# Root merge 1/3: Gaussian/plain AMM port and follow-up evidence ## Decision and scope **Port-3 is the strongest completed, repeatedly confirmed candidate in these 25 results.** Its original report supplies P, exact shipped-table Q, and F; six unchanged-source confirmation blocks cover another 1,536 disjoint real-engine seeds. Those blocks all report positive paired mean improvement over t03, zero OutOfGas, and observed afterSwap peaks below 245,000 gas. Runtime is 18,466 bytes. Its cost is a measured **0.742353627 replica edge loss** versus full Gaussian/plain tables. This is a working coarse-table port, not preservation of Taylor's full-table score. [b04-port-3; b10-a1–a4; b11-b1–b2] Other candidates leave material gaps: sym-int8 has no contract; final spline lacks final-source engine checks and uses a different F reference from Q; net-1 fails peak gas and lacks Q; net-2 has partial Q; port-2 changes filtering and reused engine seeds; port-1 fails through OOG. [corresponding inputs] All 25 inputs were present. Figures are worker observations unless labeled **merge calculation**. Complete code, payloads, harnesses, CSVs and incidental timings remain in the ledger artifacts; this synthesis is not a new engine run. ## Provenance and evidence weight Bracketed b-labels refer to the input ledger. Taylor is **gtmcknight**, [amm-sim solved at commit 4812a6382311624e1506aca131ca35c378f930d9](https://github.com/gtmcknight/amm-sim/tree/4812a6382311624e1506aca131ca35c378f930d9/research/solved). Reference filenames below are relative to that folder. Port reports say Taylor credit is in their contract headers. [b01-ref-2; b04-*; b08-spline-1; b09-*] **Requester source-recheck flags:** ref-2(1), sym-int8(1), fixed-net(1), arb-3(1), spline-1(2), net-1(3), net-2(3), a1(2), a2(2), a4(2), b1(1), b2(2). The warning identifies neither the exact passage nor whether access failure or unsupported quotation caused it. External claims from these results remain uncorroborated here. Their embedded measurements and bytes are retained as reported evidence, not independent certification. A failed external recheck does not itself refute local measurements; workers' own claims that quotations passed do not override the later warning. Unflagged inputs are not automatically certified. [task warning; named inputs] During this merge, extracting fenced Solidity bytes reproduced the stated source hashes for port-3, spline-1 final, net-1 and net-2; decoding embedded base64 reproduced stated lengths/hashes for sym-int8, fewer-cols and fixed-net. This verifies internal identity only. It does not resolve the flagged external passages, compile a strategy, rerun gates, or establish gas/score correctness. [merge checks on the named input files] The goal's **+2.83 Gaussian/plain, +3.77 Gaussian/qm, +4.26 lattice/qm, t03 about +1.45** are replica-vs-HA context, not engine differences. Plain reproduction supports +2.828; qm/lattice were not ported here. Do not subtract these figures to predict engine gains. [task goal; b01-ref-2; b04-port-3] ## Reference behavior to preserve ### P and baseline score The independent reference repeat ran unmodified Taylor files with plain `python3 -I`, before changes. Gates reproduced all printed values, also reproduced by most other replica/port items. SVD is the explicit exception: it reproduced the aggregate +2.828 but did **not** run gates.py, so its own claim of P is incomplete. Spline-2 performed no P run. [b01-ref-2; b02-*; b03-*; b04-*; b06-*; b08-*; b09-*] | Seed | Perfect information | Gaussian/plain | Gaussian/qm | Lattice/qm | |---|---:|---:|---:|---:| | 10000 | 448.261189 | 443.455598 | 443.522608 | 443.663520 | | 10001 | 576.779199 | 571.463409 | 571.623551 | 571.977293 | | 10002 | 531.886497 | 525.168028 | 525.865859 | 526.411631 | Source: [b01-ref-2], corroborated in [b02-fewer-cols; b03-fixed-net; b04-port-2–3; b06-arb-3; b08-spline-1; b09-net-2]. `real.py pol_mid.json plain: soft=1` printed edge **532.386**, retail **564.827**, arb **-32.441**, retail volume **77686**, arb volume **24942**, versus stored HA **+2.828, SE 0.041**, misclassification **16.57%**, price RMS **14.7 bps**; n=2,048, seeds 10000..12047, 10,000 steps/seed. HA pairing uses stored `calib_out_2048.json`, not freshly simulated HA. Sample SE is sd(ddof=1)/sqrt(n). The reference repeat took 68 rounded wall seconds with the stock 14-worker pool (approximately 0.0332 aggregate seconds/seed, not serial latency). Common full-table unrounded mean is **532.385686971**; fewer-cols reports **532.385688969**, a small unresolved discrepancy discussed below. [b01-ref-2; b02-fewer-cols; b04-port-3] ### Inventory, symmetry and lookup Six quote-offset arrays have 71,262 entries, not direct fee bps. A=ask offset, B=bid offset from estimated fair price. All own rows are `w=(i-160)*5`, i=0..320, spanning ±800 bps. [b01-ref-2; b02-*; b03-*] | Arrays | Shape each | Entries each | Offset min / max, bps | Rival axis | |---|---:|---:|---:|---| | A0, B0 | 321×13 | 4,173 | -24.262 / 800 | -30..30 by 5 | | A1, B1 | 321×49 | 15,729 | -31.323 / 800 | nonuniform W2L×5 | | A2, B2 | 321×49 | 15,729 | -31.333 / 800 | nonuniform W2L×5 | The rival bps axis for phases 1/2 is `[-800,-650,-500,-400,-300,-250,-200,-150,-125,-100,-85,-70,-60,-55,-50,-45,-40,-35,-30,-25,-20,-15,-10,-5,0,5,10,15,20,25,30,35,40,45,50,55,60,70,85,100,125,150,200,250,300,400,500,650,800]`. **W2L values themselves are lattice units, not bps.** Metadata: H_bps=5, W 0=160, NW=321, FB=6, N 2=49, C 2=24, lam=0.8, sigma=9.45, size=20.0, g=0.05368206098114493. These solver constants are not filter adaptation. JSON also contains VD0 321×13 and VD1 321×49, totaling 19,902 value-function entries; plain quoting does not read them. Their measured ranges are -1.075879..0.007284 and -1.045346..0.091621. Including them gives 91,164 two-dimensional entries. [b01-ref-2; b02-fewer-cols] Exhaustive reflected comparisons give **exactly zero difference**: `Aph[i][j]=Bph[320-i][ncol-1-j]`; all 71,262 forward/reverse comparisons pass, including the nonuniform symmetric axis. Thus one side can be stored and the other obtained by negating both coordinates. Input JSON: 780,061 bytes; SHA-256 `efbb5ee102db4d312f64565d9bf7232b6becbe02b624164f744a402bbbf2983b`. [b01-ref-2; b02-sym-int8] Lookup uses `s=y/x`, `w=10000*(ln(s)-m)`, `R=10000*(ln(s2h)-m)`. Phase selects a discrete table; plain has no sd or synced dimension. Clamp coordinates to the grid and bilinearly interpolate physical coordinates, including irregular R. At the upper endpoint back off the lower index and use fraction 1. Retain the original **unclamped w** when converting offsets: ask `max(0,1-exp(-(a-w)/10000))`, bid `max(0,1-exp(-(b+w)/10000))`, then effective fees clamp to [0,0.1]. Do not replace this exponential by a linear fee, clamp negative offsets, use nearest-neighbor, or confuse soft=1 with table smoothing. Store clamped fees for the next callback. Initialization calls the plain table at (0,0, phase 0), even in qm variants; qm's sd/synced logic is separate. [b01-ref-2] ### Gaussian callback semantics Reference initialization: m=ln(100), v=0, tlast=0, cnt=0, s 2 h=sp=100, synced=false; initial fb=fa=.003 is overwritten by table initialization. `SIG2=.000945²`, RHO=.26, lognormal displacement mean-log `ln(11.12e-4)`, spread 1.176, rival multiplier G=.997. Amounts, true price, true arb, rival truth and true retail index do not drive the plain decisions; truth is diagnostic only. No drift/run-specific adaptation is used. [b01-ref-2] At dt<1, including timestamp zero, skip new-step belief updates, variance growth, classification and rival band reset; treat as retail. Otherwise reset count; form the previous-quote interval `lo=ln(sp*(1-fb))`, `hi=ln(sp/(1-fa))`. For `min(dt-1,12)` silent steps add SIG2 then truncate to that fixed interval; add SIG2 once more for the current step without truncation. Skipped steps beyond the cap add no further variance. [b01-ref-2] Arb reading is `lp=ln(s*(1-fb))` on is_buy, else `ln(s/(1-fa))`. Set `Larb=phi((lp-m)/sqrt(v))/sqrt(v)`. Truncate the current prior to get Z, mt, vt. With `d=max(abs(ln(s/sp)),1e-9)`, `Lret=Z*.26*phi((ln(d)-ln(11.12e-4))/1.176)/(d*1.176)`. Posterior pi is Larb/(Larb+Lret), or zero for a nonpositive denominator. Hard arb iff pi>.5. For strict .00540 KB. Full mean 532.385686971, decoded 532.369333837. 5 bps rows still round values, so are not lossless. 20 bps falls outside mean-loss target; even its 35,964 bytes exceed the whole contract ceiling. Direct coarse lookup matched expanded reference at 10,000 random points/setting to ≤1.11e-12 bps; chosen max offset error 40.525 bps, RMS 1.200683079. P/Q measured, no Solidity F. [b02-rows-10bps] **Fewer columns.** RC01 header 16 bytes plus 11,556 payload, all 321 own rows, both sides, exactly three retained R knots/phase at -30,0,30 bps. Phase 0 indices[0,6,12], retail[18,24,30]. Per-own-row weighted least squares merges dropped columns using choose-only bilinear usage h, exponent p=1, ridge .001 toward original knots; clip fit to±800 then integer-round. Retain full scalar metadata, omit solver VD arrays. Usage totals by phase 3,668,359 /5,095,019 /2,261,541; phase 0 edges 73.208573%, retail central 13 columns 77.266840% /66.511103% of interpolation mass. Selected md 61.0 is best measured under 12,000, not proof all column schemes fail. SHA `abe2b32b5249eeec0cd619fa8499c01f7638b952828d9be10939f52245960dd8`. Reconstructed full mean 532.385688969, decoded 531.637947623; unchanged CLI prints +2.828 and +2.081 vsHA. Python byte lookup vs expanded grid:100,000 callbacks, max 1.11e-12 bps; this is **not F**. Pair lookup reads eight int16 corners=16 bytes, six multiplications and 12 adds/subtractions after weights. [b02-fewer-cols] **SVD.** Store only ask factors, use exact reflection for bid; fee-aware fit leaves fee-zero offset cells free. SVD1 is LE int16 quantized factors with per-record rank/scales. Chosen ranks 4/6/6 occupy 11,571 bytes; SHA `9090805200aa2a44cc90e7c5261db8ef6e4ae44b2d4fe1957a91678c85149941`. Central row weighting worsened the central error. Pure low rank struggles at the regime discontinuity; piecewise SVD/jump-location storage were suggested, not tested. Rank 6 on retail needs 24 int16 reads per offset at known brackets. Decoder requires NumPy, so it is not a standard-library-only decoded dependency. Claiming P from aggregate +2.828 does not meet required gates.py chronology. No F/engine. [b03-svd-lowrank] **Smooth spline.** Cubic tensor-product fit of the nonnegative fee component with reflection, reconstruct offsets at original grid nodes and then use original bilinear interpolation. Coefficients signed eighth-bps units, big-endian; phase 0 controls 26×9 (519 B), phases 1/2 controls 28×18 (1072 B each), plus 2 header bytes=2665. SHA `6749a7d91c465a96ad29fa21a4f1e116ac00aaf53233688d27e78f4b42925a18`. Visit weights from 60000..60255, h+mirror(h), weight h/mean(h)+.001; fitting therefore uses a subset of choose, report untouched. Visit-weighted mean absolute fee error .17/.26/.26 bps, RMS .50/.85/.71 bps; worst cells 23 bps phase 0 and 59 bps retail at jumps near |w|≤100. Full-grid-node reconstruction then bilinear lookup is essential: evaluating the continuous spline directly is a different unscored policy. CLI full +2.828 vs decoded +2.819 HA; no Solidity/gas/F in this item. [b03-smooth-2d] **Fixed network.** Ordinary 4→32→32→1 ReLU MLP, int16 weights/int32 biases, Q14 activations and layer shifts 12/10/12, sd weights zero. Explicit positive-w identity path is frozen within the network;150,000 decaying-LR updates with torch seed 135/NumPy 132. Export reduces shift if needed to fit, without clipping; selected output uses integer inference. Biases are in accumulator units; **signed shifts floor**, unlike signed division. Scalar/vector check 1500 cells, maximum accumulators 1,044,416,228 /2,768,536,642 /540,496,130, so signed 32-bit accumulation is unsafe. Blob SHA `9727ecc8acf6d31732793d1836c60cd637133b800bbe61b19098bfee7d17e0f6`; canonical decoded JSON 1,134,669 bytes SHA `068ee67f3219df2aaf9ab9f8e93d29dc79ad313afa17771d818f062609a1cec6`. The payload removes training dependencies for decoding. Original decoder uses NumPy. Scored policy evaluates four corners for ask and four reflected corners for bid: eight passes,9,472 multiplies plus 9,472 adds,520 rescalings,512 ReLUs,21,024 learned-byte accesses before caching. Two continuous passes are cheaper but require new Q. Source warning applies to one citation; later net ports independently report matching payload/decoded hashes, not independent confirmation of its replica score. [b03-fixed-net; b09-net-2] ## Solidity ports and version-specific gates All engine reports except spline-2 say the specified challenge setup succeeded, with wheel SHA `aedb68280aca34c110968f00148387c95885207913fb2cf19a1ab4c5cfd6887a`, challenge commit `6b6921f8d463314aa7c96234e1be25a25245f704`, solc 0.8.24, optimizer 200, viaIR, Paris, fixed 30 bps Vanilla,8 workers and bounded segmented commands. Each strategy runs **alone versus Vanilla** through `amm_sim_rs.run_batch(X, VanillaStrategy, configs, 8)`; paired comparisons subtract matching-seed submission edges from those separate matches. They are not direct port-versus-t03 pool games. [b04-*; b07-*; b08-spline-1; b09-*; b10-*; b11-*] Configurations transcribe MatchRunner._build_configs:10,000 steps; initial price 100, reserves 100/10,000, mu 0, dt 1, retail log-size sigma 1.2, buy probability.5. `default_rng(seed)` draws retail mean size U[19,21], arrival U[.6,1], then GBM sigma U[.000882,.001008]. A2, A 4 and B2 report checks against actual runner fields. Native gas means sum_ok_gas/ok, peaks maximum successful callback gas, counters reset per chunk unless port-1's stated cumulative adjustment applies. Pyrevm replay peaks cannot replace native engine peaks. [b04-port-3; b10-a2; b10-a4; b11-b2] | Port / tested version | Runtime B | F max error, bps | Engine peak / OOG | Engine port−t03 mean / SE / n | Gate assessment | |---|---:|---:|---:|---:|---| | port-1 | 22,645 | .0000657 |247,188 successful peak /290,102 | -101.726 /7.696 /64 | Q missing; gas/OOG fails | | port-2 |18,484 |.000563746 |225,081 /0 |+1.414318451 /.198959130 /64 | filter approximation; Q comparator adapted; engine seeds reused | | port-3 |18,466 |1.39892784856155e-08 |206,161 /0 |+.832385798685 /.159826531714 /64 | P, Q, F reported; fresh engine block | | spline-1 engine v24 |22,117 |not final F version |249,342 /0 |+1.314246680 /.123842280 /128 | peak fails; not final source | | spline-1 final v19 |21,106 |.008898410277 vs original spline |**not run** |**not run** | exact packed Q; F-reference mismatch remains | | spline-2 |not compiled |not run |not run |not run | handoff only | | net-1 continuous2 |16,368 |3.181e-09 |245,377 /0 |+1.228371011 /.103213544 /128 | peak fails; Q uncompleted | | net-2 continuous2 |13,568 |3.18e-09 |209,552 /0 |+1.469436515 /.137313168 /128 | good engine sample, Q partial; selection-rule deviation | Sources: each named input. OOG=0 alone does not satisfy peak≤245,000. “F passes” in a worker report means numerical replay against its stated reference, not automatic equivalence to the original goal variant. ### Port-1: parity is insufficient if state updates revert Faithful line-by-line filter with 12 individual silent truncations; signed WAD, stored log-fee L=-ln(1-fee), log-spots, stable log-domain Gaussian tails and logistic posterior. Log uses normalized atanh 7 terms, exp 10-term series, sqrt 5 Newton; Numerical Recipes erfc log-tail approximation is claimed at 1.2e-7 relative. Eight storage slots, +1 bias for zero-capable values and packed timestamp/count/synced reduce zero→nonzero SSTORE costs. No assembly. [b04-port-1] Coarse A:57 own knots,33 retail rival knots, all 13 phase 0;9,006 values,12-bit packing at.25 bps via u=4*bps+256; blob 13,717 bytes SHA `6c59b8b5ad5146e3d2d5a5d31abde17d2028c3d54384e61458c1bc18be6f5346`. Indicative choose 256 comparison was a **different .1 bps quantization**: A−full=-.103 SE .019; larger B 85×35/int16≈28 KB=-.034 SE .013. Neither is formal Q for shipped 12-bit A; report Q absent. A mistaken W2L scale lost 41 replica edge and was fixed. [b04-port-1] F used seeds 70000..70049, first 2000 swaps each plus 50 initializations, total 100,050 comparisons,50 M gas allowance. Max 6.57e-5 bps, mean 2.77e-7, median 1.05e-7, p 99.9=1.19e-5, zero>.01. With native 250 k replay limit, seed 70000 callback 707, dt 10 fails; each trunc costs roughly 15–20 k. OOG reverts timestamp/state, next callbacks see ever-larger dt and repeat the expensive 12-step path. Engine 9510000..9510063:536,430 swaps,246,328 successes,290,102 OOG (54%), mean successful 97,032 gas; init 198,194 gas. t03 had 689,546 successful swaps,0 OOG, mean 135,651; its independent peak not separable from cumulative maxima. Port edge 421.542, t03 523.268. Worst 5 differences -258.57,-205.97,-191.24,-182.92,-181.98 (seed IDs not provided). Earlier build had 23,329 failures/24,254 calls. [b04-port-1] Compiler segfaulted(-11) with 8 MB stack, worked 16 MB; large literal suspected. Full source requires generator/pinned full JSON because displayed hex is elided; stated regenerated source hash `63a8f38f1a126bee4753ba5bb4a3b0c2bc1899b1664e4737455d492ffb022ae4`. Proposed cheaper tails/exp, split literal and analytic/precomputed silent-step work were not validated fixes. Writing a timestamp before an OOG cannot by itself survive transaction reversion; that suggestion is not an established repair. [b04-port-1; merge interpretation of its reversion finding] ### Port-2: usable sample, materially changed filtering Four packed slots hold belief, variance, fees, prior/rival spots and counters. Logs via atanh, exp via halving/fourth-order series, rational CDF, truncated moments. **Silent-step variance is accumulated then truncated once**, instead of truncating after every silent step; current-step update remains full. Its own Python reference mirrors this approximation and reads shipped bytes.41 own knots,25 retail R knots, all 13 phase 0; int16 .1 bps, blob 10,464 bytes incl 132 knot bytes. Full-grid mean error.2782 bps, p 999.015, max 42.119. [b04-port-2] Paired Q isolates compression under **the aggregated filter on both sides**: choose compressed 528.943411/full 529.016498, diff-.073086 SE .013999; report 532.267743/532.351218, diff-.083475 SE .009721. This does not isolate total loss versus unmodified full real.py, although subtraction of reported absolute means suggests a larger total difference; no paired SE for that difference is supplied. F 50×2000, zero replay failures, mean max-of-two fee error.000004614 bps, max.000563746(seed 70039/callback 227/t 180). Native 697,644 successful swaps, mean 126380.99, init 72,760. Engine 9510064..9510127 final 0 OOG, but reused during debugging; final+1.4143 is not a blind fresh confirmation. Unbounded packing outside tested market ranges needs saturation. [b04-port-2] Port-2 native worst5: -2.397035298464573, -1.9183903826931896, -1.37002489728917, -1.1464575830912054, -1.0677501521778368; seed IDs not supplied. [b04-port-2] ### Port-3: complete coarse candidate Eight slots hold mean, variance, bid/ask fee, previous spot, rival estimate, timestamp, count; signed WAD. No synced dimension needed by plain lookup. Keeps individual silent-step updates/cap 12, hard/soft thresholds and tiny-mass fallback. ln atanh through 29 th degree; exp[6/6] Padé with coefficients(665280,332640,75600,10080,840,42,1); Hart rational CDF shares PDF, CDF saturation|x|>8/PDF cutoff|x|>10; sqrt 7 Newton iterations. Initialization uses precomputed .002821013442392051 fee each side and log initial reserves, avoiding table duplication. Unchecked arithmetic is for bounded simulation domain, not invalid/zero reserves. [b04-port-3] Own rows every 20 bps(81); phase 0 R -30,-20,-10,0,10,20,30; retail R[-800,-300,-125,-60,-40,-20,0,20,40,60,125,300,800]. Signed BE int16 eighth-bps values,5,346 values/10,692 payload bytes,11,055 bytes after 33-byte PUSH32 blocks andpadding. Max sampled-knot quantization.062000000000001165 bps, additional interpolation loss. Payload SHA `17dba90f09943115242f7674fdf325673620399699318b8e532981901dd5184a`. Source SHA `62930b8e68277c69fc1acfba18de781e684d7fc6daac8dc562b8210dc3932ac1`; runtime SHA `35edacf4f75cf0a9d9e5dda1cdc9421b6cdc3c1aad39c55afc457a08df556ce5`; creation 18,494 bytes. Contract/decoder/harnesses embedded in original report. [b04-port-3] P preceded changes. Q choose full 529.061693433, compressed 528.317594868, compressed−full=-.744098565 SE .017530693, n 1024; worst(60188,-3.115666539),(60809,-2.707236995),(60773,-2.696867820),(60304,-2.528082003),(60831,-2.503687832). Report full 532.385686971, compressed 531.643333344, diff=-.742353627 SE .012367923, n 2048; worst(11610,-2.810731163),(11851,-2.703957499),(10248,-2.674588068),(11005,-2.647019940),(11570,-2.539576106). Exact embedded literal is decoded for scoring. [b04-port-3] F seeds 70000..70049,100,000 swaps/200,000 fee comparisons,245 k cap/call, fresh deployment/run, no state injection. Max 1.39892784856155e-08 bps, mean 9.38885741024225e-12; init max 4.33680868994202e-15. Pyrevm replay peak 156,888, mean 61333.94154; trace hash retained in source. Native engine has different peak 206,161 andmean 95489.07264 over 686,442 swaps; init 141,957. All failure counters zero. [b04-port-3] Fresh original engine 9510128..9510191 once per strategy; port 528.667795972, t03 527.835410173, wins 46/64. Worst(port−t03):(9510147,-1.550038852),(9510189,-1.399749035),(9510154,-1.294403170),(9510157,-1.019170464),(9510177,-.863816555). Fixed before engine: repaired Yul stack issue, duplicated table causing 34,196 byte runtime, and 273,975 gas replica callback using memory work structs, precomputed initialization and shared math. Validation/P/Q/F pass as reported; native peak remains sample evidence. [b04-port-3] ### Spline follow-ups: do not transfer an old engine score to new bytes Spline-1 replaces port-3 lookup while keeping filter. Raw 2,665 byte smooth coefficients plus 8,690 byte bases/W2L=11,355 payload/11,748 prefixedbytes. Final basis records 8 bytes:5-bit first coefficient index, three 19-bit weights scale 524287, fourth residual weight. Each side evaluates at four original grid nodes, clamps node fee before bilinear interpolation, then exponential conversion using original w. Precomputed bases and shared contractions avoid dynamic Cox–de Boor divisions. Final basis SHA `deb9ad29ce5e3f51f099611b8f977c37fcf4d50f208845debb68f37727fde14f`; max grid-node deviation versus original smooth decoder.001028047 bps. [b08-spline-1] Final exact-packed Q: choose diff=-.009897890 SE .012703835 (full 529.061693433, packed 529.051795543); report=-.009736752 SE .008966433(full 532.385686971, packed 532.375950219). Report worst(10359,-1.881228032),(11826,-1.616187069),(10652,-1.503660721),(11741,-1.452449805),(11696,-1.429047246). This agrees with smooth-2d's rounded result. [b08-spline-1] Final replay F compares **original smooth-2d decoded tables**, not the exact packed-basis tables of Q:100 k swaps, noOOG, max.008898410277 bps, mean.000060668892, init 4.34e-15; pyrevm peak 193,623, mean 95754.61154. It is under tolerance against that original reference, but the exact-reference wording of gate F is not established: a direct replay against packed Q tables remains missing. Trace hash retained in source. Final source SHA `27ad2a5f3d2b218138a95f8d4ab616227ca7dfd855a9c60577f86fea09c6caed`, runtime 21,106, creation 21,134, runtime SHA `5644f5c6459497079d6d78b00d58e9b29a13f90bc703a4da7286b7079c5dfd70`. [b08-spline-1; merge gate interpretation] **Engine 9520000..9520127 tested earlier 24-bit version**, source SHA `1b27d1f99bee42d16c7b2b8989f8118f932d7d9913f644a8112798db7a90999d`, runtime 22,117. Its 1,372,160 swaps had 0 OOG, peak 249,342 (fails by 4,342), mean 144683.72628. Port−IB+2.173776453 SE .131962267; port−t03+1.314246680 SE .123842280, n 128 each. Worst tails retained in annex. Final source saves 8,752..14,374 pyrevm gas/call(mean 12,485.64) on seed 70000 replay, but **final native gas and score are unknown**. Earlier 16-bit bases failed F through filter feedback;19-bit passed stated replay. Splitting large literals fixed compilation. [b08-spline-1] Spline-2 is an uncompiled interface/decoder/patch handoff, **not another measured strategy**. Worker says exclusive-write restriction and a prior rejection prevented checkout/setup; this root's scratch permission means that explanation is not a general task blocker. It does not establish inability to implement. No P, Q, F, size, engine/gas run was done, including assigned 9521000..9521127. Dynamic basis oracle in its sketch must be replaced with precomputed packed Q30 bases; actual literal and compiled full Strategy not delivered. It repeats source-derived 2,665 bytes/.0097 loss/206,161 coarse-port gas rather than measuring spline gas. These are proposal details, not completed acceptance evidence. [b08-spline-2; task scratch rule] ### Network follow-ups: same model, different inference and packing Both use port-3 filter and continuous two-pass ask net(w, R, phase), bid net(-w,-R, phase), clamping input axes (phase 0 R±30, retail±800, w±800) but retaining original w for exponential fees. Q14 inputs nearest(16384*[w/100, R/100, phase,0]), floor shifts 12/10/12, ReLU, output z*100/16384 bps. Init offset 28.74755859375 bps→fee.002870627705508484. Thus fixed-net's eight-corner-grid Q **cannot** be inherited by two-pass continuous inference. Both replace bytes-memory loads with PUSH32 code immediates and deliver generators/model references; this differs from port-3's constant-blob layout. [b09-net-1–2] Net-1 packs 4 outputs in 64-bit lanes with 2^63 bias, skips zero ReLU inputs; layer 2 uses 16-bit lanes. Eight-pass build 17,459 bytes, pyrevm peak 357,552(seed 70000),350,947(seed 70001),≈281 kmean, max fee error 1.70e-10 over 2×2000; not shipped. Two-pass runtime 16,368, source SHA `5c40c67a9c5c3c2cb3f1aea95d6f605ef110a87b7674759b65836b5c9ec1609f`. F 100 k calls, max 3.181e-09, mean 1.910e-10, init 1.561e-13 bps; pyrevm peak 197,653, mean 105,179.6. Native 1,389,773 swaps peak 245,377, mean 141,140.96,0 OOG/other failures;**377 over limit**. Engine 9522000..9522127: port 526.439528660, t03 525.211157649, IB 524.430303270; wins 108/128 vs t03,122/128 vs IB; pairedIB+2.009225389 SE .115152456. Q 180-seed report attempt timed out before saving; only choose timing seeds 60000,60001 completed, loss.2058/.5312—not a quality statistic. [b09-net-1] Net-2 packs 6 signed 42-bit lanes per word for layer 1, extracting with signed borrow;32 inputs×6 word-MACs instead of 32×32 scalar-MACs. Max tested grid lane magnitude 2,770,833,733 below 2^41/64; generator asserts bound. Continuous inputs clamp to same box, though an exhaustive continuous-input bound is not independently proven here. Layer 0/2 weights/biases small literals, sd column skipped. Runtime 13,568, creation 13,596; source SHA `879d7ec43ff28dac022ea57e852d7c6b3724ae706415b82600c150b2624620d1`, runtime SHA `9da3ba939c355c9b131860512450bc963844743145152f5f738f70df31b77414`. [b09-net-2] Net-2 eight-pass build 14,374 bytes, pyrevm replica peak 191,177 on 3 seeds, synthetic 3000-call gaps≤40 stress 225,906; two-pass replica 156,399, synthetic 174,717. Its subtask said ship 8-pass if pyrevm≤240 k;**worker shipped 2-pass despite qualifying 8-pass** based on inferred native 8-pass≈273 k. That inference extrapolates coarse native/replay overhead and is not a measured 8-pass engine bound. No 8-pass engine seeds consumed. Net-1's much higher 8-pass cost is a different implementation, not a contradictory measurement of net-2's bytes. [b09-net-1–2] Net-2 P reproduced. Q choose only 14 timing seeds discarded, not 1024; report **10000..11280, n 1281**, full 533.477288126, net 533.457932314, loss+.019355812 SE .014184889. **11281..12047(767 seeds)missing.** Two killed 576/320-seed batches were replaced by streamed replica scoring; this does not consume native engine seeds. Worst report losses:(10822,2.222092),(11005,1.769435),(10735,1.700854),(10359,1.603411),(10122,1.455956). Partial-set mean cannot be directly compared as a paired estimate with fixed-net's 2048-seed grid loss-.039742192; the worker's “about.06 worse/3 SE” is not established on matched samples. [b09-net-2; merge statistical qualification] Net-2 F 100 k swaps/200 k fees, max 3.18e-9 bps, mean 1.91e-10, init 1.56e-13; pyrevm peak 156,399, mean 66,263. Trace hash retained in source also appears for net-1, consistent with same two-pass behavior. NumPy Q forward vs pure Python zero mismatches 20 k random inputs;8-pass 3-seed replay max 4.2e-10 bps. Solidity input rounding is half-away-from-zero, Python half-even; exact ties/float-vs-WAD can differ by one Q14 input step(≤.0061 bps input), not observed in F. [b09-net-1–2] Fresh native 9523000..9523127,1,405,854 successful swaps, peak 209,552, mean 100,510.6, all failure counters zero; port 532.015759491, t03 530.546322976, IB 529.747732439; wins 105/128 vs t03,119/128 vs IB. Paired t03+1.469436515 SE .137313168, IB+2.268027052 SE .168979495. This is promising, not a full Q pass or a fresh confirmation of a frozen net source across later blocks. [b09-net-2] ## Frozen port-3 confirmations and t03/IB baseline All six confirmation results extracted byte-identical port-3 (`62930b8e…2ac1`), t03(`cb6d9455…92d9`) and IB(`30194182…d686`); unchanged source, runtime 18,466/14,654/14,528 respectively. No tuning/reruns,256 assigned seeds per block, eight 32-seed chunks/strategy,24 calls/block,8 workers. They validate native behavior only; none reran P/Q/F. Source warnings affect a1, a2, a4, b1, b2; unflagged a3 corroborates the direction but not their exact numbers. [b10-a1–a4; b11-b1–b2] | Input | Inclusive seeds | port / t03 / IB mean edge | port−t03 / SE | port−IB / SE | Wins t03 / IB | |---|---|---|---|---|---| | a1 |9600000..9600255 |536.582364319 /535.964646020 /535.202426817 |.617718299 /.088286959 |1.379937501 /.086433425 |167 /215 | | a2 |9600256..9600511 |538.635996367 /537.890088643 /537.124958342 |.745907725 /.086837191 |1.511038025 /.090406828 |178 /216 | | a3 |9600512..9600767 |521.894351 /521.362259 /520.577859 |.532092144 /.088902584 |1.316491666 /.092166738 |162 /210 | | a4 |9600768..9601023 |534.602148660 /533.984812710 /533.247754102 |.617335950 /.084013921 |1.354394557 /.092949707 |170 /219 | | b1 |9601024..9601279 |524.519153658 /523.927618148 /523.095312735 |.591535510 /.080163631 |1.423840922 /.097610136 |169 /214 | | b2 |9601280..9601535 |530.007224059410 /529.355056261885 /528.565110600228 |.652167797525 /.089094592960 |1.442113459182 /.113628202599 |171 /215 | Each row n 256, zero ties; worst 5 s/native gas retained in annex. All port OOG/revert/other_halt/tx_error/bad_return counters zero, peak 203,501..206,326; all peers likewise successful. Different peak/mean values across different seeds are not contradictions. [same six sources] **Merge calculation from the six block means/SEs:** pooled n=1536, port−t03 mean **+.6261262375875**, SE **.03520354380573002**; port−IB **+1.404636021697**, SE **.03913350981055256**. Combine within-block SS `(n_j−1)*n_j*se_j²` and between-block SS `n_j*(m_j−m)²`, then divide total by N(N−1) for SE². Rounded inputs limit derived precision. Wins 1017/1536 vs t03,1289/1536 vs IB. Combined worst 5 t03:(9600567,-3.215 rounded),(9600978,-2.936703332),(9601517,-2.767644338529),(9600680,-2.748 rounded),(9600044,-2.677465988); IB:(9601302,-2.911625064610),(9601517,-2.442526658110),(9601471,-2.098586355913),(9601073,-1.998686022),(9601412,-1.956565082432). Original 64 port-3 seeds excluded from this confirmation pool. [merge calculation; b10-a1–a4; b11-b1–b2] The preliminary baseline t03-vs-IB result completed only 144/1024 requested seeds,9500000..9500143. Means t03 534.6038(sd 92.93, SE 7.745), IB 533.8589(sd 93.00, SE 7.750); paired t03−IB+.7448, sd 1.0027, SE .0836. Worst(9500017,-2.6096),(9500088,-1.9770),(9500117,-1.7480),(9500043,-1.2055),(9500127,-1.0736); best(9500003,5.1746),(9500004,4.7400),(9500007,3.0022),(9500079,2.9659),(9500101,2.7729). t03 swap 1,572,040 calls peak 235,394 mean 135,477; IB1,559,056 peak 234,805 mean 135,147; both zero errors; init 243,234/243,227. Early t03 probe 9500000..9500007 repeated inside 144, violating the no-reuse rule; not an additional 8 pairs. **9500144..9501023 were not run.** Worker reported roughly 12 s/seed/strategy and a time-budget constraint; these operational timings are not normalized benchmarks. [b07-t03-vs-ib] ## Classifier research: small replica gains, no engine integration Arb-2 and arb-3 investigate Taylor's claimed arb-or-retail opportunity (~2.7). They relay historical negative results: masked quote zones cost .6–2.4 and retained 13–15% errors; pin seeking above kappa 1e-5 cost more than it returned; quote-aware fill likelihood failed; a feature classifier had 18.8% error versus lattice 16.2%. Taylor's true-answer ceiling is quoted as +6.98 versus lattice/qm +4.26. These are attributed README context, not measured gains from these experiments. Arb-3 has one flagged external source; arb-2 independently relays the context. [b06-arb-2–3] **Arb-2: online sigma and rate.** Estimate sigma² from squared log-price changes between pi≥.99 arb pins divided by elapsed steps, with shrinkage toward 9.45 bps. Estimate first-retail rate from sum(1−pi)/elapsed steps, then set RHO_run=.26*(rhat/.33)^rgam. The argument for an unbiased sigma estimate assumes true pins; it does not prove the actual classifier-selected estimator is unbiased. On choose seeds, estimated rate correlates .88 with true lam, mean .330; report mean .331, correlation .873. Sigma at W=1000 has RMS error .56 bps, correlation .51, mean 9.50 versus true 9.46. Fixed 9.45 has RMS error .36 because the underlying range is narrow. Other moving-sigma variants have RMS .56–.59 and correlation about .5. [b06-arb-2] The jmax variant erroneously drops both a large jump and its elapsed time, biasing sigma downward: RMS 3.6 bps, choose loss -.7858 in candidate-minus-baseline convention, misclassification 17.90%. The prose “sigma hurts at every prior weight” is too strong: W=20000 measures +.0026 SE .0081, effectively neutral, and W 5000 plus rate also +.0026 SE .0115. Material sigma adaptation generally hurts. [b06-arb-2] Selected rate=500,r 0=.33,rgam=1.5: choose +.0219 SE .0123; report **+.0235 SE .0083, n=2048**. Worst report differences: 11701:-3.308, 11699:-2.228, 11570:-1.742, 10949:-1.198, 10866:-1.164. HA comparison rises +2.828→+2.852, both SE .041. First-trade misclassification 16.57→16.52%; all-callback accuracy 89.35→89.38%. Baseline/candidate own swaps 22,135,561/22,137,625; first trades 14,225,904/14,225,557; labeled arb 33.20/33.17%; true arb 34.38% both. Choose baseline error 16.64%, labeled arb 33.27%, true arb 34.47%, all accuracy 89.29%. [b06-arb-2] Oracle sigma+lam diagnostic report +.0133 SE .0105 (worst 11701:-3.473, 11549:-2.657, 11062:-2.341, 11225:-1.885, 11519:-1.824) is not an implementable policy or a strict ceiling: RHO*(lam/.8) is only a crude mapping. All-off/rgam=0 matches baseline exactly on 2,048 seeds. Patch SHA `341951afd984cae87648de94247ca021fcf974fb662771643216ec05339b5981`. The realizable gain is roughly 1% of the quoted opportunity; rate adds a counter/pow and is low priority, sigma is not recommended. No engine or Solidity checks were run. [b06-arb-2] **Arb-3: conditional router inverse.** Use nine midpoint Gaussian price hypotheses over the no-arb interval intersected with mean±6 sd; clip the rival estimate to each hypothesis's 30-bps band. Invert the constant-product router for implied total retail input, apply lognormal density and the Jacobian, and average under normalized Gaussian weights times truncation mass Z. Some variants also update retail posterior moments with likelihood weights. This assumes point rival state, fixed order mean 20 and effective first-retail intensity rate/2; it is not exact inference over hidden rival-only trades. Truth and per-run init parameters do not enter decisions. [b06-arb-3] For input reserve C, own multiplier g, rival G=.997, r=sqrt(g/G), and u=(C_after−C_before)/g, implied total input is u+max(0,C_after/r−C_rival)/G. Its derivative versus absolute log-spot movement is C_after/2 times [1/g plus 1/(rG) when rival fills]. X input and derivative are converted to Y under each price hypothesis. This inversion is exact conditional on the estimated state, not a claim that the estimate is exact. [b06-arb-3] Selected router=1,rate=.55,update=1 (“u 055”): choose +.049632 SE .015940; report **+.049277 SE .011965, n=2048**, baseline 532.385687, candidate 532.434964, approximate normal 95% interval [+.025825,+.072728]. Worst: 11022:-2.773506, 11701:-2.659153, 11570:-2.281416, 10212:-2.213248, 11519:-2.056842. First-decision accuracy 83.425060→83.637171%; all accuracy 89.347729→89.481151%. First mistakes 2,357,935/14,225,904→2,326,664/14,219,204; callbacks 22,135,561→22,119,046. Labeled arb across all callbacks 33.200812→33.524818%, and among first decisions 51.660590→52.150388%. [b06-arb-3] Arb-positive TP/FP/TN/FN: baseline [6301005,1048181,13476616,1309759], candidate [6344478,1070892,13447899,1255777]. Quotes change routing, so accuracy samples differ even on paired seeds. Synthetic inversion recovered 11,675 positive-own-fill cases from 20,000 generated cases, max relative error 2.7638649228662535e-13; router-off and truth-neutralized checks matched baseline on 60000..60002; 24 quadrature cases had finite likelihoods, means in bounds and nonnegative variance. These do not establish quadrature convergence or correct hidden-state conditioning. Patch SHA `ed4fde8ae42646e7d11181c257f1845f21d592338b76fca26a98d9eacaf1b1b7`. No compressed Q, Solidity F, engine, t03 or gas checks. [b06-arb-3] Do not add these two gains together or to a port's engine gain: combined policies were not tested. Whole-lattice-posterior quoting and marginalizing unseen rival-only retail state remain proposed next experiments, not results. [b06-arb-2–3; merge limitation] ## Math support: measured libraries, not strategy gates Both math items embed complete standalone libraries, harnesses and alternatives. Neither runs P/Q/F, compresses policy tables or scores versus t03. Different compiler/EVM/wrapper/gas definitions prevent direct format ranking by raw gas. Port-3 uses its own Hart/Padé routines; these libraries were not integrated into it. [b05-math-wad; b05-math-q64; b04-port-3] **WAD:** signed int256, scale 1 e 18. ln accepts every positive representable input, rejects nonpositive; exp safe through 130, zero at/below -42, rejects above 130; erf saturates at |x|≥6; pdf zero at |x|≥10; cdf saturates outside ±9 (or earlier through erf). Chosen ln is normalized atanh/Horner through z^23 versus 64 binary squarings; exp [6/6] Padé versus 16-term Taylor; erf uses six unit intervals, degree-18 erfc interpolation (114 coefficients regenerated at Decimal precision 95), versus A&S 7.1.26; pdf uses selected exp, cdf the selected erf identity. [b05-math-wad] WAD measurements: Cancun, optimizer 200, no viaIR; 20,001 points/method, ln [1e-6,1 e 6], exp [-40,40], erf [-6,6], pdf/cdf [-8,8]. All 200,010 selected/alternative EVM executions matched the integer Python model; 90 boundary, 228 join/saturation, six revert and ten bundle checks passed. Binary 64 baselines include conversion rounding; no formal global accuracy/monotonicity proof. Execution gas subtracts intrinsic/calldata gas but retains ABI/dispatch costs. [b05-math-wad] Selected max absolute errors: ln 1.77635684e-15, exp 800, erf/pdf/cdf 1.11022302e-16. Exp's 800 occurs near huge exp(+40); relative error at tiny exp(-40) reaches .197671627 (~19.8%), shared by Taylor. At x≥-20 max relative error is 4.23663826e-10. Full-range PDF/CDF max relative errors 2.32433625e-4 /7.55602298e-4, but PDF |x|≤6 gives 1.90144707e-10, CDF x≥-6 gives 1.27022085e-9. A&S max erf/CDF absolute errors 1.39374503e-7 /6.96874125e-8. Quantization and joins can introduce raw-unit artifacts; saturation intentionally drops tails. [b05-math-wad] WAD five-of-each bundle transaction gas min/mean/max 47,494/51,937.2/53,665, execution 26,230/30,673.2/32,401, all successful at 245 k. Selected wrapper runtime 6002 bytes, combined 6617. Function gas comparisons select the polynomial/Padé methods over alternatives; detailed per-function timings remain in the input. Math-only headroom is not an afterSwap measurement. [b05-math-wad] **Q64.64:** signed int128, scale 2^64, int256 intermediates; ln all positive int128; exp [-40,40], pdf [-8,8], otherwise explicit rejection; erf/cdf all int128 with saturation. Selected ln has eight atanh terms versus 48 binary squarings; exp [5/5] Padé coefficients (30240,15120,3360,420,30,1) versus 12-term Taylor; erf uses a fitted degree-16 scaled complementary tail versus 16-node Gauss–Legendre quadrature; cdf uses the complementary tail directly to avoid negative-tail subtraction. Frozen coefficients sum exactly to Q, preserving erf(0)=0, cdf(0)=.5; refits can vary by numerical-library version. Signed right shifts floor, divisions truncate. [b05-math-q64] Q64 measurements: Shanghai, viaIR, optimizer 200, 10,001 inputs/method on the same real domains; rerun reproduced 100,010 accuracy and 100 gas calls, boundary/rejection checks and runtime size. Selected ln/exp/erf/cdf were monotone on sampled grids; quadrature erf/cdf were not. Selected max absolute errors: ln 1.24345e-14, exp 192, erf 6.713962e-11, pdf 2.775558e-16, cdf 3.363132e-11. Taylor exp max 32 at greater gas cost. Quadrature erf/CDF max 1.56849e-11 /7.842429e-12, but CDF tail max relative error 11631.68 makes it unsuitable despite smaller absolute error. [b05-math-q64] At exp(-40), roughly 78 raw Q64 units cause max relative error .01263029 (~1.263%) for both exp methods. Selected PDF/CDF max relative errors 1.252073e-5 /7.014483e-5. R 8 restricts |baseline|≥1e-8: selected CDF max 6.804441e-11, exp 4.729183e-12. Near-zero floating samples explain large unfiltered log/erf relative errors; exact ln(1)=0 and erf(0)=0 hold. At |erf input|=6 omitted tail is about 2.15e-17. [b05-math-q64] Q64 selected inner maxima for ln/exp/erf/pdf/cdf are 1713/699/1404/1012/1509 gas, alternatives 6325/2211/10684/2524/11009. Inner gas includes selection/measurement, excludes ABI/transaction costs. All-alternative wrapper 2921 bytes. Sum of selected maxima 6337(one each),31685(five each) is an estimate, not a measured combined callback. Complete source and detailed benchmark tables remain in the input; no linking, assembly, external calls or storage are required. [b05-math-q64] ## Reconciliation, missing evidence and handoff 1. **A completed port need not be lossless.** Port-3's positive confirmed engine gain and .742 replica loss can both hold. The engine comparator is t03; the compression comparator is full Gaussian/plain. [b04-port-3; b10-*; b11-*] 2. **Good isolated Q does not establish a deployable strategy.** Sym-int8, smooth spline and fixed-net use different representations. Spline deployment adds basis data; continuous-net deployment changes eight grid passes to two continuous passes. Gas and F must follow exact bytes and inference semantics. [b02-sym-int8; b03-*; b08-spline-1; b09-*] 3. **Gate gaps:** SVD's aggregate score is not strict P; port-1 lacks exact Q; port-2 changes filtering and reuses engine seeds; port-3 reports P/Q/F; final spline has Q but an old engine version and a different F reference; net-1 lacks Q and fails peak gas; net-2 has partial Q and overrides its selection rule. Spline-2 is a proposal. Finite seeds establish observed gas peaks, not a universal bound. [corresponding inputs] 4. **Unresolved numerical discrepancy:** fewer-cols reports full mean 532.385688969, about .000002 above the 532.385686971 in reference/rows/port-3. It reports loss .747741346, also .747741345529 at higher precision. No explanation is supplied; retain its figures as attributed rather than normalize them. Compression loss generally means full minus decoded; port Q comparisons often use the opposite sign, explicitly labeled here. [b02-fewer-cols; b01-ref-2; b04-port-3] 5. **Source warnings remain unresolved.** Hash checks verify embedded identity, not quotations or behavior. Disjoint local confirmation reports, including unflagged a3, provide stronger additional port-3 evidence than repeated external assertions. [task warning; merge checks; b10-a3] 6. **Offline boundary:** Markdown often embeds complete contracts and data, but not the native wheel, Python environment, Taylor modules or full tables. Port-1 and rows require regeneration from pinned JSON; SVD/smooth/fixed-net decoders mention NumPy; net references require port-3's filter and model bytes. No offline end-to-end toolchain reconstruction is established. A later implementation must retain required source/dependencies rather than rely on deleted scratch or remote artifact availability. [b02-*; b03-*; b04-*; b08-*; b09-*] 7. **Consumed native seeds within this root:** baseline 9500000..9500143 (first eight t03 seeds repeated); port-1 9510000..9510063; port-2 9510064..9510127 (debug reuse); port-3 9510128..9510191; old spline-1 9520000..9520127; net-1 9522000..9522127; net-2 9523000..9523127; confirmations 9600000..9601535. Spline-2's 9521000..9521127 and baseline 9500144..9501023 were not run. This is not a global seed ledger for other roots. [b04-*; b07-*; b08-*; b09-*; b10-*; b11-*] Retain frozen port-3 as the completed baseline and net-2, final spline and sym-int8 as separate improvement paths. Fresh final-source engine checks, exact-packed-reference spline F, complete continuous-net Q and offline dependency packaging remain work, not results established here. No requester-only value is missing. [merge conclusion from all inputs] ## Numerical annex: retained trials and tail outcomes Transcribed trials retain negative results. Signs and seed conventions follow the main text; complete raw data remain in the ledger artifacts. ### b02-sym-int8 — choose trials | Setting | Retained rows | Bytes, entire blob | Choose-set loss | SE | |---|---:|---:|---:|---:| | uniform20 | 81 | 10,024 | +0.322883175 | 0.014155512 | | adaptive100 | 81 | 10,024 | -0.001281364 | 0.010820910 | | adaptive180 | 97 | 11,928 | -0.000867544 | 0.010711970 | ### b02-rows-10bps — choose trials | Row spacing | Rows kept | Bytes | Choose-set loss | SE | |---|---:|---:|---:|---:| | 5 bps | 321 | 142,524 | 0.010785683 | 0.010451818 | | 10 bps | 161 | 71,484 | 0.013334441 | 0.012074033 | | 20 bps | 81 | 35,964 | 0.323266845 | 0.014140580 | ### b03-svd-lowrank — choose trials | setting (k for A0,A1,A2; fee-aware fit) | bytes | loss | se | |---|---|---|---| | 3,3,3 | 6,463 | 1.331 | 0.026 | | 5,5,5 | 10,759 | 0.522 | 0.022 | | **4,6,6** | **11,571** | **0.468** | **0.020** | ### b03-smooth-2d — choose trials | setting | degree | knots requested | bytes | choose-set loss | se | |---|---|---|---|---|---| | s6x4 | 3 | 6x4 / 6x4 | 515 | +0.2087 | 0.0172 | | q8x6 | 2 | 8x6 / 8x6 | 599 | +0.0968 | 0.0177 | | s8x6 | 3 | 8x6 / 8x6 | 719 | +0.0867 | 0.0162 | | s12x8 | 3 | 12x8 / 12x8 | 1,091 | +0.0519 | 0.0155 | | s16x10 | 3 | 16x10 / 16x10 | 1,609 | +0.0366 | 0.0133 | | **s24x14** | 3 | 24x10 / 24x14 | **2,665** | **+0.0098** | 0.0127 | | s32x20 | 3 | 32x10 / 32x20 | 4,163 | +0.0102 | 0.0110 | ### b03-fixed-net — choose trials | Setting | Bytes, including axes/header | Choose-set loss | SE | |---|---:|---:|---:| | 16×16; 30k updates, constant LR | 906 | 0.351053120 | 0.016554603 | | 24×24; 100k updates, constant LR | 1690 | 0.281773385 | 0.014280382 | | 32×32; 150k updates, decaying LR | 2730 | 0.246807199 | 0.015961022 | | 24×24; 100k + 50k refinement | 1690 | 0.172467132 | 0.014762306 | | 32×32; 150k, fixed positive-w path **chosen** | 2730 | -0.056134271 | 0.015645435 | | 32×32; 120k, three phase output heads | 2866 | 0.019048078 | 0.014389945 | | 32×32; 120k, diagonal sampling/input basis | 2730 | 0.301765564 | 0.015335929 | ### b02-fewer-cols — every choose setting Column selections and all choose worst 5 lists remain in that input. The table below retains bytes, loss and SE for every tested setting; the chosen layout is specified above. | Setting | Bytes | Choose loss | SE | |---|---:|---:|---:| | alloca | 11572 | 0.983444637 | 0.023336074 | | allocb | 11572 | 1.111001752 | 0.025468762 | | allocc | 11572 | 1.647015026 | 0.027641001 | | allocd | 11572 | 3.456053459 | 0.032339807 | | alloce | 11572 | 1.102384774 | 0.023929140 | | allocf | 11572 | 1.063649179 | 0.023689498 | | central13 | 37272 | 0.446164205 | 0.016206938 | | central5 | 16712 | 0.729181154 | 0.020116534 | | central7 | 21852 | 0.396806283 | 0.017619461 | | central9 | 26992 | 0.525851553 | 0.016681546 | | d10 | 11572 | 2.155177298 | 0.031026489 | | d12 | 11572 | 3.014817972 | 0.035773113 | | d2 | 11572 | 4.557504254 | 0.047237780 | | d3 | 11572 | 2.180578964 | 0.034630744 | | d4 | 11572 | 1.173825281 | 0.028115312 | | d5 | 11572 | 0.949462914 | 0.025620292 | | d6 | 11572 | 0.978907182 | 0.023225898 | | d8 | 11572 | 1.368311948 | 0.026052619 | | k13 | 50122 | 0.830544558 | 0.015969491 | | k5 | 19282 | 3.205925028 | 0.034127319 | | k7 | 26992 | 1.744760748 | 0.022854633 | | k9 | 34702 | 1.300247179 | 0.019656578 | | malloca0.5 | 11572 | 1.789592695 | 0.023337329 | | malloca1.0 | 11572 | 0.772461201 | 0.019402465 | | md40.5 | 11572 | 2.160351970 | 0.025106535 | | md41.0 | 11572 | 0.919994371 | 0.021186051 | | md50.5 | 11572 | 1.900374014 | 0.023964846 | | md51.0 | 11572 | 0.774686536 | 0.019856688 | | md60.5 | 11572 | 1.866917383 | 0.022452058 | | md61.0 | 11572 | 0.743289135 | 0.020053542 | | mr10.5 | 11572 | 1.858637807 | 0.025367219 | | mr11.0 | 11572 | 0.802556668 | 0.020422064 | | mr20.5 | 11572 | 1.766029962 | 0.024363314 | | mr21.0 | 11572 | 0.760440067 | 0.019791461 | | q05 | 11572 | 2.954963352 | 0.037038175 | | q10 | 11572 | 1.344916197 | 0.027490182 | | q20 | 11572 | 0.982004345 | 0.023760793 | | quantall | 142642 | 0.010781585 | 0.010452205 | | r1 | 11572 | 0.986813183 | 0.026532624 | | r2 | 11572 | 0.976335176 | 0.025201628 | | r3 | 11572 | 1.019429773 | 0.025601529 | | r4 | 11572 | 1.253385428 | 0.028949908 | | r5 | 11572 | 1.097635502 | 0.027040405 | | r6 | 11572 | 1.059304737 | 0.025552858 | | r7 | 11572 | 1.012775993 | 0.025702096 | | r8 | 11572 | 1.545718878 | 0.028831725 | ### b06-arb-2 — every choose setting Differences are candidate minus baseline, n=1024 each. Oracle rows are diagnostics, not runnable strategies without truth. Per-setting classifier and worst 5 details remain in the source; aggregate classifier findings are retained above. | Setting | Mean diff | SE | |---|---:|---:| | `sig=200` | -0.0632 | 0.0159 | | `sig=1000` | -0.0336 | 0.0138 | | `sig=5000` | -0.0096 | 0.0102 | | `rate=500,r0=0.33` | +0.0110 | 0.0092 | | `rate=2000,r0=0.33` | +0.0066 | 0.0082 | | `rate=500,r0=0.33,rgam=0.5` | +0.0079 | 0.0077 | | `sig=1000,rate=500,r0=0.33` | -0.0036 | 0.0134 | | `sigoracle=1 (oracle diagnostic)` | +0.0131 | 0.0103 | | `lamoracle=1 (oracle diagnostic)` | +0.0284 | 0.0136 | | `lamoracle=1,rgam=2 (oracle diagnostic)` | -0.0196 | 0.0163 | | `sigoracle=1,lamoracle=1 (oracle diagnostic)` | +0.0382 | 0.0147 | | `sig=1000,pinmin=0.999,jmax=0.006` | -0.7858 | 0.0269 | | `rate=100,r0=0.33` | +0.0123 | 0.0096 | | `rate=200,r0=0.33` | +0.0156 | 0.0093 | | `rate=1000,r0=0.33` | +0.0123 | 0.0088 | | `rate=500,r0=0.33,rgam=1.5` | +0.0219 | 0.0123 | | `lamoracle=1,rgam=0.5 (oracle diagnostic)` | +0.0119 | 0.0117 | | `sig=5000,pinmin=0.999` | -0.0634 | 0.0133 | | `sig=20000` | +0.0026 | 0.0081 | | `sig=5000,rate=500,r0=0.33` | +0.0026 | 0.0115 | ### b06-arb-3 — every choose setting | Setting | Mean diff | SE | |---|---:|---:| | `r08`: `router=1,rate=0.8` | -0.209470 | 0.018294 | | `r055`: `router=1,rate=0.55` | +0.041717 | 0.015955 | | `mix`: `router=1,rate=0.8,blend=0.5` | -0.020950 | 0.015690 | | `u08`: `router=1,rate=0.8,update=1` | -0.194762 | 0.018466 | | `u055`: `router=1,rate=0.55,update=1` | +0.049632 | 0.015940 | ### Compression report worst 5 (full minus decoded) | Input | Worst five seed:loss | |---|---| | b02-sym-int8 | 10429:1.317202325; 10177:1.141135028; 10041:1.036806884; 11826:1.018917236; 10231:.987694005 | | b02-rows-10bps | 11701:6.746369940; 11699:2.081707520; 11311:1.429714550; 12009:1.280929512; 10359:1.228352947 | | b02-fewer-cols | 10633:3.148334719; 11701:3.027553196; 10942:2.934882744; 11558:2.904672827; 10103:2.771785058 | | b03-svd-lowrank | 11826:3.349; 10132:2.942; 10772:2.900; 11610:2.752; 11770:2.340 | | b03-smooth-2d | 10359:1.881; 11826:1.616; 10652:1.504; 11741:1.453; 11696:1.429 | | b03-fixed-net | 10632:3.756389833; 11701:3.199589945; 10822:2.023092355; 10735:1.665709620; 11705:1.433977657 | ### Confirmation worst 5, native engine Entries are seed:port-3−comparator, worst five of each 256-seed block. | Input | Comparator | Worst5 | |---|---|---| |b10-a1|t03|9600044:-2.677465988; 9600078:-2.631023953; 9600111:-2.483006736; 9600164:-2.450307732; 9600063:-2.271666978| |b10-a1|IB|9600118:-1.510317512; 9600078:-1.232227586; 9600044:-1.206703253; 9600070:-1.203985654; 9600063:-1.172264658| |b10-a2|t03|9600362:-2.556215850; 9600443:-2.307139497; 9600466:-2.191836339; 9600472:-2.123359035; 9600478:-1.978871573| |b10-a2|IB|9600411:-1.674005622; 9600466:-1.418228706; 9600433:-1.386732031; 9600321:-1.386459038; 9600473:-1.375364151| |b10-a3|t03|9600567:-3.215; 9600680:-2.748; 9600708:-2.161; 9600589:-2.082; 9600599:-2.008| |b10-a3|IB|9600519:-1.766; 9600567:-1.606; 9600665:-1.573; 9600708:-1.453; 9600680:-1.391| |b10-a4|t03|9600978:-2.936703332; 9600881:-2.160652881; 9601009:-2.017120337; 9600833:-1.843099339; 9600853:-1.813771889| |b10-a4|IB|9600978:-1.883011982; 9600775:-1.836834331; 9600883:-1.195045610; 9601009:-1.157687437; 9600870:-1.106520604| |b11-b1|t03|9601188:-2.337912933; 9601073:-2.329408262; 9601141:-1.954440988; 9601067:-1.843126514; 9601139:-1.663897409| |b11-b1|IB|9601073:-1.998686022; 9601170:-1.956481524; 9601188:-1.697546471; 9601139:-1.475476598; 9601120:-1.433372088| |b11-b2|t03|9601517:-2.767644338529; 9601289:-2.178502796056; 9601358:-2.103664486272; 9601325:-2.097881448749; 9601476:-1.910973878263| |b11-b2|IB|9601302:-2.911625064610; 9601517:-2.442526658110; 9601471:-2.098586355913; 9601412:-1.956565082432; 9601289:-1.647187843145| ### Confirmation native Submission afterSwap gas Every listed call succeeded and OOG=0. Each cell is calls / peak / mean gas. These values are not pyrevm replay measurements. | Input | Port-3 | t03 | IB | |---|---|---|---| | b10-a1 |2,756,066 /204,472 /95428.994888 |2,791,556 /238,223 /135485.392048 |2,770,049 /237,663 /135142.138937 | | b10-a2 |2,772,794 /204,932 /95386.547951 |2,806,292 /235,527 /135404.784329 |2,783,219 /234,958 /135078.731757 | | b10-a3 |2,719,781 /203,501 /95,569 |2,761,007 /235,527 /135,676 |2,740,575 /234,958 /135,344 | | b10-a4 |2,754,680 /206,273 /95426.647665 |2,792,831 /235,405 /135485.908565 |2,769,897 /234,977 /135150.387892 | | b11-b1 |2,741,584 /205,740 /95479.811536 |2,779,692 /237,493 /135553.079583 |2,757,297 /236,928 /135217.731441 | | b11-b2 |2,758,853 /206,326 /95426.976408674 |2,794,012 /235,488 /135477.839282723 |2,771,458 /234,885 /135138.409199418 | ### Additional follow-up native tails Each row is candidate minus comparator, from the named input; all n=128. Spline rows belong to old v 24, not the final source. | Input/pair | Worst5 seed:difference | |---|---| | b08-spline-1 v24−t03 |9520062:-1.464768; 9520067:-1.360214; 9520090:-1.319751; 9520121:-1.133484; 9520057:-.997555 | | b08-spline-1 v24−IB |9520067:-1.173812; 9520090:-1.092461; 9520075:-.641801; 9520121:-.638198; 9520062:-.431088 | | b09-net-1−t03 |9522100:-1.682760066; 9522091:-1.131634696; 9522115:-1.076202058; 9522059:-.985105541; 9522092:-.683785675 | | b09-net-1−IB |9522100:-1.104893790; 9522115:-.718558082; 9522091:-.698284662; 9522054:-.182632093; 9522042:-.079437268 | | b09-net-2−t03 |9523065:-1.841713; 9523027:-1.389749; 9523099:-1.247645; 9523017:-1.103238; 9523015:-1.084954 | | b09-net-2−IB |9523065:-1.567600; 9523053:-1.559949; 9523002:-.896067; 9523015:-.730890; 9523017:-.602271 | ## Input ledger Labels identify supplied `.imd/reads/artifacts/