**Root merge 2 of 6 — all 25 results.**

Q2, the compressed Gaussian+qm contract, reports **+3.044431807634 over IB, SE 0.032504145359, on 2,048 paired engine seeds**. This exceeds +2.7 on the measured sample but falls short of the required 4,000 seeds. Plain Gaussian P2 reports **+2.206350316657, SE 0.049738846071, on 1,024 different seeds**. These contracts/samples must stay separate. Gaussian branch blending reports **+0.319286166708, SE 0.010933121500**, on 2,048 replica holdout seeds versus Gaussian+qm, crossing +0.3 on that baseline. Its PB-1 port has no engine result and an unresolved package hash. [S04–S09; S12; S17–S25.]

Objective snapshot: IB 522.00 (20th), tenth 524.66 (+2.7 rounded), first 528.40 (+6.4). This goal enters nothing; no current-board lookup, entry or first-place result is claimed. The [pinned solved README](https://raw.githubusercontent.com/gtmcknight/amm-sim/e65314f88de945f6d055bfc2aab01c482ab6570e/research/solved/README.md) was read for context. No source experiment, setup/build, validation or engine run was repeated here.

All 25 inputs were present/nonempty and parsed. Source labels below identify their local names and original result path reconstruction; each data block and its note are attributed to that section's source. The requester flags one unsupported/unreadable cited source in each of **S04–S09, S12, S14, S17, S18, S20–S24**. Exact failed passages/reasons are unavailable. Retain these as worker reports with reduced evidentiary weight. S19 is the Q2 block without that warning. Arithmetic consistency does not repair source support.

Local checks inspected decoded payloads, hashes, engine score rows and packed oracle differences; no payload was executed. Raw code, callback logs, redundant per-seed vectors and exact timing lists are explicitly omitted; findings, sufficient statistics, worst cases and timing summaries remain. S15/S16 duplicate the same refinement measurements: count once. Columnar JSON uses named columns and ordered rows without changing values.

**Engine pools, derived from supplied sufficient statistics.**

P2 covers 12000000..12001023; Q2 covers 12001024..12003071, with separate VanillaStrategy arms; engine `6b6921f8d463314aa7c96234e1be25a25245f704`, run_batch(...,8), MatchRunner configs. Pool N=Σn, S=Σsum, Q=Σsumsq, mean=S/N, SE=sqrt((Q−S²/N)/(N−1)/N). Never average SEs or pool different contracts. [S04–S07; S17–S24.]

P2 [S04–S07]

```json
{"n":1024,"mean_diff":2.206350316657449,"se":0.04973884607103439,"sum":2259.302724257228,"sumsq":7576.407493516774,"worst5":[[12000174,-3.403795],[12000542,-2.899929968635149],[12000116,-2.48608],[12000272,-1.6806200155142506],[12000740,-1.5819277658400779]]}
```

Q2 [S17–S24]

```json
{"n":2048,"mean_diff":3.044431807633728,"se":0.0325041453591124,"sum":6234.9963420338745,"sumsq":23411.221252637326,"worst5":[[12001701,-2.224937371226247],[12002425,-1.232297],[12002099,-1.2071382251384648],[12001837,-0.4967746590483557],[12002445,-0.493482]]}
```

Q2 unflagged subset [S19]

```json
{"n":256,"mean_diff":2.934757873393122,"se":0.08265973089873915,"sum":751.2980155886393,"sumsq":2650.9119255211963,"worst5":[[12001701,-2.224937371226247],[12001746,0.2663324416890873],[12001575,0.3588694661850127],[12001542,0.5201708140612027],[12001561,0.5992509943984032]]}
```

Q2 flagged subset [S17,S18,S20–S24]

```json
{"n":1792,"mean_diff":3.0600995125252424,"se":0.03521319782015839,"sum":5483.698326445235,"sumsq":20760.30932711613,"worst5":[[12002425,-1.232297],[12002099,-1.2071382251384648],[12001837,-0.4967746590483557],[12002445,-0.493482],[12002511,-0.479787]]}
```

Maximum reported afterSwap gas: P2 193,866 / IB 234,965; Q2 204,917 / IB 237,142. Every block reports validation pass and 0 OOG/reverts; gas means cannot be pooled without callback counts. Q2 engine gas is materially above its pyrevm report (155,745 peak / 59,801 mean), still under 245,000. Compression parity is against each compressed Python twin, not original full qm. Q2 still needs 1,952 paired observations; 12003071 is this root's last represented seed, not proof that later roots have not used higher seeds. [S04–S09; S17–S24.]

Row totals reproduce paired statistics within rounding: S04 six-decimal scores; S17 rounded rows/unrounded totals differ by ~7.18e-6 in paired sum; S22 rounded paired differences/full-precision arm sums differ by ~3.44e-6. S18 repeats a two-seed probe, violating literal never-reuse despite unique reported rows; S21's recovery segmentation and S24's overlapping processes are detailed below. [S04; S17; S18; S21; S22; S24; local audit.]

**Cross-result interpretation.**

Gaussian blend is versus ref/gauss_qm.json, not the requested lattice reference. Lattice/rival blend +0.2231 is below +0.3. PB's +0.26437374328084423 is combined compression/blend versus full Gaussian+qm on 1,024 tuning seeds; subtracting Q2's same-range −0.03770027617478483 compression mean gives a derived +0.30207401945562906 versus Q2, without a contrast SE (covariance unavailable). No engine gain transfers from Q2 to PB. [S09; S11; S12; S25.]

Small first fills, particularly dt=2, explain most misreads. Amount/size features repeat existing filter evidence; hindsight fixes past labels but not the forward price estimate. S13's 13.02% is a retrospective label; first-callback error stays 16.11%, and its holdout omits 1,792 seeds. Oracle gains are hidden-information diagnostics, not attainable policy gains. S10's lattice representations fail gas; finer lattices do not improve score. [S01–S03; S10–S16.]

**Integrity caveats.** S02's declared hash matches compressed zlib bytes; its uncompressed tar is 26285316c55645f614cbcc1ab7be9ce228b1ce79c7755b821376e3c44015498f. S25's declared hash matches neither tar, compressed bytes nor any member; actual tar is 94fd852fccba78f63dab558be5f7e6267f49f1ba9bb27babdd3244ccb241d9c6, Strategy.sol 5123bb39594add4ad9a32f376ff7aa89c13428b82ea324ae4d80c474e3162d95. Its gb.py matches S12 at 20449ab0611b3467cb4db8ef06f41f2a127afb3d90bbace53b40a3045179329b. Other declared code hashes match uncompressed tars. These are identity checks, not behavioral proof. [S02; S12; S25; local hashing.]

**Consolidated engine measurements [S04–S07; S17–S24].**

Each row has n=256 per arm and paired n=256; engine="ok". All report both contracts validated, gas OOG=0/reverts=0 in each arm. P2 explicitly has valid=true, blocked=null; Q2 has validate={ib:"pass",port:"pass"}. ib_sum_check=null below means no field reported, not a failed check. Scores are edges against VanillaStrategy, gas is Submission afterSwap. This table preserves each source's unrounded/rounded sufficient statistics as provided.

```json
{"columns":["source","key","seed_start","seed_end","ib_sum","ib_sumsq","port_sum","port_sumsq","mean_diff","se","paired_sum","paired_sumsq","worst5","ib_peak_gas","ib_mean_gas","port_peak_gas","port_mean_gas","ib_sum_check"],"rows":[["S04","eng-12000000",12000000,12000255,134918.570353,73255977.56335746,135497.925331,73910779.34196405,2.263105382812,0.106348929804,579.354978,2049.46436240946,[[12000174,-3.403795],[12000116,-2.48608],[12000187,-1.562827],[12000067,-1.060272],[12000005,-1.015302]],234958,135281.010864,188851,82865.713432,"equal"],["S05","eng-12000256",12000256,12000511,136343.58940512265,74622234.70430395,136876.4184548521,75223961.44788724,2.081363475505698,0.08537029740600702,532.8290497294587,1584.7772864928336,[[12000272,-1.6806200155142506],[12000505,-1.1241006833525944],[12000264,-0.9273610959264715],[12000268,-0.694046149702217],[12000283,-0.5077152898528539]],234775,135117.32020271046,184436,82784.62251298339,"none"],["S06","eng-12000512",12000512,12000767,136071.26244280543,74524616.54432805,136617.47677273888,75149862.113184,2.133649726302621,0.10225952448218355,546.2143299334709,1848.0636910383193,[[12000542,-2.899929968635149],[12000740,-1.5819277658400779],[12000684,-1.571581031850826],[12000760,-1.4451567776682168],[12000573,-1.1128658452150262]],234743,135196.2546312131,193866,82824.29280845154,"equal"],["S07","eng-12000768",12000768,12001023,136903.69102616355,75420076.10964641,137504.59539275785,76112896.67817459,2.347282682008978,0.10233260933794379,600.9043665942984,2094.1021535761606,[[12000859,-1.3972544756191496],[12001011,-1.3577902034683689],[12000963,-1.0827266698434528],[12000846,-0.97372890604845],[12000888,-0.9653602511882013]],234965,135191.4007762263,179316,82816.6028535272,"equal"],["S17","eng2-12001024",12001024,12001279,134535.04601420724,72740247.06012902,135310.3191583854,73593461.32021116,3.02841071944588,0.08965159593982877,775.2731441781452,2872.5275372971732,[[12001072,-0.017584],[12001181,0.21166],[12001064,0.495607],[12001080,0.5631],[12001224,0.591667]],237040,135258.6,201422,92741.8,null],["S18","eng2-12001280",12001280,12001535,133290.06798785817,71243058.50380743,134075.0422040098,72092407.91462126,3.0663055318423362,0.08778990505435644,784.9742161516381,2910.0881431340563,[[12001330,-0.040143353500013745],[12001355,0.05139744335144769],[12001426,0.12384743958750732],[12001488,0.3779462163585663],[12001470,0.5565102251235885]],234825,135254.3,201771,92746.6,null],["S19","eng2-12001536",12001536,12001791,133927.0009801248,71938433.40272224,134678.29899571347,72751191.89926365,2.934757873393122,0.08265973089873915,751.2980155886393,2650.9119255211963,[[12001701,-2.224937371226247],[12001746,0.2663324416890873],[12001575,0.3588694661850127],[12001542,0.5201708140612027],[12001561,0.5992509943984032]],237142,135324.5,196089,92768.5,null],["S20","eng2-12001792",12001792,12002047,136523.1679254553,74948078.74113058,137340.006004297,75862562.18102899,3.190773745475367,0.10162575813560733,816.8380788416939,3280.543935374675,[[12001837,-0.4967746590483557],[12001888,0.16408344023579957],[12001866,0.25521026484085496],[12001982,0.44591355604143246],[12002033,0.4752781945235256]],234865,135097.2444153594,192746,92641.96188351659,null],["S21","eng2-12002048",12002048,12002303,136531.76406970056,75071834.14208785,137304.92030932324,75938206.65925084,3.020141561026157,0.09433087857921371,773.1562396226961,2915.923273033404,[[12002099,-1.2071382251384648],[12002049,0.2622814739774526],[12002129,0.3068486516315261],[12002204,0.43644596282689463],[12002074,0.6497275241475791]],234958,135152.0948740215,200090,92636.30391585782,null],["S22","eng2-12002304",12002304,12002559,134966.06795031478,72992525.49347362,135709.79265987553,73811446.5332479,2.9051746601562476,0.09205500269678532,743.7247129999994,2713.8409738234323,[[12002425,-1.232297],[12002445,-0.493482],[12002511,-0.479787],[12002538,-0.425494],[12002549,-0.298269]],234849,135172.4,204917,92697.6,null],["S23","eng2-12002560",12002560,12002815,134426.2506365197,72597097.98044562,135214.27164723852,73459608.96681729,3.078207073120458,0.09473943358590119,788.0210107188373,3011.6164237911166,[[12002686,-0.3016351237820345],[12002789,-0.21948262528189844],[12002685,-0.15768852965493352],[12002705,-0.0898765632501295],[12002673,-0.02408907445828845]],234724,135273.71815616175,200493,92735.23216075297,null],["S24","eng2-12002816",12002816,12003071,133720.1691743112,72053658.54587439,134521.88009824342,72934912.554013,3.1316832966102535,0.09137635049980958,801.7109239322249,3055.769040662272,[[12003053,-0.32324448579453247],[12002884,-0.2718951660700668],[12002989,-0.0040246230287834805],[12002923,0.3112862524220077],[12002908,0.40647341886676713]],237142,135290.1407339996,189776,92763.5009344473,null]]}
```

**Per-result evidence.** Complete/valid/validation/gas fields are the worker's reports. Notes are summarized; code hashes are in the register. Engine rows are replaced by sufficient statistics, timings by summaries. S16's duplicated fields reference S15 explicitly.

<a id="s01"></a>
**S01 — b09-misread-amounts** — no requester source warning.

All 512 training seeds 60000..60511 and 256 diagnostic test seeds 61000..61255; 10000..12047 untouched. Numpy-only depth-6 CART, 64 training-quantile bins, minimum leaf 1,000, threshold .5; scalar direction chosen by training AUC, no holdout tuning. Trade-weighted arb AUC. Errors favor retail→arb (172,957) over arb→retail (115,400). Misread median displacement 6.95 versus 11.94 bps. Exact amounts mostly repeat spot movement; router equality needs an interior split, all-own fills merely bound the rival. Rival-input ratio training AUC .507. Added features cut errors by only .101 percentage points; no closed-loop fee-policy gain measured. Retain both hypotheses for ambiguous small fills. Breakdown quartiles are test own-fill input-notional; archive contains definitions/counts/cutoffs/trees/replay and dataset manifest.

```json
{"key":"misread-amounts","family":"amounts","rows":{"train":3567655,"test":1776190},"lattice_misread_test":0.16234580759941222,"auc":{"filter_only":0.9226584279206888,"mine_only":0.8959820252998588,"filter_plus_mine":0.9234240714525783},"misread_test":{"filter_only":0.16103401100107534,"filter_plus_mine":0.16002792494046245},"by_dt":{"1":0.1516549618892447,"2":0.205838512700964,"3-5":0.18173892129156394,"6+":0.15517049872264801},"by_size_quartile":[0.24656118257485676,0.20618988530493393,0.14243312081829176,0.05419909559326919],"top_feature":{"name":"neg_arb_price_innovation_abs_bps","definition":"Negative of: Absolute value of 10000*ln(p_arb/exp(previous price log-mean)); p_arb=s*(1-fb) for pool buying X, s/(1-fa) otherwise.","auc_alone":0.8739056596699507}}
```

<a id="s02"></a>
**S02 — b09-misread-hindsight** — no requester source warning.

All 60000..60511 training and 61000..61255 diagnostic test seeds; label reconstruction matches latt.py on every row (archive: 0 mismatches / 5,343,845 rows). Standardized logistic features plus squares, L2=1e-3, Newton. Fork arb/retail lattices and score the next three own callbacks. Future observations improve retrospective labels, not first-callback decisions. Hard relabel worsens current RMS; posterior-weighted branch mixture nearly equals the ordinary lattice. Soft three-callback revision removes 44% of the past estimate's true-flag error gap, not the forward estimate's gap. Monotone-rival router bound is false and worsens classification. Rival clip/sync and phase are unrevised state; revision of those was then untested here (later S13 tests it). Two extra 121-cell lattices for three callbacks are not viable in-contract; Gaussian form needs two (m,v) pairs. Small dt≥2 first trades / informative next quotes are proposed levers. Full medians, AUCs and all one/two/three-callback error comparisons follow.

```json
{"key":"misread-hindsight","family":"hindsight","rows":{"train":3567655,"test":1776190},"lattice_misread_test":0.16235,"auc":{"filter_only":0.92007,"mine_only":0.88652,"filter_plus_mine":0.95393},"misread_test":{"filter_only":0.16403,"filter_plus_mine":0.11772},"by_dt":{"1":0.15165,"2":0.20584,"3-5":0.18174,"6+":0.15517},"by_size_quartile":[0.24655,0.20596,0.14248,0.0544],"top_feature":{"name":"llr3","definition":"Fork the lattice at the first trade into an arb branch (pin at the arb reading) and a retail branch (cut to the no-arb interval); run both through the next 3 own callbacks with the quotes in force (latt.py updates); llr3 = sum of log predictive-likelihood differences, arb minus retail.","auc_alone":0.8627},"detail":{"seeds":{"train":"replica 60000..60511","test":"replica 61000..61255"},"label_check":"latt.py label recomputed on every row: 0 mismatches","lattice_misread_train":0.16209,"auc_lattice_pi_alone":0.92124,"models":{"columns":["label","auc_test","misread_test","misread_train"],"rows":[["filter_only",0.92007,0.16403,0.16372],["mine_only",0.88652,0.18717,0.18672],["filter_plus_mine",0.95393,0.11772,0.11803]]},"model_form":"logistic regression, standardized features + squares, L2 1e-3, Newton; AUC and misread for the true arb flag","revision_bayes_test":{"columns":["label","misread_llr","misread_llr_plus_bound","auc_llr"],"rows":[["1",0.14886,0.16676,0.93241],["2",0.13109,0.15371,0.94472],["3",0.11972,0.1428,0.95218]]},"price_error_bps_test":{"1":{"columns":["label","rms_main","rms_revised","rms_true_flag","rms_main_misread","rms_revised_misread","rms_true_flag_misread","share_of_flag_gap_removed","rms_soft_revised","rms_soft_revised_misread","share_of_flag_gap_removed_soft"],"rows":[["at_first_trade",13.793,14.778,8.819,24.651,28.999,9.77,-0.2502,13.154,22.989,0.1532],["at_kth_later_callback",14.043,15.397,10.423,23.192,28.68,null,-0.4501,14.043,23.194,0.0001]]},"2":{"columns":["label","rms_main","rms_revised","rms_true_flag","rms_main_misread","rms_revised_misread","rms_true_flag_misread","share_of_flag_gap_removed","rms_soft_revised","rms_soft_revised_misread","share_of_flag_gap_removed_soft"],"rows":[["at_first_trade",13.793,13.815,8.819,24.651,25.111,9.77,-0.0054,12.408,20.988,0.3227],["at_kth_later_callback",13.957,14.87,11.567,20.945,24.085,null,-0.4317,13.982,20.947,-0.0116]]},"3":{"columns":["label","rms_main","rms_revised","rms_true_flag","rms_main_misread","rms_revised_misread","rms_true_flag_misread","share_of_flag_gap_removed","rms_soft_revised","rms_soft_revised_misread","share_of_flag_gap_removed_soft"],"rows":[["at_first_trade",13.793,13.123,8.819,24.651,22.956,9.77,0.1603,11.846,19.591,0.4438],["at_kth_later_callback",13.867,14.462,12.299,19.117,21.053,null,-0.4107,13.887,19.118,-0.0134]]},"0":{"rms_soft_pi0_at_first_trade":13.793}},"by_dt_share":{"1":0.75157,"2":0.14466,"3-5":0.09363,"6+":0.01014},"by_dt_revised3":{"1":0.11373,"2":0.14337,"3-5":0.13162,"6+":0.11635},"size_quartile_edges_Y":[2.5845,5.4245,10.0475],"by_size_quartile_revised3":[0.19199,0.15297,0.0989,0.03502],"by_size_quartile_arb_share":[0.58296,0.53754,0.6015,0.44613],"misread_split":{"arb_read_as_retail":0.06497,"retail_read_as_arb":0.09738,"arb_share":0.54203},"single_feature_auc_mine":{"llr3":0.8627,"llr2":0.82818,"llr1":0.74327,"absza1":0.67314,"side1":0.64934,"side2":0.6492,"side3":0.63156,"absza2":0.62125,"absza3":0.60376,"nsame":0.58604,"kind1":0.57999,"dtn1":0.57093,"sla1":0.56101,"abszr1":0.54618,"abszr3":0.54347,"dtn2":0.54025,"abszr2":0.52975,"sla2":0.52813,"dtn3":0.52298,"kind2":0.51377,"logdn2":0.51284,"logdn3":0.51216,"sla3":0.50761,"logdn1":0.5053,"kind3":0.50484,"blr3":0.50458,"zr3":0.50068,"za2":0.50053,"zr2":0.50037,"za3":0.50033,"zr1":0.50021,"za1":0.50016},"single_feature_auc_filter":{"pi_logit":0.92124,"off":0.87081,"other":0.85705,"sd":0.71815,"logdt":0.65878,"d":0.56009,"logd":0.56009,"logsize":0.5597,"adv":0.54125,"isbuy":0.50069},"auc_predict_misread":{"abs_pi_logit":0.8144,"filter_plus_mine_disagrees":0.90494},"median_misread_vs_correct":{"pi_logit":[0.248,0.715],"off":[12.764,7.03],"sd":[19.212,16.019],"d":[6.951,11.941],"dt":[1.0,1.0],"adv":[1.749,3.093],"size":[3.479,5.971],"llr3":[0.0,0.0],"nsame":[0.0,0.0]},"breakdowns":"by_dt and by_size_quartile: lattice misread rate on the test set; size = order Y amount on own pool","gas":"llr needs two extra 121-cell lattices for 3 callbacks: not viable in a contract; a Gaussian form needs two (m,v) pairs"}}
```

<a id="s03"></a>
**S03 — b09-misread-size** — no requester source warning.

Full universe 3,567,655 train rows (60000..60511), 1,776,190 test (61000..61255); fit40,000 from8% sample, test12%=213,080. Ridge logistic IRLS. Blind lattice+qm reproduces reference seeds10000/10001 edge/first/misread counts. Added size/timing does not improve; 2D logit(pi)×feature tables flip no majority. Size is already in L_ret. Displacement <4.2bps error24.8%, >23bps4.5%; dt2 arb share26%. Keep d<~8bps first trades ambiguous; seek informative phase quotes rather than redundant features. Full fixes/breaks, features and breakdowns follow.

```json
{"key":"misread-size","family":"size","rows":{"train":40000,"test":213080},"lattice_misread_test":0.1623,"auc":{"filter_only":0.922,"mine_only":0.8711,"filter_plus_mine":0.922},"misread_test":{"filter_only":0.1624,"filter_plus_mine":0.1621},"by_dt":{"1":0.1512,"2":0.2091,"3-5":0.1813,"6+":0.1592},"by_size_quartile":[0.2465,0.2082,0.1417,0.0544],"top_feature":{"name":"fee_cross","definition":"Fee in force on the side the first trade crossed, -ln(1-fb) on a pool buy or -ln(1-fa) on a sell, in bps, set by our last quote before the trade. A higher fee means arb (the arb crosses wide quotes). Next best: d_over_fee = |ln(s/sp)|/fee_cross, AUC 0.773.","auc_alone":0.8256},"extra":{"lattice_pi_auc_test":0.9213,"lattice_misread_test_sample":0.1625,"mine_only_misread":0.2054,"by_dt_share":{"1":0.7508,"2":0.1443,"3-5":0.0946,"6+":0.0103},"by_dt_arb_share":{"1":0.6469,"2":0.2583,"3-5":0.1889,"6+":0.1542},"size_quartile_edges_Y":[2.5708,5.3926,10.0732],"by_size_quartile_arb_share":[0.5824,0.5389,0.6009,0.4475],"d_quintile_edges_bps":[4.163,8.2592,13.7386,23.0505],"by_d_quintile_misread":[0.2476,0.2241,0.1787,0.1197,0.0447],"single_auc_mine":{"fee_cross":0.8256,"fee_other":0.8035,"ln_d_over_fee":0.7732,"d_over_fee":0.7731,"ln_d_x_ln_dt":0.666,"ln_since_pin":0.6617,"ln_dt":0.6604,"dt_eq1":0.658,"dt_eq2":0.5826,"prev_first_read_arb":0.5806,"ln_d":0.5596,"zret":0.5591,"ln_sizeY":0.5591,"ln_d_sq":0.5551,"zret_sq":0.5368,"prev_cnt":0.5274,"isbuy":0.5014,"prev_cnt0":0.5},"misread_types":{"arb_read_retail":0.065,"retail_read_arb":0.0975},"combo_fixes_breaks_of_misreads":[1677,1594,34623],"rows_all":{"train":3567655,"test":1776190},"code_ext":"code_zb64: base64(zlib(tar)) of mkdata.py, analyze.py, final.json; run python3 -I mkdata.py <solved> <seed0> <n> <out> <frac>, then analyze.py <train,...> <test,...> <out> 40000"}}
```

<a id="s04"></a>
**S04 — b10-eng-12000000** — requester source warning applies.

Both decoded artifacts validated; four64-seed blocks per arm, gas reset each block. Six-decimal scores/statistics. Gas afterSwap means successful calls; errors all Submission callbacks. P2 below+2.7 on this range.

Statistics: consolidated engine table above, row S04.

Reported run_batch time summary: 8 segments, min 8.56 s, max 14.402 s, total 89.378 s. Different seats/overlap make these unsuitable for cross-seat throughput comparisons.

<a id="s05"></a>
**S05 — b10-eng-12000256** — requester source warning applies.

Setup/compilation/validation succeeded unchanged. Full-precision scores. Config draw order follows MatchRunner. Gas call-weighted, errors all callbacks. ib_sum_check="none", not an equality claim. No outside seeds.

Statistics: consolidated engine table above, row S05.

Reported run_batch time summary: 8 segments, min 80.299972 s, max 159.580527 s, total 950.759951 s. Different seats/overlap make these unsuitable for cross-seat throughput comparisons.

<a id="s06"></a>
**S06 — b10-eng-12000512** — requester source warning applies.

Both validated unchanged. Full precision, IB sum check on individually rounded six-decimal scores. Startup attempt failed before run_batch during compilation, consumed no seeds. Gas resets each block, weighted successful-call mean.

Statistics: consolidated engine table above, row S06.

Reported run_batch time summary: 8 segments, min 16.491312 s, max 48.459799 s, total 257.775184 s. Different seats/overlap make these unsuitable for cross-seat throughput comparisons.

<a id="s07"></a>
**S07 — b10-eng-12000768** — requester source warning applies.

Both decoded hashes verified; unchanged pinned compiler/validation. RNG default_rng(seed), size/rate/sigma draws in order. Full precision, rounded-row IB check. Other halt/tx_error/bad_return zero. No board entry.

Statistics: consolidated engine table above, row S07.

Reported run_batch time summary: 8 segments, min 29.70708 s, max 73.828235 s, total 418.628622 s. Different seats/overlap make these unsuitable for cross-seat throughput comparisons.

<a id="s08"></a>
**S08 — b11-qm-1** — requester source warning applies.

P2 math unchanged; s-only cost−.036, c-only−.54 versus full; three/four sd levels ~−.30/−.18 on this sample. Compression/parity details below. Init plain pol_mid as real.py. Tune60000..60255 only, no holdout/engine seeds. Repeated words triggered copied data/validator failure; position/word XOR makes words unique and passes.697 runtime bytes spare. pyrevm includes21,000 intrinsic(getName21,521); Foundry not rerun. Archived stress report:16,384calls,0OOG/reverts,mean74,014,no failures. Proposed8bps sixth sd/finer|w|≤50knots/9-bit encoding, unmeasured.

```json
{"key":"qm-1","tables":"qm","knots":{"w_knots":[-800,-400,-250,-200,-150,-125,-100,-90,-80,-70,-60,-50,-40,-30,-20,-10,0,10,20,30,40,50,60,70,80,90,100,125,150,200,250,400,800],"sd_levels_bps":[0,4,12,20,34],"sd_level_indices":[0,1,3,5,7],"rival_modes":"s only (A0c phase 0; A1s,A2s used for both synced and clipped)","rival_cols_phase12_W2L_indices":[0,1,2,3,4,5,6,7,8,9,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,39,40,41,42,43,44,45,46,47,48],"rival_cols_phase0_indices":[0,2,4,6,8,10,12],"mirror":"bid(w,R)=ask(-w,-R)","residual_format":"ask residual a-max(w,0) rounded to 1 bps, unsigned byte = residual+60, XOR-scrambled per position/word, 32 cells per PUSH32 word, 12705 cells in 398 words"},"knot_cost_paired":{"n":256,"mean_diff":-0.11017381459311437,"se":0.025161103115988247,"sum":-28.20449653583728,"sumsq":44.4349318337058,"seeds":"60000..60255","worst5":[[60218,-1.3357098298187111],[60199,-1.1955930748700325],[60137,-1.0684920017023387],[60065,-1.0462858651748093],[60104,-1.041194573390385]]},"parity":{"callbacks":214731,"max_fee_diff_bps":8.5021e-10,"over_0_01_bps":0,"seeds":"10000..10019"},"gas":{"peak_afterswap":165011,"mean_afterswap":64980,"oog":0,"reverts":0},"validate":"pass","stress":{"sequences":64,"calls":16384,"reverts":0,"peak_gas":163949},"bytes_runtime":23879,"complete":true,"engine_seeds_consumed":0}
```

<a id="s09"></a>
**S09 — b11-qm-2** — requester source warning applies.

P2 Gaussian math unchanged. Retains s/c via synced slot7; phase2 shares phase1, measured +.015 SE .011 on 1,024 tuning seeds. One-bps 8-bit residual quantization -.002 SE .008. Costs on 60000..61023: four sd levels -.038, three -.098, s-only -.021, P2 71 knots -.001. Full chosen cost -.038 SE .012 (256-seed estimate -.063 SE .027). 33 w knots, seven band/41 rival columns, mirror bid; source mirror exact except 1c max .71 bps asymmetry (archive README). Init 28.276 bps from plain pol_mid, not qm. Runtime spare 847 bytes; 43-rival-column b3 was 115 bytes over limit. pyrevm gas includes intrinsic (getName 21,521). No engine seeds consumed here. About 15 spare words cannot fit ~92 words for sixth sd level; denser encoding/drop columns proposed. Archive: solc .8.24/viaIR/Paris/optimizer200/no assembly, slots0..7.

```json
{"tables":"qm","knots":{"w_knots_bps":[-800,-400,-250,-200,-150,-120,-100,-90,-80,-70,-60,-50,-40,-30,-20,-10,0,10,20,30,40,50,60,70,80,90,100,120,150,200,250,400,800],"sd_levels_bps":[0,4,12,20,34],"sd_level_indices":[0,1,3,5,7],"phase0_band_cols_bps":[-30,-20,-10,0,10,20,30],"rival_cols_W2L_indices":[0,2,4,6,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,42,44,46,48],"rival_modes":"s and c both kept; phases 1 and 2 share one s/c pair (phase 2 := phase 1)","residual_format":"ask only, bid by mirror bid(w,R)=ask(-w,-R); unsigned 8-bit u, residual a-max(w,0)=(u-64)*1 bps; 14685 cells, 459 PUSH32 words in a binary tree"},"knot_cost_paired":{"n":256,"mean_diff":-0.06255076500684909,"se":0.026639635332316065,"sum":-16.012995841753366,"sumsq":47.32889387925289},"knot_cost_paired_1024":{"n":1024,"mean_diff":-0.03770027617478483,"se":0.01244791052979014,"sum":-38.605082802979666,"sumsq":163.77410390235286},"parity":{"callbacks":214310,"max_fee_diff_bps":6.79479e-09,"over_0_01_bps":0},"gas":{"peak_afterswap":155745,"mean_afterswap":59801,"oog":0,"reverts":0},"validate":"pass","stress":{"sequences":64,"reverts":0,"peak_gas":158473},"bytes_runtime":23729,"complete":true}
```

<a id="s10"></a>
**S10 — b12-latt-gas-1** — no requester source warning.

Filter only, no fee tables. Cold-storage pyrevm tx includes21,000 intrinsic+six-word calldata.13 slots fit, convolution does not. Cap12 versus lattice cap30:11 callbacks with dt−1>12 plus follow-ons cause17.9bps max error despite99.8% within.01. Cap30 improves parity, raises gas. Looped memory multiply-add~160gas(b2.sol); symmetric unrolling halves total, ~23support×33taps still>100k. dt1 typical210k/peak518k; silent step+75–100k; dt0peak25.1k. Sentinel avoids22.1k zero→nonzero SSTORE. Variants reduce gas2–4×, break parity; none≤170k. Need≤~120k filter plus tables. Packed4×64-bit lanes, multistep operators, quasi-stationary shape/Gaussian are untested alternatives. First three lattice edges match gates.py, not a full merger rerun.

```json
{"key":"latt-gas-1","slots_used":13,"slots_detail":"2 state slots (centre, ln own spot, timestamp, support bounds; mean, sd) + 11 mass slots (12 cells x 21 bits + sentinel bit each); only the slots covering the support are read and written","bits_per_cell":21,"kernel_width_cells":33,"cell_bps":2.5,"cells":121,"silent_cap":12,"gas":{"peak":4545807,"mean":189976,"by_dt":{"1":517882,"2":822142,"3-5":1624892,"6+":4545807},"dt0_peak":25103,"callbacks":53268,"first_trades":34603},"variants":{"columns":["name","desc","bits","kernel","peak_gas","mean_gas","by_dt","max_err_mean_bps","max_err_sd_bps","max_err_pi","share_mean_err_le_0_01bps","bytes_runtime","reverts"],"rows":[["base","2.5bps x121, exact kernel, silent cap 12 (brief)",21,33,4545807,189976,{"1":517882,"2":822142,"3-5":1624892,"6+":4545807},17.9149,9.8588,0.180818,0.9979,6370,0],["cap30","2.5bps x121, exact kernel, silent cap 30 (latt.py own cap; parity check)",21,33,9040497,190312,{"1":517882,"2":822142,"3-5":1624892,"6+":9040497},0.0234,0.0892,0.000357,0.9999,6370,0],["b16","2.5bps x121, exact kernel, cap 12, 16-bit cells",16,33,4467289,187870,{"1":510577,"2":817570,"3-5":1561801,"6+":4467289},17.917,9.8603,0.180838,0.9884,6371,0],["trunc8","2.5bps x121, kernel truncated to +-8 cells (2.1 sigma), cap 12",21,17,2607348,142034,{"1":370344,"2":550609,"3-5":976037,"6+":2607348},90.7513,42.1998,0.99831,0.0172,6002,0],["cap1","2.5bps x121, exact kernel, coarse silent update: cap 1",21,33,819795,175269,{"1":507944,"2":819795,"3-5":729452,"6+":733942},51.4099,25.793,0.712627,0.4944,6370,0],["c5","5bps x61, kernel +-8, cap 12",21,17,1487304,96915,{"1":227044,"2":323571,"3-5":569246,"6+":1487304},24.4182,13.5131,0.591935,0.246,6003,0],["c5cap4","5bps x61, kernel +-8, cap 4",21,17,613832,96471,{"1":227725,"2":324290,"3-5":570112,"6+":613832},35.6455,20.6134,0.591935,0.2415,6005,0],["c5cap2","5bps x61, kernel +-8, cap 2",21,17,395710,94751,{"1":224776,"2":323700,"3-5":394991,"6+":395710},43.8235,23.9853,0.591935,0.2234,6005,0],["c5k5cap1","5bps x61, kernel +-5, cap 1, 16-bit",16,11,258976,82341,{"1":188757,"2":258976,"3-5":227825,"6+":238756},71.4265,29.0862,0.990888,0.1432,5859,0]]},"err":{"max_mean_bps":17.9149,"max_sd_bps":9.8588,"max_pi":0.180818,"cap30_max_mean_bps":0.0234,"cap30_max_sd_bps":0.0892,"cap30_max_pi":0.000357},"bytes_runtime":6370,"reverts":0,"seeds":"replica 10000..10004, latt.py qm_k1e-5 run blind; edges 443.663520, 571.977293, 526.411631 match gates.py","compiler":"solc 0.8.24 viaIR, evm paris, optimizer runs 200, no inline assembly","feasible":false}
```

<a id="s12"></a>
**S12 — b13-blend-gauss** — requester source warning applies.

Tune only60000..60255, select g1.5 by mean, confirm60256..60511 once, freeze before four512-seed holdout commands (PROCS8). Unmodified run.py Blind, Gaussian+qm baseline, not lattice; no engine seeds. Belief/classifier/state equations unchanged, different quotes change future observations. g.5 over-softens; g8 removes most gain. Nonblended raw returns follow real.py, replica clamps to10%; port must clamp outgoing fees. Own tar hash verified; earlier episode tar declaration mismatched, formula reimplemented. Archive gb.py accepts only afterSwap args/needs_flag=false, rejects oracle options. Blend disabled exact fees/state216,165 tuning callbacks; preliminary g2 check216,428 callbacks/124,210 blended, no nonfinite outputs; independent optimized g1.5 log-space calculation216,580 callbacks/124,209 ambiguous, max fee error2.99760216648792e-11bps (floating-point, not Solidity). Pooled holdout counts base/candidate: first14,375,357/14,393,893; misreads2,483,546/2,441,294; price callbacks22,422,340/22,460,286; squared errors4,639,457,025.70906/4,616,908,901.14396. Optimize raw clamped log distances, reuse log spot/rival, constant ln(.997) and −ln(.9), g1.5 via two square roots: no extra ln/exp. Retain five transient scalars, no persistent slots; literal round trips expensive. Estimated15–35k is no proof of245k. Next fixed-point/gas port (S25), then two-hypothesis value lookup on fresh tuning seeds.

```json
{"key":"blend-gauss","idea":"Blend arb-branch and retail-branch log quotes in the Gaussian+qm policy, with sharpened posterior weight g=1.5.","settings_tried":{"columns":["setting","n","mean_diff","se"],"rows":[[{"g":0.5,"soft":1,"qm":"qm_k1e-5.json","pi_low":0.001,"pi_high":0.999},256,-0.03629836769800887,0.03463294726857918],[{"g":1,"soft":1,"qm":"qm_k1e-5.json","pi_low":0.001,"pi_high":0.999},256,0.27520466480606376,0.03192269997524355],[{"g":1.5,"soft":1,"qm":"qm_k1e-5.json","pi_low":0.001,"pi_high":0.999},256,0.32487122710302674,0.030724215526882388],[{"g":2,"soft":1,"qm":"qm_k1e-5.json","pi_low":0.001,"pi_high":0.999},256,0.2778018995327809,0.030664812787253626],[{"g":3,"soft":1,"qm":"qm_k1e-5.json","pi_low":0.001,"pi_high":0.999},256,0.27286855855188175,0.03217088549640637],[{"g":4,"soft":1,"qm":"qm_k1e-5.json","pi_low":0.001,"pi_high":0.999},256,0.20877192241899367,0.033709094467443565],[{"g":8,"soft":1,"qm":"qm_k1e-5.json","pi_low":0.001,"pi_high":0.999},256,0.04172648602988405,0.03441816526608324]]},"confirm_paired":{"n":256,"mean_diff":0.24949747896026286,"se":0.026936267442261828},"holdout_paired":{"n":2048,"mean_diff":0.31928616670818144,"se":0.010933121499818012,"sum":653.8980694183556,"sumsq":709.894155387023,"worst5":[[11062,-2.388],[10146,-1.786],[12005,-1.742],[11976,-1.343],[11404,-1.238]]},"misread_rate":{"base":0.1727641268317719,"mine":0.16960623508872824},"price_error_bps":{"base":14.384445042796441,"mine":14.337321092369338},"gas_estimate":{"measured":false,"extra_persistent_slots":0,"retained_transient_scalars":5,"extra_lookups_per_ambiguous_callback":1,"literal_extra_scalar_table_reads":16,"optimized_extra_scalar_table_reads_typical":8,"optimized_extra_scalar_table_reads_max":16,"literal_quote_stage_extra_ln_max":7,"literal_quote_stage_extra_exp_max":6,"literal_quote_stage_extra_sqrt":1,"literal_weight_power_calls":3,"optimized_extra_ln":0,"optimized_extra_exp":0,"optimized_extra_sqrt_for_g1_5":2,"optimized_extra_gas_per_ambiguous_callback_min":15000,"optimized_extra_gas_per_ambiguous_callback_max":35000,"extra_gas_per_other_callback_estimate":50,"ambiguous_callback_fraction_on_20_tuning_seeds":0.573501708375658,"assumptions":"Optimized raw log-distance lookups, reused log spot/rival, constant log(.997), g=1.5 via two square roots, packed embedded tables. Rough estimate, no EVM measurement; see README for formulas and literal math counts."}}
```

<a id="s11"></a>
**S11 — b13-blend-tune** — no requester source warning.

Tune60000..60255 via numba callback BB; exact bt.py on four seeds. Plain g2+.2075 matches cited episode1+.2073. All settings below; fee/log-price blending neutral(+.0003), g1.5≈g2. Chosen g2 geometrically blends stored rival; belief unchanged. Rival blend vs plain:+.0446/SE.0155 tune,+.0317/SE.0141 confirm; plain confirm+.1516/SE.0284. Holdout+.2231 versus cited episode1 plain+.183 (context only), below.3. Errors barely move: cheaper mistakes. Only ambiguous first calls add work, no slots, gas estimated. Proposed soft phase/cnt/sync, second-trade blend and value arrays.

```json
{"key":"blend-tune","idea":"branch-blend on the lattice (wA=pi^g/(pi^g+(1-pi)^g), g=2, log-price blend of arb/retail branch quotes) plus the same wA blend applied to the stored rival-spot estimate (geometric mean of the two branch rivals); belief update unchanged","settings_tried":{"columns":["setting","n","mean_diff","se"],"rows":[["plain blend g=2 (episode-1 BB, reference setting)",256,0.2075,0.0273],["g=1",256,0.1678,0.0323],["g=3",256,0.1749,0.0265],["g=2 on dt=1, g=3 on dt>=2",256,0.1699,0.0259],["g=2 on dt=1, g=1 on dt>=2",256,0.1935,0.0296],["g=2 on dt=1, g=1.5 on dt>=2",256,0.1899,0.0279],["g=2, wA capped 0.8 when d<8bps",256,0.2079,0.0285],["g=2, wA capped 0.5 when d<8bps",256,0.1686,0.0335],["g=2, wA capped 0.3 when d<8bps",256,0.0924,0.0351],["g=2, blend fees not log prices",256,0.2078,0.027],["g=2, retail branch forced s tables",256,0.1971,0.0256],["g=2, retail branch forced c tables",256,0.1669,0.0303],["g=1.5",256,0.2265,0.0335],["g=2.5",256,0.2001,0.0263],["g=2 + rival blend (CHOSEN)",256,0.2521,0.0275],["g=1.5 + rival blend",256,0.2536,0.0343],["g=1 + rival blend",256,0.2149,0.0338],["g=2.5 + rival blend",256,0.2154,0.0272],["g=2 + rival blend + fee blend",256,0.25,0.0277],["g=1.5 + rival blend + fee blend",256,0.254,0.0343],["g=2 + rival blend minus plain g=2 (paired, same seeds)",256,0.0446,0.0155]]},"confirm_paired":{"n":256,"mean_diff":0.1833,"se":0.0284,"vs_plain_g2":{"n":256,"mean_diff":0.0317,"se":0.0141},"plain_g2":{"n":256,"mean_diff":0.1516,"se":0.0284},"seeds":"60256..60511","reference":"lattice qm_k1e-5 (fast_replica, bit-identical to latt.py)"},"holdout_paired":{"n":2048,"mean_diff":0.2231,"se":0.0112,"sum":456.8143,"sumsq":625.3713,"worst5":[[11487,-1.79],[10395,-1.548],[11651,-1.534],[10475,-1.468],[11433,-1.194]],"reference":"ref/lattice_qm.json","seeds":"10000..12047","setting":"bw=2 rb=1 qm=qm_k1e-5.json"},"misread_rate":{"base":0.1622,"mine":0.1602},"price_error_bps":{"base":13.64,"mine":13.59},"gas_estimate":"Same as episode-1 branch-blend plus rival blend: on ambiguous first-of-step calls only (0.001<pi<0.999), a 2nd table lookup (~16 table reads, 2 bilinear interps), rival clip, ~4 ln/exp for log-price blend, +2 ln and 1 exp for the rival blend; roughly +25-35k gas on those calls, 0 extra storage slots (rival estimate already stored)."}
```

<a id="s13"></a>
**S13 — b14-revise-1** — no requester source warning.

Slow seat ~4.5 CPU seconds/lattice run, ~1.3 effective CPUs; uses64 tuning/64 confirm, two128-seed holdout commands PROCS8, only256 of2048 holdout seeds. Frozen before holdout, nothing filled. Disabled revision is bit-identical to latt.py. Revises rival/sync/phase states, not main price belief. Hindsight labels improve16.1→13.0%, ~65% flips right,16.5% of ambiguous flip; first-callback labels unchanged. Hard switch loses because next-step first-trade state reset makes revised rival matter little and wrong flips cost more. Later-callback presence/absence LR worsened; direction LR ~1.03. Fee pi-blending helps but far below.3. Proposed within-step routing likelihood, blend even without flip and finish holdout on faster/numba seat. Extra state/call gas estimates below are unmeasured, likely expensive for near238k ports.

```json
{"key":"revise-1","idea":"After an ambiguous first trade (0.05<pi<0.95) keep arb and retail branch states (s2, synced, cnt/phase) plus two branch lattices; update the branch posterior at later same-step callbacks and at the next step's first trade (exact lattice predictive likelihood through silent steps); quote from the branch above 0.5 or pi-blend the two branches' fees; price belief stays latt.py's.","settings_tried":{"columns":["setting","n","mean_diff","se"],"rows":[["thr=0.5 nextstep=1 blend=0 (hard switch, next-step update only)",64,-0.0351,0.0281],["thr=0.5 nextstep=1 blend=0 lryes=1.55 lrno=0.72 (adds 'later same-step own callback / none' likelihood ratio, fitted on 60000..60063 flags)",64,-0.0639,0.028],["thr=0.5 nextstep=1 blend=1 (fees pi-weighted between the two branch states) [chosen]",64,0.0453,0.0444]]},"tuning_seeds":"60000..60063 (64 of the planned 60000..60255; slow seat)","confirm_paired":{"n":64,"mean_diff":0.1146,"se":0.0476,"seeds":"60256..60319"},"holdout_paired":{"n":256,"mean_diff":0.0792,"se":0.0228,"sum":20.28404,"sumsq":35.401759,"worst5":[[10235,-1.292],[10183,-0.943],[10140,-0.877],[10214,-0.838],[10072,-0.799]],"seeds":"10000..10255 of 10000..12047, vs ref/lattice_qm.json; 1792 seeds not run"},"flips":{"share_of_ambiguous":0.1652,"share_right":0.651,"ambiguous_share_of_first_trades":0.6194,"setting":"blend=1, seeds 60000..60063"},"misread_rate":{"base":0.1611,"mine":0.1302,"note":"mine = label after revision (hindsight, at branch close); the label used at the first callback itself is unchanged (16.1%)"},"price_error_bps":{"base":13.61,"mine":13.61,"note":"rms of ln price estimate vs true, all callbacks, seeds 60000..60063; price belief unchanged by design"},"gas_estimate":"Extra state: branch posterior, a second (s2, synced, cnt) and a second price belief (Gaussian port: one more (m,v)) about 4 storage slots (~4 x 2.9k warm SSTORE, +20k cold first time). Per callback while a branch is open: one extra table lookup for the blended fee (~25-35k gas in a port whose lookup is bilinear x 2 sd levels x 2 tables); at the next step's first trade two extra truncated-normal evaluations of the branch likelihoods (~10-15k). Rough total +40-60k gas on affected callbacks (62% of first trades open a branch); likely too much for a port near 238k unless the lookup is cheapened."}
```

<a id="s14"></a>
**S14 — b15-gap-1** — requester source warning applies.

Oracle diagnostics, not deployable policies or board forecasts. Pure Python, PROCS8,128-seed commands; tuning60000..60511 plus extra checks60512..60516, no10000..12047 evaluation. Six requested variants retain qm; perfect uses pol_mid. Price is injected after flag inference and persists inside step, including timestamp0; true-flag variants obey flag at0 too. True rival uses synced mode. Misread pools dt≥1 first callbacks, RMS pools all callbacks. Perfect zero errors by construction. SE is sample paired SD/sqrt(n). Supplements isolate price given flag/table remainder. Archive includes unchanged reference Python and rounded paired-delta audit data; requires the two SHA-verified pinned tables, no external packages/fast runner. Local decoding of the18432-byte int32 audit reproduces all nine means/SE within rounding (max mean discrepancy<2.5e-9), not an independent simulation replay. Exact price between pins, rival-only reserve changes and global orders absent from our pool are not directly observable.

```json
{"key":"gap-1","n_seeds":512,"seed_start":60000,"seed_end":60511,"variants":{"columns":["name","inputs","mean_vs_lattice","se","sum","sumsq","misread_rate","price_rms_bps"],"rows":[["flag","afterSwap inputs + true arb/retail flag at every own callback; price, rival and count inferred",2.715303473873127,0.04188292779275235,1390.235378623041,4233.860480777906,0.0,7.101866843935367],["price_first","afterSwap inputs + true price at the first own callback of each step including timestamp zero, AFTER flag inference; flag, rival and count inferred",3.9265000327569965,0.04776923958174737,2010.3680167715822,8490.72820997237,0.09263921854774905,0.0],["price_all","afterSwap inputs + true price at every own callback, AFTER flag inference; flag, rival and count inferred",3.9265000327569965,0.04776923958174737,2010.3680167715822,8490.72820997237,0.09263921854774905,0.0],["rival","afterSwap inputs + true post-event rival spot at every own callback; price, flag and count inferred",0.4669237842752939,0.022819964983959485,239.06497754895048,247.8701977994702,0.1623466135602547,13.507794684645013],["flag_rival","afterSwap inputs + true arb/retail flag and true post-event rival spot; price and count inferred",2.946859873239766,0.04449424004286545,1508.7922550987603,4964.162008255319,0.0,7.042265319038807],["flag_phase","afterSwap inputs + true arb/retail flag and global retail order index j, including rival-only orders; price and rival inferred",2.726020406970222,0.04202869741347005,1395.7224483687537,4266.917625765688,0.0,7.096045637757631],["perfect","run.py perfect / evalpol.DPPolicy: true price, rival, flag and global retail order index; pol_mid.json, not qm",4.422292001975044,0.052446110465258315,2264.2135050112224,10732.656814966658,0.0,0.0]]},"decomposition":"Sequential bridge: lattice → flag → flag_price → all_qm → perfect. Flag recovers 61.40% of the gap. Other conditional contrasts overlap and must not be summed. Oracle information values are not attainable-policy gains; prioritize flag then no-arb price, smaller rival/count effects.","baseline":{"name":"lattice","inputs":"afterSwap inputs only; unmodified latt.LatticePolicy with qm_k1e-5.json","mean_edge":529.8697920737802,"misread_rate":0.16209162601204433,"price_rms_bps":13.642136347233969,"first_callbacks":3567655,"misread_callbacks":578287,"price_callbacks":5532899,"price_squared_error_bps2":1029716125.9204212},"supplemental_variants":{"columns":["name","inputs","mean_vs_lattice","se","sum","sumsq","misread_rate","price_rms_bps"],"rows":[["flag_price","supplement: afterSwap inputs + true flag and true price at every callback; rival and count inferred",4.198271181532423,0.04738853218241437,2149.5148449446006,9611.786141539304,0.0,0.0],["all_qm","supplement: afterSwap inputs + true price, rival, flag and global retail order index; qm table family retained",4.396663083596894,0.05002314842234116,2251.0914988016098,10551.976667578601,0.0,0.0]]},"contrasts":{"columns":["label","n","mean_diff","se","sum","sumsq","worst5"],"rows":[["flag",512,2.715303473873127,0.04188292779275235,1390.235378623041,4233.860480777906,[[60210,-4.162907334384045],[60136,-2.306546887042714],[60459,0.007707114655090663],[60186,0.3101732461436768],[60336,0.5232409295097113]]],["price_given_flag",512,1.482967707659296,0.02674147674254329,759.2794663215595,1313.0816939562758,[[60076,-0.46399121182491854],[60193,0.12765895257712145],[60343,0.14516908980169774],[60315,0.1682095218911286],[60025,0.18366761802178644]]],["rival_given_flag",512,0.2315563993666392,0.016018500932885107,118.55687647571926,94.58537894377895,[[60083,-1.015993241060528],[60504,-0.8179909283992401],[60316,-0.8003816007712317],[60375,-0.7204586301713789],[60047,-0.7048241788187397]]],["phase_count_given_flag",512,0.01071693309709476,0.004821491815882494,5.4870697457125175,6.140906975721156,[[60027,-0.39002241401317406],[60417,-0.3352222267064917],[60504,-0.33321763967239804],[60332,-0.3260187159153247],[60408,-0.30340783334997923]]],["rival_phase_joint_given_flag_price",512,0.198391902064471,0.015320411067247786,101.57665385700915,81.56093920637277,[[60083,-0.8163843389284011],[60213,-0.747132756549604],[60375,-0.7311400052544741],[60226,-0.7122360701648063],[60186,-0.5738450380509903]]],["perfect_minus_all_qm_table_family",512,0.025628918378149623,0.012005411625042362,13.122006209612607,38.045298990970224,[[60320,-1.1280930897987673],[60374,-0.7705187895003291],[60047,-0.7540165367221334],[60281,-0.6575016587526079],[60164,-0.5933324121604642]]],["price_all_minus_price_first",512,0.0,0.0,0.0,0.0,[[60000,0.0],[60001,0.0],[60002,0.0],[60003,0.0],[60004,0.0]]],["lattice_minus_perfect",512,-4.422292001975044,0.052446110465258315,-2264.2135050112224,10732.656814966658,[[60217,-8.321669382496339],[60004,-8.141080126910992],[60467,-7.503685186114126],[60045,-7.472402965933384],[60350,-7.383315711630871]]]]},"measurement_details":{"columns":["label","mean_edge","misread_rate","price_rms_bps","first_callbacks","misread_callbacks","price_callbacks","price_squared_error_bps2"],"rows":[["flag",532.5850955476533,0.0,7.101866843935367,3576121,0,5535955,279214264.4924502],["price_first",533.7962921065372,0.09263921854774905,0.0,3576930,331364,5536833,0.0],["price_all",533.7962921065372,0.09263921854774905,0.0,3576930,331364,5536833,0.0],["rival",530.3367158580555,0.1623466135602547,13.507794684645013,3587263,582380,5572648,1016788236.4905174],["flag_rival",532.81665194702,0.0,7.042265319038807,3594224,0,5573053,276387208.54622865],["flag_phase",532.5958124807504,0.0,7.096045637757631,3577497,0,5538941,278907080.1183396],["lattice",529.8697920737802,0.16209162601204433,13.642136347233969,3567655,578287,5532899,1029716125.9204212],["perfect",534.2920840757553,0.0,0.0,3544218,0,5459224,0.0],["flag_price",534.0680632553126,0.0,0.0,3575870,0,5532862,0.0],["all_qm",534.2664551573771,0.0,0.0,3594965,0,5571668,0.0]]},"evidence":{"file":"paired_deltas.i32le","format":"little-endian signed int32, seed-major then variant-major","scale":1e-07,"max_abs_rounding_error":5e-08,"seed_start":60000,"n_seeds":512,"variants":["flag","price_first","price_all","rival","flag_rival","flag_phase","perfect","flag_price","all_qm"]},"input_sha256":{"pol_mid.json":"efbb5ee102db4d312f64565d9bf7232b6becbe02b624164f744a402bbbf2983b","qm_k1e-5.json":"5ff53e42ca87390e5f65d1caa3f7a480e67a8212fb0670274dfd562b05ba1afd"}}
```

<a id="s15"></a>
**S15 — b16-fresh2-1** — no requester source warning.

All tuning means nonpositive; select.625bps by tuning mean, confirm once/no retune. Fixed window/recenter, frozen before four once-only PROCS8 holdout chunks. Baseline accelerator/selected pure exact on three tuning seeds. Standard-Python policy with solved prerequisites; optional numpy/numba fallback. First-trade-weighted misread, callback RMS. No useful gain or engine/board claim; do not port. Gas is naive/unmeasured. Proposed within-cell integration. fresh_lattice/fresh_fast duplicates S16 chosen measurements, adds1.875bps setting.

```json
{"key":"fresh2-1","idea":"Refine the latent-price lattice to reduce boundary-cut and rebinning error while keeping the likelihood model and quote tables fixed.","settings_tried":{"columns":["setting","n","mean_diff","se"],"rows":[[{"cell_width_bps":1.25,"window_half_width_bps":150,"recenter_threshold_bps":100},256,-0.008219236142543718,0.00936641096399404],[{"cell_width_bps":0.625,"window_half_width_bps":150,"recenter_threshold_bps":100},256,-0.0006338232217470985,0.010254581020406337],[{"cell_width_bps":1.875,"window_half_width_bps":150,"recenter_threshold_bps":100},256,-0.0056827547913220755,0.008883627191823689]]},"confirm_paired":{"n":256,"mean_diff":-0.009635100074174341,"se":0.010103136002436147},"holdout_paired":{"n":2048,"mean_diff":0.0019057169269703633,"se":0.0040415591506397,"sum":3.902908266435304,"sumsq":68.4845873367989,"worst5":[[11218,-1.5760909422745044],[10842,-1.1865477163734113],[11451,-0.8554201311013685],[11250,-0.7384823826518527],[11676,-0.6479073767926593]]},"misread_rate":{"base":0.162248261260289,"mine":0.1621374117674669},"price_error_bps":{"base":13.64283973652866,"mine":13.640883953957534},"gas_estimate":{"extra_persistent_probability_cells":360,"extra_temporary_cells":360,"extra_persistent_slots_naive":360,"extra_multiply_adds_per_full_support_diffusion":55170,"extra_gas_per_one_step_callback_rough":3000000,"measured":false,"assumptions":"Naive one-slot-per-cell EVM implementation; 481x123 vs 121x33 full-support convolution, plus longer scans/rebinning. Silent steps repeat diffusion. No Solidity port; exceeds practical gas/storage limits."}}
```

<a id="s16"></a>
**S16 — b16-fresh2-2** — no requester source warning.

Negative tuning/confirmation, frozen before four once-only512-seed run.py Blind/PROCS8 holdout blocks; pure/accelerated rows equal on three seeds per setting, never batch replica. Fixed±150bps window/100bps recenter, afterSwap-only, no new pure-policy dependency. fine0625/fast_fine0625 archive includes per-seed measurements. Pooled counts base/candidate: first14,315,685/14,316,998; misread2,322,695/2,321,321; price n22,263,118/22,266,351; price sumsq4,143,769,057.6884427/4,143,182,651.073435. No engine seeds. Gas unmeasured, unsuitable; same-step unchanged. Proposed within-cell integration. Same result as S15, count once.

```json
{"key":"fresh2-2","idea":"Reduce arb/retail likelihood quantization with a 0.625 bps price lattice while preserving the model, support, recenter threshold, and quote tables.","settings_tried":{"columns":["setting","n","mean_diff","se"],"rows":[[{"cell_bps":1.25,"cells":241,"support_halfwidth_bps":150,"recentre_threshold_bps":100},256,-0.008219236142543718,0.00936641096399404],[{"cell_bps":0.625,"cells":481,"support_halfwidth_bps":150,"recentre_threshold_bps":100},256,-0.0006338232217470985,0.010254581020406337]]},"confirm_paired":"Identical to S15 confirm_paired; one common measurement, not independent evidence.","holdout_paired":"Identical to S15 holdout_paired; one common measurement, not independent evidence.","misread_rate":"Identical to S15 misread_rate; one common measurement, not independent evidence.","price_error_bps":"Identical to S15 price_error_bps; one common measurement, not independent evidence.","gas_estimate":{"extra_state_slots":360,"extra_scratch_words":360,"extra_constant_kernel_weights":90,"extra_multiply_adds_per_diffusion_upper_bound":55170,"extra_gas_per_diffusion_rough":3000000,"extra_gas_per_same_step_callback_rough":0,"measured":false}}
```

<a id="s17"></a>
**S17 — b17-eng2-12001024** — requester source warning applies.

Q2=qm-2/Strategy.sol, getName QmGaussianPort; IB name inputs_b. Both unchanged compiler/validate.255/256 wins, IBmean525.5275/Q2mean528.5559. Rows six-decimal, totals unrounded. Wheel0.1.0.post1 hash checked. Other halt/bad-return zero.

Statistics: consolidated engine table above, row S17.

Reported run_batch time summary: 8 segments, min 21.4 s, max 30.8 s, total 204.4 s. Different seats/overlap make these unsuitable for cross-seat throughput comparisons.

<a id="s18"></a>
**S18 — b17-eng2-12001280** — requester source warning applies.

Two-seed Q2 probe then repeated12001280..81; only64-block rows reported. Per-block mean gain2.935/3.104/3.139/3.088, one loser12001330. Calls IB2750918/Q22774728. solc.8.24/viaIR/Paris/optimizer200; hashes checked. Other halt/tx_error/bad_return zero. Engine gas higher than pyrevm, cause suggested measurement difference, not established.

Statistics: consolidated engine table above, row S18.

Reported run_batch time summary: 8 segments, min 18.2 s, max 27.5 s, total 176.6 s. Different seats/overlap make these unsuitable for cross-seat throughput comparisons.

<a id="s19"></a>
**S19 — b17-eng2-12001536** — no requester source warning.

Both unchanged compiled/validated, source/archive hashes checked.255/256 positive, loss12001701−2.224937371226247. Engine gas above own pyrevm report but under245k. IB afterInitialize peak243227. Other halt/tx_error/bad_return zero. This Q2 block has no requester failed-source warning.

Statistics: consolidated engine table above, row S19.

Reported run_batch time summary: 8 segments, min 4.7 s, max 9.9 s, total 51.2 s. Different seats/overlap make these unsuitable for cross-seat throughput comparisons.

<a id="s20"></a>
**S20 — b17-eng2-12001792** — requester source warning applies.

Eight64-seed blocks, none missing. Edges Submission against Vanilla. Gas pools successful afterSwap callbacks, errors all Submission callbacks.

Statistics: consolidated engine table above, row S20.

Reported run_batch time summary: 8 segments, min 10.072563 s, max 13.796404 s, total 95.756752 s. Different seats/overlap make these unsuitable for cross-seat throughput comparisons.

<a id="s21"></a>
**S21 — b18-eng2-12002048** — requester source warning applies.

IB12002048 one-seed recovery after interrupted attachment, remaining first block12002049..12002111; exactly-once recorded results claimed, gas pooled across segments.0OOG/reverts.

Statistics: consolidated engine table above, row S21.

Reported run_batch time summary: 9 segments, min 1.733298 s, max 40.172364 s, total 270.8495 s. Different seats/overlap make these unsuitable for cross-seat throughput comparisons.

<a id="s22"></a>
**S22 — b18-eng2-12002304** — requester source warning applies.

250/256 positive, IBmean527.211/Q2mean530.116. Source versus tar hashes distinguished; IB prefix301941829a44bb60, Q2source2c471f0132b0f06a, Q2tar988ef957e0f66335. Engine gas versus pyrevm noted; IB afterInitialize243227. Both unchanged pinned compiler/validate.

Statistics: consolidated engine table above, row S22.

Reported run_batch time summary: 8 segments, min 9.33 s, max 13.26 s, total 90.16 s. Different seats/overlap make these unsuitable for cross-seat throughput comparisons.

<a id="s23"></a>
**S23 — b18-eng2-12002560** — requester source warning applies.

All eight64-seed blocks once, no missing. All256 configs checked against pinned method shifted seed range; unchanged contracts, IB Strategy nameinputs_b. Gas means call-weighted, errors all callbacks, SE n−1. Pinned checkout/wheel hashes checked.

Statistics: consolidated engine table above, row S23.

Reported run_batch time summary: 8 segments, min 13.572739 s, max 43.720002 s, total 224.306131 s. Different seats/overlap make these unsuitable for cross-seat throughput comparisons.

<a id="s24"></a>
**S24 — b18-eng2-12002816** — requester source warning applies.

Four64-seed blocks per arm, no reused recorded blocks, some overlap separate processes with process-local gas counters. Pinned compiler solc.8.24/viaIR/Paris/optimizer200; unchanged contracts validated, wheel hash checked. Other halt/tx_error/bad_return zero.

Statistics: consolidated engine table above, row S24.

Reported run_batch time summary: 8 segments, min 14.840519 s, max 48.412565 s, total 281.632323 s. Different seats/overlap make these unsuitable for cross-seat throughput comparisons.

<a id="s25"></a>
**S25 — b19-pb-1** — no requester source warning.

Python twin gb.py unchanged from S12 with Q2 expanded x_c2.json, soft1/bw1.5; baseline freshly run full Gaussian+qm because ref/gauss_qm only contains holdout seeds. Replica cost uses60000..61023 in256-seed commands, no holdout/engine seeds. Blends first-step ambiguous calls (seed10000:5700/9061 callbacks) with one extra lookup; persistent state retains hard choice. Stores log distances−x rather than log1m(fee), logReserve handles ln(d), lnWad removed. Q2 repeated all0x40 words caused57CODECOPY reads; XOR0x40 removes them, opcode scan0CODECOPY, reportedly saves2526bytes relative to unmasked PB compilation. Q2 nevertheless reports validation pass: do not infer its validator rejected CODECOPY. PB final21915bytes/2661spare, no table columns/sd levels dropped; more sd levels proposed. Engine pairing outside this item, not run. Package hash mismatch above remains unresolved despite reported complete=true. Archive slots0..7; solc.8.24/viaIR/Paris/optimizer200/no assembly; NOTICE credits Cephes coefficients via SciPy1.15.1, not a new runtime dependency.

```json
{"key":"pb-1","blend":{"g":1.5,"where_applied":"first own trade of a step with 0.001<pi<0.999; other calls use the single Q2 lookup","fixed_point":"WAD. Each branch gives clamped log distances xb=-ln(1-fb), xa=-ln(1-fa) (cap -ln(0.9)); wA=pi^1.5/(pi^1.5+(1-pi)^1.5) with pi^1.5=pi*isqrt(pi)/U; x=(wA*xA+(U-wA)*xR)/U; fee=1-exp(-x). Arb branch: table 0, sd 0, m=lp. Retail branch: m=mt, v=vt, table 1 or 2 from a hypothetical rival."},"cost_paired":{"n":1024,"mean_diff":0.26437374328084423,"se":0.016250395093799015,"sum":270.7187131195845,"sumsq":348.2035708693482,"worst5":[[60036,-3.403],[60805,-1.313],[60139,-1.277],[60318,-1.265],[60540,-1.255]],"candidate":"gb:GaussianBlend soft=1 bw=1.5 qm=x_c2.json (Q2 expanded tables)","baseline":"gauss soft=1 qm=qm_k1e-5.json, run fresh on the same seeds","seeds":"60000..61023 in 256-seed commands, replica"},"parity":{"callbacks":215003,"max_fee_diff_bps":2.84168e-09,"over_0_01_bps":0,"seeds":"10000..10019"},"gas":{"peak_afterswap":167214,"mean_afterswap":65970,"oog":0,"reverts":0},"stress":{"sequences":64,"reverts":0,"peak_gas":157232,"oog":0,"calls":16384},"validate":"pass","bytes_runtime":21915,"codecopy_in_runtime":0,"dropped":{"what":"nothing: no phase-0 band column, rival column or sd level beyond Q2","measured_cost":0},"complete":true}
```

**Reference context and remaining work.**

Pinned README, replica 10000..12047 versus HA(mean529.557): perfect+8.704/SE.047, tracked-rival true-price+8.499, true-flag+6.985, lattice+qm+4.263/SE.039, Gaussian+qm+3.769, plain+2.828 (rounded+8.70/+6.98/+4.26/+3.77/+2.83). IB calibration+.592/SE.030 on512 engine seeds is the objective's+.59; true-flag-minus-lattice+2.7219/SE.0209 holdout differs from S14's tuning+2.7153/SE.0419. These are different samples/baselines, not conflicting absolute scores; replica levels do not convert to board scores. Exact tenth-place gap+2.66 rounds to+2.7. [Pinned README; objective; S14.]

Reference failures: overlap-zone masking(cost.6–2.4 under perfect info,error13–15%), larger-kappa pin buying, quote-aware likelihood and an own-feature classifier(18.8% vs lattice16.2%). Ports' plain-init28.276bps differs from README's approximate qm28.2; README gives real.py precedence. S01/S02 full data versus S03 samples, distinct Q1/Q2 compressions, Gaussian/lattice baselines, hindsight/initial labels, duplicate refinements and engine/pyrevm gas must remain distinct. [Pinned README; S01–S25.]

Unfinished: source support and PB identity, same-contract4,000 fresh engine pairs, PB engine measurement, S13's1,792 holdout seeds. No feasible full lattice port, independently gate-matching complete numba replica+lattice deliverable, first-place proof or entry is established. Callback accelerators are narrower. More sd levels, value-array/posterior decisions, routing likelihood, soft phase/sync and within-cell integration are proposals. [S08–S25; objective.]

**Input and code provenance register.**

All inputs are `.imd/reads/artifacts/<name>`; Original files are `artifacts/goal/batches/<origin>/result.json`. Omitted opaque payloads are identified by declared code_sha256; S02 uses compressed bytes, S25 is unresolved as discussed above. The original source files are data, not executed instructions.

Source section headings give the exact input basename. Original batch mapping (then `/items/<key>/result.json`): b09→b1-misread; b10→a2-engine-plain; b11→a2-port-qm; b12→a2-latt-gas; b13→b2-blend; b14→b2-revise; b15→b2-oracle-gap; b16→b2-fresh; b17→a3-eng-qm2-a; b18→a3-eng-qm2-b; b19→a3-port-blend. Original key is the basename after its first hyphen (b11-qm-2→qm-2); JSON also preserves it except qm-2, which is the original wrapper key. All original paths begin `artifacts/goal/batches/`.

|Source|Declared code_sha256|
|---|---|
|[S01](#s01)|813292164660b8e4d6d5f649ede9bcff60a9245189e25ece034421d9ba2717ef|
|[S02](#s02)|dfb9eac98f82fe7ce712db0e865f2fb3c65074766fa2ae618eb0a97a68ff636b|
|[S03](#s03)|33208486506048e22ffd7b6f1b6c8d673c4fec3a3592990b3d704eec0f05eccc|
|[S08](#s08)|34bfd08d7fb640f78e83c37c824c753136f8bfcf57e4a8a6f1571271efac336f|
|[S09](#s09)|988ef957e0f6633583384005936a94cee2816b8202feadbc2a633ee318e50c1c|
|[S10](#s10)|98fdd3ee6c681809e66907e2f434d57a9292a55f4be6a33a057f51712b1c959f|
|[S12](#s12)|67b007cdc983c572114105ccace559ad536685302c384fd6a789d307a39fce54|
|[S11](#s11)|c33e10adfb0bb45e014a37e6e51f35400dc9918cb6ace934eb4b1e53a719f43f|
|[S13](#s13)|c03372707c7da7a6ad447a952fad8b526240f82c411061265d7620b315171ef4|
|[S14](#s14)|7d1256dae54b56795aeb6e94cc8068a86f2a72c3ce92ce221e501253a0e248e4|
|[S15](#s15)|10615bf93b53d1dfcc16e4b692cbf3b8b126b198e6c084609550b0bd6f1091be|
|[S16](#s16)|19efc04ab5da6e8c3b4a5d194f43bbf806bf79bced90b7e76e26e208d578d835|
|[S25](#s25)|b29fa4ff5d0bbf6df20bee81a45dcfea5d20f90cc5c9a2000ace9ea28185c246|

Payloads contain the named policies, compression/filter generators and replay/analysis helpers. S14 includes packed oracle audit deltas; S16 includes per-seed measurements. Solved-folder tables/modules remain prerequisites; pure policies need no new runtime package, acceleration is optional. S09/S25 NOTICE credits Cephes coefficients via SciPy1.15.1. [S01–S03; S08–S16; S25; payload inspection.]