None of these 25 results establishes the gain needed for the top 10. The useful positive result is **V10, the censored-price posterior**, with a small, repeatedly positive gain over STACK and lower measured afterSwap gas. Pooling its original confirmation and the two unchanged-code replications gives **n=6000, mean_diff=0.13620592019836136, SE=0.007911506404821834**, with an approximate 95% interval [0.12069936764491057, 0.15171247275181216]. Its reconstructed retail delta is -1.1354274182279815 (SE 0.00814959664897208), offset by an arbitrage-loss reduction of +1.2716333384263434 (SE 0.004253055813377944). Across those three blocks, max successful afterSwap gas is 194882 and OOG=0. These are pooled calculations from reported summaries, not a new simulation or board result.[^02c][^08a][^08b]

The assignment's starting figures are STACK **519.71 (29th)**, tenth **524.66**, first **528.40**, last observed board/local transfer **0.65**, and a requested local target of approximately **+5.0**. Its prior findings were a ceiling near **+1.2** for tuning STACK's fee family and losses of **18 to 83** for rules written from scratch, mostly to arbitrage. Preserve those as requester context, not measurements made here. There is an arithmetic tension in the target: the rounded board gap is 4.95; a literal 0.65 transfer would require about 7.615 local gain, rather than 5.0. Applying that ratio to the pooled V10 point estimate gives only 0.08853384812893489 board edge, as an illustrative extrapolation, not a forecast. None reports submitting a new candidate or achieving the goal.

STACK is the `code` field of the [specified baseline artifact](https://api.imdnotdead.fun/artifacts/c33720874aa1285e9dc766540b8a90410118a8e8b840d949ea0dcc211bdb27c1), code hash `2eaafc7960e520b8a7fd322ec8afaff8fa677c730228a961efb509cc30b25719`.[^01]

All 25 inputs were present and parseable. Footnotes identify their exact paths and JSON hashes. This merge separates confirmation, held-out tuning, incomplete work and proposals. Principal figures retain source precision; compact tables explicitly round. Repeated counters, helper scripts, full optimizer trajectories and losing contracts are summarized; all families, outcomes, nulls and material caveats are retained. V10's exact source is preserved losslessly below; other contracts have a hash ledger.

**Source reliability matters.** The requester flags failed source rechecks for calibration-and-harness (1), censored-price (1), liquidity-shadow (2), policy-distillation (4), both confirm-v10 items (1 each), both v10-floor items (1 each), v10-blend/a (1), and both v10-retail items (1 each). Which URLs/passages failed and the failure responses were not supplied. Earlier claims that all quotations matched do not resolve this warning. Statistics remain attributed worker measurements, not independently replicated evidence; unsupported design arguments remain hypotheses. A citation failure alone does not disprove an included code string or total-edge record, but the two design-only items have no performance evidence to compensate.[^01][^02c][^03l][^03p][^08a][^08b][^09a][^09b][^10a][^11a][^11b]

The common protocol is separate markets `run_batch(X, VanillaStrategy, configs, 8)`, pairing candidate and STACK by seed. Simulator revision: `6b6921f8d463314aa7c96234e1be25a25245f704`; wheel `amm_sim_rs-0.1.0.post1`, SHA-256 `aedb68280aca34c110968f00148387c95885207913fb2cf19a1ab4c5cfd6887a`. `default_rng(seed)` draws size U(19,21), rate U(.6,1), sigma U(.000882,.001008), in that order. Defaults: 10000 steps; initial price/X/Y=100/100/10000; mu=0, dt=1; retail size sigma 1.2, buy probability .5; price seed=seed, retail seed=seed+1. Compiler: solc 0.8.24, optimizer 200, viaIR, Paris. Paired SE=sample SD(d)/sqrt(n).[^01][^02c][^02n]

**Evaluator findings.** CH's [pinned source](https://raw.githubusercontent.com/ChengHua926/amm-challenge/d34a5199c59a4c44d5336a25f1436b4ae1f21104/contracts/src/Strategy.sol), code SHA-256 `6dd68f77a70900319460215374143dbd5d924198450a46d9f5f91f6b1e3d35c3`, reproduced 525.5258821780834 on seeds 0..999 in b01 and the completed calibrated experiments. b01: CH max 151647/OOG 0; 12000 config fields matched MatchRunner; STACK [10000,10256) was bit-identical under 1x256 versus 4x64 chunking, including gas totals, paired mean=SE=0, arm mean 523.2683436100853, afterSwap max 211576, initialization 220962. Later peaks near 214k are consistent. CH was not repeated through MatchRunner.run_match; equivalence rests on configs and the same batch call.[^01]

afterSwap errors can leave old fees active while `run_batch` returns scores. b01's OOG fixture counted 18102 callbacks (8930+9172), yet returned apparent edge 331.6740290718941; the harness rejected it, paired n=0. Initialization failures propagate as exceptions. Gas counters are process-global: isolate each arm/chunk, reset before, snapshot after, sum counts/errors and maximize gas; no concurrent batches within a process. Submission/Baseline mean first/second argument, empirically checked by swapping them. The instrumented wheel's gas_stats/reset reportedly do not exist in the pinned Rust tree; their behavior was tested empirically. Any OOG invalidates a candidate regardless of max_ok.[^01]

**Edge attribution is reconstructed.** The wheel exposes total edge and flow volumes, not trade-type edges. Workers reconstruct arb from prior end-step spot/fees, current fair price and constant k=1000000, then retail=total-arb. Summing back to total is not independent validation. b01's volume check max absolute/relative errors were 8.360799620277248e-05 Y/4.174892464576448e-09; using current-step fees instead caused relative errors .794/.812. b02c independently matched arb to total on 10 zero-retail diagnostic paths, max error 5.179509598995935e-11. b02n failed its declared absolute 1e-6 volume tolerance (max .00011530283882166259): split unvalidated, overall valid=false, native paired totals/gas still valid. Other workers accepted scale-aware tolerances, sometimes after failed assertions; these differing standards must remain visible.[^01][^02c][^02n]

Where workers returned null edge components, null means unmeasured, not zero or mathematically impossible to reconstruct. The value-policy and REGIME results supply volume proxies in some cases; these must not be relabeled as edge deltas. No component covariance is supplied, so retail and arb SEs cannot be added.[^02v][^07a][^07b]

The following confirmation ledger gives candidate minus STACK. Intervals are half-open. R and A are reconstructed retail and arb **mean** deltas unless marked unvalidated; their reported SEs are retained in the component ledger below. Gas is Submission.afterSwap, explicitly separated from initialization or all-callback maxima. All rows with measured gas report OOG=0; missing gas is never converted to zero.

| Candidate/source | Seeds | n | Mean (SE) vs STACK | Worst d @seed | max_ok / OOG | Status |
| --- | --- | --- | --- | --- | --- | --- |
| V10[^02c] | [5000000,5002000) | 2000 | 0.1372320888962471 (0.01363197915215346) | -2.8918776608781513 @5000729 | 194870 / 0 | compliant reported |
| STACK floor .3[^02v] | [5002000,5004000) | 2000 | 0.016364585156108575 (0.0014646489212871408) | -0.3435225786081446 @5002929 | 214561 / 0 | native total valid; split absent |
| Normalizer D[^02n] | [5004000,5006000) | 2000 | -1.3756328776753344 (0.04581777642932667) | -38.527588753941245 @5005166 | 240832 / 0 | paired valid; attribution invalid |
| V10 unchanged[^08a] | [5008000,5010000) | 2000 | 0.1445235199955189 (0.013566995142630451) | -2.8005274638323954 @5009936 | 194139 / 0 | compliant reported |
| V10 unchanged[^08b] | [5010000,5012000) | 2000 | 0.12686215170331805 (0.01391194627355796) | -3.0704815268940138 @5011040 | 194882 / 0 | compliant reported |
| V10 floor .3[^09a] | [5012000,5012400) | 400 | 0.10625783844175331 (0.030470134095873534) | -2.3219612485586367 @5012382 | 195078 / 0 | 400 only; no CH calibration |
| V10 floor .8[^09b] | [5014000,5016000) | 2000 | 0.18236411895041843 (0.04046483899952992) | -2.6497201519990767 @5014736 | 196935 / 0 | compliant reported |
| V10 tox-fast .15[^10a] | [5016000,5018000) | 2000 | 0.12460569065899302 (0.015224278147697251) | -5.85720967046916 @5016325 | 194605 / 0 | compliant reported |
| V10 tox-fast .30[^10b] | [5018000,5020000) | 2000 | 0.12127452784554266 (0.014781738671253325) | -4.027840212609362 @5018921 | 195466 / 0 | compliant reported |
| V10 fallback[^11a] | [5020000,5022000) | 2000 | 0.11694634196808508 (0.015019437639641227) | -4.7258718593310505 @5020480 | 194858 / 0 | compliant reported |
| V10 lower cap 28[^11b] | [5022000,5024000) | 2000 | 0.1533725409948071 (0.014114041643220169) | -2.451734575486512 @5022131 | 194759 / 0 | strict single-pass INVALID |

| Source/candidate | R mean (SE) vs STACK | A mean (SE) vs STACK |
| --- | --- | --- |
| V10[^02c] | -1.135804799066593 (0.013995494300275123) | 1.2730368879628409 (0.006935307161576756) |
| STACK floor .3[^02v] | null | null |
| Normalizer D (unvalidated split)[^02n] | 0.9901206884883873 (0.015004238905845327) | -2.365753566163721 (0.044013453624143405) |
| V10 unchanged[^08a] | -1.1223103927737246 (0.014348940541849816) | 1.266833912769243 (0.007477615744106323) |
| V10 unchanged[^08b] | -1.1481670628436267 (0.014000335339235418) | 1.2750292145469462 (0.007669364144180742) |
| V10 floor .3[^09a] | -1.1452807751856757 (0.030869516296926396) | 1.2515386136274305 (0.01469765665275859) |
| V10 floor .8[^09b] | -1.1130631958592558 (0.015034961232340124) | 1.2954273148096744 (0.03362897368865322) |
| V10 tox-fast .15[^10a] | -1.3418763081451621 (0.015475179587860469) | 1.4664819988041546 (0.008693455396430579) |
| V10 tox-fast .30[^10b] | -1.5549805367427008 (0.015598759711610274) | 1.6762550645882444 (0.008890728197677372) |
| V10 fallback[^11a] | -1.1413228240546553 (0.014977487518045528) | 1.25826916602274 (0.008088000581186904) |
| V10 lower cap 28[^11b] | -1.1112437414958636 (0.014442400417784723) | 1.2646162824906715 (0.0075057654723846) |

The additional comparisons below isolate whether a V10 modification helped beyond V10. Positive candidate-minus-STACK values cannot be credited to the modification when candidate-minus-V10 is flat. b09b's V10 reference on the same block was +0.18218552332110957 (SE 0.04046641997245594) over STACK; b10b's was +0.1260454276671473 (SE 0.013769778616960491). b11a's candidate and V10 arms are aliases of the same source/run, not two independent replications.[^09b][^10b][^11a]

| Source | n | Mean (SE) vs V10 | Worst d @seed |
| --- | --- | --- | --- |
| [^09a] | 400 | 0.0 (0.0) | 0.0 @5012000 |
| [^09b] | 2000 | 0.00017859562930885886 (0.0001604829557281328) | -0.05188252397704218 @5014623 |
| [^10a] | 2000 | 0.0009965097343738306 (0.003808122297323427) | -1.508370362889309 @5016734 |
| [^10b] | 2000 | -0.004770899821604644 (0.005900741325340772) | -3.2007638086773227 @5018921 |
| [^11a] | 2000 | 0.0 (0.0) | 0.0 @5020000 |
| [^11b] | 2000 | 0.0018787537178642708 (0.002555729916292394) | -1.9318602942065581 @5022990 |

The primary V10 pool above uses only b02c, b08a, and b08b, with disjoint 2000-seed blocks and the same code hash. Its five worst differences, in order, are 5011040:-3.0704815268940138; 5000729:-2.8918776608781513; 5009936:-2.8005274638323954; 5011985:-2.6964666407454843; 5011677:-2.521602122643003. Pooling uses N=sum(n), weighted mean, and pooled sum of squares `sum((n-1)*n*SE^2+n*(mean-pooled_mean)^2)`; pooled SE is sqrt(SS/((N-1)*N)). This does not treat raw candidate means from different seed blocks as paired evidence.[^02c][^08a][^08b]

The ledgers retain each experiment's minimum paired difference; five-element source tail lists are condensed, except the primary V10 pool above. Missing seed IDs are not inferred. b09a/b11a have identically zero V10 differences on their respective 400/2000 seeds.

b02-event-hazard is **not an evaluated mechanism**: setup failed at `uv venv --python 3.12 .venv` with exit 133, `rosetta error: failed to open elf at /lib64/ld-linux-x86-64.so.2` and `Trace/breakpoint trap`. Wheel checksum passed, but calibration, strategy checksum, search and gas checks did not run; n=0, code/hash=null, gas=null. Its reserved [5006000,5008000) was not executed. b09a reports only [5012000,5012400), not a completed 2000-seed confirmation, and skipped CH calibration and its shortlist. b12a/b left [5024000,5026000) and [5026000,5028000) untouched after failing their held-out gates. These unused/incomplete blocks are not evidence of performance.[^02e][^09a][^12a][^12b]

**Censored-price mechanism and discarded alternatives.** V10 replaces STACK's fastHat update and uncertainty propagation with a soft, seven-node assumed-density posterior. Silence contributes interval evidence; a first print mixes an arb endpoint with a censored retail observation; later retail routing contributes a noisy inferred-opponent observation. The distinguishing V10 change divides arb/retail odds by an approximate no-trade-interval probability mass. It retains slow pHat and the original fee family, including the 70/30 slow/fast stale-price blend. This is approximate inference, not an exact likelihood of hidden routed flow. Code inspection of the supplied contract also resolves a loose phrase in several narratives: 30/36 bps are caps on the **strictly lower side**, not universal caps on both fees.[^02c]

All twelve pilot variants were screened on [10000,10100); compact values below are paired mean (SE), rounded to six decimals. V3/V4 came close to the gas ceiling despite zero observed OOG. Shortlisting V2/V5/V10 on [12000,12500) gave respectively -0.2712646059675557 (0.026701612136902815), -0.1758400035030543 (0.027760545025748644), and +0.1175435755474075 (0.026433708870234766). V10 was frozen before confirmation. Full replacement of the slow toxicity or stale-price signal lost substantially; it is not supported by the successful partial posterior.[^02c]

| Variant | Structure | Mean (SE) vs STACK | afterSwap max; OOG |
| --- | --- | --- | --- |
| V0 | Uniform interval moments + arb/retail mixture | -0.711254 (0.078549) | 190072; 0 |
| V1 | Hard truncated seven-node Gaussian | -1.288062 (0.089434) | 190072; 0 |
| V2 | Soft rational interval likelihood | -0.239898 (0.053408) | 190076; 0 |
| V3 | Sequential silence updates, capped eight | -0.159791 (0.051852) | 249302; 0 |
| V4 | V3 also replaces slow toxicity price | -7.048421 (0.217132) | 249267; 0 |
| V5 | Single aggregate silence update | -0.153062 (0.052155) | 190538; 0 |
| V6 | V5 replaces complete stale signal | -6.085631 (0.200665) | 190311; 0 |
| V7 | V5 replaces both slow/stale signals | -7.345482 (0.214518) | 190408; 0 |
| V8 | V5 later routing prints as intervals | -0.194047 (0.049656) | 190538; 0 |
| V9 | V5 excludes later routing observations | -0.237745 (0.066193) | 190519; 0 |
| V10 | V5 odds / no-trade-interval mass | 0.176799 (0.058067) | 192838; 0 |
| V11 | V5 rejects direction-disagreeing routing | -0.144493 (0.053197) | 190538; 0 |

b02c initially hit memory termination on a 1000-config calibration; disjoint 20-config chunks completed the same check. V2/V3 Yul stack-depth errors were repaired before measurement. A failed freeze-helper import was followed by an immediately interrupted compilation launch (exit 130), with no confirmation results/counters returned; the completed confirmation followed a corrected freeze. Its 20-seed harness smoke test on 13000..13019 was +0.47913799465114837 (SE 0.16437950174434668), explicitly not a selection criterion. These repairs do not supply additional independent holdouts.[^02c]

**Value-policy pilot.** Alternative fMid families were L2 composition `base+sqrt(cA^2+cT^2+2*kappa*cA*cT)`, max composition `base+max(cA,cT)+kappa*min(cA,cT)`, four eight-bucket tox/sigma lookup multipliers, and `max(STACK fMid,kappa*sqrt(uncertainty))`. Every non-floor family lost in the tested blocks; only floor kappa=0.3 was positive. The code adds that floor after fMid and uses no new state. Its confirmation gain +0.016364585156108575 is real evidence of a tiny effect, not a structural route to +5. Retail/arb edges are null; the reported Y-volume deltas are retail -17.030342651700778 (SE 0.4379089975069555), arb -6.487618318704174 (SE 0.296204349956911). On confirmation 35.7% of seeds were negative, none exactly zero, and the largest gain was 0.6282823405387603.[^02v]

| Block / n | Variant | Mean (SE) vs STACK |
| --- | --- | --- |
| 10000:10250 / 250 | floor 1.0 | -61.445906 (0.471358) |
| 10000:10250 / 250 | floor 1.5 | -113.070804 (0.936219) |
| 10000:10250 / 250 | L2 0 | -0.339464 (0.045794) |
| 10000:10250 / 250 | L2 .5 | -0.056820 (0.025778) |
| 10000:10250 / 250 | LUT a | -0.874217 (0.044378) |
| 10000:10250 / 250 | LUT b | -0.152866 (0.034606) |
| 10000:10250 / 250 | max .3 | -0.316119 (0.044550) |
| 10000:10250 / 250 | max .6 | -0.061000 (0.033394) |
| 10250:10500 / 250 | floor .3 | 0.008726 (0.004403) |
| 10250:10500 / 250 | floor .6 | -20.464431 (0.215213) |
| 10250:10500 / 250 | L2 1.5 | -0.132897 (0.026306) |
| 10250:10500 / 250 | LUT c | -0.022779 (0.026576) |
| 10250:10500 / 250 | LUT d | -0.095783 (0.036662) |
| 10250:10500 / 250 | max 1.2 | -0.082792 (0.022584) |
| 10500:10900 / 400 | floor .3 | 0.010935 (0.003278) |
| 10500:10900 / 400 | floor .45 | -4.511034 (0.063098) |
| 10500:10900 / 400 | LUT c | -0.059864 (0.018174) |

All value-policy tuning OOG=0; successful afterSwap maxima span 211714..212828 for the variants.[^02v]

**Normalizer-state pilot.** A replaced routing blends with one-sided interval projection; B added size/depth split-likelihood weighting; C used interval endpoints for protective quote targets and retained low-fee routing observations; D retained STACK routing and added a midpoint update only for narrow opposing bounds. Buys/sells do not identify interior splits because total orders and Vanilla-only trades are hidden. Bounds and inferred Vanilla depth are heuristic, even with constant k. Screen means (SE) on [10000,10120) were A -17.34930027361266 (0.5915777292355843), B -28.77672598474418 (0.8759949089289745), C -1.7099022201156535 (0.08624643818887999), D -1.480752068556618 (0.24004535280983302), all OOG=0, afterSwap maxima 238781/238968/238978/238905. Their reconstructed R/A means were respectively +8.92843783691549/-26.277738110528155, +9.308825060763724/-38.08555104550791, +1.6975029925770595/-3.4074052126927197, +0.9664689101409659/-2.4472209786975907.[^02n]

C/D shortlisted on [20000,20200) at -1.751862882173036 (SE 0.06752382757671126) and -1.5334296493028046 (0.16997084065763401), afterSwap maxima 238966/238893; R/A were +1.6725103510459707/-3.4243732332190073 and +1.013534540414112/-2.5469641897169173. D was frozen solely to report the least-bad pilot, with an explicit prior recommendation to retain STACK. Confirmation means were candidate 524.2014888693757, STACK 525.5771217470509; D initialization was 243077 and afterSwap 240832, both below 250000 with no halts. The total loss is usable negative evidence; failed attribution is not validated merely because its components sum to the total. A screen shell wrapper exited 127 after requested runs completed; the worker checked full seed coverage and did not rerun those chunks.[^02n]

**Designs that remain proposals.** Liquidity-shadow proposes optimizing a fee-dependent continuation value rather than adding an inventory penalty. Its reported simulator reading is trade-time edge without an inventory penalty, fees outside reserves, one arb phase before sequential retail, and a fixed-product reserve curve. Those claims come through the source-warning-affected design; they are also consistent with the empirical workers' reconstruction methods, but this merge did not independently retrieve the engine. Let `g=ln(fair/spot)~N(ghat,s^2)`, `s^2=sigmaHat^2+uncertainty`, `y=reserveY`, and `fa/fb` be ask/bid fees. The proposed small-fill approximation is `A=(y/4){E[(g-fa)_+^2]+E[(-g-fb)_+^2]}`, `Ua=E[(g-fa)_+]`, `Ub=E[(-g-fb)_+]`, `tau=ghat-Ua+Ub=E[clip(g,-fb,fa)]`; continuation-adjusted side values are `Va=fa-tau`, `Vb=fb+tau`, hence `Va+Vb=fa+fb`. The designer reports an arithmetic-only <1% agreement of the quadratic loss approximation in its checked m>f cases, not a simulator validation.[^03l]

With tail probabilities `pia=P(g>fa)`, `pib=P(g<-fb)`, the proposal uses `dVa/dfa=1-pia`, `dVb/dfa=pia`, `dVb/dfb=1-pib`, `dVa/dfb=pib`, and arb gradients `dA/dfs=-(y/2)Us`. Retail value is `Rs=ms*share_s*Vs`, `ms=lambdaHat*Qbar/2`, `share=clip(1/2+k*(c-V),0,1)`, `k=y/(4Q)`, `c=30bps`; optimize `J=Ra+Rb-A`. Its uncoupled stationary scale is `V*=15bps+Q/y`, a hypothesis only. Candidate families are (A) a clipped-tau residual with strength 0/.25/.5/1; (B) small pairwise argmax offsets of 3/6 bps; (C) zero-residual band targets with widths 2s times .5/1/1.5; (D) online share-slope learning clamped to .5..2 times its initial value; (E) fixed pair-sum, optimized split replacing skew/stale heuristics; (F) an inventory-penalty control.[^03l]

Liquidity-shadow's inputs are own spot/depth, prior fees, pHat/fastHat, uncertainty, silence, size/rate estimates and first-print arb probability; hidden fair price and rival reserves are unavailable. Vanilla's 30bps fee is assumed. Proposed slots 16..19 hold tau/rho/calibration/share state, optionally 20..23; fixed tail approximations and 3..5 quote pairs aim to fit roughly 30k gas headroom, simplifying above 235k. These are unmeasured estimates. Falsifiers: fill-identity slope outside [.5,1.5], reversed tau not losing, no benefit from clip/coupling/arb-gradient/share ablations, or any OOG; proposed A+B survival threshold+.3. Risks: large/heavy-tailed fills, correlated rival residual, noisy fair-price/variance estimates, and caps. None of its six families was simulated; later SHARE/ARB searches do not exhaust the coupled objective.[^03l]

Policy-distillation proposes a **42-cell** residual table over seven signed-gap/sigma bins, first-print flag and three elapsed-time buckets, with two signed int8 half-bps outputs or a mirrored symmetry constraint. Keep STACK's estimators, optimize full rollout edge minus an L1 penalty, then deploy constants; zero residual should reproduce STACK. This is black-box policy search rather than fitting privileged fair-price/trade-type labels. Alternatives are GATED-MIX (STACK/wide-quote mixture), HORIZON-FEE (protect-side uncertainty over a callback horizon), ASYM-ROUTER-TARGET (attract-side inferred-rival target), and TREE-DISTILL (depth-3/eight-leaf distillation of the searched policy). The suggested horizon `1/(1-exp(-lambdaHat))` is approximate because callbacks also depend on arbitrage and routing.[^03p]

The distillation proposal specifies 42..84 parameters, populations 16..24, 1000-seed batches rising to 4000 and 20..40 generations, all unexecuted. Constants need no new storage; packed tables could use slots 16..19. Claimed <1k extra gas/~216k peak is an estimate. Diagnostics: zero-table equivalence, callback-feature parity, shuffle, residual/STACK-term correlation and stale-block removal. Proposed rejection: held-out gain<+.3 or within 2SE of zero; shuffle retaining>=80% suggests a level effect; |correlation|>.9 suggests rescaling. Risks are overfit, leakage, fee jumps and negative residuals inviting arb. Fees persist between callbacks, not throughout a timestamp; the claim that jumps are harmless was untested. b12 tests only a smaller 20-parameter family with shorter batches.[^03p][^12a][^12b]

**Board observations, not recovered contracts.** b04a analyzes ranks 1..5/STACK; b04b ranks 1,6..10,20/STACK, 1000 records each from the [leaderboard API](https://www.paradigm.xyz/puzzles/api/v1/leaderboard?limit=100) and linked submissions. Overlapping fits agree to rounding. Fields include edge/PnL, average bid/ask fees and flow volumes; seeds, shared market IDs, paths, code, fair price, fee weighting and trade-type edge are absent. Entry-specific match UUIDs cannot certify the requested index pairing. These are fetched snapshots, not current-board claims.[^04a][^04b]

Fee units below are bps; mid=(bid+ask)/2, lean=bid-ask. Each fee regression uses `[1, retail/10000, arb/10000, their product]`, once with own arb and once with STACK arb as a common-market proxy. Residual SD is sqrt(SSE/(n-rank)); all designs have rank 4. b04a reports uncentered fee-design condition numbers 1515.59..2276.15, so partial slopes at sample means are more interpretable than individual coefficients at zero volume. Compact table values other than board scores are rounded to six decimals. O/P mean own-arb/proxy-arb designs; slope units are bps per 10000 Y. Every row is sourced to that entry plus STACK where required.[^04a][^04b]

| Rank / entry / source | Board edge | Mean mid / lean | Mid R2 O/P | Mid residual SD O/P | Proxy retail / arb slope | Retail-only slope |
| --- | --- | --- | --- | --- | --- | --- |
| 1: [Curvance.com](https://www.paradigm.xyz/puzzles/api/v1/submissions/52df1e42-4219-4931-b014-586d91720298)[^04a] | 528.3992637207801 | 35.415784/-0.131328 | 0.427942/0.577477 | 1.220106/1.048582 | 0.198141/6.758539 | 0.935883 |
| 2: [Tide v9](https://www.paradigm.xyz/puzzles/api/v1/submissions/c38c9e1f-f458-4d13-963c-9294c3b91c62)[^04a] | 528.3616108587111 | 35.372205/-0.021999 | 0.439103/0.551979 | 1.075563/0.961266 | 0.189353/5.732266 | 0.800059 |
| 3: [MixZ11](https://www.paradigm.xyz/puzzles/api/v1/submissions/e9362d90-f6b7-4497-835c-dd538d030160)[^04a] | 528.2575935240054 | 35.406885/-0.088382 | 0.617595/0.669247 | 0.886107/0.824094 | 0.254707/6.058186 | 0.913466 |
| 4: [BranchedReturn](https://www.paradigm.xyz/puzzles/api/v1/submissions/4aad6b12-aa93-4593-98cc-642e0b9fd006)[^04a] | 528.2060777235199 | 35.529458/-0.054578 | 0.632393/0.674598 | 0.784711/0.738292 | 0.209198/5.579466 | 0.804985 |
| 5: [OpalG5W6I0](https://www.paradigm.xyz/puzzles/api/v1/submissions/88bbb4b5-6f95-45a4-8fd0-d1d8ba43ca52)[^04a] | 528.0990216289588 | 35.245193/-0.053278 | 0.645456/0.678748 | 0.852858/0.811829 | 0.301904/5.745027 | 0.923299 |
| 6: [Morphogenetically Mogged](https://www.paradigm.xyz/puzzles/api/v1/submissions/8cdc1896-049f-4ae0-8e54-022d4866e4a1)[^04b] | 526.7765462821 | 34.957745/-0.135118 | 0.530102/0.620553 | 1.054343/0.947448 | 0.435516/5.108733 | 0.993812 |
| 7: [v5](https://www.paradigm.xyz/puzzles/api/v1/submissions/4f22492b-9643-47d2-82a2-bddb09edaccc)[^04b] | 525.3769622854 | 35.569604/-0.019673 | 0.531005/0.614050 | 1.063336/0.964610 | 0.219357/6.514150 | 0.914977 |
| 8: [call me bot](https://www.paradigm.xyz/puzzles/api/v1/submissions/ff0c41ee-376c-4869-90e4-52b937a2fb61)[^04b] | 524.8989933796 | 34.753391/-0.571426 | 0.640515/0.682828 | 0.849452/0.797895 | 0.324444/5.518877 | 0.913102 |
| 9: [lvw](https://www.paradigm.xyz/puzzles/api/v1/submissions/2c1b6306-9e54-4f26-9c03-8c2f009ee6db)[^04b] | 524.8101934238 | 35.439070/-0.228850 | 0.504759/0.569795 | 1.171956/1.092297 | 0.251199/6.600942 | 0.960621 |
| 10: [mevmogger](https://www.paradigm.xyz/puzzles/api/v1/submissions/65573db6-d287-47b1-928a-bd65eaa4b7a9)[^04b] | 524.6579809037 | 35.294927/-0.206795 | 0.702780/0.738911 | 0.725147/0.679644 | 0.251026/5.926323 | 0.891129 |
| 20: [CONTRO.APP - waitlist open now](https://www.paradigm.xyz/puzzles/api/v1/submissions/a1731836-9dbc-4c10-8290-6631285030c6)[^04b] | 521.977743321 | 33.993598/0.571342 | 0.436582/0.532003 | 1.115237/1.016420 | 0.339119/4.847856 | 0.856752 |
| 29: [imd.fun swarm (seat #1443)](https://www.paradigm.xyz/puzzles/api/v1/submissions/6438e80d-34a7-42f5-890f-b7d21f0ade41)[^04a] | 519.7095370332961 | 36.715588/1.010437 | 0.841348/0.841348 | 0.786955/0.786955 | 1.415132/3.214048 | 1.816653 |

All sampled average fees vary. b04b STACK mean mid 36.7155877151bps exceeds competitors 33.9935982943..35.5696044422; lean +1.0104367071 versus about-.571..+.571. Top-five mean lean is-.131..-.022, but lean residuals remain roughly 3.3..3.5bps (b 04b3.26..3.77), with low R2. Rank 9 mid/ask maxima 50.3513563766/69.0374595675 and lean minimum -37.3722063818 allow, but do not prove, shock/path explanations. Fetched retail-only slopes 1.8166532281 STACK versus .8000592770..9358833771 top-five disagree with the brief's approximate 1.76/1.1; neither is the multivariable partial slope.[^04a][^04b]

Tercile conventions differ: b04a assigns equality downward (n 334/333/333), b04b upward (333/333/334), using cutpoints 69854.14367876071/81018.66749099916 Y. Tables preserve their mid-fee deltas and label a's SE versus b's SD. Bid/ask/lean analyses agree that competitors become cheaper on both sides with higher flow and keep lean below STACK; low-flow asks for ranks 2/4/7 can exceed STACK despite cheaper mids. STACK self-differences are zero.[^04a][^04b]

| Rank/source (a: SE; b: SD) | Low | Middle | High |
| --- | --- | --- | --- |
| 1[^04a] | -0.505466 (0.037383) | -1.284915 (0.037313) | -2.111417 (0.037411) |
| 2[^04a] | -0.396726 (0.036349) | -1.314428 (0.034679) | -2.321837 (0.032531) |
| 3[^04a] | -0.468588 (0.037706) | -1.258931 (0.035560) | -2.201113 (0.032672) |
| 4[^04a] | -0.211704 (0.036276) | -1.163403 (0.033467) | -2.186209 (0.030493) |
| 5[^04a] | -0.632576 (0.033194) | -1.454140 (0.032091) | -2.326984 (0.035447) |
| 1[^04b] | -0.506015 (0.684149) | -1.280133 (0.682059) | -2.110829 (0.681744) |
| 6[^04b] | -1.038676 (0.598874) | -1.706139 (0.531371) | -2.526406 (0.547848) |
| 7[^04b] | -0.314552 (0.753861) | -1.127795 (0.709897) | -1.993059 (0.632618) |
| 8[^04b] | -1.132523 (0.608994) | -1.936250 (0.593258) | -2.815255 (0.539761) |
| 9[^04b] | -0.515783 (0.812861) | -1.238447 (1.110898) | -2.072932 (0.690741) |
| 10[^04b] | -0.585990 (0.687982) | -1.308621 (0.729910) | -2.364537 (0.635231) |
| 20[^04b] | -1.872217 (0.667345) | -2.695499 (0.554850) | -3.595629 (0.563361) |

Gap regressions use `[1, delta_retail/10000, delta_arb/10000, delta_mid_bps]`. The next table retains paired gap, SE, explained variance and algebraic mean contributions. These are **volume/fee associations**, not retail/arb edge attribution; an intercept makes fitted mean equal observed mean automatically. Endogenous volume and missing/nonlinear state prevent causal interpretation or unique rule recovery. b04a's top-five gap fits explain 41.7..45.5% of variance and covariates 20.1..30.4% of mean gap; b04b spans 13.6..42.5% of variance and includes a negative mean covariate share for rank 20.[^04a][^04b]

| Rank/source | Gap mean (SE) | Worst (not supplied = null) | Gap R2 | Mean components: intercept, retail-volume, arb-volume, fee | Covariate share of mean |
| --- | --- | --- | --- | --- | --- |
| 1[^04a] | 8.689727 (0.062592) | 2.065332 | 0.425091 | 6.217912, 3.454839, -0.871817, -0.111208 | 0.284452 |
| 2[^04a] | 8.652074 (0.062087) | 1.917195 | 0.454548 | 6.616394, 2.986311, -1.219676, 0.269045 | 0.235282 |
| 3[^04a] | 8.548056 (0.062711) | 2.446438 | 0.417585 | 6.150662, 3.394575, -0.918822, -0.078359 | 0.280461 |
| 4[^04a] | 8.496541 (0.061984) | 2.068824 | 0.448503 | 6.788137, 2.301905, -0.848050, 0.254550 | 0.201071 |
| 5[^04a] | 8.389485 (0.062640) | -11.276427 | 0.417280 | 5.837077, 3.065798, -1.768904, 1.255513 | 0.304239 |
| 6[^04b] | 7.067009 (0.052600) | null | 0.356943 | 4.734162, 2.500633, -0.858318, 0.690532 | 0.330104 |
| 7[^04b] | 5.667425 (0.055040) | null | 0.306453 | 4.616047, 3.226850, -2.410730, 0.235258 | 0.185512 |
| 8[^04b] | 5.189456 (0.050270) | null | 0.259287 | 4.090130, 3.314633, -2.132394, -0.082913 | 0.211838 |
| 9[^04b] | 5.100656 (0.052376) | null | 0.309869 | 4.214442, 2.871243, -2.475261, 0.490232 | 0.173745 |
| 10[^04b] | 4.948444 (0.046694) | null | 0.283764 | 4.213022, 2.631744, -1.968946, 0.072623 | 0.148617 |
| 20[^04b] | 2.268206 (0.047528) | null | 0.135841 | 2.603298, 2.767821, -4.447415, 1.344503 | -0.147734 |

b04a's testable proposals: ranks 3..5 use `mid=clip(35.4+6*v+q,5,100)`, bid=ask; ranks 1..2 add transient lean `ell=1.5*h`, also testing its negative sign. v/q compare fast/slow EMAs of completed timestamp-bucket price variance and **total** Y flow; h is a bounded signed standardized prior-bucket return. Its fast/slow decays are 63/64 and 1023/1024 per elapsed step, elapsed capped at 4096, v/q clipped to [-1,1], h to [-2,2], with warmup 35.4bps. These are own-history proxies, not the API's hidden retail/arb volumes. b04b instead proposes (i) ranks 8/10: `mid=b+.15*sqrtV`, (b,lean)=(33,-.57)/(33.5,-.21); (ii) ranks 1/6/7/9: `mid=b+.12*sqrtV+.10*shock`, b=33.5 (rank 6 33), lean=0, shock=max(0,sqrt(Vfast)-sqrtV); (iii) rank 20: `mid=32+.08*sqrtV+.05*Q`, lean=.57. Here V/Q update by 1/64 per completed bucket, Vfast by 1/8, price changes are in bps and Q is own Y flow per elapsed time; fees clamp [1,100]bps. Different proxy definitions make these coefficients non-interchangeable.[^04a][^04b]

Board hypotheses require removal/sign ablations, fresh paired gains and zero OOG; aggregate fee fit is insufficient. b04b suggests shock gains 0/.05/.10/.20, decays 8/16/32, flow gains 0/.025/.05/.10. b04a mid-residual order is rank 4, STACK,5,3,2,1 in both designs; b04b's most predictable competitors are 10/8, while 1/6/7/9 retain about .95..1.09bps proxy residuals. Aggregate predictability does not establish a simple swap rule, and the cheaper-side experiments below reject an automatic inference that fee cuts help.[^04a][^04b][^11a][^11b]

**Whole-family searches on tuning seeds.** The following paired results are held-out tuning evaluations, not the >=2000-seed confirmations required by the overall goal. R/A are reconstructed mean deltas where supplied. b06b is just two partial 16-seed evaluations: its missing source code, gas and seed IDs preclude a validated candidate claim.

| Source | Seeds | n | Mean (SE) vs STACK | Worst | R mean (SE if supplied) | A mean (SE if supplied) | max_ok/OOG |
| --- | --- | --- | --- | --- | --- | --- | --- |
| [^05a] | [110000,110512) | 512 | -0.23209032978094113 (0.01286418335743683) | -1.222266458397712 @110087 | -0.20501330487098002 (0.011181100972920417) | -0.027077024909960744 (0.009531165273663529) | 219173/0 |
| [^05b] | [130000,130512) | 512 | -0.0021 (0.00326) | -0.7708 @130035 | 0.00168 (0.00312) | -0.00378 (0.00121) | 214228/0 |
| [^06a] | [210000,210512) | 512 | -0.024030512124604497 (0.007341628224760941) | -1.8040017088139848 @210020 | -0.121960624576537 (0.007312988999897048) | 0.0979301124519322 (0.002529238391012772) | 235618/0 |
| [^06b] | seed IDs absent | 16 | -0.3220185810505072 (0.09743781201664779) | -0.9247207857755484 | null (null) | null (null) | null/null |
| [^07a] | [310000,310512) | 512 | -0.5755359342052299 (0.04581899892333414) | -16.679177770917022 @310360 | null | null | 212641/0 |
| [^07b] | [330000,330384) | 384 | -0.8356353613735269 (0.039605775302907854) | -8.30029587990407 | null | null | 212934/0 |
| [^12a] | [530000,530512) | 512 | -0.23909302626313944 (0.02151175567562457) | -5.601101562553367 | 0.41435160324045084 | -0.6534446295035903 | 214057/0 |
| [^12b] | [550000,550512) | 512 | -0.0389275909257184 (0.018164335492415494) | -1.1780050356396714 | 0.8823457127385416 | -0.921273303664259 | 213812/0 |

SHARE-a changes only the attracting-side quote after STACK's caps, identifying that side against the 70/30 stale estimate. It maximizes supplied `fee*share` over eight grid fees `scale*j/7`, uses t=c1*sizeHat, subtracts `c2*sigmaHat*sqrt(1+silent)*(1+c3*probArb)`, and blends with STACK at separate first/later weights. CMA seed 1, population 10, 108 objective evaluations on [100000,100064), plus three rechecks on [100512,100768), produced 111 logged evaluations, all negative (range -8.052146531904931 to -0.22341425131036452; merge calculation from its log). Recheck gains were -0.24922139427232226, -0.3594747917519838, -0.3213993838462863. The winner retained roughly .93/.96 of STACK on first/later prints. It passed 151 share-math cases and 120 exposed-side/estimator checks, but lost on [110000,110512). Its initially rounded volatility bounds yielded CH 525.5234442071842; that pilot was discarded before the corrected search. One STACK held-out baseline chunk was repeated after an overly strict volume assertion, before candidate evaluation; the candidate itself was frozen and run once. Final decomposition checks allow near-zero-profit volume plus 1e-5 and stable/direct arb-edge discrepancy <1e-5.[^05a]

SHARE-b uses an approximate closed-form argmax `fStar=.0015+t-1.2*t^2`, t capped .04, with claimed fit error <.04bps for t<=.016; signed risk subtraction permits either direction. This is not the same implementation as a's eight-point grid. Seed 2/population 10 completed 600 evaluations in 60 generations on [120000,120064), stopping after 1125 seconds when the mean collapsed toward STACK weights 1. Only 200 evaluation records are retained in that source (60 best plus 140 spread through the run). By minimum first/later STACK weight, evaluation counts/mean fitness were: <.7: 8/-5.4105; .7..85:39/-1.0391; .85..95:84/-.2794; .95..999:264/-.0662; .999..1:205/-.0725. The best recorded gain was .0159. Rechecks on [121000,121256) were .00399(SE .00478), .00384(.00477), .00101(.00443). Winner WF=WL=1, LO=0, HI=.003229: only the 32.29bps attracting-side clamp remains active; share/risk parameters are inert. It was flat/negative on [130000,130512). Both SHARE implementations calibrated exactly, had identity variants equal to STACK, and passed final validation; neither supports adopting the share-revenue model.[^05b]

ARB-side-a saves first-print classifier probability and `sigmaHat*sqrt(1+silent)` in slot 16, then adds `a1*scale*(b0+b1*probArbLast)*d^(prints_since_first)` to the numerically higher STACK fee, with total cap hiCap (ties choose bid). Lower fee is unchanged for the same prior state, but future paths change. Seed 1/population 10 used 190 objectives on [200000,200032) and three rechecks on [201000,201256), all negative; log range -335.29509136398565..-0.014066744646287432 (merge calculation). Rechecks were -.01716144308883072, -.051101532517776826, -.05777924052701655. The winner nearly removes the surcharge (a1=.0017873909184314295), but held-out [210000,210512) loses. Zero surcharge reproduced STACK, CH calibrated exactly using 16-seed chunks after memory termination, and final validation passed. A partial STACK baseline was repeated after changing decomposition tolerance from 1e-5 to 1e-3; final max volume error was 7.967033161548898e-05. ARB-side-b stopped at two 16-seed trials, -75.0 and -0.3220185810505072, latter params (.1,.1,.1,.9,.08); no CMA population, recheck, held-out run or gas instrumentation was completed.[^06a][^06b]

REGIME splits 15 fee coefficients/caps into four regimes: first after silence>1, first after silence=1, second print, later prints. It changes the quote **after** the observed print, not the fee that print paid. a used 60 log multipliers, seed 1, population 12, initial sigma .12, 1152 evaluations/96 generations in 1817.9 seconds, rotating 64-seed blocks in [300000,309744). No evaluation was positive; best -.5018, median -2.6979. Rechecks on [309744,310000) were top0 -.8162(SE .0677), top1 -.7751(.0439), selected top2 -.6261(.0455), final CMA mean -.6666(.052). It lost -.5755359342052299 on [310000,310512). Four coordinates are inert because CAP30 only acts on first prints and CAP16 only on later prints. Initial perturbations were large in 60 dimensions; the note's suggested smaller sigma .02 or fewer active parameters remains untested. A prose mention of 97 generations conflicts with the structured count and 96 log rows; use 96. Full search gas max 214731/OOG 0; held-out added about 1.1k gas to STACK. Edge splits are null; held-out volume deltas were retail +708.7900073218756(SE7.650333165271051), arb +604.6656641766716(SE5.2097525318654165).[^07a]

REGIME-b used seed 2/population 8/sigma .12, only 24 evaluations on [320000,320032); every candidate lost (recorded range -18.1376..-.9595). Three rechecks were reduced to 64 seeds [320100,320164): -.9044282699815174(SE .10092459983339352), -2.5287593344205437(.13373665132211643), -3.0658306180670394(.20067573522098203). Held-out was reduced from 512 to 384 seeds [330000,330384), with no internal omissions. Two exit 137 terminations and command/runtime ceilings limited the search; CH calibration and separate amm-match validation were skipped, though the contract compiled through SolidityCompiler and had zero measured OOG. Suggested sigma .03 or about 20 selected parameters is future work, not a gain. Both regime outcomes reject these sampled policies, not every possible state-dependent split.[^07b]

**V10 refinements.** Floor-a measured mean sqrt(uncertainty) .0015998 for V10 versus .0063816 for STACK on eight diagnostic seeds; V10 median/p99/max were .0015835/.0032862/.0046498, STACK .0065692/.0092261/.0095917. Floor-b measured .0015919872381250002 versus .006404864248875, ratio 4.023188186117922, on a separate eight-seed diagnostic. These use fee-encoded end-step probes, not direct native uncertainty fields. The first grids kappa=.1/.2/.3/.45/.6 were exactly equal to V10 on a's 64 seeds [400000,400064) and b's 128 [420000,420128). a froze .3 and found zero incremental effect in its incomplete 400-seed confirmation. Its speculation that binding would begin around kappa 2..3 is superseded by b's extension, which observed binding near .8 and severe losses at larger values.[^09a][^09b]

Floor-b extension means (SE) versus V10 on 128 seeds: .7=0(0); .8=.0005036065588561733(.00031558127772019997); .9=-.004652517342058626(.00646802310597134); 1.2=-1.0140940239171563(.055832647040764544); 1.8=-35.40587510776863(.48326383876975654); 2.4=-139.21181415888586(1.6621109615886158); 4=-274.9109547744654(3.722487984857109). On the next 384 seeds, .6 was +.00003530030074442673(SE same), .8 +.00006160340376378433(SE .00024395949408807918); the selection difference was not significant. Thus “floor<=.6 never binds” is only true of the initial screens: .6 changed a later path. Frozen .8 differed on 972/2000 confirmation seeds but added only .00017859562930885886(SE .0001604829557281328) over V10, with incremental R +.000043641685714732145(SE .00014350548048563033), A +.00013495394359404678(SE .00004212133078356895). Its top-level gas 220975 is an all-callback maximum; afterSwap is196935. Initial screen JSON was overwritten by the extension; b preserved those screen figures from console output.[^09b]

Blend A weights fastHat into the toxicity price; B raises fastHat weight in staleHat from .3; C multiplies STALE_DIR_COEF. Independent screens agree that B and C lose, while A differences are within noise. Each screen uses 128 seeds (a [440000,440128), b [460000,460128)); table values are mean(SE), six decimals, versus V10.[^10a][^10b]

| Axis | Value | 10a[^10a] | 10b[^10b] |
| --- | --- | --- | --- |
| A | 0.15 | 0.002905 (0.012793) | -0.008791 (0.013232) |
| A | 0.3 | 0.004876 (0.019578) | 0.002821 (0.018224) |
| A | 0.5 | -0.017532 (0.026145) | 0.004462 (0.024331) |
| B | 0.5 | -0.876281 (0.066710) | -0.738883 (0.058486) |
| B | 0.7 | -2.530026 (0.112196) | -2.403090 (0.102827) |
| C | 0.8 | -4.984393 (0.216399) | -4.829381 (0.192255) |
| C | 1.25 | -18.545154 (0.579520) | -18.940137 (0.807327) |
| C | 1.5 | -112.339784 (3.115121) | -107.421476 (2.909497) |

On the next 384 seeds a shortlisted A=.15 at +.01286338784136302(SE .008790333834224354) versus .30 at +.00996774907163959(.012773361101042836); the combined-axis candidate was byte-identical to .30 and simulated once. b shortlisted .50=+.00022222045979821323(.013813698638517818), .30=+.001696776676511913(.011292604680675602), .15=-.013310791143336331(.008735985424277158), freezing .30. Neither selection establishes a preference that survives confirmation. b's incremental R/A over V10 were -.4062639260105809(SE .006065795693408151)/+.4014930261889765(.0026952972411231767): another near-canceling tradeoff. Both calibrated CH and passed final validation; a repeated V10 on 16 diagnostic seeds with identical records/counters, b's baseline-settings generator reproduced V10 exactly on 16 seeds. b's reported overall initialization max 220981 is separate from afterSwap 195466.[^10a][^10b]

Retail A increases STALE_ATTRACT_FRAC from 1.05; B multiplies only the strictly lower fee immediately after stale adjustment, before trade boost/compression/caps (ties unchanged); C reduces only the first-print lower-side cap from 30bps, leaving routing thresholds and final lower-side 36bps cap intact. Every screen lost in both copies. Screen seeds are a [480000,480128), b [500000,500128); the next 384 seeds are their shortlists. Compact values are mean(SE) versus V10.[^11a][^11b]

| Axis | Value | 11a[^11a] | 11b[^11b] |
| --- | --- | --- | --- |
| A | 1.3 | -0.556554 (0.044509) | -0.606183 (0.048418) |
| A | 1.6 | -1.546846 (0.057438) | -1.542185 (0.060252) |
| A | 2.0 | -2.718191 (0.064776) | -2.665011 (0.075287) |
| B | 0.85 | -0.421240 (0.041809) | -0.469401 (0.047484) |
| B | 0.9 | -0.248287 (0.028946) | -0.284771 (0.041589) |
| B | 0.95 | -0.101534 (0.025010) | -0.141861 (0.040312) |
| C | 26 | -0.018527 (0.013328) | -0.021152 (0.013337) |
| C | 28 | -0.002166 (0.009171) | -0.006343 (0.007573) |

On a's shortlist unchanged V10=0, C28=-.004624628836863505(SE .004758737198527349), C26=-.01750121227830374(.006875077128051377): it correctly retained V10. b's C28=+.003219679518680074(.004668441032463799), C26=-.003061487874739773(.00703579660260766), unchanged=0: it froze C28 despite the insignificant advantage. a's screen decomposition found both R and A negative for all A/B settings; C28 had small positive R (+.004469713904005079) outweighed by negative A (-.0066357512598367585). Its C28/C26 shortlist R/A were +.0018993150405721693/-.006523943877434268 and +.0011414016427557054/-.018642613921058326. Cheaper quotes therefore did not reliably increase retail edge, even before considering arb loss.[^11a][^11b]

b11b reports `simulator_valid=true` but **`valid=false`, `strict_protocol_compliant=false`**. Its first 64 frozen-arm confirmation seeds [5022000,5022064) completed with no callback failures but were rerun after an 8.922609049477614e-05 Y reconstruction assertion fired before saving per-seed edges. The recovery had identical mean 539.4746559204559 and callback counters; no policy selection/change used confirmation data, but the strict single-pass rule was violated. All 2000 seeds were retained. Its stable-formula arb-edge audit differed by at most roughly 1.94e-9, and max volume error across arms was .00017910497626871802. The C28 increment remains statistically unresolved: +.0018787537178642708(SE .002555729916292394), with R +.0067699861335822785(.0023090360838334727), A -.004891232415715501(.001227762128175223). Do not silently promote this to a compliant confirmation.[^11b]

**RES-TABLE implementation and controls.** Both b12 copies add pre-compression residuals to STACK with five signed gap/sigma bins at -2,-.75,.75,2, first-print flag and elapsed>1 flag: `cell=ig*4+(first?2:0)+(silent>1?1:0)`. Values are signed int8 half-bps; ask row is the mirrored bid gap row, **without sign inversion**. Negative fees saturate at zero; no storage is added. The non-first/elapsed>1 state is unreachable, leaving five inert entries in the 20-dimensional design. Both seed 1/seed 2 searches use population 12, mu 6, initial sigma 4 half-bps units, rounded/clipped latent tables. Zero-table equivalence passed 32 seeds each, with b reporting all 320000 step records identical.[^12a][^12b]

a evaluated 53 tables (52 finished within 1800s, one in-flight completion at 1833.6689471961s; top3 unaffected) on [520000,520064), rechecking [521000,521256). Every logged mean was negative, range -2.0020644065830826..-.3146506376940037 (merge calculation); selected evaluation 1 rechecked -.2890065370527688(SE .025422559963927806), versus evaluation 22 -.31485672337071957(.02286999906437466) and 40 -.44345704223272886(.038650347784695234). b evaluated 111 tables in 1801.9533251959947s, nine full generations on [540000,540064); all were negative, range -1.8486511847475553..-.03943390745307607 (merge calculation). Rechecks [541000,541256) were evaluation 55 -.040801335433506614(.02701109591412885),84 -.11490132183064139(.028381435006178096),65 -.11197763151534312(.02130632242269024). Source gas checks report no failed candidate in either search; both calibrated CH and validated the frozen contract.[^12a][^12b]

The frozen bid tables in half-bps units (gap-major, state order same-step/elapsed<=1, same-step/elapsed>1, first/elapsed<=1, first/elapsed>1) are a `[1,3,1,-5,4,2,-2,2,1,1,0,2,-3,-1,-2,2,0,-1,-3,-1]`; b `[-6,-4,-3,-4,0,4,-5,0,0,5,-1,-1,-2,-7,-3,-2,-2,-4,-1,-1]`. b's packed hex is `fafcfdfc0004fb000005fffffef9fdfefefcffff`. Their held-out gates required mean>2SE and >=.05, with no failures; both failed. Shuffle and exact-mean level controls use the same 512 held-out seeds and were not used for selection.[^12a][^12b]

| Source/control; n=512 | Mean (SE) vs STACK | Worst | R mean | A mean | max_ok/OOG |
| --- | --- | --- | --- | --- | --- |
| 12a shuffled[^12a] | -0.5196971455567604 (0.021461834174538393) | -2.570266522453835 | -0.17010450774045208 | -0.34959263781630834 | 213886/0 |
| 12a level[^12a] | -0.01444637617505029 (0.003403153167373188) | -0.4630215585827955 | -0.04143431273949018 | 0.026987936564439888 | 212740/0 |
| 12b shuffled[^12b] | -0.7492254310664952 (0.025458721840595243) | -4.2221637708092885 | 0.49648612047180074 | -1.245711551538296 | 213800/0 |
| 12b level[^12b] | -0.11796333883680243 (0.018582634471081456) | -1.4211329290418462 | 0.8330835229561656 | -0.9510468617929674 | 211995/0 |

a's exact level is .05 half-bps (.025bps); its initial rounded-zero diagnostic equaled STACK and is separate. b's level is -1.85 half-bps (-.925bps). a repeated only a STACK recheck baseline after relaxing an absolute-only volume assertion, before candidate rechecks. b repeated only the identical level control when error .0001024852754198946 exceeded .0001, adopting `.0001+1e-8*volume`; no held-out candidate/STACK rerun or selection change. Both learned tables outperform shuffles but lose to STACK; a loses more than its level control, b less. That suggests cell dependence within losing policies, not a positive policy. Full 42-cell, gated/horizon/tree searches and the full design ablations remain untested.[^12a][^12b][^03p]

**Handoff.** Preserve V10 as a small-gain, lower-gas candidate, not a top-10 solution. Its refinements show no reliable incremental gain. Liquidity-shadow and untested structural families remain proposals; bounded negative searches do not prove a global ceiling. Carry all protocol-invalid/incomplete flags and source warnings into the final write-up.

The ledger hashes exact UTF-8 code strings: all 17 supplied strings match (16 distinct, since b11a repeats V10). b06b's hash is unverifiable because code is missing. Only V10 is encoded below; losing contracts are identified by source/hash, not claimed reconstructible from rounded parameters.

| Source | Solidity code SHA-256 | Merge byte check |
| --- | --- | --- |
| [^01] | 2eaafc7960e520b8a7fd322ec8afaff8fa677c730228a961efb509cc30b25719 | verified |
| [^02c] | cd9bdb45ab0537ad25c4fd90ca8618898cef6386de3df65f192eaef65fbca245 | verified |
| [^02v] | 5c36797359a0e8e724f703fca6bdb3c152cbd73134f2085c8f47f00c6a3fabcf | verified |
| [^02n] | ce7e02d29699d5b55794e9abf57ff3d3e4833af24534a412b9387fda0f43d2ff | verified |
| [^05a] | 35085dc82d2aeb7a34b077e30b5a408a2aed909cafc23253270eeb9756e90681 | verified |
| [^05b] | 43a13736a1c976ab0137014596f4b96c95ff5a4fb2e8c5e799939491ffbed28b | verified |
| [^06a] | 8274e4615c33bba86ebdaf44d8b1b2ba8895c2548fd52c85272453f0a5eae3ec | verified |
| [^06b] | 3ccedbcff3d4faf7951a6f5a32729c0b31598198c32d5c4fd6de7523ee314be8 | code missing |
| [^07a] | 10a8c19ce4cc5b3c5f32d701fbdb4f794d35ab212a30918cec72b931ea2e408a | verified |
| [^07b] | 88e49bf983166c778dd9db709cdff2da51a3a4bacca1361027a23248fcfc0b70 | verified |
| [^09a] | 0d6e3a4c0a839baaf3e07667c6f0b22afb30c0699ab928910b793c0ce9b194ee | verified |
| [^09b] | ebf9989a4a3d47e0473411213da63bc683598ffb1e855891e7ba2e8fec952e6e | verified |
| [^10a] | c46e2b14968bd78fae31442877437e7a906d61c8707056ee488139955bc8ab9d | verified |
| [^10b] | c8a4b226079a7c993e0ff91f64cc1c3253a60738fda85f56dd4a800ce725a8eb | verified |
| [^11a] | cd9bdb45ab0537ad25c4fd90ca8618898cef6386de3df65f192eaef65fbca245 | verified |
| [^11b] | 6694d28afa215f0851a46e87b645bbf05fd298df564278b0af3666b589b53cb6 | verified |
| [^12a] | 1f9110932896a7eda224defaab58e86e3929ed611964cd65a91e203370266751 | verified |
| [^12b] | 6e167c114b1c373c05db4020fbb76c516b57418cd89d90515e0ea2f58c83f8d9 | verified |

This **zlib/base64 payload** preserves b02c/b11a's exact UTF-8 code, including whitespace: 17834 bytes, SHA-256 `cd9bdb45ab0537ad25c4fd90ca8618898cef6386de3df65f192eaef65fbca245`. Decode with Python standard library `zlib.decompress(base64.b64decode(payload)).decode("utf-8")`. Decoding and hashing were verified here; encoding saves space for evidence. Its imports are the challenge-standard `AMMStrategyBase.sol` and `IAMMStrategy.sol` at the [pinned checkout](https://github.com/benedictbrady/amm-challenge/tree/6b6921f8d463314aa7c96234e1be25a25245f704/contracts/src); neither binaries nor runtime dependencies are bundled in this Markdown. This merge checked source bytes, not a fresh compilation or simulator execution.[^02c][^11a]

```text
eNrdO11X4ziy7/wKzTzMiRsHSAI0EBJOgDDDuXwtSe/M7D19c0zigLud2GM7BHqa/36r9C1ZDundt+UhxHJVqb5VKinb22Rwd/5H
/Soah/M8rF9OwnkRTaMwOyLXl8ONNAseZwHJkziaRMUr+b+drYOt5m57YyOapUlWkL9719eDIguK8PH1NMjDNzLNkhn5eWvberEF
NH5uS7RhFkzCy/k0UQiXGgaH3hgncxgZF0S8IFFOLMrk7w0Cf4toXjT39gmg5EUwL0j/qnc36J+Pznp3pEMO2m6oweWvN72r0fC3
+/7gt9urcwD9vXdOtsn+3k4Fyvnl/ei8f9b7E2A/7u1YfxVIvbOhRNprrYk0uPxX/8exhrd/SKTDxtpT/Xrdk2j7h2uiXfWuT88V
3uHhj0h2etW/OZe4B82D9eUzUfca6xvBxDw82FsP9e633nDUu4J/gNX8uPOjWKP7/rB3eYW6/XHcQe+6D5iN5pqo9/1/9u+Ho0+A
s8qJL6/vUCPXn66GALlbAflrb9gfMffgkB9LfFTNcn15M0J8QFrbe+975/0RtfHt7QCna4GJyAdyejeocqZh76o/6g0BFcS5gE9U
1s7e2toa8iTBIr+1t8LxdMC9lSLc94aXt2vB8yBioI2Dda08GPbvRme3n24E+/tVJjztDfqjiz6aYXe1JqmZz277F+jlOz+WBDha
o7lyiour299HNPw5/O7B7moMqnYG23rHExD0H59655IX9IGVGOgxHPiwcbDbWgmMQSOV8y6kxndjLQ/WyO9DXlrDUtoU7wUJZp/R
/9xQJ9hfB3ZwdXvXH93d3w77ZxiGh2tHsMLmMYnZ3RGNFH26mI+LKJmTYFqE2eU8KqIgjr6FNUE3YkN/+MQa+dMj4QvgzIOYJM9h
lkWTkGRhscjmORHoEsvjdQL+5XFS5P+78xn4EqHR5oMN12DrM4vhNrFI7OELwSDpkh1yQpaT6LkmWPTlW48ccTdAQhadfaRzaKpI
TP4R3x2UY5G/PvjsCFWbfoNKcLjT3LPRG7v4RkijCcgUSWpCF77Uiseov7nsN1gGaU0Wd2QcxPEkKAJS4NC/bS9h+DQLn0+jyUUY
Sp53gGf9dS//qr9uaK/jIC+GObzCgf3dGoNofvbapYkmUTYooMKUdFoaHahH+1AUi1e7Oge/BYWuzhLhPIJ6WgfaNxicPUyMtx+1
tzkEhf7uwEG+SF501g519CJMqWHOksVcUWmAAjfcbF4k2QWoDEE51+UJpwCgMwXuVAYCDwmzIoCnVwXYcmonDilv1F22imgWQkqZ
pRBZ3HgnpVd18eqIaG7/kCQx7BVuwuUA5F5BUaFEU1KTGLr36RyGcZDm4eQ+WDpo1ks0HaiApxHpGtuUE+PpSAM0SWruOZqE4+D1
DNQWZuGkJt74apfiCzKeSUS68XK2iGvsySejNFn+Hkxqcr+i0C185Y+UAH9UFNTmpZKEdFdKgT0pAnIfU4mPJrP8GnPwL7+YKt6x
rambhQXd5Zw6uk2NJmsPSrcqSwguNCpdrZ7zTPLqRZmKHvxUHXLAN3ZZHtnU3yNln9aXdRPMnOHNeNreJndZMg7znDwHWRRAfOLG
Os2SNHgE75nAqpVjdi6eQoKdgQR8iyxSSOXhVskCenh3ibUV86zob1QtVNSl7BylvX/bMIN7GmV5cTnn4W1jGqgyvaSJSi5ZmIfZ
c2is2sYbcDsTElfw1MiCKD1L+R10MpH9YRpHTp2G4ScW/4xslJ8uXmFmbVE70pawMvuPwYwGiyB0TM1+wo0vRo9ckqeXszSOwonJ
OidIef9bwggJNJ8JY2yz6BCmCCwBAJLPmERNUYXqY23dhOCBgyPwsmmYkYDMkzolCAOQxp6hPhjThMZ6Q+iEmFhJmuQFB0TCW+71
YxwOoOwIecVmCKz7DCQJvtyAz1al+0e6+FGoOmmUk88jQ0awhjvrL/nury7TWrmUAw15buxcrpfNz6TblTs8IwFyCyKs6Z5GdY7Z
jdlJpAvleDSnUHvp75gfepj+muVpx6TL5uN+v8nT1tKHuersjYlFvYjC1hUseybj0uJiWxKAKE7Lob6l58wTbptCxck1pYd+8JCf
R9NpTfg448zj/8yo0qkFcfpEg1JzrRO9W3RUbgK5KYHK756CwmlFmvHyELZYk7ssoqnxJ9uZSxmw4aIRzMLzKEOGAzBHm40maaoP
ll1MzexaSimNRdpE1kWK6LrEMErshqoFm5/dqyrA/MSzk2IcR1E6Dz9h1rZi/ycO1bbWOykHe+1Jq1k9tlVrppU/KE+Ce4fg4AQN
sH3TZkX6zSRIi+g5/JWVcbyEYmW2b/fcHGWPgX8sG22eTVi8sGo/Lj8r/fDBN0LOLhUdzOpjK5GRW4y5445Bhy45WrmTUsFZoFOO
ZKGjQpKNtx2W4S4AcSB7n+/UfRBuygEbFQ6IQF09y3HjCs7gNUt1vpzXK1NSia8j8p0UmeXaSvT3fNDJNWq7K9qaHk94/LHMnbYl
tbxQOxOQxgBiwkr66yquzSX/Ehf1KMmQpSCKSZYsimj+SMK/FrTnk2MVUw8mXxY5FqGzIHuMsFuQZhEUq1tG9bB4yAENV2ZaEeRQ
O0AlgVmQjIP5HCq9B1a/8gYMCbKHCEAfQ168GJWBnU1FNdUlLd4qw1FWMDkcSzjVPMlmQTzQ60xaI5WUzledUuEkdCtm9UqYRyVM
C0PWWwZqOdQOeS8KlvcG1uncwEoEH8YNkHaFWc/EDiHI88UsnNQnMIDHhViwUYNvkZtEa/5AYbcoFhnaJl0UmmFrvJ/g67sGD7e5
klTtTny7ROSaIaUL3ScqgTCFmRWVb+pJFkW+VgT5+vrOs13h6xY2qZgbifIWI5jhIm3vMP6kdYYkpO929TaeJlsJgHbzuBYc7SGc
7T4oosTeBv0gj2ZZrZHt2ocfnjmp9VbbJpUI2SezVTGXLvInIQ7D/WCfbpUXJIrU5b696wki/LkMr1nalXW1hoz8ukmJVhQ2AqhL
mqLJoNFo2l1iGnN8G7bW5F0m0YkaqbORUiH7tkZTyDo3lblC6VxkIRtyvZaRfRq8YoISqKNa5bN0mV7VpIZK3zZc3dMVOwOWO6p3
BdRNgEJXnBN6nCB/VJDutpd1rK10kLwI4W0QZwN4ED1iku3wabQQK7eAzYFNfcvg7K2ZR45emaIJ4OqCxMlyABZxdLm4oRxSTek9
D3UuI1Sjjit9WcZIvWmnkr7qrsnX5jmkLxlzTX8dYc+DccHRxcmbr1Tu6e/kKaQvrcigdASJIU4hfR6AGhNicvxniK04sMo2jb7D
AJAQzsPntthSxvEdZPYcFmPT9CqTdFgY/c0x9SxTZ4diOhWaiReho3sk0ZkvCyIldL4xdcRn/jVcCr8RZ6U+pytVqZ+8ctUwCEsx
NvEHuuhrpzxWF47tgxWn9jLwIM6luKWQWWsXJY6mKESXiXPCnursyUgozoRvEnFNYzKyxjTvNUzyAjZ1pfKxdAVGlZCyCiv3aqr6
XDjFkGZLR84VDHjaV9dBAKehJWCNbikLl+YfPEVTKaR5Iu9LShUCBAW9p2aQUER9xxUVlwS4b+h2lJCuXouwLv+yqfFerjWkt/Av
XZPREzFeN8fLRUJV7WFPsJodm/kSO3z8XXbed16a3mj50Ish5mnTWq/iaMuM65st5t+/wzbQBXHsambKopCTX9VqoHCnSZJL17Au
Ofla9eq5K0Z7g6rR7JLGPt/7mVOJYTdFs6R1OJai5eqn8axedoBqNKfZ1s6rY1jDU/haG42TWYrQwyCKf4+Kp0GcpGHtge/aypdX
7GNDyfN7JAO++SvfaNFJrs7T/8YcFWz/J5pwsb2xsWZfifYbRQLhXIALyiCWrQfFYmvH5XpUT0hNxT8jC9TkBIqa1GGZmsX8Snr7
ZXqlwJCcrZRzvyynpOS6XCRqCu1qES8rVA+iw/P+8THZh43od/tWgb3Fpy0ArXjid1CQNC0d2/rVJLMFr901kjc69EtGskg27hbx
stwmc/hZbTHUXZLSpkJTjLhRJNyTqULdJtpge5RsMS6I2eRxXAWaLVSp9qy+Jg952zjvNW4HDejJrKPyM8q+tnbA3LbupvB3D4tX
82KOEMK4EVXZtCKzcJZkr+TFYyeetMeJ1e9at6HQTV+2Zgt+7se+wr/ckI5BPcPq1dBLNBwyTuHL+ogTsUy9bHGl0Q1o/WULVOUo
nZ8iUbeVEECdns0TVawrx4AIVMndBoadpNVd2cfnNzM6Evng5OBIPJRXPpR/07wK96HGadQbjsWXqtoHNK8zegjmE/nsx4n/FK3s
1wPUZse4dbdRBpQefc+a4h1q0bKWn7X3z229/9ojsyjPsZGeLOf1VFY+eKcjUPc3aBcdIwSP2NFBt1yGMVTv6nXXOKO+4Ij2Z5Vq
kD3YcCc+MbTzZvbbK9yAShMHePaf0oNOEIEdFmyR23n8SoI5Cxl6ipAsCibkXBfLJphM6TlAAh9AFbQD20JYK0n4ksbROCp0XL25
uuWs9Kc070Mo0JsPNCSgQKWeXq4Qp9aKVu3GrO2uUVZdfhbZLKKAIDsjqLd2PjjPB+wzAh2b4yhS1cezUFVEkwXlyNiRMT596qOe
Tz/bK4SKaN/G+FWIOkliM/jiywpCX1WCecbY26SUHQhijRk9QPhzh+Qsf/V8rpBnH+X/6q2MXUFJhv+q2A1tNb1sgVP5TiWpVh3s
lP7Jwhl00nRlCQH7TdCnEoQ+OAFDd91d5YcvHXnjhGKCyCj25rftA8/fdeAlk0kuEj+n4VPMXfu+xSjK/8qKGuPgA3aHvFKSx0WR
ur7dI9Cnwe/+LlLYbuw4SKjrOSWsgzKWXI9xr8d5BN1SBstQy2hSPAm1PkV1SFgVhppEeK99HPLMfBwnJ3FSx69H1Du6T9EJ/q8/
RUcufr6dQm6UBsQPQZE7SO5aVGdBLiVmmMiwTz83mx8AxzfsSgHpVD79LNmEEqR3xrbpeR+fgD8rWKFqJE5VjYAOBtMseTDgkAv8
4oD92sse9BCudN4ZA9TDl4USUnDBh5D7JbxYmZCKj/y5+nxhsQzDuR2wFIXjV4asvCwpOmEwhZVUkE3ZCMMRCiNWSiPDcCBFSYDz
RMU59fl/256SGcedS43RyguXIsGhCn2B4LkLWaZeWSP7Upny27KilK05Clg+80O3E2ALirnuQz3wl3ikEtTZ2VMdVFvFTqBxAy4i
7eOrGtZX1ek7rPkOFvGGQTIt1IXEOPoaxtFTkkxI8JxEECIQxPjz1Gi+iIoohCKkIH8tgkkW0JPteTLRbyyo3NSh0gHTPEOx/FR2
Ntzo8W1CMB5r5wGwmxDUvnR22l+OP7a/bG5Wno52vhy3QMswY7fWqn/xIHOcQK4S348aqPPZYrP2pd7CgYprjJ0vnU7r5GPzqAZf
mt+/w+fuye4efWzQx72Txg5Qq2iQipTXSXn+TI9qKWbO1E6bRurs5JPuzomZNCH1HVXcxlNW6hjJcfcDS4/+N5u/JbPHEq3gK/TS
Xfkx7Og3O8s2fmvgN3Z/Ma0St2PeKvQxz5dqJSTWVMTY0utPvJW1xoyTpoz4jDPXqmVyMCtzIH2S3u5TVJsOqixkgAoD7kpWT9hA
XQ4cNUrdfvuXQhVtK63CVr/1yvHVe9tke3eMZfdxR/3kzRMpZ6qf6PAx9cM4cXxBO9By2Oc8VMjCayz1g5hxkWSKf9hkvMc91tqL
uDAzJI7YR9bLJ6iISA1IuvZlKDi++oWI+5GSiqi34cnnHFqeyAYFqBDCCUqn7xrXS9/cqjF/miJvIvMBX6tYAe7CVhz/yccPmF6t
CLFW/RrUyz+FoXeZOUvaYSd3Djyr5G4hgeipp4+zGFcRVYERJ/KYkp11KpEpUmkhpBjmdX42Jk8+uG5hWbqIEzBUDksNaiNJiiPy
EBV1WKAfoZ7NQ7yGBeVmHj5DkXMTLguwBLbkQGVBHMt+0pZlK7YYCRle3vfZUkeKt6O4TPomZj5+CsdfgTH3EoVblRfTx7LyhXqc
4xVvbDWaB/JW8CvzxOZBG1COj+lPoB13Jl/ZTS8LDWEZVqtZidVqWlgIy7Aa+5VYjX2bxX2BdVCJZIslhaqWyRZJSlQtkJKHS2FD
0jvWGXj9C9lGM6M47f9aCLkbRydEoLarCwCqIng+mUFEFu/lvcewuAlmYU37uSuNofJvXvMiw64dK/O0rPOz6NmNGjs/o33eNv4f
LfArSw==
```

Provenance: paths below are relative to `artifacts/goal/batches/`; staged inputs are `.imd/reads/artifacts/<input-name>`. Each hash identifies the input JSON, not its Solidity. Original files are not promised to accompany this merge. No external page was newly retrieved.

[^01]: `b01-calibration-and-harness`; `e01-evaluator-check/items/calibration-and-harness/result.json`; input SHA-256 `b93b7ddd2c7bc0a2894328169619b8a7f48285ca6a531880c1b604c8726650aa`.

[^02c]: `b02-censored-price`; `e01-mechanism-pilots/items/censored-price/result.json`; input SHA-256 `33f54e81f3372088e9eced52c0174f310ac348af98ef5727b73785c088679194`.

[^02v]: `b02-value-policy`; `e01-mechanism-pilots/items/value-policy/result.json`; input SHA-256 `f68b806e83fd1e392242efc2abcf09b9271b74e93fcf9eba699efcca11dd6e4e`.

[^02n]: `b02-normalizer-state`; `e01-mechanism-pilots/items/normalizer-state/result.json`; input SHA-256 `8b70ee003e3aefb48e4562dd909d60c0a1dc6dd6a67fb128709371d2d4851c79`.

[^02e]: `b02-event-hazard`; `e01-mechanism-pilots/items/event-hazard/result.json`; input SHA-256 `123d1f05288c3ac72fe986f64e9c197a40c2c2df306a761c84ddf38d4c4f6091`.

[^03l]: `b03-liquidity-shadow`; `e01-mechanism-designs/items/liquidity-shadow/result.json`; input SHA-256 `19605acdf177f96d708cab1fa7cec6a0b909a5024bea25ffe1c8b22aca2e912e`.

[^03p]: `b03-policy-distillation`; `e01-mechanism-designs/items/policy-distillation/result.json`; input SHA-256 `dcfeb5f5837ef2707b3949b09bd866f8dd1f8cf5e3c2ec71e86488d66d709e7b`.

[^04a]: `b04-fit-a`; `e02-board-fit/items/fit-a/result.json`; input SHA-256 `6851aa886b12c4f7e0cfff202ce20f8f8de17cf364ccea3beccfaffbb4916fbc`.

[^04b]: `b04-fit-b`; `e02-board-fit/items/fit-b/result.json`; input SHA-256 `d7c0d000c3424a6e8fa53a4d6b93e96ced19a470cde228c5fd0c36111d60849d`.

[^05a]: `b05-a`; `e02-search-share/items/a/result.json`; input SHA-256 `03da9f9e554abdedaedfad0bfaec737591c0e35ce2d2b12830d061a9c7f44ce1`.

[^05b]: `b05-b`; `e02-search-share/items/b/result.json`; input SHA-256 `00f46f734ceb30bb02e5dcc8359c3647b73b5fcc66a0fcde3b37002af60770eb`.

[^06a]: `b06-a`; `e02-search-arbside/items/a/result.json`; input SHA-256 `8b5397430b4b82bd6d9082181afdd94ed971118769724db522d31f5ba0726827`.

[^06b]: `b06-b`; `e02-search-arbside/items/b/result.json`; input SHA-256 `23fe325164208469130356d04ac01e63984b8c5342612f11fea8bbc5f9eb72a6`.

[^07a]: `b07-a`; `e02-search-regime/items/a/result.json`; input SHA-256 `f7304d7633252c361d816df42b511b28de3efa6ddfa395b9dda728faa89a48db`.

[^07b]: `b07-b`; `e02-search-regime/items/b/result.json`; input SHA-256 `7ff902af5003074be9b18ef2c875198ccd2892f95a230f6aa8afda396841bd75`.

[^08a]: `b08-a`; `e03-confirm-v10/items/a/result.json`; input SHA-256 `33b2eaa4d851f9fd0efd0d49f371568e482329f40b1f630f42f0c770222b9ea3`.

[^08b]: `b08-b`; `e03-confirm-v10/items/b/result.json`; input SHA-256 `81c23f1d4bc0f81ae375182d733f9a78cb06924e6ca33f262e7ebcb913be21a4`.

[^09a]: `b09-a`; `e03-v10-floor/items/a/result.json`; input SHA-256 `9e843df2dbb4045dc377215155c2c3954ee6b23dc0c700c9bd06d0175621f8ca`.

[^09b]: `b09-b`; `e03-v10-floor/items/b/result.json`; input SHA-256 `f6be9de4933d0cd9a95c03ebad937f3a1f643a0fdf937ab50b321651f0bbbf7f`.

[^10a]: `b10-a`; `e03-v10-blend/items/a/result.json`; input SHA-256 `bf8035ded3fcd9819c87f0cd7199e9c396714f2337fcd86524323adb597f19bc`.

[^10b]: `b10-b`; `e03-v10-blend/items/b/result.json`; input SHA-256 `ef86b523a51644542fdbac2470e6d3205f744656230a0771a66a88af54bb8361`.

[^11a]: `b11-a`; `e03-v10-retail/items/a/result.json`; input SHA-256 `39f1cf65ced0e274838409dd423cbb2a5006c69d9b7cb8655d3f0029ab54b0dc`.

[^11b]: `b11-b`; `e03-v10-retail/items/b/result.json`; input SHA-256 `067627a024da160124a2029a20370025d13f3ef8ea4c8d7a93c2bcaa70d2f4b8`.

[^12a]: `b12-a`; `e03-restable/items/a/result.json`; input SHA-256 `b6454450c9e7951d940bd7066a1826cc6073ae98fbd560e47d9e896630cbb7b2`.

[^12b]: `b12-b`; `e03-restable/items/b/result.json`; input SHA-256 `9e413c23c0cf996711bab9ae02c8767e001f708d052fdefb5264d91a703eb508`.