# Root merge 5 of 6: engine confirmation and replica inference experiments This merge covers all 25 supplied JSON results; none is missing or empty. Statements below are attributed worker reports unless explicitly labelled merge calculations. This is a synthesis, not a new engine run, contract, deployment, or leaderboard entry. Source IDs resolve in the register at the end; a section and its tables inherit the source named in its heading. The embedded tg-2 tellpol.py was inspected as data; no embedded code was executed. Raw score rows, full gas snapshots, encoded code payloads, repetitive audit hashes and routine run logs are condensed rather than reproduced; scientific outcomes, negative findings, missing coverage and qualifications are retained. ## Outcome and evidence boundaries The ten engine results cover consecutive fresh seeds **12007680..12009855**, **2,176 pairs per comparison**. Pooling their reported sums and sumsquares gives **M09−IB +3.84316931613 (SE 0.0317036258107)**, **PB−IB +3.37233779213 (SE 0.0315825269442)** and **M09−PB +0.470831523997 (SE 0.011780212595)**. These are merge calculations from the ten engine sources in the next section. Every source reports completed allocations, valid frozen contracts, successful unchanged-source validation, and zero callback faults on its measured runs. The largest measured Submission.afterSwap gas across them is M09 **219,533**, IB **237,040**, PB **215,849**. The local set exceeds the +2.7 mean-gain target but contains fewer than the required 4,000 seeds; evidence in other roots must be combined before declaring that goal-wide gate passed. No source here establishes a new board rank or the requested 20-run Python/Solidity fee parity gate. On the replica, the strongest new mechanism is the **IEEE-754 reserve-grid tell**: TellGauss is +3.19108085952 versus Gaussian+qm, +2.69618031874 versus lattice+qm, and −0.0257565316385 versus the true-flag Gaussian diagnostic on 2,048 holdout seeds. Both Gaussian tell workers agree on the aggregate rows at their reported precision. TellLattice adds only +0.0110978313138 (SE 0.00414716034935) over TellGauss and remains −0.0146587003247 versus flag Gaussian. Sources: [b54-tg-1](#b54-tg-1), [b54-tg-2](#b54-tg-2), [b55-tl-1](#b55-tl-1). These are blind callback policies on floating-point replica inputs, but the tell has **no measured engine or Solidity transfer**; preserve that as the key unresolved issue rather than treating near-oracle replica performance as a shipped solution. The other Track B ideas do not clear +0.3 incremental edge over their inherited controls. Posterior-value integration loses; hidden-rival transition helps but by less than +0.3; its survivor-mass accounting needs repair but the repair has no clear incremental score gain. Tail-risk quotes, uniform-cell density, integer recentering, ambiguity-aware IG2, depth rescaling and silent-step no-fill evidence show negligible or negative isolated gains. Individual sources and coverage follow; gains versus lattice that include thr=.9 are not credited to the new mechanism. The requester supplied the board context: IB **522.00**, 20th; 10th **524.66** (+2.7 over IB in the task wording); first **528.40** (+6.4). The [pinned solved README](https://raw.githubusercontent.com/gtmcknight/amm-sim/e65314f88de945f6d055bfc2aab01c482ab6570e/research/solved/README.md), read for context, instead writes the top-ten gap as +2.66 and the engineering bar as +2.7. Its replica anchors are solved perfect information +8.70 over HA, Gaussian+qm +3.77, lattice+qm +4.26, true-flag +6.98; IB is +0.592 SE .030 over HA on 512 earlier engine seeds. Those historical numbers are context, not new observations in this root; replica absolute scores cannot be converted into board scores. The README also distinguishes the 250,000 call limit, the 245,000 afterSwap acceptance margin and 32 storage slots. ## Source reliability The requester reports failed source rechecks for the items below. The supplied task gives counts, but not the precise failed URL/passage or which of unreadable, changed, or unsupported quotation applied. Accordingly, the numerical observations remain attributed to the supplied JSON, while external quotations, source-provenance assertions and code-level explanations in these flagged results are **not independently corroborated by those citations**. Agreement between workers on the same deterministic seeds is reproducibility evidence, not an independent statistical sample. Worker claims such as “all source quotes verified” do not override the later requester warning. No missing passage is reconstructed. | Result | Failed source rechecks reported by requester | |---|---:| | [b45-eng-12008704](#b45-eng-12008704) | 1 | | [b49-fresh7-1](#b49-fresh7-1) | 1 | | [b49-fresh7-2](#b49-fresh7-2) | 2 | | [b51-mass8-2](#b51-mass8-2) | 1 | | [b52-ig8-1](#b52-ig8-1) | 3 | | [b52-ig8-2](#b52-ig8-2) | 2 | | [b53-fresh8-1](#b53-fresh8-1) | 1 | | [b53-fresh8-2](#b53-fresh8-2) | 1 | | [b54-tg-1](#b54-tg-1) | 1 | | [b54-tg-2](#b54-tg-2) | 1 | For engine eng-12008704, this particularly limits external method/source provenance, not the fact that its JSON contains local edge rows and live-gas claims. For the fresh, mass, IG2 and tell items it limits the claimed external grounding of mechanisms or validation. Full and partial observations are retained with those qualifications. No external failed citation is used here as independent proof of engine implementation details. ## Track A: frozen-contract engine evidence All ten workers identify engine commit `6b6921f8d463314aa7c96234e1be25a25245f704`, released wheel SHA-256 `aedb68280aca34c110968f00148387c95885207913fb2cf19a1ab4c5cfd6887a`, and native extension `ff88b6353c2aad7342f0a26194f7347e78b9c3b38d683a7449006d423978cd14`. Frozen compiler settings are Solidity 0.8.24, optimizer enabled/runs=200, viaIR, Paris. Each submission runs alone against VanillaStrategy via `run_batch(X,VanillaStrategy,configs,8)`; pairing is on identical absolute-seed configurations, not a direct M09-versus-IB match. Config construction uses NumPy default_rng(absolute_seed), draws retail size, arrival rate, then sigma, and checks the 12 fields against MatchRunner construction at indices 0..7. Reported base settings include 10,000 steps, initial price=100, X=100, Y=10,000; size mean varies 19..21, arrival .6..1.0, sigma .000882.. .001008, retail log-sd 1.2, buy probability .5, dt=1, drift=0. Source: all ten engine JSONs, with detailed baseline values in b44-eng-12007680 and b44-eng-12007936. The source/runtime SHA-256 identities agree across the engine sources: | Arm | Source SHA-256 | Runtime SHA-256 | Runtime bytes | |---|---|---|---:| | ib | `301941829a44bb60977bb28181a5e4bfa6544f3696763f987054789b2611d686` | `a4f70330df778c7a4c916170f1b3036522789028ae2dfd8f715b93ff7f24ab2d` | 14528 | | m09 | `a44cca4b0155ca3953f27ccb6b43ac8ee74b1cdc4f27022ee5fc51d8ab67ed5f` | `e95f44d13169747859490d2f5fb627a55b4a9cb3c84dae0c95a6b37e0023e281` | 21984 | | pb | `5123bb39594add4ad9a32f376ff7aa89c13428b82ea324ae4d80c474e3162d95` | `86e512cfac1eca4852143103c93a73ac52c96bb0bf70b82af90f7f1e80935e0e` | 21915 | | vanilla | `62d40ccef0de99a385e332a36ddf9cbd1da73eda06bd7954492a9880cb91a572` | `69ed6562da17454439bad4fbe3690194539001ef8ca2068ea50d937ee50c670b` | 640 | Frozen source provenance reported by the workers: [IB/PB audit artifact](https://api.imdnotdead.fun/artifacts/db67c7a0e8fa98f87b069766df6e213890861b15f967695293f519ecc46eeeb6), uncompressed tar `feb0611a5da584cde367b0018b4cc5a3e36e0ee40523bb5c0d2da0f5ade3b551`; [M09 artifact](https://api.imdnotdead.fun/artifacts/063270cb0315aaded47d5cb3f437297bf8e9aac55e179fc4d26534a6d0e269d3), tar `974c242b4c8d250bdd33d2de9e842cdcc2b517e715f665a32c306cc72c002e73`. PB is reconstructed from audit Q2.sol plus PB.lines.json and hash-checked; those are reported provenance checks, not a new source audit in this merge. The original IB entry in the task is [afd5183e…](https://api.imdnotdead.fun/artifacts/afd5183e9ff3235334e041c00b0cd19079599a2c6464db95f8c1586712105557). ### Per-block paired statistics Differences are submission edge A minus submission edge B on common seeds. All sums and sumsquares below are copied from the corresponding result; decimal display uses 12 significant digits. Each block from b44/b45 has 256 seeds, each b50 block 64. Allocations are non-overlapping and contiguous in the listed order. | Source / seed start | Pair | n | Mean | SE | Sum | Sumsq | |---|---|---:|---:|---:|---:|---:| | [b44-eng-12007680](#b44-eng-12007680) | M09-IB | 256 | 3.95633890443 | 0.0915080860987 | 1012.82275953 | 4553.70716958 | | [b44-eng-12007680](#b44-eng-12007680) | M09-PB | 256 | 0.449451105454 | 0.0346976413927 | 115.059482996 | 130.305921879 | | [b44-eng-12007680](#b44-eng-12007680) | PB-IB | 256 | 3.50688779898 | 0.0908410571681 | 897.763276538 | 3687.05201659 | | [b44-eng-12007936](#b44-eng-12007936) | M09-IB | 256 | 3.83932345854 | 0.0857149472274 | 982.866805387 | 4253.15914874 | | [b44-eng-12007936](#b44-eng-12007936) | M09-PB | 256 | 0.49379395932 | 0.033144643596 | 126.411253586 | 134.135593225 | | [b44-eng-12007936](#b44-eng-12007936) | PB-IB | 256 | 3.34552949922 | 0.0862434073171 | 856.455551801 | 3350.84507728 | | [b44-eng-12008192](#b44-eng-12008192) | M09-IB | 256 | 3.73803633697 | 0.10232882391 | 956.937302265 | 4260.62557395 | | [b44-eng-12008192](#b44-eng-12008192) | PB-IB | 256 | 3.28006741362 | 0.101685427644 | 839.697257888 | 3429.25399493 | | [b44-eng-12008192](#b44-eng-12008192) | M09-PB | 256 | 0.457968923347 | 0.0333126227262 | 117.240044377 | 126.135525668 | | [b44-eng-12008448](#b44-eng-12008448) | M09-IB | 256 | 3.88680857219 | 0.103324329742 | 995.022994481 | 4564.38777385 | | [b44-eng-12008448](#b44-eng-12008448) | PB-IB | 256 | 3.35928768136 | 0.103758339003 | 859.977646428 | 3591.70327523 | | [b44-eng-12008448](#b44-eng-12008448) | M09-PB | 256 | 0.527520890831 | 0.0336089905091 | 135.045348053 | 144.977196093 | | [b45-eng-12008704](#b45-eng-12008704) | M09-IB | 256 | 3.85428162154 | 0.0794144083392 | 986.696095115 | 4214.70262333 | | [b45-eng-12008704](#b45-eng-12008704) | M09-PB | 256 | 0.469333681053 | 0.033822776378 | 120.14942235 | 131.069198249 | | [b45-eng-12008704](#b45-eng-12008704) | PB-IB | 256 | 3.38494794049 | 0.0781993098953 | 866.546672765 | 3332.41119672 | | [b45-eng-12008960](#b45-eng-12008960) | M09-IB | 256 | 3.87763748643 | 0.0879914089344 | 992.675196525 | 4354.66417355 | | [b45-eng-12008960](#b45-eng-12008960) | PB-IB | 256 | 3.42454974565 | 0.0895974984739 | 876.684734887 | 3526.29950779 | | [b45-eng-12008960](#b45-eng-12008960) | M09-PB | 256 | 0.453087740775 | 0.0352347006908 | 115.990461638 | 133.597940402 | | [b45-eng-12009216](#b45-eng-12009216) | M09-IB | 256 | 3.91443519933 | 0.0950922787811 | 1002.09541103 | 4512.93465809 | | [b45-eng-12009216](#b45-eng-12009216) | M09-PB | 256 | 0.438077228427 | 0.038003996534 | 112.147770477 | 143.413533432 | | [b45-eng-12009216](#b45-eng-12009216) | PB-IB | 256 | 3.47635797091 | 0.0945516044207 | 889.947640552 | 3677.38015898 | | [b45-eng-12009472](#b45-eng-12009472) | M09-IB | 256 | 3.65199918153 | 0.0861019809512 | 934.911790472 | 3898.25371097 | | [b45-eng-12009472](#b45-eng-12009472) | PB-IB | 256 | 3.19917085835 | 0.0846088914891 | 818.987739737 | 3087.39933011 | | [b45-eng-12009472](#b45-eng-12009472) | M09-PB | 256 | 0.452828323183 | 0.0331446286484 | 115.924050735 | 124.208108641 | | [b50-eng-12009728](#b50-eng-12009728) | M09-IB | 64 | 4.00257713111 | 0.200858851064 | 256.164936391 | 1187.98804529 | | [b50-eng-12009728](#b50-eng-12009728) | PB-IB | 64 | 3.47710823309 | 0.196286633901 | 222.534926918 | 929.124707291 | | [b50-eng-12009728](#b50-eng-12009728) | M09-PB | 64 | 0.525468898025 | 0.0673154493327 | 33.6300094736 | 35.9420067251 | | [b50-eng-12009792](#b50-eng-12009792) | M09-IB | 64 | 3.78973657332 | 0.209429247832 | 242.543140693 | 1096.0205898 | | [b50-eng-12009792](#b50-eng-12009792) | PB-IB | 64 | 3.27518106503 | 0.205146784299 | 209.611588162 | 856.203443493 | | [b50-eng-12009792](#b50-eng-12009792) | M09-PB | 64 | 0.51455550829 | 0.0676222255202 | 32.9315525306 | 35.3825017807 | ### Pooled results (merge calculations) For each comparison, N=Σn, S=Σsum, Q=Σsumsq, mean=S/N and sample SE=sqrt((Q−S²/N)/(N(N−1))). Standard errors are not averaged. Pooled worst five are obtained by sorting the union of block worst-five lists; no omitted within-block observation can precede its block’s fifth worst. **M09-IB**: `{"n":2176,"mean_diff":3.8431693161273444,"se":0.03170362581071535,"sum":8362.736431893101,"sumsq":36896.443467154466,"worst5":[[12007853,-0.872755470428956],[12008281,-0.3216320394054719],[12009106,-0.0017517041642349795],[12007809,0.05443217501726849],[12009464,0.26215069093666443]]}`. **M09-PB**: `{"n":2176,"mean_diff":0.4708315239965146,"se":0.011780212595039934,"sum":1024.5293962164158,"sumsq":1139.1675260943164,"worst5":[[12009248,-2.4796735218765207],[12009060,-2.072566805819065],[12008309,-1.742841409490211],[12008559,-1.664329682518428],[12009660,-1.4914933180672278]]}`. **PB-IB**: `{"n":2176,"mean_diff":3.37233779213083,"se":0.03158252694421053,"sum":7338.207035676686,"sumsq":29467.672708415586,"worst5":[[12007853,-1.9254545963386818],[12008281,-0.6564487235871184],[12009035,-0.6492450027079713],[12007809,-0.557561709182778],[12009250,-0.004159861412745158]]}`. ### Live gas, validity and allocation caveats | Source | IB afterSwap peak | PB afterSwap peak | M09 afterSwap peak | |---|---:|---:|---:| | [b44-eng-12007680](#b44-eng-12007680) | 234957 | 215849 | 201948 | | [b44-eng-12007936](#b44-eng-12007936) | 234825 | 211077 | 201784 | | [b44-eng-12008192](#b44-eng-12008192) | 234937 | 201971 | 202178 | | [b44-eng-12008448](#b44-eng-12008448) | 237040 | 200684 | 202489 | | [b45-eng-12008704](#b45-eng-12008704) | 234736 | 206001 | 202431 | | [b45-eng-12008960](#b45-eng-12008960) | 234865 | 200429 | 201956 | | [b45-eng-12009216](#b45-eng-12009216) | 234852 | 200594 | 202161 | | [b45-eng-12009472](#b45-eng-12009472) | 234724 | 212865 | 219533 | | [b50-eng-12009728](#b50-eng-12009728) | 234849 | 215621 | 201302 | | [b50-eng-12009792](#b50-eng-12009792) | 234837 | 199547 | 201730 | Across these reports, initialization peaks are IB 243,227, PB 140,992, M09 229,594; Vanilla baseline afterInitialize 21,796, afterSwap 22,592, getName 21,526. Submission getName is IB 21,735, PB 21,521, M09 21,530. Gas comes from reset immediately before each engine call and snapshot immediately afterward, with separate Submission/Baseline roles; validation gas is not substituted. Reports give zero oog, revert, bad_return, tx_error and other_halt, and all measured calls successful. This validates those observed callbacks only. Sources: gas snapshots and summaries in all ten engine results. All allocated engine rows are reported complete: no missing intersection seeds, no imputation, no retries of consumed engine blocks. Commands use 64-seed slices and eight workers; eng-12008960 used separate concurrent arm processes, not a common gas counter. eng-12008192’s interrupted tool output was resolved from on-disk completion outputs without rerunning. The unusually high M09 peak 219,533 in eng-12009472 is retained, not replaced with the more typical ~202k. Sources: corresponding allocation, gas and limitations fields. ## Track B: per-item findings and measurements B tables retain each principal n/mean/SE (12 significant digits); sums/sumsquares, individual worst-five and duplicate rounded printouts are condensed here, while engine pooling statistics are retained. Unless a section says otherwise, base means unchanged ThrLattice `thr=.9`, `qm_k1e-5.json`, source SHA-256 `663965707d4f20f81ab0ade8ed510742a330605fd9e889d426db1d920abc8958`; reference lattice is threshold .5. Full holdout is replica 10000..12047. The common full-holdout inherited threshold gain is **+0.3228919198104912 (SE .01069159987528695)**; it must be separated from a new idea’s gain. Different classifier thresholds make ~29% base misread and ~16.2% lattice misread different decision diagnostics, not evidence that the latter necessarily earns more edge. All B gas figures are explicitly unmeasured estimates or operation counts. No B item establishes Solidity validity or a board position. Sources: B items below, especially values7-2, rival7-1, mass8-1, ig8-1/2 and fresh8-2. ### b46-values7-1 Source: [b46-values7-1](#b46-values7-1). Status: candidate_not_better_than_base_on_partial_tuning; confirmation and holdout NOT run (time). This is a partial negative experiment, not a completed replication or promotion. VN was regenerated in float64 for 110 sweeps (gain 536.8206, final change 1.88e-7). A 25-action, up-to-five-node shortlist changes quote lookup only. Posterior and matched-Gaussian modes were run only on 64000..64063; confirmation 64256..64511 and holdout were NOT RUN. The posterior-minus-Gaussian mean there is −0.07464104822100825, with no paired SE provided for that contrast. Earlier grid/endpoint branch checks were inherited, not rerun. On seed 64000, disabled policy and causal six-input replay both have zero fee drift; Python and JIT action values match exactly. Reduced versus full posterior over 6,624 calls agrees on actions 94.6%, maximum chosen-value difference .0008059793582, mean regret 6.45297e-7, max regret 6.97062e-5, maximum fee difference 4.9985763 bps. Fee drift versus base reaches 4.9987502 bps. The old off-grid commutation requirement was rejected as mathematically inappropriate, not declared passed. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `tuning/post_vs_base` | 64 | -0.107410594743 | 0.0602685897154 | | `tuning/gauss_ablation_vs_base` | 64 | -0.032769546522 | 0.0507186051957 | Diagnostics (misread fraction; RMS bps): `tuning/arm_stats/post`: misread 0.287320409, RMS 13.1779272 bps; `tuning/arm_stats/gauss`: misread 0.287246977, RMS 13.2048496 bps; `tuning/arm_stats/base`: misread 0.290890205, RMS 13.3422935 bps. Cost / feasibility: No added persistent slots; at most 5 nodes ×25 actions ×16 corners=2,000 value reads plus IG. The source’s ~6,000-gas read-only lower component omits arithmetic. Approximately 0.5-GB VN makes a direct port infeasible; no EVM measurement. Additional limitations: Only 64 tuning seeds; no confirmation or holdout. Rival uncertainty remains scalar, IG offsets moment-based. The ~29% misread uses thr=.9 and is not the reference’s 16% at .5. ### b46-values7-2 Source: [b46-values7-2](#b46-values7-2). Status: completed_negative_incremental_result. Completed negative result with a different nine-action shortlist (baseline and ±5 bps/side), five posterior conditional-mean bins, ambiguous first callbacks pi in (.05,.95), and a five-bin matched Gaussian control. Tune 64000..64255, confirm 64256..64511, full holdout once. Posterior-over-base is −.08763557 on holdout, Gaussian-kernel-over-base −.00924492; the posterior-specific loss is −.07839065. The prior 6.9090026463-bps off-grid fee discrepancy remains a diagnostic: interpolating values then optimizing need not equal interpolating optimized offsets; no false parity pass is claimed. Grid tests: 682 cases, zero non-tie argmax mismatch/value error/fee error. Independent same-action off-grid test: 720 actions, max value error 1.66533e-16. Endpoint and causal-state checks pass; disabled/active hidden-scramble checks cover 62,345 callbacks. Five-bin posterior versus full lattice changes 8/200 audited argmaxes, max value error .00051827818, max regret 1.87904e-5 and fee drift 4.98651976 bps; five-bin versus 101-bin Gaussian changes 9/200, max value error .00183944104 and fee drift 2.2379015 bps. Exact diffusion acceleration matches full belief/fees/state on 20 seeds 63000..63019 (210,961 callbacks), fastmath off; pure active replay matches accelerated digest on 64002. The stdlib solver matches only a reduced grid (4,995 VN values, zero error; 735 QMDP values, 1.39e-17); full-grid fallback runtime is unmeasured. Float64 VN is 527,054,136 bytes, regenerated over 110 sweeps; table arrays are excluded from the source archive. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `tuning/paired/posterior_over_base` | 256 | -0.102395944278 | 0.0215695539737 | | `tuning/paired/gaussian_kernel_over_base` | 256 | -0.00666624599472 | 0.00826615733782 | | `tuning/paired/posterior_over_gaussian` | 256 | -0.0957296982836 | 0.0200930151955 | | `confirmation/paired/posterior_over_base` | 256 | -0.0954254588625 | 0.0210395830937 | | `confirmation/paired/gaussian_kernel_over_base` | 256 | -0.0138117697549 | 0.00864018586578 | | `confirmation/paired/posterior_over_gaussian` | 256 | -0.0816136891076 | 0.0205573645478 | | `holdout/paired/posterior_over_base` | 2048 | -0.087635570325 | 0.00779521691964 | | `holdout/paired/gaussian_kernel_over_base` | 2048 | -0.00924491746978 | 0.00325051939144 | | `holdout/paired/posterior_over_gaussian` | 2048 | -0.0783906528552 | 0.00767268600885 | | `holdout/paired/posterior_over_lattice_reference` | 2048 | 0.235256349486 | 0.0124170702804 | | `holdout/paired/gaussian_over_lattice_reference` | 2048 | 0.313647002341 | 0.0107728447548 | | `holdout/paired/base_over_lattice_reference` | 2048 | 0.32289191981 | 0.0106915998753 | Diagnostics (misread fraction; RMS bps): `holdout/posterior`: misread 0.29051613, RMS 13.0813421 bps; `holdout/gaussian`: misread 0.288831263, RMS 13.2660785 bps; `holdout/base`: misread 0.288834005, RMS 13.2771528 bps. Cost / feasibility: No extra quote-mechanism slots; <=720 corner reads, ~2,880 corner multiplies, 720 additions, 72 IG reads and 121-cell scan per changed callback. Illustrative incremental range 120,000..2,000,000 gas assumes table storage; excludes base/deployment and is not measured. Additional limitations: Local shortlist/five-node quadrature approximate the actual posterior. Off-grid 6.909-bps drift is diagnostic. Reduced-grid stdlib parity does not establish full-grid runtime/convergence. No .3 incremental gain. ### b47-rival7-1 Source: [b47-rival7-1](#b47-rival7-1). Status: complete; holdout scored once, n=2048 per arm; incremental gain over joint K1 below the +0.3 port bar. Hidden-rival transition on joint K1: Poisson(.8) rival-only orders truncated to two, three lognormal size nodes, three compressed rival nodes, one preceding rival-only order at first fill. Renormalize the rival marginal without multiplying price-hypothesis mass to avoid double-counting no-arb evidence. One setting, tune 64000..64255, confirmation 64256..64511, full holdout once. Candidate beats fresh threshold base by .20841244 and lattice by .53130436. Relative to inherited joint K1 its mean gain is .1286 (=.2084−.0798); no holdout SE is available because K1 rows were not rerun or supplied. Direct tuning/confirmation comparisons over K1 are +.1012 SE .0306 and +.12074307 SE .02928360. Silent and in-step parts were not ablated separately. Per-run HA lines: candidate +4.795 SE .037, base +4.586 SE .038, lattice +4.263 SE .039. Runtime about 5 s/seed versus 3 for K1, one core. Subsequent mass8 workers found and repaired an intermediate survivor-mass normalization bug: retain these as the original implementation’s performance, not a proof of correct transition accounting. The neglected Poisson N>=3 mass is .0474. Checks report zero-transition/rt=0 equivalence to K1, joint=0 equivalence to threshold base, causal replay and hidden-input independence. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `tuning/cand_vs_base` | 256 | 0.191456906724 | 0.0358434561679 | | `tuning/cand_vs_jointK1` | 256 | 0.101166454968 | 0.0305911254314 | | `tuning/jointK1_vs_base` | 256 | 0.0902904517564 | 0.0245152399395 | | `confirmation/cand_vs_base` | 256 | 0.206998746195 | 0.0336329774423 | | `confirmation/cand_vs_jointK1` | 256 | 0.120743070943 | 0.0292835958228 | | `confirmation/jointK1_vs_base` | 256 | 0.086255675252 | 0.0239328452674 | | `holdout/vs_fresh_unchanged_base` | 2048 | 0.208412436442 | 0.0120827215955 | | `holdout/vs_ref_lattice_qm` | 2048 | 0.531304356252 | 0.0148468612585 | | `holdout/fresh_base_vs_ref_lattice_qm` | 2048 | 0.32289191981 | 0.0106915998753 | Diagnostics (misread fraction; RMS bps): `holdout/candidate`: misread 0.280365587, RMS 13.0113075 bps; `holdout/fresh_base_thr09`: misread 0.288834005, RMS 13.2771528 bps; `holdout/ref_lattice_qm`: misread 0.162248261, RMS 13.6428397 bps. Cost / feasibility: No added slots. First fill uses 1+min(dt−1,30) kernels, <=54 router tests/kernel plus compression; retail likelihood up to 6× K1 per-cell work. Rough unmeasured direct-port estimate ~200k gas/kernel, already too expensive across silent gaps; requires tabulation/closed form. Runtime ~1.7× K1 (5 versus 3 s/seed). Additional limitations: No holdout K1 rerun/rows, so no incremental-over-K1 SE. Only rt=2 tested; silent/in-step effects inseparable. Silent arb clip and k=1 order use hypothesis mean; sell conversion also uses mean price. ### b48-tails7-1 Source: [b48-tails7-1](#b48-tails7-1). Status: see finding. Keep belief/mean/classifier/phase/rival unchanged; use separate ask/bid tail standard deviations from upper/lower partial second moments. Strengths .25,.5,.75,1 tried on 64000..64255, select .25 before 64256..64511 confirmation; full holdout once. No useful gain: retail +.24, arb −.23, net +.00876964 over base. Larger strengths lose monotonically in tuning (.75 −.016, 1 −.028). Reflection tests find zero counterexamples/200, symmetric-prior lookup matches and SD difference <=1.42e-14; disabled, hidden-scramble and six-input replay pass on 64001/64300. O(1) two-Gaussian tail moments are analytically possible, but not implemented: for component d=mu_i−m, sd=s, upper moment=(s²+d²)Phi(d/s)+d*s*phi(d/s), sum by weight, lower=v−upper; a point mass contributes to its side. This is exact for that mixture, not a claim that real.py’s collapsed/truncated posterior retains such a representation. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `tuning_paired_vs_base/tailk=0.25` | 256 | 0.00834650024658 | 0.0169340627685 | | `tuning_paired_vs_base/tailk=0.50` | 256 | 0.00181008579514 | 0.0203196805936 | | `tuning_paired_vs_base/tailk=0.75` | 256 | -0.0155078166906 | 0.0215502495849 | | `tuning_paired_vs_base/tailk=1.0` | 256 | -0.0283284725937 | 0.0231940112646 | | `confirmation_paired_vs_base` | 256 | 0.0120694903159 | 0.0159778843946 | | `holdout_paired_candidate_vs_fresh_base` | 2048 | 0.00876964167771 | 0.00552311687371 | | `holdout_paired_candidate_vs_ref_lattice_qm` | 2048 | 0.331661561488 | 0.0111113007545 | | `holdout_paired_fresh_base_vs_ref_lattice_qm` | 2048 | 0.32289191981 | 0.0106915998753 | Diagnostics (misread fraction; RMS bps): `candidate_holdout`: misread 0.28807108, RMS 13.2151995 bps; `fresh_base_holdout`: misread 0.288834005, RMS 13.2771528 bps. Cost / feasibility: Extra <=121-cell moment pass, two sqrt and separate side lookups; no required new state, or 0..2 cache slots. Source’s hypothetical two-Gaussian incremental estimate 15,000..60,000 gas is unmeasured; lattice is not directly portable. Additional limitations: Selection among small, noisy tuning differences; moments recomputed every callback. Two-Gaussian reduction is analytic only, not an implemented experiment. ### b49-fresh7-1 Source: [b49-fresh7-1](#b49-fresh7-1). Status: completed experiment; no promotion. **Source caveat:** requester reports 1 failed citation recheck(s); external support remains qualified as described above. Fresh idea: uniform-cell arb density consistent with the existing cut support, replacing linear/triangular interpolation. Single preregistered setting; tune 64000..64255, confirmation 64256..64511. Only holdout 10000..10511 completed (512/2048); 10512..12047 is unmeasured, not imputed. Incremental −.01284130, while +.30931547 over lattice includes +.32215677 inherited threshold gain on exactly these 512 seeds. A centre-cell point distribution queried .75 cells away has base density*HC=.25 but candidate 0; cut [.6,.9] has zero mass, showing a representation inconsistency, not proof of a correct Bayesian filter. Density integrates to 1; disabled and scrambled callbacks match over 20,742 calls on two seeds, exact acceleration parity over 21,419 more calls. Uses stdlib tuple-index acceleration, not a C dependency (C compiler attempt failed and was unused). Reference gates matched all four columns at six decimals; not candidate tuning. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `measurements/tune/candidate_vs_base` | 256 | -0.0192254752781 | 0.0174657759027 | | `measurements/confirm/candidate_vs_base` | 256 | 0.00220608327559 | 0.0170100638316 | | `measurements/hold/candidate_vs_base` | 512 | -0.0128412964515 | 0.0118770360839 | | `measurements/hold/candidate_vs_lattice` | 512 | 0.309315474211 | 0.0219759725547 | | `measurements/hold/inherited_base_vs_lattice` | 512 | 0.322156770662 | 0.0218054723984 | Diagnostics (misread fraction; RMS bps): `hold/candidate_diagnostics`: misread 0.288094021, RMS 13.2735364 bps; `hold/base_diagnostics`: misread 0.28800126, RMS 13.2598274 bps; `hold/lattice_diagnostics`: misread 0.161961712, RMS 13.6285525 bps. Cost / feasibility: No added state/slots: one index rounding, bounds check, probability read and division replaces two-cell interpolation. Rough incremental −300..200 gas per new-step callback assumes accessible cells; not measured and excludes lattice cost. Additional limitations: Only 512 holdout seeds; remaining 1,536 absent. Calibration gates are reference checks, not candidate selection. Diagnostic error reductions are not themselves an economic objective. ### b49-fresh7-2 Source: [b49-fresh7-2](#b49-fresh7-2). Status: completed; no promotion. **Source caveat:** requester reports 2 failed citation recheck(s); external support remains qualified as described above. Fresh idea: integer-cell recenter shifts remove fractional-rebin numerical heating. No parameter fitting; tune 64000..64255, confirmation 64256..64511, full holdout. Incremental .00115202 is null; total .32404394 over lattice is mainly inherited .32289192. Compact-support moments are preserved, but global preservation is false: a distribution of .01 at cell 0 and .99 at 60 shifted one cell loses the .01 tail, mass becomes .99 and variance 0. On actual tuning paths 33,020 recenters lose total mass 9.89147e-6, max per-call 4.43952e-6, max mean error 6.74324e-7 log units and variance error 5.45042e-9 log². The compact-support 1,000-trial gate has <=8.88e-16 mean and <=1.27e-21 variance error. Single-pin half-cell base heating is 1.5625e-8 log², candidate 0. Disabled and hidden-scramble checks match 31,443 callbacks; exact accelerated comparison 62,885, fallback 21,456. Boundary failures are reported; the active candidate was not silently replaced by unchanged base. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `splits/tuning/candidate_vs_base` | 256 | 0.00424149145703 | 0.0095447239015 | | `splits/confirmation/candidate_vs_base` | 256 | 0.00567154227817 | 0.011376411548 | | `splits/holdout/candidate_vs_base` | 2048 | 0.00115201895854 | 0.00301466734997 | | `splits/holdout/candidate_vs_ref_lattice_qm` | 2048 | 0.324043938769 | 0.0106325033606 | | `splits/holdout/base_vs_ref_lattice_qm` | 2048 | 0.32289191981 | 0.0106915998753 | Diagnostics (misread fraction; RMS bps): `holdout/candidate`: misread 0.289402113, RMS 13.2802357 bps; `holdout/base`: misread 0.288834005, RMS 13.2771528 bps; `holdout/reference_diagnostics_on_candidate_seeds`: misread 0.162248261, RMS 13.6428397 bps. Cost / feasibility: No extra slots/cells; rounded integer shift plus O(active support) copies replaces fractional deposits. Hypothetical incremental −2,000..100 gas/recenter is an arithmetic-only estimate, not measured absolute feasibility. Additional limitations: Finite-window clipping remains; no claim of global moment conservation or elimination of kernel/cut approximations. No new external observation signal. ### b51-mass8-1 Source: [b51-mass8-1](#b51-mass8-1). Status: complete. Mandatory survivor-mass accounting repair to original RivalT, without changing any other policy math: after each intermediate compress(cur,2Q), restore the incoming unnormalized survivor mass; only final compression conditions the Poisson mixture. Finite-law tests pass. Tune 65000..65255, confirmation 65256..65511, complete holdout. Repair-minus-original tuning .0011208134 and confirmation .0049330958 are indistinguishable from zero at their SEs. Holdout repaired-minus-threshold is .21516975; this is NOT the repair alone. Original holdout rows are absent; repaired-minus-original paired holdout and SE are null. Historical original over threshold .20841244 remains a separate aggregate, not a new control. Approximation audit on 24 tuning states with 20,000 MC trials/state/law separates accounting, compression, quadrature and truncation: max original-minus-repaired mass .04053946, mean 1.50643 bps, SD 1.35220 bps; repaired-versus-exact compression mass error 0, mean <=1.78e-11 bps, SD up to 1.31954 bps. Continuous-MC quadrature discrepancies reach mass .0509859, mean 2.25128 bps, SD 2.95471 bps; MC maximum mass SE .0035334, mean SE .0855686 bps. Count tail .0474225961. These sampled errors are not global bounds. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `tuning/repaired_minus_thr09` | 256 | 0.160754972479 | 0.0318303584743 | | `tuning/repaired_minus_original` | 256 | 0.0011208134039 | 0.00766903236985 | | `confirmation/repaired_minus_thr09` | 256 | 0.159117266345 | 0.0357052421429 | | `confirmation/repaired_minus_original` | 256 | 0.00493309575185 | 0.00774981880781 | | `holdout/repaired_minus_thr09` | 2048 | 0.215169751577 | 0.0119307267863 | | `holdout/repaired_minus_ref_lattice` | 2048 | 0.538061671388 | 0.014746584075 | | `holdout/thr09_minus_ref_lattice` | 2048 | 0.32289191981 | 0.0106915998753 | Diagnostics (misread fraction; RMS bps): `holdout/repaired`: misread 0.280589638, RMS 13.0177074 bps; `holdout/fresh_thr09`: misread 0.288834005, RMS 13.2771528 bps; `holdout/ref_lattice_on_paired_seeds`: misread 0.162248261, RMS 13.6428397 bps. Cost / feasibility: No added persistent state: <=54 survivor additions and <=12 rescalings per kernel. 222 gas counts only 54 ADD and 12 MUL opcodes; actual repair and total policy gas are unknown. Additional limitations: Historical original holdout evidence is aggregate-only. Approximation checks cover 24 states, not all states; fixing recurrence need not improve policy edge. ### b51-mass8-2 Source: [b51-mass8-2](#b51-mass8-2). Status: checked survivor-mass repair; partial holdout, missing ranges explicit. **Source caveat:** requester reports 1 failed citation recheck(s); external support remains qualified as described above. Same accounting repair independently checked; one setting, no performance selection. Tune 65000..65255 and confirmation 65256..65511 reproduce mass8-1’s paired repair-minus-original numbers. Holdout is PARTIAL: 10000..11535 completed for both repaired and fresh threshold arms (1536); 11536..12047 missing. Mean .21775572 over threshold and .53828070 over lattice apply only to that range; inherited threshold gain there is .32052499, not the full-range .32289192. No original holdout rows, no repaired-minus-original paired result or SE. Do not pool this subset with mass8-1’s overlapping full holdout. Finite exact enumeration over 24 cases/order counts 0..3 with nonunit initial masses has max weight error 1.0842e-19. Counterexample: repaired unnormalized order masses [.5,.15,.05], original [.5,.15,.1]; conditioning early doubles the final order branch incorrectly. Zero arrivals, all rejection, empty/zero mass and zero-total graceful behavior pass. Source-boundary AST check says only step_kernel changed; callback off/scramble/replay and joint=0/rt=0 controls pass. Six sampled tuning states, 80,000 MC trials/state/law separate discrete GH validation from continuous-size quadrature; compression and size/count truncations remain approximations. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `tuning/paired` | 256 | 0.0011208134039 | 0.00766903236985 | | `tuning_vs_fresh_thr09/paired` | 256 | 0.160754972479 | 0.0318303584743 | | `confirmation/paired` | 256 | 0.00493309575185 | 0.00774981880781 | | `holdout/vs_fresh_thr09/paired` | 1536 | 0.217755719397 | 0.0139748178693 | | `holdout/vs_ref_lattice_qm/paired` | 1536 | 0.538280704514 | 0.0171913447186 | | `holdout/fresh_thr09_vs_ref_lattice_qm/paired` | 1536 | 0.320524985117 | 0.0125273017498 | Diagnostics (misread fraction; RMS bps): `holdout/vs_fresh_thr09/repaired`: misread 0.280538709, RMS 13.0145716 bps; `holdout/vs_fresh_thr09/fresh_thr09`: misread 0.288813392, RMS 13.2759204 bps; `holdout/vs_ref_lattice_qm/ref_lattice_qm`: misread 0.162275541, RMS 13.6447747 bps. Cost / feasibility: No added persistent state; same <=54 survivor additions and <=12 rescalings/kernel. Hypothetical fixed-point 500..3,000 extra gas/kernel, up to 15,500..93,000 over 31 kernels; unmeasured, excluding inherited routing/lattice. Additional limitations: Only 1,536 holdout seeds. No original holdout rows or repair-minus-original SE; six-state MC diagnostics retain sampling error and are not global bounds. ### b52-ig8-1 Source: [b52-ig8-1](#b52-ig8-1). Status: complete. **Source caveat:** requester reports 3 failed citation recheck(s); external support remains qualified as described above. One preregistered offline-only switch qmdp.igtab -> existing igtab2, kappa=1e-5, same thr=.9. Tune 65000..65255, confirmation 65256..65511 and full holdout. Incremental .0045746993 is null; total .3274666191 includes .3228919198 threshold gain. Passive solve reaches 110 sweeps, 1.88e-7 change; last persisted checkpoint is 105. A timed-out invocation reached transient sweep 40 but retained only checkpoint 5; resumed bounded work then generated complete tables, never a partial-table fallback. Both control/IG2 have all 40 slices. Regenerated six solved tables and 1,073,424 control quote offsets match exactly; IG2 changes 544,992 scalar offsets, max 49.36 bps, RMS .86249796. Control JSON hash can differ on unused metadata despite identical actions. IG2 subtracts overlap ambiguity from ideal-pin information, but continuous truncated-normal moments differ from original discrete 5-bps igtab even at zero retail overlap: this is not an isolated ambiguity-subtraction ablation. Independent integral cases have errors up to .1353296864 bps² on its cases; scalar selected matrices <=8.53e-13 bps², zero-overlap versus original discrete differences up to 51.79231315 bps². Control and candidate replay/full-state/hidden-scramble checks match on 46,240 and 46,304 tuning callbacks respectively; stdlib-vs-optional backend parity is only on a small grid, not a full-grid proof. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `paired/tuning_vs_thr09/paired` | 256 | -0.0231829636536 | 0.0209375098827 | | `paired/confirmation_vs_thr09/paired` | 256 | 0.0122510527533 | 0.0200290990402 | | `paired/holdout_vs_fresh_thr09/paired` | 2048 | 0.00457469931605 | 0.00695757600583 | | `paired/holdout_vs_lattice_qm/paired` | 2048 | 0.327466619127 | 0.0112465873863 | | `attribution/fresh_thr09_vs_lattice_qm` | 2048 | 0.32289191981 | 0.0106915998753 | Diagnostics (misread fraction; RMS bps): `holdout_vs_fresh_thr09/arms_on_common_seeds/candidate`: misread 0.285570209, RMS 13.2339108 bps; `holdout_vs_fresh_thr09/arms_on_common_seeds/reference`: misread 0.288834005, RMS 13.2771528 bps; `holdout_vs_lattice_qm/arms_on_common_seeds/reference`: misread 0.162248261, RMS 13.6428397 bps. Cost / feasibility: IG2 adds offline overlap quadrature only (two 400-node integrals/quote pair/sd). No incremental callback operations, table cells or slots; structural incremental gas 0 assumes the same branch path and encoding, not measured absolute gas. Additional limitations: Gaussian IG prior, fixed .26 retail intensity, finite 200-bps integration, moment/rival/phase approximations remain. Small-backend checks do not prove every full-grid fallback argmax. ### b52-ig8-2 Source: [b52-ig8-2](#b52-ig8-2). Status: completed. **Source caveat:** requester reports 2 failed citation recheck(s); external support remains qualified as described above. Replicates the same IG2 switch with all holdout aggregates exactly agreeing with ig8-1; same deterministic seeds must not double N. Passive solve converges at 110 sweeps, checkpoint 85 last saved, 1.88e-7 change; all 40 control and 40 active slabs complete. Float32 VN export/reload matches upstream CLI and solved tables exactly; gain discrepancy 1.7486e-11/run. Table byte hashes differ from ig8-1 while quote outcomes and comparison statistics agree; metadata/serialization and manifests are not evidence of a scoring conflict. Generator resume validation was hardened to hash JSON metadata, with no math change. Independent overlap cases differ from midpoint by at most .1203690274 bps² on this different case set, scalar parity <=2.20939e-9 bps², continuous-ideal minus discrete up to 37.74554 bps². These error maxima do not contradict ig8-1 because test cases differ. Zero-overlap matches continuous ideal, not necessarily original discrete igtab. Callback scramble/off/replay, diffusion parity, reference gates and unpacked stdlib checks pass on their stated samples. Full arbitrary-input or full-grid formal parity is not claimed. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `tuning/IG2_vs_fresh_thr09` | 256 | -0.0231829636536 | 0.0209375098827 | | `confirmation/IG2_vs_fresh_thr09` | 256 | 0.0122510527533 | 0.0200290990402 | | `holdout/IG2_vs_fresh_thr09` | 2048 | 0.00457469931605 | 0.00695757600583 | | `holdout/IG2_vs_ref_lattice_qm` | 2048 | 0.327466619127 | 0.0112465873863 | | `holdout/fresh_thr09_vs_ref_lattice_qm` | 2048 | 0.32289191981 | 0.0106915998753 | Diagnostics (misread fraction; RMS bps): `holdout/IG2`: misread 0.285570209, RMS 13.2339108 bps; `holdout/fresh_thr09`: misread 0.288834005, RMS 13.2771528 bps; `holdout/ref_lattice_qm_on_candidate_seeds`: misread 0.162248261, RMS 13.6428397 bps. Cost / feasibility: Offline IG2 table generation only; zero incremental callback operations, cells or slots. Structural incremental gas 0 is not an EVM benchmark; byte content, paths and compression may change gas. Additional limitations: Optional NumPy/Numba generation/acceleration has a shipped stdlib fallback; full score reproduction still requires pinned external tables/rows. No missing/fallback runs imputed. ### b53-fresh8-1 Source: [b53-fresh8-1](#b53-fresh8-1). Status: complete. **Source caveat:** requester reports 1 failed citation recheck(s); external support remains qualified as described above. Fresh idea: scale table impact units by reserve depth, using k=10000/sqrt(X*Y*exp(m)); lookup at (w/k,R/k,sd/k), transform each fee to clamp(1−(1−f)^k,0,.1). One active formula, no search; tune 65000..65255, confirmation 65256..65511, full holdout. Incremental −.0492503681 rejects this formula, not every depth-aware strategy; +.2736415517 over lattice is inherited threshold benefit minus that loss. Four tuning diagnostic seeds (65004..65007), 49,737 callbacks: k .9249481..1.0356810; relative k estimate RMS .0006592973, max .0036995753; fee changes RMS .4218448 bps, max 17.87518474 bps; unclamped conversion roundtrip <=1.38778e-13 bps. No freshly depth-solved DP comparison, so exact scale covariance is not established. Rival fixed 30-bps fees, fixed volatility and uncertain m break the proposed similarity. Disabled/scramble/replay/reference gates are reported passing; filter code equality does not imply identical trajectories after changed fees. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `tuning_incremental/paired` | 256 | -0.0573904488551 | 0.0218526296832 | | `confirmation_incremental/paired` | 256 | -0.0409893290526 | 0.0226057661195 | | `holdout_incremental_vs_fresh_thr09/paired` | 2048 | -0.0492503681384 | 0.00889976171451 | | `holdout_vs_reference_lattice_qm/paired` | 2048 | 0.273641551672 | 0.0117368556224 | | `inherited_threshold_gain/fresh_measurement/paired` | 2048 | 0.32289191981 | 0.0106915998753 | Diagnostics (misread fraction; RMS bps): `holdout_incremental_vs_fresh_thr09/candidate`: misread 0.289140654, RMS 13.3081187 bps; `holdout_incremental_vs_fresh_thr09/reference`: misread 0.288834005, RMS 13.2771528 bps; `holdout_vs_reference_lattice_qm/reference`: misread 0.162248261, RMS 13.6428397 bps. Cost / feasibility: Zero extra persistent slots. O(1) added work: 1 sqrt, 1 exp, 2 log1p, 2 expm1, 4 multiplies and 4 divides; no new table/cell reads. Unmeasured engineering estimate +15,000..45,000 gas, excludes base. Additional limitations: Rejects only this approximate scaling formula. No depth-specific DP, exact covariance or gain claim. Rival fees, volatility, table clipping and uncertain m remain mismatched. ### b53-fresh8-2 Source: [b53-fresh8-2](#b53-fresh8-2). Status: completed; null result, no promotion. **Source caveat:** requester reports 1 failed citation recheck(s); external support remains qualified as described above. Fresh idea: on silent steps, weight each cut lattice cell by exp(−k*.8*q_i), where q_i is router-implied chance a retail order would fill us with rival clipped to the cell’s price band. The active first-trade likelihood/classifier remains unchanged. k=1,2,4,8 tried on 65000..65255; k=1 selected before confirmation 65256..65511 and full holdout. Incremental −.002761049 is null; .320130871 over lattice is .322891920 inherited threshold minus .002761049. No-fill factor is almost flat where belief matters: most mass within ~20 bps whereas rival clip matters beyond ~30 bps. For seed 65000 max/min weight ratio median 1.0003, p90 1.159, p99 1.285; 34% silent steps >1.05. Static rival clipping, omitted rival-only orders/within-step absence, midpoint rate/size, ignored 1e-4 threshold and drift limit the likelihood. A first large base chunk timed out at 480 s and wrote no output; it was split and rerun, leaving no missing holdout seed. This is replica rerunning, not engine seed reuse. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `tuning_vs_base_65000_65255/1` | 256 | 0.0136443886048 | 0.00805031020711 | | `tuning_vs_base_65000_65255/2` | 256 | 0.0078132969893 | 0.0104440820612 | | `tuning_vs_base_65000_65255/4` | 256 | -0.0270898146174 | 0.0138143973482 | | `tuning_vs_base_65000_65255/8` | 256 | -0.146678364482 | 0.0202722465489 | | `confirmation_vs_base_65256_65511` | 256 | -0.0220943890809 | 0.00904688116178 | | `holdout/candidate_vs_fresh_base` | 2048 | -0.00276104898418 | 0.00334998825114 | | `holdout/candidate_vs_ref_lattice_qm` | 2048 | 0.320130870826 | 0.0108817200818 | | `holdout/fresh_base_vs_ref_lattice_qm` | 2048 | 0.32289191981 | 0.0106915998753 | Diagnostics (misread fraction; RMS bps): `first_trade_misread_and_RMS_by_arm_holdout/candidate`: misread 0.28899641, RMS 13.2793313 bps; `first_trade_misread_and_RMS_by_arm_holdout/base_thr09`: misread 0.288834005, RMS 13.2771528 bps; `first_trade_misread_and_RMS_by_arm_holdout/lattice_qm`: misread 0.162248261, RMS 13.6428397 bps. Cost / feasibility: Adds two reserve slots (one packed), per active cell/silent step: 1 exp, 1..2 sqrt, 2 log+erfc tails, 1 weight exp. Rough unmeasured 1,500..3,000 gas/cell/step implies >245k for several silent steps without tabulation/capping; no contract exists. Additional limitations: Static rival clipping omits path dependence/rival-only orders; midpoint constants and omission of minimum-fill/drift/late-step absence remain. No gain and no promotion. ### b54-tg-1 Source: [b54-tg-1](#b54-tg-1). Status: done. **Source caveat:** requester reports 1 failed citation recheck(s); external support remains qualified as described above. TellGauss: multiply retail likelihood by r when an input lies on the reserve floating-point grid, zero retail likelihood off that grid; do not force the posterior to one when tell is present because retail can share it. Tune 66000..66255 among C=1/r, C=2/r, C=1/r/2; C=1 with original r wins and freezes before confirmation 66256..66511 and full holdout. Holdout agrees with tg-2 at reported rounding; candidate mean edge 536.5171, +6.960 SE .044 over HA reported by run.py. Tuning audit flags all 970,221 arbs and 6,035 of 1,819,678 retail callbacks; first-trade-only retail flags 2,717/826,104. False-positive size quartiles differ from tg-2’s because size definitions differ: here net input in Y, quartile flag counts [211,586,1159,4079], rates [.000463819,.001288139,.002547706,.008966412]. Those are not contradictory to tg-2’s gross amountY quartiles. The worker says actual engine reserve-to-WAD conversion and floating operation order could destroy the signature; this is an untested transfer hypothesis, and the requester flagged a supporting source recheck. Its gas estimate is a hypothetical 300..1500 additional gas/callback, not measured. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `tuning/vs_gauss_qm_local/C1/paired` | 256 | 3.0702 | 0.0626 | | `tuning/vs_gauss_qm_local/C2/paired` | 256 | 3.0379 | 0.0623 | | `tuning/vs_gauss_qm_local/C1h/paired` | 256 | 3.0604 | 0.0628 | | `tuning/C1_minus_C2` | 256 | 0.0324 | 0.0102 | | `tuning/C1_minus_C1_r_half` | 256 | 0.0099 | 0.0072 | | `confirmation/vs_gauss_qm_local/paired` | 256 | 3.131 | 0.062 | | `confirmation/vs_flag_qm_local/paired` | 256 | -0.0319 | 0.009 | | `holdout_vs_flag_qm/paired` | 2048 | -0.0258 | 0.0027 | | `holdout_vs_lattice_qm/paired` | 2048 | 2.6962 | 0.0209 | | `holdout_vs_gauss_qm/paired` | 2048 | 3.1911 | 0.0221 | | `holdout_vs_HA` | 2048 | 6.96 | 0.044 | Cost / feasibility: Source estimates 300..1,500 extra gas/afterSwap for arithmetic/grid operations; not measured. WAD signal survival and actual integer implementation remain untested. Additional limitations: Floating-point signature, engine operation order and WAD conversion untested. Worker reports local Gaussian/reference differences up to .002/seed. Size-dependent retail false flags justify retaining soft r. ### b54-tg-2 Source: [b54-tg-2](#b54-tg-2). Status: complete. **Source caveat:** requester reports 1 failed citation recheck(s); external support remains qualified as described above. TellGauss fixed C=1, soft=1, same qm; one prescribed setting, no sweep. tol=C*2.220446e-16*v and r=min(1,2*tol/g) when tell holds, otherwise 0. Here v is the net trade input, not posterior variance. In the embedded tellpol.py: v=amountX*(1−fb) and z=X on isBuy, otherwise v=amountY*(1−fa) and z=Y; g=min(ulp(z),ulp(z−v)); tell means abs(v/g−round(v/g))*g<=tol. Only L_ret is multiplied by r. ULP is computed with frexp/ldexp; nonpositive/nonfinite reserves are rejected. Tune 66000..66255, confirm 66256..66511, full holdout once. Same all-arb/all-retail totals as tg-1; retail quartiles use callback amountY (both sides), counts [209,589,1161,4076], false rates [.000459422,.001294733,.002552103,.008959817]. Overall retail false flag .003316520835; first-trade retail 2717/826104. Tell-off matches Gaussian bitwise, active hidden-scramble and offline-truth audit matches blind runner; six tuning tapes give 63,504 callbacks and 123 ULP boundary cases, not a proof for all callbacks. First-96 local audit differs from prior orchestrator: 456/672163 versus 459/672194 misreads, gain +2.9788276 over Gaussian and −.0312077 over flag; no tuning toward those earlier figures. It declines a numerical EVM gas estimate because no defined integer-callback tell exists. Engine/board transfer remains unestablished. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `tuning/vs_gauss_qm/paired` | 256 | 3.07023194626 | 0.0625833643135 | | `tuning/vs_flag_qm_diagnostic/paired` | 256 | -0.0164394230532 | 0.0094936450937 | | `confirmation/vs_gauss_qm/paired` | 256 | 3.13099539192 | 0.0620162485644 | | `confirmation/vs_flag_qm_diagnostic/paired` | 256 | -0.0318984016567 | 0.0089988512742 | | `holdout_vs_flag_qm/paired` | 2048 | -0.0257565316385 | 0.0027241635817 | | `holdout_vs_lattice_qm/paired` | 2048 | 2.69618031874 | 0.0208616865004 | | `holdout_vs_gauss_qm/paired` | 2048 | 3.19108085952 | 0.022113405967 | Diagnostics (misread fraction; RMS bps): `holdout_vs_flag_qm`: misread 0.000652486763, RMS 7.20786394 bps; `holdout_vs_flag_qm/reference`: misread 0, RMS 7.11191196 bps; `holdout_vs_lattice_qm/reference`: misread 0.162248261, RMS 13.6428397 bps; `holdout_vs_gauss_qm/reference`: misread 0.172764127, RMS 14.384445 bps. Cost / feasibility: O(1) first-new-timestamp tell: two ULP/exponent extractions, net-input arithmetic, division/rounding, distance and likelihood rescale; no extra dynamic state. No defensible numeric EVM gas estimate asserted. Additional limitations: Only fixed C=1 setting; blindness/bitwise tests cover six tapes, not all inputs. First-96 audit differs slightly from orchestrator and was not fitted. No reliable integer tell or EVM gas estimate. ### b55-tl-1 Source: [b55-tl-1](#b55-tl-1). Status: complete. TellLattice versus independently run TellGauss with fixed C=1, no fitting. Tune 66000..66255 is −.0161936732 SE .0116353034 over Gaussian tell; full holdout reverses to +.0110978313 SE .00414716035. This is the small benefit of lattice given the same blind tell, not a new true-flag lattice experiment. Mean edge 536.528202524, misread 7327/14350137=.00051058746 and price RMS 7.16007936 bps, versus TellGauss 9364/14351249=.00065248676 and 7.20786394 bps. Against true-flag Gaussian it is −.0146587003, and versus ordinary lattice +2.70727815005. Tell-off equivalence to both references and Poison hidden-input scramble pass on 66000/66001 with six-argument callbacks; oraclearb is rejected. Original dead-belief restart, zero-retail-mass pin fallbacks, hard thresholds and tables remain. No Solidity/gas measurement: 121 lattice cells, 33 kernel taps, up to 3993 inner iterations/diffusion and 31 diffusions/first trade; naive 121 storage slots exceed the 32-slot budget. Paired observations: | Field / comparison | n | Mean | SE | |---|---:|---:|---:| | `tuning_candidate_vs_control/paired` | 256 | -0.0161936732332 | 0.0116353034259 | | `holdout_candidate_vs_control/paired` | 2048 | 0.0110978313138 | 0.00414716034935 | | `holdout_candidate_vs_flag_qm/paired` | 2048 | -0.0146587003247 | 0.00377188778647 | | `holdout_control_vs_flag_qm/paired` | 2048 | -0.0257565316385 | 0.0027241635817 | | `holdout_candidate_vs_lattice_qm/paired` | 2048 | 2.70727815005 | 0.0207122002824 | Diagnostics (misread fraction; RMS bps): `holdout_candidate_vs_control/candidate`: misread 0.00051058746, RMS 7.16007936 bps; `holdout_candidate_vs_control/control`: misread 0.000652486763, RMS 7.20786394 bps; `holdout_candidate_vs_flag_qm/control`: misread 0, RMS 7.11191196 bps; `holdout_candidate_vs_lattice_qm/control`: misread 0.162248261, RMS 13.6428397 bps. Cost / feasibility: O(1) tell overhead plus 121-cell lattice, 33-tap kernel and up to 31 diffusions. A one-cell-per-slot implementation needs 121 probability slots before other state. No numerical gas claim or measured afterSwap peak. Additional limitations: True-flag control is Gaussian, not flag-aware lattice. Dead-belief/restart/zero-mass fallbacks remain. Replica tell has no engine/WAD transfer validation. ## Reconciliation and next handoff **Do not pool duplicate or overlapping replica work.** values7-1 is a different 25-action/fewer-seed design from values7-2’s nine-action experiment; both are negative, but they are not duplicate estimates of an identical policy. fresh7-1 is only 512 holdout seeds; its base gain .32215677 is a subset statistic, not a conflict with full-range .32289192. mass8-2 covers 1536 seeds inside mass8-1’s 2048; use the full result for full-range reporting and the partial one as corroboration of shared seeds. IG2 copies give identical pair aggregates on identical splits; count the experiment once. tg-1/tg-2 describe different tuning searches but freeze the same C=1/r policy and agree on final observations. Their different retail-size quartiles reflect net versus callback-amount definitions, and their gas estimates disagree in certainty, not measured runtime. Prefer “unmeasured/no integer implementation” as the gas conclusion. Sources: corresponding per-item sections. **Keep correctness separate from economics.** Repairing RivalT survivor normalization is required by its finite recurrence; small and noisy repair-minus-original scores do not invalidate the accounting fix. Conversely original RivalT’s .2084 over threshold includes inherited joint-router work, and repaired .2152 is still below .3 over threshold; neither establishes .3 for the repair alone. Integer recentering fixes one source of numerical diffusion but finite-window clipping disproves unconditional moment conservation. Posterior integration needs same-action/grid tests, not off-grid commutation of argmax and interpolation. IG2’s continuous-versus-discrete prior change prevents attributing its null result exclusively to ambiguity subtraction. Sources: values7-1/2, rival7-1, fresh7-2, mass8-1/2, ig8-1/2. **Priority supported by this root:** retain M09 as the empirically stronger frozen engine arm, merge these disjoint engine sufficient statistics with the other roots without reusing consumed seeds, and assess the requested 4,000-seed bar there. Treat the reserve-grid tell as the strongest replica lead requiring a new callback-level engine/WAD investigation before any port or performance extrapolation. TellLattice’s tiny advantage does not justify pretending its 121-cell filter already fits contract limits. No engine-validated tell, 20-run <=.01-bps Solidity parity evidence, new board score, or completed first-place solution is present in these 25 sources. This priority is a merge inference from the measured comparisons and stated limitations; no new experiment has been run. ## Source register Every listed JSON was present and parsed. Its exact original path is `artifacts/goal/batches//items//result.json`, with `` the source ID after its bNN- prefix. Supplied read copy is `.imd/reads/artifacts/`. SHA-256 below is over those supplied bytes, calculated by this merge; it identifies data, not behavioral certification. Encoded payloads, raw engine rows, command snapshots and repetitive provenance hashes are condensed, not additional delivered files. | Prefix | Original batch | |---|---| | b44 | a7-confirm-m09-a | | b45 | a7-confirm-m09-b | | b46 | b7-posterior-values | | b47 | b7-rival-transition | | b48 | b7-tail-risk | | b49 | b7-fresh | | b50 | a8-close-m09 | | b51 | b8-rival-mass | | b52 | b8-ambiguity-bonus | | b53 | b8-fresh | | b54 | b9-tell-gauss | | b55 | b9-tell-latt | | Input ID | SHA-256 of supplied result JSON | |---|---| | b44-eng-12007680 | `1f5d6a32741ac8d81d4cf5b132929204f6a0ef05208311e0dec081dd2fcb450a` | | b44-eng-12007936 | `80d6783a3353501d8328336cdb91dc16234737f52bc1a181dcfbb36a661f2496` | | b44-eng-12008192 | `b9aae47ca55936204c0411e0527338fa8f1efb4478547fdc006f6bfe628e462f` | | b44-eng-12008448 | `78abda6cf4cef610fe2075bc2374fbd0813f875230f2ba708f3289e96bf1766a` | | b45-eng-12008704 | `ff7af9c83a3310a5d68f1ce0ead595921cdde8b7b1489dbd6bce8a4292300bc4` | | b45-eng-12008960 | `f5fe1e0459b482d086a239978064b8c6736846a3fbcc30f46794faa5f6c2fe70` | | b45-eng-12009216 | `e947793e4387d03df93e1f33d725adae767028262b2b8aec04020423d8b7baeb` | | b45-eng-12009472 | `4220ddf0a4afa20734df659df15359977175f350399f9592c7680e3622330f71` | | b46-values7-1 | `3c96f125bad036f00fddf696c5dd8ad3d2d3a816d3c3f8a8c8b6d778cc6058eb` | | b46-values7-2 | `a7a3042c87321955ec960f8ccb9432a0e7dd45b93ce611faf7fd848ad3409ea4` | | b47-rival7-1 | `4bff6228dad5f9d86ce37985cfd0f4653753ebfb690756c547cd9ac88efb1f0b` | | b48-tails7-1 | `fed1d10e650e399ecf9ed86800c004cb1557ec7290c32a584d893c383c295fd0` | | b49-fresh7-1 | `1a3fe828b2b96e7d923ce49b08c5bcacd9886a133c9ac4cff23a8b86dba6fd7d` | | b49-fresh7-2 | `0caf5f316d0c511c1b8748c8c19052626e0d9d4aabdc3160f27cea3398308fb9` | | b50-eng-12009728 | `37096a34a13345a470a4fbfdca0be064dba0c81da6c06afc47b5af3a725111da` | | b50-eng-12009792 | `7a8f37e2dc7d4d6ecab3ffa60bf2f5e6a1dcf7a95c1b3c7b0b2395694c2dc85e` | | b51-mass8-1 | `4a0e539935a07e8ab8dddd427095efc5b0b34ee8bd252b66f841ad8c331b4737` | | b51-mass8-2 | `a8b005c66f39f6bfa36e127b5bb8b321701162805c22da43a9271960c28f063e` | | b52-ig8-1 | `bc08c5b7158c7922a4a39a29c129ed59e53d05da9a9f11b462bc2f6826249c2d` | | b52-ig8-2 | `c0db3f01f2808d1865093e05ffbf65394c749fa041a5eae825082f2d6b39c39d` | | b53-fresh8-1 | `b9aa2156600d7122e5ad0ea7ca44a8e2f4e258f7bd9b5be8f1403f82212c125b` | | b53-fresh8-2 | `2e4346889f1e938304957abcc33d6b3a547ac32051f82a209071a90928a95774` | | b54-tg-1 | `bb9ea8d48c176d7b07b7378680c0fb4e2d853882ec2070fc82179b7412b4e5c1` | | b54-tg-2 | `291066f0057c627e719fb643ab7f553f8efe3a0e5e323ec02ae5096be86d028c` | | b55-tl-1 | `f00ec0d31fae4bf74aaa1a3dce06e2b68e0081920c5495b09f58ea38ea880c9d` |