**Root merge 2/4 — Taylor (gtmcknight) AMM policy research, 25 results**

Credit: Taylor (gtmcknight) authored the replica, Gaussian filter, dynamic program, fee policy and qm tables at [amm-sim, pinned solved research](https://github.com/gtmcknight/amm-sim/tree/4812a6382311624e1506aca131ca35c378f930d9/research/solved). IdentityMD's qt1/TwoComp/COMBO, classifiers, residuals and tells extend that work. This merge reports the supplied experiments; it does not independently rerun simulations, certify contracts or submit a leaderboard entry.

The strongest delivered result in this group is the lean rounding tell: la-1 and la-2 report complete label/fee checks and final contracts within size and gas limits, with fresh-engine paired gains over qt1 of **+3.143933 (SE 0.046044, n=448)** and **+3.190869 (SE 0.045594, n=448)** respectively. Their identical Python report result is **+3.159895 (SE 0.022226, n=2048)**; this is the same policy result, not two statistically independent samples. Neither submitted to the leaderboard. These findings warrant retaining both final implementations for downstream review, not claiming the goal achieved. [b26-la-1: gates.Q/G/E; b26-la-2: Q/G/E]

The requester supplied the starting leaderboard figures: leader **528.42**, qt1 **524.94**, **8th**, about **+3.5** short; Taylor's baseline about **+3.8** over HA, first-trade call wrong about **16%**, and true-flag code **+6.98** over HA. These are task context, not a fresh leaderboard observation. Replica gains and new engine-seed means cannot simply be added to 524.94 to assert a new ranking. The lean results are roughly the scale of the stated gap but do not establish #1. [Task brief; b26-la-1: limits; b26-la-2: limitations]

All 25 named input files were present, nonempty and valid JSON. Their SHA-256 values were checked against the request; all matched. The source register below preserves their original artifact paths and digests. A bracket such as `[b23-tca-path: F]` cites that input's JSON field, not an independently verified external source. Exact paired statistics and worst-five tails are retained in the statistical annex. Figures in prose/tables are rounded where indicated; annex values come directly from the inputs.

**Evidence weights and source warnings — Taylor-derived results**

The task flags failed source rechecks for the following items. It does not identify the individual failed URLs/passages, so this merge cannot attribute a specific failure to a particular citation or claim. Input hash agreement establishes receipt of the result, not truth of its upstream citations. Retain local measurements as attributed reports; downgrade assertions that depend on those external citations. Several items themselves say quotes were refetched successfully, but that does not override the later warning supplied to this merge.

| Input | Failed sources reported by task | Consequence here |
|---|---:|---|
| b18-cost8 | 1 | Numeric/gas observations retained as local reports; inherited G4 report qualification is not a new head score or independently reverified source fact. |
| b19-o1-a | 1 | Port classification evidence retained with port-only scope; source/README assertions are not independent proof of scoring-engine behavior or organizer acceptance. |
| b19-o1-b | 2 | Same caution; its partial-window Bayes gains and catastrophic hard-label tails are reported, not promoted to complete Q. |
| b20-o2-a | 2 | Instrumented-source engine replay is reported evidence; it does not establish release-wheel or leaderboard equivalence. |
| b20-o2-c | 2 | Same caution despite its internal statement that all quoted passages were rechecked. |
| b22-net-head | 1 | Local stacking statistics retained; exact HEAD weights were already unavailable and baseline reconstruction is approximate. |
| b24-buy-grid | 2 | Source-build grid verification remains reported, sample-specific evidence, explicitly not release-wheel evidence. |
| b26-la-3 | 4 | Only a reported setup block and Gate P reproduction; no independent candidate or performance evidence. |

Unflagged downstream results provide additional measurements, not retroactive validation of every upstream citation. The engine, port and replica distinctions remain important because the tell depends on exact floating-point arithmetic. No newly browsed evidence is used in this merge. Public upstream references carried by these results include [Taylor README](https://raw.githubusercontent.com/gtmcknight/amm-sim/4812a6382311624e1506aca131ca35c378f930d9/research/solved/README.md), [Taylor replica](https://raw.githubusercontent.com/gtmcknight/amm-sim/4812a6382311624e1506aca131ca35c378f930d9/research/solved/replica2.py), and [pinned challenge engine](https://github.com/benedictbrady/amm-challenge/tree/6b6921f8d463314aa7c96234e1be25a25245f704); these links identify provenance, not fresh verification. [b19-o1-a/b; b20-o2-a/c; b24-buy-grid]

**Comparability and gate accounting — Taylor replica versus engine**

Unless a row says otherwise, full replica windows mean choose **64000..65023 (1024 seeds)** and report **10000..12047 (2048 seeds)**, with **10,000 steps** per seed. A paired mean is the candidate's edge minus its named reference on the same seeds, and SE is the sample standard deviation of those differences divided by sqrt(n). COMBO, TwoComp/TC-A, gated ComboG/G4 and qt1 are different bases; a gain over one is not a gain over the others. [b13-res-e/f; b14-bs-a/b; b16-smoothing-b2; b21-qt1; b22-net-head; b23-tca-path; b25-visit-refit]

Most completed experiments report Gate P's unchanged twelve README values matching at six decimals. The reusable rows are:

| Seed | Perfect info | 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 |

These Gate P columns are not qt1 or COMBO scores. Their separate baseline checks commonly give qt1 **443.101741 / 571.454483 / 525.908441** and COMBO **444.195330 / 572.172682 / 526.274760**. Gate P is not established in numba-tape (nothing run), o2-b (setup block only), or independently rerun in o1-b (explicitly inherited). o2-a supplies engine tracing rather than a P result. la-3 reports P but nothing beyond setup. [b13-res-e: gate_P; b14-bs-b: gates; b17-numba-tape; b19-o1-b: limitations; b20-o2-a/b; b26-la-3: P]

Gate Q for deployable policies requires observable inputs, separate selection/report windows and pairing against the actual base; true-price and true-flag interventions below are diagnostics, not live Q passes. Tell-policy Q uses the engine-order Python port on Taylor's random streams: source-order WAD arithmetic is needed and is not supplied by unmodified replica2 callbacks. Baseline parity checks support that harness but do not make its RNG the Rust RNG. Gate F means fee agreement within **0.01 bps on 50 × 2,000 callbacks**; high-cap numeric F is distinct from the **245,000 gas / zero-OOG** condition. T is label agreement; G is size/gas feasibility; E is paired engine evaluation in the relevant child results. [b15-loss-a; b21-combo/qt1; b23-tca-path; b25-visit-refit: caveats; b26-la-1/2]

**Learned residuals, belief solves and smoothing — Taylor-derived replica findings**

| Item | Full choose gain vs base (SE) | Full report gain vs base (SE) | Full report vs qt1 (SE) | Disposition |
|---|---:|---:|---:|---|
| res-e, COMBO residual | -0.182056 (0.015336) | -0.184300 (0.010696) | +0.412956 (0.014586) | Negative incremental result; below +0.5 qt1 contract threshold. |
| res-f, COMBO residual | -1.407254 (0.024138) | -1.339070 (0.017958) | -0.741679 (0.020433) | Negative; strict raw-return zero-head identity also fails. |
| bs-a, extended belief DP | -7.211455 (0.059104) | -7.162411 (0.040675) | -6.564993 (0.042098) | Pipeline/control failure, not evidence belief axes are intrinsically bad. |
| bs-b | Not run | Not run | Not run | Census and baselines only; no candidate. |
| smoothing-b2, COMBO smoother | +0.034512 (0.009384) | +0.040699 (0.006635) | +0.638090 (0.013823) | Small incremental gain; no contract/engine evidence. |

Each row is sourced to its named input's choose/report fields (bs-a uses choose_64000_65023/report_10000_12047; smoothing-b2 uses pairs). None of these items implements or measures a contract.

res-e used **14 RMS-scaled observable inputs, 8 tanh units, 138 parameters**, bounded **3 bps** bid/ask residuals and diagonal CMA-ES. It completed **102 generations** in about **2100 s**, selecting generation **12** by validation, not report. Training used 64000..64511 and validation 64512..64767 at 2,000 steps; full-window results used 10,000. Zero-head applied fees and all replica metrics matched COMBO on four choose seeds; randomized hidden-input replay also passed. Raw COMBO returns can exceed 0.1 before replica clipping, so its identity statement is about effective fees. Its measured runtime was **0.0153208 s/1,000 steps**. [b13-res-e: frozen_spec, training_spec, zero_head_identity_check, gate_Q, seconds_per_1000_steps]

res-f used **18 standardized inputs, 16 tanh units, 338 parameters**, **5 bps** bounds, population **12**, **13 completed generations** in **2099.25 s**, and selected generation **12**. It rotated 256-seed minibatches within the same training half and validated every fourth generation. Applied zero-head fees were identical, but raw returns differed twice: COMBO returned about **0.150729 / 0.153602**, the clipped head **0.1 / 0.1**. This is a real strict-identity failure, not an economic difference after clipping. Both nonzero and selected-head hidden-input replays passed. Report throughput was **0.130188 s/1,000 steps**; short-horizon validation did not generalize to a useful residual. [b13-res-f: frozen_spec, training, zero_head_identity_check, hidden_input_invariance, selected_policy_hidden_input_invariance, report]

bs-a's collapsed control lost **-7.230291 (SE 0.041030)** versus COMBO on report. Adding belief axes inside that pipeline recovered only **+0.067880 (SE 0.010662)**; choose recovery was **+0.064024 (SE 0.014105)**. The extended solve had **198,671,352 values** over eight populated levels; control had **24,833,919**. Control converged after **102 sweeps** to max change **1.968389e-7**; extended stopped at **23 sweeps**, max change **8.578324e-6**, span **1.294586e-5**, with oscillation and no convergence to tolerance. Reduced displacement range **±300 bps**, **121 rather than 321 rows**, and mean-pi mixing at all phase-0/1 decisions are unisolated suspected causes of the large baseline loss. Fill buckets were marginalized in the DP. Observable replay passed. Calibration examples mean-pi/true-arb were **0.9335/0.9275**, **0.5594/0.5579**, **0.3305/0.3219**; four of twelve proposed belief cells did not occur. Do not credit the +0.068 internal recovery as a deployable gain. [b14-bs-a: solve, control_comparison, findings, limitations]

bs-b ran no RVI sweeps, created no extended/collapsed tables and scored no candidate (all candidate n=0). It collected **692,152** COMBO callbacks on **64000..64063**, with **30,323** zero-sd quotes; sd quartiles were **5.500282 / 11.695291 / 17.046391 bps**. There were **95/180** observed conditional rows and **987** nonzero entries; sd-level0/pi-middle was unvisited. Randomized hidden-input replay matched every callback. Its census omits intermediate silent-step updates, measures unchanged-COMBO transitions only, has no action-dependent counterfactual model or unvisited-row treatment, and defines later same-step pi as zero. It nonetheless reproduced COMBO−qt1 **+0.609604647 (SE 0.019400523)** choose and **+0.597390963 (SE 0.013387994)** report. These are reusable baseline measurements, not a solved belief policy. [b14-bs-b: gates, transition_counts, solve, choose, report, limitations]

smoothing-b2 carries up to **four** components, prunes below tau, merges closest same-label pairs by moment matching, propagates silent steps up to **12**, and uses a later-fill approximate size likelihood with fixed mean **20** and log-sd **1.2**. Choose tau/gain-vs-COMBO: **0.001/+0.034512**, **0.01/+0.028620**, **0.03/-0.065760**, **0.1/-0.729842**; tau **0.001** was frozen. Report first-call error was **16.036770%** versus COMBO **16.118613%**; delayed error **16.038341%**, slightly worse than its own first call. Average retained components **3.612026**, mean pruned mass **0.000195908**, maximum **0.003695040**. Thus its later-fill update did not materially improve delayed hard classification. Error counts are on each policy's own trajectory; pruning also affects the delayed diagnostic. [b16-smoothing-b2: tau_grid_gains, report.diagnostics, method]

With the true flag injected, smoother and ComboFlag were identical on every seed: report **Cflag−ComboFlag=0**, while ComboFlag−COMBO was **+2.570119360 (SE 0.020762657)** and ComboFlag−qt1 **+3.167510323 (SE 0.022385774)**. The oracle collapses branch ambiguity to a singleton. Per-component rival estimates reset from shared COMBO state each first callback; the likelihood does not integrate rival/price uncertainty or callback selection. Neither the modest smoothing gain nor the oracle equality proves engine performance. [b16-smoothing-b2: ComboFlag_check, method, limitations]

**Where the remaining loss sits — Taylor price and flag diagnostics**

loss-a's PricePol reads true price at every callback, sets zero belief variance and therefore also obtains an effectively exact first-trade flag. It is a price-plus-flag oracle, not merely a better classifier. On report it scored **+8.021774 (SE 0.045883)** over HA, **+3.718987 (SE 0.024144)** over COMBO, and **+4.316378 (SE 0.025937)** over qt1. Corresponding baseline gains were COMBO−HA **+4.302787 (SE 0.037598)** and qt1−HA **+3.705396 (SE 0.038086)**. The expected roughly +8.50 used plain pol_mid tables; PricePol used QMS1-compressed qm tables at sd=0. A table explanation for the roughly 0.48 shortfall is plausible but was not isolated or rerun through track.py. [b15-loss-a: frozen_spec, report, findings, limitations]

On diagnostic seeds **64000..64255**, **2,560,000 steps**, PricePol−COMBO totaled **945.343**, or **3.692748 per seed**: arb edge **581.387** and retail edge **363.957** (rounding explains the last decimal); fee difference **941.794** is a separate accounting view, not a third additive component. COMBO miscalled **289,088/1,793,071** first trades (**16.12%**). PricePol's inferred call agreed with all **1,787,791** true flags; |inferred log-price minus true log-price| was at most **8.88e-16** on arb, at least **2.724e-5** on retail. [b15-loss-a: diagnostic]

| Diagnostic partition | Shares of total PricePol−COMBO gap | Source |
|---|---|---|
| First-call correctness | wrong 22.35%; right 72.63%; no own trade 5.02% | loss-a diagnostic.call |
| Belief error at last quote | <2.5 bps -5.44%; 2.5–5 1.73%; 5–10 8.23%; 10+ 95.48%; none 0.01% | loss-a diagnostic.err |
| Quote phase | hard arb -1.58%; hard retail 29.30%; ambiguous 72.27%; none 0.01% | loss-a diagnostic.phase |
| Silent steps | 0: 76.94%; 1: 11.48%; 2–3: 7.56%; 4–9: 3.06%; 10+: 0.96% | loss-a diagnostic.silent |
| Fill-size | absolute z<1: 45.72%; 1–2: 46.38%; >2: 2.88%; no own trade: 5.02% | loss-a diagnostic.size |

The size row uses absolute standardized size z. These are overlapping partitions, not additive across rows. Ambiguous/zero-silent and hard-retail/zero-silent cells account for **53.25%** and **25.01%**. Wrong-call steps have about **1.6×** the per-step gap (**0.000731 vs 0.000457**) but right calls are much more frequent. Error ≥10 bps accounts for **902.579** gap, including **633.870** arb loss. Long silence has high per-step loss (**0.003886**) but low frequency. [b15-loss-a: diagnostic, findings]

Crucial correction to loss-a's interpretation: its statement that fixing the call addresses “at most about a quarter” is **not a causal upper bound**. Its own limitations say pool states diverge and earlier quotes affect later inventory; buckets are labels on COMBO's trajectory at the same seed/step. The explicitly intervened flag experiments below recover much more than a quarter of PricePol's advantage. Preserve the descriptive attribution, reject that causal inference. [b15-loss-a: findings versus limitations; b21-combo/qt1]

| Oracle intervention | Choose gain over its own base (SE), n=1024 | Report gain over its own base (SE), n=2048 |
|---|---:|---:|
| COMBO FlagSell | +1.590665 (0.023481) | +1.622094 (0.017508) |
| COMBO FlagBuy | +1.570674 (0.023613) | +1.601987 (0.017350) |
| COMBO FlagAll | +2.503358 (0.028557) | +2.570198 (0.020763) |
| COMBO FlagAll−FlagSell | +0.912693 (0.019764) | +0.948104 (0.013811) |
| qt1 FlagSell | +2.023970 (0.025661) | +2.037043 (0.018857) |
| qt1 FlagBuy | +1.985588 (0.025518) | +2.018623 (0.018712) |
| qt1 FlagAll | +3.112983 (0.030987) | +3.167510 (0.022386) |
| qt1 FlagAll−FlagSell | +1.089013 (0.020582) | +1.130467 (0.014258) |

Sources: b21-combo.windows; b21-qt1.paired. Sell means **pool sells X, isBuy=false**; buy means **pool buys X, isBuy=true**. Both items ran full windows, with no tuning; combo ran report first, so do not describe that diagnostic as choose-before-report selection. Oracle flag-only gains do not add: COMBO's sum of one-side report gains **3.224081** exceeds all-side **2.570198** by **0.653883**; qt1's **4.055666** exceeds **3.167510** by **0.888156 (SE 0.015827)**. Choose gaps are **0.657981** and **0.896575 (SE 0.022204)**. qt1 sell-only recovers **64.31%** of all-side report gain. [b21-combo: additivity; b21-qt1: additivity, interpretation]

qt1 report base errors are **1,242,721/7,181,967 sells (17.303352%)** and **1,245,109/7,192,478 buys (17.311266%)**; choose rates are **17.224559% / 17.214827%**. COMBO combined-window rates are **16.0919% sell / 16.0846% buy**. These are compatible with improved COMBO versus qt1 and differing samples, not a contradiction to the approximate 16% premise. COMBO's all-flag gain differs slightly between b21-combo (**2.570198**) and b16 (**2.570119**), with reported base means **533.860143** versus **533.860222**; preserve each source value rather than silently equate them. The b21-combo limitation that only sell-side truth might be recoverable is overtaken by later buy-grid measurements, not by the oracle experiment itself. [b21-qt1: first_trades; b21-combo: first_trades, limitations; b16-smoothing-b2: report]

**Acceleration and stacked corrections — Taylor-derived tools and policies**

numba-tape is a non-delivery: no source fetch/inspection, gates, kernel, parity, hook or timing work; all requested measures are null. It contributes no negative evidence about whether acceleration can work. numba-rng delivered a scalar Numba port with CPython RNG behavior: **384/384** baseline rows across QMRef/TwoComp/COMBO on **10000..10063 and 64000..64063**, maximum edge error **1.19e-11**; **128/128** rounded HEAD8 rows, maximum **1.53e-11**, both below **1e-6**. Zero head equals accelerated COMBO exactly; hidden-input and observable replay checks passed. [b17-numba-tape; b17-numba-rng: verdicts, parity, invariance]

Two concrete parity repairs matter for reuse: CPython 3.12 compensated sum required matching eight-term head summation; residuals had to be added to **unclipped** base fees before final clipping. An earlier seed-10039 divergence of **4.05e-4** disappeared after these fixes. Final parity is close, not universally bit-exact; only default configurations/10,000 steps were tested and qcap was not ported. Warm runs/s were **16.4123 QMRef**, **14.4037 TwoComp**, **12.4093 COMBO**, **8.4584 HEAD8**, **8.9006 zero head**; cold first-run JIT **74.5817 s**, log/replay first compilation **17.6599 s**, imports about **1.83 s**. The **50 runs/s/core target failed**. The source explicitly retracts an inflated roughly 200× contention-based comparison in favor of roughly **20–25×** from uncontended runs; >50/s elsewhere is unmeasured. HEAD's allocation overhead remains. [b17-numba-rng: divergences_found_and_fixed_during_development, timings, limitations]

net-head stacks a frozen **12-input/16-tanh NET** classifier (quantized weights, logit guard 3) and rounded HEAD8 output weights without retraining. All five policies have full coverage. On report, NET−COMBO **+0.078039 (SE 0.009144)**; HEAD−COMBO **+0.076607 (0.008472)**; STACK−COMBO **+0.148967 (0.010412)**; STACK−qt1 **+0.746358 (0.014554)**. Choose STACK−COMBO **+0.127323 (0.014661)** and STACK−qt1 **+0.736927 (0.020197)**. Report interaction STACK−NET−HEAD+COMBO **-0.005679 (SE 0.010402)**; choose **-0.009924 (0.015023)**: compatible with additive gains at this precision, not proof of exact additivity. [b22-net-head: paired, additivity; one source flagged]

NET reproduces its source at six decimals. HEAD does not: published choose/report gains **0.080720 / 0.076499**, rebuilt **0.080902 / 0.076607**; exact frozen weights were unavailable, and 9-decimal rounding versus platform differences were not isolated. Degenerate STACK/part identities and hidden-input tests passed. The net guard follows the actual probability-to-logit code, not the abbreviated one-line specification. Report first-call error rates: qt1 **17.3073%**, COMBO **16.1186%**, NET **15.7247%**, HEAD **16.1198%**, STACK **15.7202%**. The net removes about **0.39 percentage points**; the head adds score without an appreciable classification improvement. [b22-net-head: rebuild_checks, first_trade_error]

Tail risk is retained: STACK loses **-4.416473** versus COMBO on seed **10632**, versus NET **-4.571141**, versus HEAD **-4.094168**, worse than either part's worst standalone report tail. Between steps **9925..9975**, STACK makes **20** first-call errors versus NET **1**, COMBO **2**, and loses about **4.0** edge. The detailed mechanism is unexamined; do not remove that seed. Excluding it only as sensitivity gives interaction **-0.003606 (SE 0.010198)**. Parts share runs and comparisons are correlated. No engine/contract/gas evidence for STACK exists; choose parts were trained/selected there, so report is the clean comparison. [b22-net-head: tail_seed_10632, additivity, limitations]

cost8 separately measured rounded HEAD8 arithmetic on **gated ComboG GATE=4**, not ungated COMBO or the whole STACK. Numeric F passed at a **500,000** measurement cap over **100,000 callbacks**, maximum head/hypothetical fee error **3.104958e-6 bps**, but gas failed: base peak **215,608**, overhead-only **215,676**, head **246,963**; maximum added **51,564** (interface **75**, arithmetic **51,496**), mean added **30,980.595**, versus target **12,000**. Actual **245,000-cap** replay had **2 OOGs**, at **70030/166** and **70046/1790**, stopping those tails; **97,956 successful callbacks**, not a full successful 100,000. Runtime sizes base **22,370**, overhead **22,399**, head **23,801**, qt1 **20,605 bytes**. This rejects this implementation's feasibility, not every possible optimized head. [b18-cost8: F, gas, sizes; one source flagged]

cost8's engine row is only unchanged cp-c/G4 versus qt1, each independently vs 30-bps Vanilla: **n=16**, seeds **9900000..9900015**, **+0.507088 (SE 0.175472)**, candidate peak **228,340**, qt1 **208,997**, zero OOG. Its inherited G4 report qualification **+0.601755886 (SE 0.013533815, n=2048)** was not remeasured there and says nothing about a head-corrected policy. No head engine run, new fee-policy score, submission or deployment occurred. [b18-cost8: Q, engine, limitations]

**Rounding tell: port evidence, compiled-source checks, and disagreements — Taylor-derived classifiers**

The tell uses only current TradeInfo, previous own reserve WADs and previously returned fee WADs. Let gamma be the engine's binary64 `1-fee`. On sells, the arb path implies amountY is the WAD conversion of rounded `(newY-oldY)/gamma`; on buys, arb amountX lies on the grid of rounded `(virtualX-oldX)/gamma`. All operations and conversions must preserve binary64 round-to-nearest-even and engine order; ordinary real-number algebra or simple fixed-point arithmetic is not equivalent. Retail can accidentally lie on the same grid, particularly for large amounts. “Both paths compatible” is not an impossible state and must not be read as a certain arb. [b19-o1-a/b; b20-o2-c: method; b24-buy-grid: method; flagged upstream sources as above]

| Item and evidence | Seeds / callbacks | Sell arb matches; retail false positives (all / first) | Buy grid matches / retail false positives | Scope |
|---|---|---|---|---|
| o1-a Python port | 74000..74007; 82,788 | 15,336/15,336; 39/26,448 and 19/12,288 | 14,812/14,812; 44/26,192 all, 13/12,107 first | Port vs replica edge max error 2.2584e-10; no Rust run. |
| o1-b Python port | 74100..74107; 91,908 | 16,591/16,591; 31/30,352 and 15/13,131 | 15,172/15,172; 43/29,793 all, 16/12,889 first | Port vs replica within about 1.7e-10; no Rust run. |
| o2-a traced source engine | 9950000..9950007; 84,954 | 15,038/15,038; 36/27,532 and 18/12,698 | Grid not tested; shared buy path indistinguishable | Actual-path replay matches all reported WADs. |
| o2-c traced source engine | 9950016..9950023; 91,299 | 15,726/15,726; 31/29,743 and 9/12,885 | Grid not tested; shared buy path indistinguishable | All actual-path bits and WADs match. |
| buy-grid traced source engine | 9950024..9950031; 93,187 | 15,100/15,100; 47/30,777 and 20/13,338 | 16,441/16,441; 41/30,869 all, 19/13,128 first | Compiled pinned source, expressly NOT release wheel. |

Sources: each item's matrices/port_matches_engine/validation; port totals in o1-b are the sum of its buy/sell totals. Zero arb false negatives applies to these samples only. First-trade diagnostics in the engine tables include t=0; some policy diagnostics require dt>=1 and therefore exclude timestamp zero. These populations should not be pooled blindly.

o1-a's tell errs on **32/54,543 first trades (0.0587%)** versus COMBO **8,821 (16.1726%)**; it corrects **8,818** of COMBO's mistakes. Sell strict exclusive labels identify **15,436** correctly but leave **12,188** ambiguous; **12,169** of those are arb, **19** retail. Sell reserve preimages number 1–2 but amountX preimages reached **591**; collapsed intermediate replay averaged **2.775** path evaluations, max **8**, with brute-force checks on two seeds agreeing. Approximate grid widths **1.82e6 WAD Y** and **14,250 WAD X** are local reserve-scale examples, not constants. Buy execution redo matched every buy, router trailing-zero test was unhelpful, and MIN_AMOUNT picked out only **12** buy arbs with no retail. No fee policy was rescored. [b19-o1-a: matrix, candidates, combo_errors, buy_side]

o1-b's first-trade tell makes **31/57,783** errors versus COMBO **8,930 (15.454%)**, correcting **8,926** of COMBO's errors. Buy MIN_AMOUNT identified just **14** arbs; same-function buy replay still supplies no discrimination. Its Bayes tell uses sell-both likelihoods **13069/16591** (arb) versus **15/13131** (retail), buy-positive **1** versus **16/12889**, calibrated on 74100..74107. On 256-seed choose/report subsets, Bayes−COMBO is **+2.530980 (SE 0.058713)** / **+2.543263 (0.057067)**; true-flag gains **+2.543321 / +2.554086**, differences from Bayes **0.012341 / 0.010823**. The summary's “within 0.011” holds on report, not exactly on choose. This is partial-window port evidence, not full Q or qt1 pairing. [b19-o1-b: rule_error_rate, combo_errors, edge_with_tell]

Hard 0/1 labels are unsafe despite low classification error: choose **-0.966195 (SE 0.715807)** versus COMBO, worst **64120/-116.003659**; report **-0.346020 (0.817314)**, worst **10146/-158.769603** (also **10208/-87.540656**, **10220/-60.373974**). Bayes report errors were **469/1,790,420**, hard-label **1,059/1,790,139**, on different closed-loop trajectories. This is strong evidence to retain uncertainty/retail mass rather than hard-pin every positive match. [b19-o1-b: edge_with_tell]

o2-a's sell matrices give **3,024 arb-only, 12,014 both** among arbs; retail **27,496 retail-only, 36 both**. Its inverse enumeration started from captured valid floats, useful for exhaustive diagnostic replay but not itself a public-input implementation. o2-c instead inverted WADs with monotone binary searches, checked all preimage boundaries and used hidden bits only for validation: first sells **16,557 exclusive / 28,611 total (57.8694%)**, **12,054 ambiguous**, of which **12,045 arb / 9 retail**. It reports **100** mutation rejections and **861** enumeration self-checks. A tiny amount had **9,444** float preimages, illustrating why unbounded full-path enumeration is unsuitable for a gas promise. Neither item measures a score gain or contract. o2-b contributes only a setup failure at rustup installation: the installer reported it could not locate its working directory; no simulations or findings beyond the block. [b20-o2-a: limitations; b20-o2-c: finding, method, validation; b20-o2-b: blocked]

The apparent buy-side disagreement is resolved by the predicate: **same execute_buy_x path ≠ same upstream amount generation**. o2-a/c's “no tell” is correct for execution-path compatibility only; o1-a/b and buy-grid test the upstream grid. buy-grid's independent complete public-WAD search agrees with original o1-b on all **47,310 buys**, and full versus one-division sell tests agree on all **45,877 sells**. Its **93,187** actual execution bit/WAD replays, **186,374** fee conversions and **16,441 buy / 15,100 sell** arb-sizing replays have zero mismatches. That supports the mechanism on the pinned source build, not every build or the release wheel. The source-warning downgrade still applies. [b20-o2-a/c: buy_side; b24-buy-grid: answer, validation]

buy-grid's **standalone** budget-4 buy grid and sell one-division predicates have zero fallbacks in sample. At budget 2, buy fallbacks are **4,340 arb +10 retail**; sell one-division **978 arb +2,175 retail**. A joint A/R compatibility matrix uses different costs: even budget-16 full sell still falls back on **111 arb +80 retail**; one-division plus retail replay still has **1 arb** fallback. Do not confuse standalone with combined costs. Budgets limit candidate evaluations, not preimage-search work or EVM gas; unsupported/capped outcomes are unknown, never negatives. buy-grid also checked **125** virtual-grid cases/mutations and **60** fee-endpoint cases; zero observed false negatives remains empirical. [b24-buy-grid: method.cap_definitions, matrices, standalone_bounded, validation]

**Contract progression — Taylor-derived tells from infeasible to lean**

| Item | Replica report vs qt1 | Label/fee checks | Runtime bytes | Gas/engine outcome |
|---|---:|---|---:|---|
| tca-path | +3.133460 (SE 0.022264), n=2048 | T 350,662 first trades, 0 mismatches; F 100,000 callbacks, max 4.82993e-7 bps | 24,552 | Replay peak 298,073; pilot 2,579 OOG; E not run. |
| tca-short | +3.157779 (SE 0.022523), n=2048 | Python comparison only; no Solidity T/F | None | Report worst -2.383452 violates its -2 tail condition; no contract/engine. |
| qtr2-path | +3.149894 (SE 0.022363), n=2048 | Python checks, limited standalone soft-float tell; integrated T/F incomplete | 25,564 | 988 bytes over EIP-170; deployment fails; no engine seeds. |
| la-1 final | +3.159895 (SE 0.022226), n=2048 | T 351,787, 0 mismatches; F 100,000, max 4.82996e-7 bps | 24,383 | Final E 448 seeds, peak 232,160, zero OOG. |
| la-2 final | +3.159895 (SE 0.022226), n=2048 | T 351,787, 0 mismatches; F 100,000, max 4.82996e-7 bps | 23,960 | Final E 448 seeds, peak 239,497, zero OOG. |
| la-3 | Not run | Only P reported; setup failed | None | No implementation, scoring or engine seeds. |

Sources: b23-tca-path Q/T/F/G/E; b23-tca-short Q/T_python/F/G/E; b23-qtr2-path gates/replica_results; b26-la-1 gates; b26-la-2 Q/T/F/G/E; b26-la-3. Numeric F for tca-path was measured at **2,000,000 gas**; it does not excuse gas failure.

tca-path's bounded tell Q choose gains were **+3.077736966 vs qt1 / +2.669511044 vs base**; report **+3.133460282 / +2.708529950**. It fell back **145,781** times in report. T used **74200..74249**, 10,000 steps; independent-filter F used exact contract-driven fee WADs. Replay had **131** callbacks over 245,000. Its engine pilot **9963072..9963079** yielded **-1.497195 (SE 1.196684, n=8)** versus qt1, worst **-8.375487**, with **83,962 calls, 81,383 successful, 2,579 OOG**. Highest successful gas **249,996** omits halted calls at the engine's **250,000** cap and already exceeds the task's 245,000 threshold. Planned E ranges **9960000..9961023** were not run/reserved. Keep the source and passing arithmetic checks, reject deployment of this version. [b23-tca-path: Q/T/F/G/E, engine_seed_ledger]

tca-short selected **EPS=0.003**, **CAP=16** on full choose only; tested EPS **0 / 0.003 / 0.01 / 0.03**, with choose vs-qt1 means **3.103252 / 3.104106 / 3.086704 / 3.049387**. Its sell shortcut/full-redo comparison had **625/351,773** differing labels, all full-redo cap fallbacks; **zero supported disagreements**. Report tell errors **3,953/14,353,157 (about 0.028%)**, filter errors **1,709,493**, fallbacks **213**. Report gain versus exact TwoComp base **+2.732981 (SE 0.020932)**; worst qt1 seed **10481/-2.383452**, worst versus base **-2.430504**. The failed tail criterion cannot be removed because mean passes. No soft-float Solidity was delivered. [b23-tca-short: policy, T_python, Q, F]

qtr2-path froze **cap=8, floor=0.0001, budget=128, dual_cuts=2**, choosing among floors **1e-5 / 1e-4 / 1e-3** by choose mean subject to min≥-2. Report worst **11062/-1.397359**, **49,808** fallbacks, coverage **0.996531823**; exact-base comparison equals qt1. Python labels match **351,717** first trades; standalone soft tell matches **5,678** comparisons on four shorter streams, with **25** fallbacks. Soft64 tests covered **100,000 each add/sub/mul/div/WAD plus 102,682 uint** cases; unsupported/prior-preservation tests passed. Lossless signed-delta LEB128 compaction preserved table bytes, but integrated runtime **25,564** failed CreateContractSizeLimit, so integrated T/F and engine tests remain absent. Planned pilot **9963088..9963095** and E **9962048..9963071** were unused. [b23-qtr2-path: replica_results, python_label_check, soft64-test, soft-tell-test, edgecases, fallback-test, candidate, engine_seed_ledger]

la-1/2 replace expensive full-path work with a fixed-budget lean grid: sell cost is previousY×currentY preimage count; buy cost is previousX count plus virtualX candidates, budget **4**. Positive updates use retail likelihood **15/13131 sell, 16/12889 buy**, with posterior capped at **0.9999** and mixture retained; negative sets pi=0; unsupported/budget exhaustion preserves prior. la-1 uses at most **6-ULP** monotone walks; at most **1** step was needed on its observed streams. Its comparison against buy-grid had **56,445** first trades and zero disagreements; **20,000** random preimage checks had zero supported disagreements, **2,406** over-budget outcomes. This is not a claim that every arbitrary WAD is supported. Report and debug had no fallback in la-1. [b26-la-1: reference, gates.tell_exactness, gates.Q; b26-la-2: bundled NOTES.md]

la-1's final tell-only cost is **18,190 peak / 11,054.847 mean gas**, measured probe-minus-noop. It is not directly comparable to la-2's **26,459**, measured internal gasleft with different overhead. la-1 final F replay peak **195,813**; final G pilot **9970016..9970023** gives **+3.947216 (SE 0.284037, n=8)** and **216,022** peak, zero errors. Its final E on **9970164..9970355** and **9970400..9970655** gives **+3.143932838 (SE 0.046044361, n=448)**, all positive; min **9970281/+0.427909874**; **4,825,492** candidate callbacks and **232,160** peak. Runtime headroom is **193 bytes**; observed gas headroom **12,840**. Larger silent gaps, rather than tell cost alone, drove the remaining peak. [b26-la-1: T/F/G/E, gas_profile_pyrevm, limits; headrooms also stated/calculated from reported limits]

la-1's headline E **+3.165517119 (SE 0.042864962, n=512)** pools **64** earlier-build-v2 seeds **9970100..9970163** with **448 final-build** seeds. v2 reports **+3.316607090 (SE 0.116256557)**, peak **239,374**. The item supports pooling edges by **547,289** equal pyrevm callback outputs and equal fee hashes between v2/final, but this is not final-build engine execution on those 64 seeds and gas is not pooled. Preserve the 512 statistic as qualified historical evidence; use 448 for final-artifact evidence. Earlier pilots v1/v2 used **9970000..9970007 / 9970008..9970015**, runtime **24,124 / 24,202**, peak **241,348 / 218,149**. [b26-la-1: builds, gates.G/E, gas_profile_pyrevm]

la-2's first version passed pilot but reached **246,180 gas** with zero OOG on **9971100..9971163**: this **fails 245,000 despite no engine OOG**. A specialized exact RN-even WAD multiply then reduced cost; F outputs matched v1 over 100,000 callbacks. Final pilot **9971008..9971015** reports **+3.395409 (SE 0.446632, n=8)**, peak **234,067**. Final E excludes all superseded 64 seeds: **9971164..9971355 (192)** and **9971400..9971655 (256)**, combined **+3.190869021 (SE 0.045593915, n=448)**, peak **239,497**, zero OOG/other errors over **4,858,855** candidate calls. Worst **9971438/-1.268483989**. Final size **23,960** leaves **616 bytes**; observed gas margin **5,503**. The two lean variants have different fresh seeds and implementations; this mean difference alone does not establish la-2 better. la-1 has more observed gas margin, la-2 more size margin. [b26-la-2: F/G/E, checks, bundled NOTES.md; margins calculated from stated limits]

la-1 and la-2 used only **64000..64063** debug coverage and no reported tuning before their full 2048 report runs; they do not provide a fresh full 1024-seed choose screen. Retain their local Q passes with that coverage caveat under the broader goal's choose/report discipline. Their report result is identical, and parent tell policies already exposed report seeds; freeze assertions support no within-item report selection, not unlimited independence for future adaptations. Each reports separate candidate/qt1 arms versus Vanilla with MatchRunner-equivalent configurations; no contract was submitted. [b26-la-1: gates.Q, seeds_consumed; b26-la-2: checks, seeds_consumed, bundled NOTES.md; task gate definition]

la-3 ran P and checked input/wheel hashes but setup stopped at `uv venv --python 3.12 .venv`, exit **126**, “cannot execute binary file.” No candidate, Q/T/F/G/E or engine seed consumption followed. Its supplied historical artifacts are not la-3 achievements. o2-b likewise stopped at required setup and adds no measured tell evidence. These are reported execution blocks, not requester-choice gaps; no host inspection is part of this merge. [b26-la-3: blocked, setup, implementation_status; b20-o2-b: blocked]

**Visit-weighted table refit — Taylor-derived incremental gain, separate from lean validation**

visit-refit freezes a different TwoComp+Bayes-tell BASE, **floor=0.001, cap=64**, then changes only QMS1 coefficients. Floor/cap selection used **64000..64255**; the cap acceptance rule moved from intended 1-SE to **2-SE after seeing the grid**, still before fit/report. Floors were within about **1.5 SE**, so selection is weakly resolved. Training visits used **64000..64511**, **5,572,165** lookups; validation **64512..65023**, **5,601,643**. Visits concentrated in tables 0c **348,388 ppm**, 1s **399,543**, 1c **63,103**, 2s **184,677**, 2c **4,290**. [b25-visit-refit: base, visits]

Regularization lambda **10** won edge selection on the validation half: **+0.058324 (SE 0.007068, n=512)** versus BASE. Lambda **0.03 / 0.3 / 3 / 30 / 300** gains were **0.047187 / 0.040894 / 0.057253 / 0.055207 / 0.035022**; **3000 / 30000** were fitted but not edge-scored. Lambdas 3/10/30 were within **0.5 SE**. Full choose gain C−BASE **+0.047886 (SE 0.005075)** is partially selection-biased; training half alone **+0.037449 (0.007262)**. Report **C−BASE +0.049063 (SE 0.003681)**, **BASE−qt1 +3.156474 (0.022252)**, **C−qt1 +3.205537 (0.022437)**; worst C−BASE **11701/-1.402185**, worst C−qt1 **11062/-1.338114**. [b25-visit-refit: fit, scores]

Original/new blobs are each **10,167 bytes**; **375** header/knot bytes and layout unchanged; **1,599/4,896** coefficients change, max **11 bps**, no int16 clipping. Validation weighted RMS fee error improves **0.8346→0.7567 bps**; unvisited RMS **5.0202→5.0083**, max unvisited node movement **2.2623 bps**. Nodes remain below **817.72 bps**; mirrored bid/ask checks, decoder equality and **4,096** clamped out-of-grid tests pass. Full tables improve BASE by **+0.108129 (SE 0.012871, n=512)** on the validation half, a useful compression comparator; treating this as a universal maximum recoverable score is not proved. [b25-visit-refit: fit, checks, full_table_diagnostic]

Report BASE/C tell fallbacks **6,190 / 6,179**, errors **4,894 / 4,962**; original pi error on those trajectories about **11.888% / 11.910%**, after tell **0.03409% / 0.03457%**. Own-WAD bookkeeping/hidden-input invariance matched all **46,429 BASE / 46,413 C** callbacks on four check seeds. PyPy was rejected after per-seed differences up to **0.14** on 16 seeds; all results use CPython and explicit left-to-right sums, which can differ from Python 3.12 builtin sum. Baseline port parity's largest error was **1.26622e-7** for TwoComp seed 74102. On unmodified replica2 without WADs the policy falls back to TwoComp. Knots were inherited from original Gaussian-filter visits; one full-table bid mirror irregularity remains represented by the mirror. No contract, engine or gas measurement for this refit exists. **Do not add +0.049063 to either lean engine score:** the BASE, tell, floor and execution differ, and their composition has not been measured. [b25-visit-refit: hidden_input_invariance, harness_check, scores, caveats]

**Downstream decisions and preserved uncertainty — Taylor-derived synthesis**

1. Retain both lean final artifacts as the strongest engineering outcome; state final E **n=448** for each, distinguish la-1's historical mixed-build pool, retain tails and margins. No evidence here establishes leaderboard submission or rank. [b26-la-1/2]
2. Retain the tell mechanism and its rare retail false positives. The decisive progress is combining exact observable arithmetic with Bayesian uncertainty and cheap predicates; full-path agreement alone was too expensive. Build changes or different floating arithmetic require renewed checks. [b19-o1-b; b23-tca-path; b24-buy-grid; b26-la-1/2]
3. Keep table refitting as a measured incremental replica result, and smoothing/NET/HEAD as smaller alternatives. Do not assume their gains add to the lean tell. STACK's own additive test is limited to NET+HEAD on COMBO and has a new tail episode. [b16-smoothing-b2; b22-net-head; b25-visit-refit]
4. Retain negative results: res-e/f regressions; bs-a broken control/nonconvergence; cost8 and tca-path gas failures; qtr2-path size failure; tca-short missing contract and failed tail threshold. Retain partial/non-deliveries: bs-b, numba-tape, o2-b, la-3. None is silently dropped or represented as a successful policy. [Named inputs above]
5. Preserve differing baselines/rounding: COMBO−qt1 report **0.597391** in several items, **0.597418** in bs-a, roughly **0.597256** implicit in res-e, and **0.597368** in net-head's cited older HEAD run. These are small implementation/platform/reference differences, not interchangeable exact values. Do not pool their row-level evidence without recovering the actual streams and code. [b14-bs-a findings; b14-bs-b report; b13-res-e report; b22-net-head rebuild_checks]
6. Classification-oracle gains are subadditive and price-loss buckets are associational. Neither summing sell+buy gains nor converting loss-bucket shares to causal ceilings is justified. [b15-loss-a limitations; b21-combo/qt1 additivity]

No raw source bundles, large traces, trained weight arrays, callback lists or full candidate histograms are reproduced here. They are implementation/audit data in the source artifacts, not silently rejected findings. This bounded merge retains their existence, meaningful checks, policy definitions, failures, quantitative outcomes and artifact identifiers. Missing exact HEAD weights and unembedded raw traces are explicitly identified; the supplied results generally include regeneration code but not a self-contained offline engine/compiler environment. The two lean bundle hashes and delivered Solidity hashes below were additionally checked in memory during this merge; no embedded code was executed. Other simulation/test claims remain source reports.

| Lean artifact identity | la-1 | la-2 |
|---|---|---|
| Embedded bundle field | bundle.base64 (tar.xz) | bundle.data (tar.xz) |
| Bundle SHA-256 | f8c97bd765720970af5b856c73852cac1860e8f3e9d0f598c43b4be5f5a54a47 | 59279fdaf37d7b68b2d1bafb28dbd529e3e48ee6461bb5332fa76d661129df8f |
| Final source in bundle | Strategy.sol | scratch/Strategy.sol |
| Final source SHA-256 | 0d002e11b60bd7f8c966711e5cd1ce03ed92d63226e082a1c0ff790348c0b350 | 1754f2f94751d08623ffdb1127162542e7261cb9db155989e8ae2848d3bfa14e |

la-1's reference/test layout depends on tca-path bundle **546b440855726c3cd2231e4879e0002dda182dc609ee9c5155c04e4790ca6fdb** for base files. la-2 includes Solidity imports and active reference dependencies with offline materialization; neither bundles its engine/compiler packages. The refit replacement blob is **ca5a606d074224ccc72fb4999c1d01a677ba00440ba39df2f99eba4913a72e1d**, original QMS1 **3ac7cb8811ecfb6a17f499aa4e9b70dd55f7fb6432b3aec3b904e11a3dece38e**. These identifiers support recovery from the cited inputs; they are not additional deliverables. [b26-la-1 bundle/NOTES-la-1.md; b26-la-2 bundle/NOTES.md; b23-tca-path source_bundle; b25-visit-refit checks]

**Statistical annex — Taylor-derived paired report results and engine tails**

The following report rows all use 10000..12047, n=2048. Values are copied from the named input/field; worst5 lists seed: difference, in reported order. Oracle rows remain diagnostics. Common full-window conventions and differences in harness above apply. This annex preserves principal tails without reproducing duplicate summaries, training curves or every candidate screen.

| Source field / comparison | n | mean_diff | SE | worst5 (seed: difference) |
|---|---:|---:|---:|---|
| b13-res-e: report/C-COMBO | 2048 | -0.18430004138662953 | 0.01069557809204496 | 11701: -3.467628013990975; 10942: -2.8887970854052583; 11696: -2.1945839047502886; 11693: -2.141005153671813; 11918: -1.8927538945172273 |
| b13-res-e: report/C-qt1 | 2048 | 0.41295621629031487 | 0.014586029716095518 | 10146: -2.7242323301666147; 10942: -2.3517784158157156; 10903: -2.041208296265836; 11062: -1.8278881431466516; 11487: -1.7941479076247333 |
| b13-res-f: report/C-COMBO | 2048 | -1.33906979170075 | 0.01795801422330611 | 11232: -4.482227879214861; 11161: -4.04478009030538; 11921: -4.020521538546973; 11284: -3.8859917259519534; 10668: -3.690183944349428 |
| b13-res-f: report/C-qt1 | 2048 | -0.7416788283150937 | 0.02043325733330658 | 11062: -4.958966713007385; 10162: -4.171042004326978; 10668: -4.144609526021213; 11808: -3.9636216322834343; 10850: -3.853950169939594 |
| b14-bs-a: report_10000_12047/C-COMBO | 2048 | -7.162411 | 0.040675 | 10335: -14.648244; 10765: -14.448413; 11322: -14.132123; 11191: -13.764133; 11692: -13.063115 |
| b14-bs-a: report_10000_12047/C-qt1 | 2048 | -6.564993 | 0.042098 | 10146: -14.680724; 10765: -13.638787; 11322: -13.076416; 11692: -12.498726; 10335: -12.424902 |
| b14-bs-a: report_10000_12047/C-control | 2048 | 0.06788 | 0.010662 | 10358: -3.298932; 11701: -3.210156; 11311: -2.622633; 11827: -2.210379; 10353: -1.868814 |
| b14-bs-b: report/COMBO-qt1 | 2048 | 0.5973909633856562 | 0.013387993974623986 | 10146: -3.552496929485528; 11062: -2.3348491437070606; 11808: -1.8227381435300458; 10519: -1.782691465320454; 11376: -1.7502956875490554 |
| b15-loss-a: report/PricePol-HA | 2048 | 8.021774 | 0.045883 | 10481: 3.072917; 10905: 3.324188; 10618: 3.353896; 11902: 3.40083; 10458: 3.410866 |
| b15-loss-a: report/PricePol-COMBO | 2048 | 3.718987 | 0.024144 | 11701: -1.411865; 11519: -0.802234; 10923: 0.333407; 11530: 0.729521; 10125: 0.919702 |
| b15-loss-a: report/PricePol-qt1 | 2048 | 4.316378 | 0.025937 | 11376: 0.050789; 11519: 0.103583; 11808: 0.686741; 11062: 0.740355; 11701: 0.947237 |
| b16-smoothing-b2: report/pairs/C-COMBO | 2048 | 0.04069944777966861 | 0.00663531544632357 | 11918: -0.9505559476841654; 10772: -0.9194077289280358; 10461: -0.9078301972947997; 10459: -0.8725258900223025; 10699: -0.859192634585952 |
| b16-smoothing-b2: report/pairs/C-qt1 | 2048 | 0.6380904111653247 | 0.013823226333168948 | 10146: -3.220655130285195; 11062: -2.4683113959351886; 10519: -1.6271008378673741; 11746: -1.5603073642805612; 11808: -1.4347261916885827 |
| b21-combo: report/FlagSell-base | 2048 | 1.622094 | 0.017508 | 11701: -3.84117; 10256: -1.426053; 10498: -1.127832; 10578: -0.831714; 10359: -0.692266 |
| b21-combo: report/FlagBuy-base | 2048 | 1.601987 | 0.01735 | 11701: -4.31632; 11530: -1.564423; 11519: -1.150452; 10998: -0.922527; 11233: -0.833044 |
| b21-combo: report/FlagAll-base | 2048 | 2.570198 | 0.020763 | 11701: -1.138705; 11519: -0.810231; 11911: -0.365037; 10578: -0.297158; 10367: -0.244856 |
| b21-qt1: report/FlagSell-base | 2048 | 2.0370433285168046 | 0.018857482700314156 | 11701: -1.8932826913702456; 10146: -1.4052865297487642; 11062: -1.289270529935834; 10476: -1.0855912484274768; 10498: -0.788584514701256 |
| b21-qt1: report/FlagBuy-base | 2048 | 2.018622977058917 | 0.018711825159457175 | 11062: -1.8188211029238346; 10146: -0.9966197449285801; 10256: -0.9113130344431966; 11376: -0.9075469354098686; 11224: -0.8184560309782682 |
| b21-qt1: report/FlagAll-base | 2048 | 3.167510323083941 | 0.022385773867452264 | 11062: -1.3007952811217365; 10146: -0.01615525243778393; 11519: 0.09558604262929293; 11376: 0.2773640732128797; 10141: 0.29737815162343395 |
| b22-net-head: paired/report/NET-COMBO | 2048 | 0.078039 | 0.009144 | 10335: -1.674976; 11696: -1.672931; 11530: -1.526035; 11191: -1.121414; 10436: -1.099873 |
| b22-net-head: paired/report/HEAD-COMBO | 2048 | 0.076607 | 0.008472 | 11311: -1.833278; 12015: -1.779181; 10841: -1.331887; 10256: -1.202205; 10401: -1.186789 |
| b22-net-head: paired/report/STACK-COMBO | 2048 | 0.148967 | 0.010412 | 10632: -4.416473; 11311: -3.01643; 11530: -1.921214; 11696: -1.514754; 10222: -1.462945 |
| b22-net-head: paired/report/STACK-qt1 | 2048 | 0.746358 | 0.014554 | 10632: -3.570241; 10146: -2.478542; 11062: -1.699438; 10401: -1.651281; 10930: -1.488914 |
| b23-tca-path: Q/report/base | 2048 | 2.7085299501905924 | 0.020633408969569385 | 10578: -0.8093814750295678; 10141: -0.08104406803897746; 10403: -0.07310549074918526; 10923: -0.06804735057534117; 10449: 0.0204613125187052 |
| b23-tca-path: Q/report/qt1 | 2048 | 3.133460282026983 | 0.02226357822411278 | 11062: -1.3959077621872211; 11701: -0.42019020228906356; 10146: 0.08090180609713116; 11376: 0.1012977086736555; 11519: 0.16334820503948322 |
| b23-tca-short: Q/report_10000_12047/tell_vs_qt1 | 2048 | 3.15777887503495 | 0.022523380035794013 | 10481: -2.3834520078763717; 11062: -1.3642312923469717; 11376: -0.3633058744172786; 10256: 0.12827039339481416; 11519: 0.1800823649667791 |
| b23-qtr2-path: replica_results/report_vs_qt1 | 2048 | 3.1498943513647712 | 0.0223633021238525 | 11062: -1.397359046747738; 11701: -0.2859600185729505; 11376: -0.036299684025721035; 10146: 0.11670847958089325; 11519: 0.14240717451173168 |
| b25-visit-refit: scores/report/paired/C-BASE | 2048 | 0.049063 | 0.003681 | 11701: -1.402185; 11455: -0.911079; 10728: -0.859496; 11998: -0.76263; 10357: -0.591789 |
| b25-visit-refit: scores/report/paired/C-qt1 | 2048 | 3.205537 | 0.022437 | 11062: -1.338114; 11701: -0.493342; 11376: -0.255387; 11519: 0.207946; 10642: 0.276573 |
| b25-visit-refit: scores/report/paired/BASE-qt1 | 2048 | 3.156474 | 0.022252 | 11062: -1.380901; 11376: -0.101975; 11519: 0.094928; 10141: 0.148161; 10642: 0.34402 |
| b26-la-1: Q report vs qt1 (also identical b26-la-2 Q.qt1) | 2048 | 3.1598948140998226 | 0.0222262108662516 | 11062: -1.3596829746003323; 11376: 0.012546694402089997; 11519: 0.16165349985828925; 10642: 0.225518981451728; 10146: 0.25751453048019357 |

Engine rows below pair each candidate and qt1 on the same configs in separate runs against 30-bps Vanilla. Historical controls, failed pilots and mixed versions are marked; no row is a leaderboard result.

| Source field / comparison | n | mean_diff | SE | worst5 (seed: difference) |
|---|---:|---:|---:|---|
| b18-cost8: engine control, 9900000..9900015, NOT head | 16 | 0.5070876322232678 | 0.1754716088056967 | 9900002: -0.6056395420580998; 9900015: -0.42835789565822324; 9900011: -0.2209867544794406; 9900007: -0.1961727637108197; 9900001: 0.033896842077524525 |
| b23-tca-path: failed engine pilot, 9963072..9963079 | 8 | -1.4971951358683881 | 1.19668431887727 | 9963072: -8.375487159474005; 9963075: -2.534036491975712; 9963074: -2.4123258315458997; 9963078: -1.4829465124904573; 9963079: -1.0807645093178735 |
| b26-la-1: final G, 9970016..9970023 | 8 | 3.947216228461592 | 0.28403721952629163 | 9970019: 2.5274340660932353; 9970021: 3.215820866163824; 9970017: 3.6650725687185286; 9970018: 3.982215567406911; 9970023: 4.041622473611312 |
| b26-la-1: final E, 9970164..9970355 + 9970400..9970655 | 448 | 3.143932837854402 | 0.046044360817685696 | 9970281: 0.427909874493821; 9970194: 0.6868779762933173; 9970190: 0.7036156443575123; 9970264: 0.9803579456873877; 9970506: 1.1459802696848556 |
| b26-la-1: mixed-build E historical pool, 512; NOT final-build-only | 512 | 3.165517119347506 | 0.0428649621114079 | 9970281: 0.427909874493821; 9970194: 0.6868779762933173; 9970190: 0.7036156443575123; 9970264: 0.9803579456873877; 9970506: 1.1459802696848556 |
| b26-la-2: final G, 9971008..9971015 | 8 | 3.395408776187459 | 0.4466319657533357 | 9971009: 2.0501670918985155; 9971010: 2.2913536252637527; 9971008: 2.5671574132402952; 9971014: 2.8134041046074003; 9971015: 3.2507186319057837 |
| b26-la-2: final E, 9971164..9971355 + 9971400..9971655 | 448 | 3.1908690213305713 | 0.04559391548818873 | 9971438: -1.2684839890202966; 9971523: 0.752369573923886; 9971297: 0.9475907235146224; 9971586: 1.0374947344848806; 9971348: 1.0399854012783862 |
| b26-la-2: superseded E, 9971100..9971163; gas failure, excluded | 64 | 3.15979529293263 | 0.1417646841861695 | 9971124: 0.026274626315057503; 9971154: 1.1424981008339046; 9971146: 1.1431939822709865; 9971139: 1.444132544277295; 9971101: 1.4466127472338712 |

Engine seed reservations must be treated conservatively downstream: consumed ranges in this group also include o2-a **9950000..9950007**, o2-c **9950016..9950023**, buy-grid **9950024..9950031**, la-1 old pilots **9970000..9970015** and old E **9970100..9970163**, and la-2 old pilot **9971000..9971007**. Use the respective input ledgers before allocating future seeds; do not infer that ranges absent from this root are globally unused. Planned-but-unrun ranges are identified in the prose, not counted as consumed. [b20-o2-a/c; b24-buy-grid measurement; b26-la-1/2 seeds_consumed]

**Source register — Taylor-derived input provenance, all 25 checked**

Each ID refers to `.imd/reads/artifacts/<ID>` supplied to this merge. Original artifact location is the following `artifacts/goal/batches/<batch>/items/<item>/result.json`. Digests below are of the complete input JSON bytes and match the requester. Paths identify the original artifacts; availability to the final reader depends on the network retaining them.

| Input ID | Original batch/item result | Verified SHA-256 |
|---|---|---|
| b13-res-e | `artifacts/goal/batches/e04-residual-b/items/res-e/result.json` | `a098c1ea59d99711159eadc469ec72aaad042a01d1a5d73b91d552e346c6c594` |
| b13-res-f | `artifacts/goal/batches/e04-residual-b/items/res-f/result.json` | `93cdb55e4136f5044a35558276a260dc4bc6e542e6857ca578f874557f806abd` |
| b14-bs-a | `artifacts/goal/batches/e04-belief-solve/items/bs-a/result.json` | `a4400aa17333ac67b8a4e285f29d857593fb98077e942a393502925ff80bee4e` |
| b14-bs-b | `artifacts/goal/batches/e04-belief-solve/items/bs-b/result.json` | `ef39a75952bff8edbe5f0fc2d238eb238107d32c98c06a08c3e05caa73fb6f15` |
| b15-loss-a | `artifacts/goal/batches/e04-price-loss/items/loss-a/result.json` | `86e867cb85dd1da260d2657ae105921b77616c07deb699896f33ec47aece1ba6` |
| b16-smoothing-b2 | `artifacts/goal/batches/e04-smoothing-b2/items/smoothing-b2/result.json` | `56ad08f8f641973dbd356593e8137d471403731e27c8fd22f959f584e8ba5d5c` |
| b17-numba-tape | `artifacts/goal/batches/e05-fast-sim/items/numba-tape/result.json` | `c8630c13c60fedecb072e3ee45872960775dac11d580d4620388cabf3a88f194` |
| b17-numba-rng | `artifacts/goal/batches/e05-fast-sim/items/numba-rng/result.json` | `b9c917ebb99152aafd12993b327f0a82529efe414a4a8fea8f4a0e3493cc1f8f` |
| b18-cost8 | `artifacts/goal/batches/e05-head-cost/items/cost8/result.json` | `3385a6e029d6bea8c32e24205e2207988e325b8b70ad133c8b2083a052e3ac3e` |
| b19-o1-a | `artifacts/goal/batches/e06-tell-port/items/o1-a/result.json` | `9242f3ebbe1ef8d23e590c9784f147655714aaf8a6da7a841273170f20c9ad9b` |
| b19-o1-b | `artifacts/goal/batches/e06-tell-port/items/o1-b/result.json` | `d56e3af5cc1e32dcc52be27046df89fb403d770a7fabf39e1a0ef94827349a5a` |
| b20-o2-a | `artifacts/goal/batches/e06-tell-engine/items/o2-a/result.json` | `71067c00571bc062b25540e9d75ef8eec895512f676a898950a114cdef15a64e` |
| b20-o2-b | `artifacts/goal/batches/e06-tell-engine/items/o2-b/result.json` | `1f96ad790f76ecb8943d8823e54e8138f735fae34529e1b2c47e6ee2ee2cecef` |
| b20-o2-c | `artifacts/goal/batches/e06-tell-engine/items/o2-c/result.json` | `e714896bd9db6c7290b97fbeed9b16f5fb9d8d58d63e7a83376c3b3f560c48f2` |
| b21-combo | `artifacts/goal/batches/e06-side-flag/items/combo/result.json` | `3f2100314d249370523545d444c11063bc2ae6a01dd1f6d1c33450536357f823` |
| b21-qt1 | `artifacts/goal/batches/e06-side-flag/items/qt1/result.json` | `9d6540b8d956ac9fdae42af7764d41345f9e5dc027693b47be93055579883578` |
| b22-net-head | `artifacts/goal/batches/e06-stack/items/net-head/result.json` | `ec7b69dc097cbd69a9d8159a9a085070cec8827e672fd0ebb4b88fc1e04ea4db` |
| b23-tca-path | `artifacts/goal/batches/e07-tell-build/items/tca-path/result.json` | `dc359635b5851de9f5b3ecf81dc2107d46acd5dbe67e0be234bcfbe0c8a488e7` |
| b23-tca-short | `artifacts/goal/batches/e07-tell-build/items/tca-short/result.json` | `27b1249d19af14698bccada40f75fe3fbd2b34f7fcae89a291b046392935a9d1` |
| b23-qtr2-path | `artifacts/goal/batches/e07-tell-build/items/qtr2-path/result.json` | `f969ac56c8b0adc16ff7598a9340c4f5aab9a8304f892bd9d7fbad7c2037fbbd` |
| b24-buy-grid | `artifacts/goal/batches/e07-buy-grid-check/items/buy-grid/result.json` | `a162a591237e25c49d30b98c30c1c1fdc23746ead5ff3371a460a77357825624` |
| b25-visit-refit | `artifacts/goal/batches/e07-tell-tables/items/visit-refit/result.json` | `5893446c82f8bc1473468bba6b156ec2a96ce3df5ddfc897c05c63cf545e1162` |
| b26-la-1 | `artifacts/goal/batches/e08-lean-a/items/la-1/result.json` | `1f399ab09bf68c0f88d9849a12eafd611176253dd98a27c05057f8df58e817ed` |
| b26-la-2 | `artifacts/goal/batches/e08-lean-a/items/la-2/result.json` | `b27ba5d1ec7d2a8ffb80fd39c02b63fa1e282fcb793906fe5f81861de5749f3c` |
| b26-la-3 | `artifacts/goal/batches/e08-lean-a/items/la-3/result.json` | `4d24f1117ca9c41a0fc3dcba28739db257f3af72fec73ea7654f71de80187502` |

Merge validation: all named JSON inputs parsed; all 25 expected digests matched; both lean archive digests and final source digests checked without executing payloads; output UTF-8 and byte limit checked. These checks establish document integrity only, not behavioral certification.