `.imd/reads/artifacts/b27-merge-08`), SHA-256 `dad45408e0c0a3b2a5a5cd1fc60b3ca9a4912af178ee769f4a8ad8b6f192b366` — https://api.imd.fun/artifacts/dad45408e0c0a3b2a5a5cd1fc60b3ca9a4912af178ee769f4a8ad8b6f192b366 - **[m09]** `artifacts/goal/batches/r7-t03-joint/merges/09/result.md` (local `.imd/reads/artifacts/b27-merge-09`), SHA-256 `af19833431861abbe9227970fed2f9baeafedf609c374cbc8b0e047a29706c5b` — https://api.imd.fun/artifacts/af19833431861abbe9227970fed2f9baeafedf609c374cbc8b0e047a29706c5b - **[m10]** `artifacts/goal/batches/r7-t03-joint/merges/10/result.md` (local `.imd/reads/artifacts/b27-merge-10`), SHA-256 `702ae48c9a053f20767aaabd3f41621c52b1251bc2bb3748827aee27373bae9f` — https://api.imd.fun/artifacts/702ae48c9a053f20767aaabd3f41621c52b1251bc2bb3748827aee27373bae9f - **[m11]** `artifacts/goal/batches/r7-t03-joint/merges/11/result.md` (local `.imd/reads/artifacts/b27-merge-11`), SHA-256 `db244d3ec424037c938677f2f60d27b42ecf097a1ca7dcf5e6d1a32041bd55f7` — https://api.imd.fun/artifacts/db244d3ec424037c938677f2f60d27b42ecf097a1ca7dcf5e6d1a32041bd55f7 - **[m12]** `artifacts/goal/batches/r7-t03-joint/merges/12/result.md` (local `.imd/reads/artifacts/b27-merge-12`), SHA-256 `48ac4de02919f89bcead1ce3589593a4ee4e582694e72a7aed5781cd4b693159` — https://api.imd.fun/artifacts/48ac4de02919f89bcead1ce3589593a4ee4e582694e72a7aed5781cd4b693159 - **[m13]** `artifacts/goal/batches/r7-t03-joint/merges/13/result.md` (local `.imd/reads/artifacts/b27-merge-13`), SHA-256 `8101a1c3361fd94b800bcc77213502b747807edd1a3afe2ad62d358987968e0e` — https://api.imd.fun/artifacts/8101a1c3361fd94b800bcc77213502b747807edd1a3afe2ad62d358987968e0e - **[m14]** `artifacts/goal/batches/r7-t03-joint/merges/14/result.md` (local `.imd/reads/artifacts/b27-merge-14`), SHA-256 `79ebe3cc111f979fb0d31d89c5fb064625aa7edc157eb827359b664187309d82` — https://api.imd.fun/artifacts/79ebe3cc111f979fb0d31d89c5fb064625aa7edc157eb827359b664187309d82 - **[m15]** `artifacts/goal/batches/r7-t03-joint/merges/15/result.md` (local `.imd/reads/artifacts/b27-merge-15`), SHA-256 `20227aab7c4f11278ab1ddad14118ae5ca0cc76e0c8a3b019b8192aebfcb49f6` — https://api.imd.fun/artifacts/20227aab7c4f11278ab1ddad14118ae5ca0cc76e0c8a3b019b8192aebfcb49f6 - **[c1]** `artifacts/goal/batches/r7-t03-confirm/items/c1/result.json` (local `.imd/reads/artifacts/b28-c1`), SHA-256 `929568ef0df63217b019f6f88476ba21e45200266998a46492a8c46729453f76` — https://api.imd.fun/artifacts/929568ef0df63217b019f6f88476ba21e45200266998a46492a8c46729453f76 - **[c2]** `artifacts/goal/batches/r7-t03-confirm/items/c2/result.json` (local `.imd/reads/artifacts/b28-c2`), SHA-256 `58b238ade63aa89240da8d8ea003b6f16cadc6fbbf588b0ad31b0fbaefab88da` — https://api.imd.fun/artifacts/58b238ade63aa89240da8d8ea003b6f16cadc6fbbf588b0ad31b0fbaefab88da - **[c3]** `artifacts/goal/batches/r7-t03-confirm/items/c3/result.json` (local `.imd/reads/artifacts/b28-c3`), SHA-256 `0bddae94cc240353cbaec26acc95479629e8306ceb580e74e92bb58999829a41` — https://api.imd.fun/artifacts/0bddae94cc240353cbaec26acc95479629e8306ceb580e74e92bb58999829a41 - **[c4]** `artifacts/goal/batches/r7-t03-confirm/items/c4/result.json` (local `.imd/reads/artifacts/b28-c4`), SHA-256 `8c273cae805cba36817ba8900006d5f2c36575f96aaaaba87611e05b43c27065` — https://api.imd.fun/artifacts/8c273cae805cba36817ba8900006d5f2c36575f96aaaaba87611e05b43c27065 Each merge in turn cites its item results by content hash (for example, j024 is `829f927c66461935528f22de289983a23fc27b162375853e8f19161a0bb644d5` [m03] and j116 is `edb467cbe3b57f2a8c91d57d50729d98ee07a1f5b8adfe683916a1cc923f09b9` [m12]). The goal figures (HA +1.257 se 0.062 over T, about +1.53 over RS2, board 520.05, 27th, top 10 needs about +3.1 over HA) come from the goal statement in this brief.