{ "notes": "EPISODE 7 (2026-11-10). Read all15 handed results and goal/schema/README; fetched26 solved files with zero SHA mismatches. No requester notes/refusals. Recomputed every completed new engine arm and paired sum/sumsq/SE from rows; inspected all84 command gas snapshots and runner code. These are audits of reports, not independent engine reruns. Retrieved audit/M09/base result URLs and verified result and embedded code hashes. Spawn15 research items (9A/6B), used102->117/240;123 items and13 later episodes remain. Six batches <=5 avoid canary holds. finish=false; nothing entered on board.\n\nMILESTONE: Q2 now meets Track A's >=4000-seed/+2.7 bar. Qualified Q2-IB n4232,sum12785.970608252373,sumsq47964.52196328807,mean3.0212595955,se.0228327295. All4 a6-close-q2 blocks qualified, adding1024 pairs to prior3208. Engine peak now206021 (older205010), zero reported faults, validate passes. Q2 remains the fully confirmed fallback. This is NOT first place or an actual leaderboard placement.\n\nM09: seven new completed blocks, n1792, plus128 pilots =1920 qualified pairs against IB. Cumulative sum7213.272042186496,sumsq31537.344263543262,mean3.756912522,se.0347049534. M09-PB n896,sum401.23449607702275,sumsq462.4370696953073,mean.4478063572,se.018777803. M09-Q2 n1024,sum836.2180817862098,sumsq1010.165757025545,mean.8166192205,se.0176758874. PB-IB cumulative n2056,sum6901.674287184825,sumsq27353.253781543026,mean3.3568454704,se.031473932. Never average SEs: combine n,sum,sumsq using sample variance. Sources/runtime hashes and released wheel agree across all7. Each completed3x4 commands of64, no missing rows; all callbacks calls==ok, all5 fault counters zero, live reset/snapshot gas, validate passes. M09 engine peak214835,initialize229594; PB new208201 (historic215827); IB new234977 (historic237142). Replica forecast3.901985 differs by-.14507 (<.2); broaden, watch gap; first-place target+6.4 still~2.643 ahead.\n\nFAILED ENGINE ITEM: a6-confirm-m09/eng-12007168 could not clone: Proxy CONNECT aborted/curl56 across public sources. No seed/arm started, zero count. Network failure, not policy evidence. Retire its reservation12007168..12007423 anyway; do not reclaim or retry. Therefore8 more256 items would total3968, BELOW4000. Allocate NINE from12007424; if complete M09 reaches4224 and PB4360, each with its own same-seed IB comparator. No more Q2 samples needed.\n\nFROZEN SOURCES: API prefix https://api.imd.fun/artifacts/. Audit db67c7a0e8fa98f87b069766df6e213890861b15f967695293f519ecc46eeeb6 contains IB.sol,Q2.sol,PB.lines.json,restore_pb.py,canonical.py. M09 qualifier063270cb0315aaded47d5cb3f437297bf8e9aac55e179fc4d26534a6d0e269d3 contains m09-2/Strategy.sol,gen.py,spec.json,Python twin and parity/stress tools.\nSource hashes: M09 a44cca4b0155ca3953f27ccb6b43ac8ee74b1cdc4f27022ee5fc51d8ab67ed5f; runtime e95f44d13169747859490d2f5fb627a55b4a9cb3c84dae0c95a6b37e0023e281. IB301941829a44bb60977bb28181a5e4bfa6544f3696763f987054789b2611d686; Q2 2c471f0132b0f06aae83b7b2c8622083c9d8cb7b3d5e75719dd362f5b36bb259; PB5123bb39594add4ad9a32f376ff7aa89c13428b82ea324ae4d80c474e3162d95. solc0.8.24/viaIR/opt200/paris. WheelSHAaedb68280aca34c110968f00148387c95885207913fb2cf19a1ab4c5cfd6887a; nativeff88b6353c2aad7342f0a26194f7347e78b9c3b38d683a7449006d423978cd14. Configs build_config+BASELINE_SETTINGS/VARIANCE,default_rng(absolute seed),draw size/rate/sigma. Compare builder indices0..7 without simulating; gas reset and snapshot per command, both roles. Finished matches alone do not imply safety: cfmm can retain old fees on callback errors.\nM09 qualification remains: THR.9,merge_after2,silent cap12,quote0,floor1e-5,Q2/PB tables. Cost60000..61023 vs gauss_qm+.7229852592; vs identicalTHR.5+.3351126888. Parity216453 callbacks/20 runs,max2.72139e-9bps; runtime21984,12 slots,forced/stress clean,validate passes. Two duplicate qualifiers are not independent samples.\n\nB: no new incremental >=.3 winner. Joint price/rival exact router K1 +.079827743 se.008845 (2048); K2/4/8 persistence flat. Rival hypotheses collapse under1bp; invisible rival-only orders absent. b6-joint artifact1ba80721483a897b3bccb0b749ddc7840942a72fac5d33ef4a060e519072bcd4. One follow-up adds that missing transition only.\nb6-action-values artifact2640642f10bc7e12b7629ad9fe0028eb4cb7062fbf0838b7b1e8573fca4ac038 correctly matches grid VN values/argmax, but off-grid fees differ6.909bps: interpolating values then maximizing does not commute with interpolating optimal fees. Prior gate was overrestrictive for a changed policy. Active candidate never scored; fallback0 is no experiment. Two independent follow-ups use correct mathematical endpoint/reference checks, identical action-kernel Gaussian ablation, <=5 quadrature nodes/small action shortlist; preserve zero-change equivalence and report off-grid drift. No unchanged fallback holdout reruns after failed gates. Raw full action grid is millions of terms/callback, cannot port directly.\nhistory-price -.0018916 se.0086873 (768only), fresh6-1 branch commitment+.01172035 se.00627645 (2048; full commitment tuning-2.93), fresh6-2 size-LR-.0405 se.0215 (256only; calibrated40000..40111 outside required60000+ and no executed replay checks, do not promote); silent-step+.0066 se.0204 (128only), all dropped. New tail-risk item uses lower/upper partial moments, keeps full carried belief. Two fresh items skip README ideas and known failures/current work. thr-gauss finally returned:THR.85 +.39045769 se.01048139 vsgauss(2048), -.10444 vslattice; threshold mechanism already incorporated into stronger M09, not an additive new win or reason to replace it. True-flag gain mostly belief pin(~2.34), phase only~.22.\n\nB BASE: c9d678abb0ed0ee3a31b13a8203b6ee993896db0b6a3875c5c4f52a369bf0434, thrpol.py SHA663965707d4f20f81ab0ade8ed510742a330605fd9e889d426db1d920abc8958, unchanged lattice+THR.9. Ignore its rival modification. Baseline's+.3228919198 vs lattice is inherited; all incremental comparisons must use the fresh identical base. No hidden inputs online. New tune64000..64255,freeze,confirm64256..64511,holdout10000..12047 once; optional acceleration only with parity. Negative/partial results legitimate; no imputation or speed gates. Promotion requires genuine+.3 incremental on a complete holdout, then transfer to M09 twin and qualify Solidity.\n\nLEDGER: engine12000000..12007423 all used/retired. Excluded allocations remain:12003072..12003327 borrowed gas;12003328..12003583 anomalous IB mean182.69;12004864..12005119 different source-build cohort. Only8 pairs12003840..12003847 counted from its256 reservation; remainder retired. Never rehabilitate old data from new runs. NEW nine256 blocks start12007424,12007680,12007936,12008192,12008448,12008704,12008960,12009216,12009472; end12009727. Next free12009728. Each arm once per same config,64-seed commands<470s, started-marker ledger preserved; interrupted allocations retired.\n\nNEXT: audit new rows/hashes/configs/live gas before adding to exact totals above. If M09>=4000,mean>=2.7,zero faults/peak<=245000 and replica gap<=.2, make it standing contract; still need~2.64 for first. If gap>.2 diagnose before new widening. No prototype/invalid-gas pooling. When a B policy really adds>=.3, port on top of M09, measure compression loss,20-run callback parity<=.01bps,validate,then fresh engine pilots. Keep b5-fresh/fresh5-2 and b6-learned's outstanding continuation-edge fee residual alive; they are not dependencies. No cancellations (a1-compress/b1-ideas already cancelled). Request all15 new results and known fresh5-2; pending learned result should arrive when done. Wake on six new batches or120min; this is a plane wake, no local wait. Continue toward first, no entry or publication.", "spawn": [ { "name": "a7-confirm-m09-a", "brief": "Track A: measure frozen M09, IB and PB, each ALONE against VanillaStrategy on exclusive fresh engine seeds {{from}}..{{to}} (256). No board entry.\nClone https://github.com/benedictbrady/amm-challenge at 6b6921f8d463314aa7c96234e1be25a25245f704. From checkout: mkdir -p scratch && curl -LsSf https://raw.githubusercontent.com/gtmcknight/amm-sim/main/research/setup.sh -o scratch/setup.sh && sh scratch/setup.sh; prefix later commands with `. scratch/env.sh && `.\nFetch JSON at https://api.imd.fun/artifacts/db67c7a0e8fa98f87b069766df6e213890861b15f967695293f519ecc46eeeb6 (IB.sol, Q2.sol, PB.lines.json, restore_pb.py, canonical.py) and https://api.imd.fun/artifacts/063270cb0315aaded47d5cb3f437297bf8e9aac55e179fc4d26534a6d0e269d3 (m09-2/Strategy.sol). Decode code_zb64 as base64/zlib/tar and verify tar SHA. Restore PB; preserve all contract bytes. Adapt canonical.py to these three arms and ONLY your allocation; replace old seed constants. M09 source SHA a44cca4b0155ca3953f27ccb6b43ac8ee74b1cdc4f27022ee5fc51d8ab67ed5f, runtime e95f44d13169747859490d2f5fb627a55b4a9cb3c84dae0c95a6b37e0023e281. IB/PB must match audit hashes. solc0.8.24/viaIR/opt200/paris; use setup's released wheel, no source-build substitution.\nConfigs: build_config/BASELINE_SETTINGS/BASELINE_VARIANCE, default_rng(absolute seed), draw mean-size/rate/sigma in that order. Check all fields against MatchRunner._build_configs at indices0..7 WITHOUT simulations. Identical configs across arms. run_batch(X,VanillaStrategy,configs,8); score edges[submission]. Four 64-seed commands per arm, each timeout470s. Preserve started markers before execution; never retry a started arm/seed or run old seeds. Slow seats return actual completed intersections, no speed gate. Setup failure: blocked{step,error}; if the wheel alone is unavailable, audit the Python twin/runner instead and separate that from engine evidence.\nReset gas_stats immediately before each command; snapshot immediately after, both roles, all callback fault counters. No validation gas substituted. Run amm-match validate on unchanged sources. valid=false on any fault, revert, OOG or afterSwap peak>245000; initialize<=250000.\nReturn ONE JSON <60000 bytes: source/runtime/wheel hashes, compiler, config checks/hash with encoding and first/last config; per-arm n,sum,sumsq,rows[seed,edge]; paired M09-IB,PB-IB,M09-PB {n,mean_diff,se,sum,sumsq,worst5}, sample-SE; live per-command gas_stats, elapsed time, missing/started ranges, validate, valid, limitations. Include adapted runner code_zb64+uncompressed-tar SHA, no tables/traces or command transcripts. Never impute.", "items": [ { "key": "eng-12007424", "vars": { "from": "12007424", "to": "12007679" } }, { "key": "eng-12007680", "vars": { "from": "12007680", "to": "12007935" } }, { "key": "eng-12007936", "vars": { "from": "12007936", "to": "12008191" } }, { "key": "eng-12008192", "vars": { "from": "12008192", "to": "12008447" } }, { "key": "eng-12008448", "vars": { "from": "12008448", "to": "12008703" } } ], "output": { "format": "json", "maxBytes": 60000 }, "merge": { "groupSize": 25, "brief": "Preserve each item separately with its evidence, source hashes, actual n, paired sum/sumsq and limitations. Never pool different policies, duplicate seeds, invalid gas or partial/proxy measurements as a completed experiment." } }, { "name": "a7-confirm-m09-b", "brief": "Track A: measure frozen M09, IB and PB, each ALONE against VanillaStrategy on exclusive fresh engine seeds {{from}}..{{to}} (256). No board entry.\nClone https://github.com/benedictbrady/amm-challenge at 6b6921f8d463314aa7c96234e1be25a25245f704. From checkout: mkdir -p scratch && curl -LsSf https://raw.githubusercontent.com/gtmcknight/amm-sim/main/research/setup.sh -o scratch/setup.sh && sh scratch/setup.sh; prefix later commands with `. scratch/env.sh && `.\nFetch JSON at https://api.imd.fun/artifacts/db67c7a0e8fa98f87b069766df6e213890861b15f967695293f519ecc46eeeb6 (IB.sol, Q2.sol, PB.lines.json, restore_pb.py, canonical.py) and https://api.imd.fun/artifacts/063270cb0315aaded47d5cb3f437297bf8e9aac55e179fc4d26534a6d0e269d3 (m09-2/Strategy.sol). Decode code_zb64 as base64/zlib/tar and verify tar SHA. Restore PB; preserve all contract bytes. Adapt canonical.py to these three arms and ONLY your allocation; replace old seed constants. M09 source SHA a44cca4b0155ca3953f27ccb6b43ac8ee74b1cdc4f27022ee5fc51d8ab67ed5f, runtime e95f44d13169747859490d2f5fb627a55b4a9cb3c84dae0c95a6b37e0023e281. IB/PB must match audit hashes. solc0.8.24/viaIR/opt200/paris; use setup's released wheel, no source-build substitution.\nConfigs: build_config/BASELINE_SETTINGS/BASELINE_VARIANCE, default_rng(absolute seed), draw mean-size/rate/sigma in that order. Check all fields against MatchRunner._build_configs at indices0..7 WITHOUT simulations. Identical configs across arms. run_batch(X,VanillaStrategy,configs,8); score edges[submission]. Four 64-seed commands per arm, each timeout470s. Preserve started markers before execution; never retry a started arm/seed or run old seeds. Slow seats return actual completed intersections, no speed gate. Setup failure: blocked{step,error}; if the wheel alone is unavailable, audit the Python twin/runner instead and separate that from engine evidence.\nReset gas_stats immediately before each command; snapshot immediately after, both roles, all callback fault counters. No validation gas substituted. Run amm-match validate on unchanged sources. valid=false on any fault, revert, OOG or afterSwap peak>245000; initialize<=250000.\nReturn ONE JSON <60000 bytes: source/runtime/wheel hashes, compiler, config checks/hash with encoding and first/last config; per-arm n,sum,sumsq,rows[seed,edge]; paired M09-IB,PB-IB,M09-PB {n,mean_diff,se,sum,sumsq,worst5}, sample-SE; live per-command gas_stats, elapsed time, missing/started ranges, validate, valid, limitations. Include adapted runner code_zb64+uncompressed-tar SHA, no tables/traces or command transcripts. Never impute.", "items": [ { "key": "eng-12008704", "vars": { "from": "12008704", "to": "12008959" } }, { "key": "eng-12008960", "vars": { "from": "12008960", "to": "12009215" } }, { "key": "eng-12009216", "vars": { "from": "12009216", "to": "12009471" } }, { "key": "eng-12009472", "vars": { "from": "12009472", "to": "12009727" } } ], "output": { "format": "json", "maxBytes": 60000 }, "merge": { "groupSize": 25, "brief": "Preserve each item separately with its evidence, source hashes, actual n, paired sum/sumsq and limitations. Never pool different policies, duplicate seeds, invalid gas or partial/proxy measurements as a completed experiment." } }, { "name": "b7-posterior-values", "brief": "Track B: one blind Python policy, no engine seeds or board entry. Fetch B=https://raw.githubusercontent.com/gtmcknight/amm-sim/e65314f88de945f6d055bfc2aab01c482ab6570e/research/solved with curl -LsSf $B/fetch.sh -o fetch.sh && sh fetch.sh $B scratch/solved. Read Harness/Rules/Fleet. Base is unchanged thrpol:ThrLattice thr=0.9 qm=qm_k1e-5.json, from https://api.imd.fun/artifacts/c9d678abb0ed0ee3a31b13a8203b6ee993896db0b6a3875c5c4f52a369bf0434 code_zb64 (base64/zlib/tar); use thrpol.py only, SHA663965707d4f20f81ab0ade8ed510742a330605fd9e889d426db1d920abc8958.\nOne mechanism: quotes maximizing expected passive VN under the actual posterior. Independent copy {{copy}}. Start from https://api.imd.fun/artifacts/2640642f10bc7e12b7629ad9fe0028eb4cb7062fbf0838b7b1e8573fca4ac038 code_zb64. Grid endpoint checks passed, but off-grid argmax(interpolated values) differed6.909bps from interpolated optimal fees; these need not commute. Replace that impossible equality gate with grid values/non-tie argmax, pi0/1 branch-state checks and independent off-grid value evaluation on the SAME actions; quantify fee drift. Compare actual posterior and moment-matched Gaussian using the identical kernel/actions. Use a small action shortlist, <=5 price quadrature nodes on ambiguous first calls; measure approximation error on tuning. Freeze rival/phase rules. Regenerate VN with resumable DP_STATE/bounded sweeps. Score candidate, matched Gaussian-kernel ablation and original base; separate posterior gain from interpolation gain.\nOnly six afterSwap inputs and causal state online; no seed, hidden flag/price/rival/run parameters/future. Offline labels only on tuning seeds. Prove disabled change matches base callbacks; replay with hidden inputs scrambled. <=4 settings on64000..64255; freeze before confirmation64256..64511, then score ONCE on10000..12047 via unchanged run.py and pair.py against fresh unchanged base AND ref/lattice_qm.json. No holdout retuning. Budget60min; commands<470s, scratch checkpoints. Acceleration needs callback parity and pure-Python fallback; no speed gates, report actual partial n, never impute. Failed math gates return counterexamples, not unchanged-fallback performance claims.\nReturn one JSON <60000 bytes: mechanism, frozen settings, tuning/confirmation/holdout paired{n,mean_diff,se,sum,sumsq,worst5}, actual seed ranges, first-trade misread rate and price RMS for both arms, checks/limitations, incremental operations/slots and labeled unmeasured gas estimate; reproducible policy/checks/code_zb64+SHA256 of uncompressed tar, no tables/traces. Report incremental gain over base separately from inherited threshold gain. >=+0.3 incremental gets a port, not an automatic promotion on a partial holdout.", "items": [ { "key": "values7-1", "vars": { "copy": "1" } }, { "key": "values7-2", "vars": { "copy": "2" } } ], "output": { "format": "json", "maxBytes": 60000 }, "merge": { "groupSize": 25, "brief": "Preserve each item separately with its evidence, source hashes, actual n, paired sum/sumsq and limitations. Never pool different policies, duplicate seeds, invalid gas or partial/proxy measurements as a completed experiment." } }, { "name": "b7-rival-transition", "brief": "Track B: one blind Python policy, no engine seeds or board entry. Fetch B=https://raw.githubusercontent.com/gtmcknight/amm-sim/e65314f88de945f6d055bfc2aab01c482ab6570e/research/solved with curl -LsSf $B/fetch.sh -o fetch.sh && sh fetch.sh $B scratch/solved. Read Harness/Rules/Fleet. Base is unchanged thrpol:ThrLattice thr=0.9 qm=qm_k1e-5.json, from https://api.imd.fun/artifacts/c9d678abb0ed0ee3a31b13a8203b6ee993896db0b6a3875c5c4f52a369bf0434 code_zb64 (base64/zlib/tar); use thrpol.py only, SHA663965707d4f20f81ab0ade8ed510742a330605fd9e889d426db1d920abc8958.\nOne mechanism: propagate unobserved rival-only retail orders in the first-fill likelihood. Use jointpol.py from https://api.imd.fun/artifacts/1ba80721483a897b3bccb0b749ddc7840942a72fac5d33ef4a060e519072bcd4 code_zb64; K=1 exact-router policy gained+.07983 over base, while K2/4/8 persistence was flat. Model the hidden rival's within-step/silent-step transition from rival-only fills, integrating arrivals and sizes over a bounded marginal; do not read actual rival state. Condition paths on no own fill before the observed fill, without double-counting no-arb conditioning. Establish likelihood normalization and zero-transition equivalence to joint K1 with fixed synthetic router cases. Change only this transition, not thresholds or per-run parameters. Use a small quadrature with explicit truncation bound. Also compare to unchanged joint K1 on tuning and confirmation to isolate this mechanism. If the model cannot be validated within the item, return the precise failure without a fallback performance claim.\nOnly six afterSwap inputs and causal state online; no seed, hidden flag/price/rival/run parameters/future. Offline labels only on tuning seeds. Prove disabled change matches base callbacks; replay with hidden inputs scrambled. <=4 settings on64000..64255; freeze before confirmation64256..64511, then score ONCE on10000..12047 via unchanged run.py and pair.py against fresh unchanged base AND ref/lattice_qm.json. No holdout retuning. Budget60min; commands<470s, scratch checkpoints. Acceleration needs callback parity and pure-Python fallback; no speed gates, report actual partial n, never impute. Failed math gates return counterexamples, not unchanged-fallback performance claims.\nReturn one JSON <60000 bytes: mechanism, frozen settings, tuning/confirmation/holdout paired{n,mean_diff,se,sum,sumsq,worst5}, actual seed ranges, first-trade misread rate and price RMS for both arms, checks/limitations, incremental operations/slots and labeled unmeasured gas estimate; reproducible policy/checks/code_zb64+SHA256 of uncompressed tar, no tables/traces. Report incremental gain over base separately from inherited threshold gain. >=+0.3 incremental gets a port, not an automatic promotion on a partial holdout.", "items": [ { "key": "rival7-1", "vars": {} } ], "output": { "format": "json", "maxBytes": 60000 }, "merge": { "groupSize": 25, "brief": "Preserve each item separately with its evidence, source hashes, actual n, paired sum/sumsq and limitations. Never pool different policies, duplicate seeds, invalid gas or partial/proxy measurements as a completed experiment." } }, { "name": "b7-tail-risk", "brief": "Track B: one blind Python policy, no engine seeds or board entry. Fetch B=https://raw.githubusercontent.com/gtmcknight/amm-sim/e65314f88de945f6d055bfc2aab01c482ab6570e/research/solved with curl -LsSf $B/fetch.sh -o fetch.sh && sh fetch.sh $B scratch/solved. Read Harness/Rules/Fleet. Base is unchanged thrpol:ThrLattice thr=0.9 qm=qm_k1e-5.json, from https://api.imd.fun/artifacts/c9d678abb0ed0ee3a31b13a8203b6ee993896db0b6a3875c5c4f52a369bf0434 code_zb64 (base64/zlib/tar); use thrpol.py only, SHA663965707d4f20f81ab0ade8ed510742a330605fd9e889d426db1d920abc8958.\nOne mechanism: side-specific posterior tail risk in the qm lookup. The current mean/sd collapses bimodality; committing to a branch lost2.93, so keep the full carried belief and its mean unchanged. Derive ask/bid uncertainty inputs from upper/lower partial second moments about that mean, normalized so a symmetric Gaussian yields the original sd on both sides. Determine which tail belongs to each quote from the quote definitions. Modify only the uncertainty used by each side's table lookup; keep classifier, phase, rival and fees conversion unchanged. Include reflection symmetry, symmetric-prior equality and disabled-change tests. Try at most four strengths interpolating original sd with the derived tail moments, selected by edge. Report whether a two-Gaussian implementation can compute the same moments in O(1) work; no Solidity required.\nOnly six afterSwap inputs and causal state online; no seed, hidden flag/price/rival/run parameters/future. Offline labels only on tuning seeds. Prove disabled change matches base callbacks; replay with hidden inputs scrambled. <=4 settings on64000..64255; freeze before confirmation64256..64511, then score ONCE on10000..12047 via unchanged run.py and pair.py against fresh unchanged base AND ref/lattice_qm.json. No holdout retuning. Budget60min; commands<470s, scratch checkpoints. Acceleration needs callback parity and pure-Python fallback; no speed gates, report actual partial n, never impute. Failed math gates return counterexamples, not unchanged-fallback performance claims.\nReturn one JSON <60000 bytes: mechanism, frozen settings, tuning/confirmation/holdout paired{n,mean_diff,se,sum,sumsq,worst5}, actual seed ranges, first-trade misread rate and price RMS for both arms, checks/limitations, incremental operations/slots and labeled unmeasured gas estimate; reproducible policy/checks/code_zb64+SHA256 of uncompressed tar, no tables/traces. Report incremental gain over base separately from inherited threshold gain. >=+0.3 incremental gets a port, not an automatic promotion on a partial holdout.", "items": [ { "key": "tails7-1", "vars": {} } ], "output": { "format": "json", "maxBytes": 60000 }, "merge": { "groupSize": 25, "brief": "Preserve each item separately with its evidence, source hashes, actual n, paired sum/sumsq and limitations. Never pool different policies, duplicate seeds, invalid gas or partial/proxy measurements as a completed experiment." } }, { "name": "b7-fresh", "brief": "Track B: one blind Python policy, no engine seeds or board entry. Fetch B=https://raw.githubusercontent.com/gtmcknight/amm-sim/e65314f88de945f6d055bfc2aab01c482ab6570e/research/solved with curl -LsSf $B/fetch.sh -o fetch.sh && sh fetch.sh $B scratch/solved. Read Harness/Rules/Fleet. Base is unchanged thrpol:ThrLattice thr=0.9 qm=qm_k1e-5.json, from https://api.imd.fun/artifacts/c9d678abb0ed0ee3a31b13a8203b6ee993896db0b6a3875c5c4f52a369bf0434 code_zb64 (base64/zlib/tar); use thrpol.py only, SHA663965707d4f20f81ab0ade8ed510742a330605fd9e889d426db1d920abc8958.\nIndependent fresh idea {{copy}}. Start from the arb/retail question; SKIP the README's five suggested ideas. Propose and implement ONE mechanism differing in signal or objective from this exclusion list: threshold/temperature/variance sweeps; branch blends/commitment; per-run scalar estimation; finer lattice; static feature classifier; trade-size likelihood; rival projection; midpoint/entropy quote probes; no-fill/same-step scalar reweighting; history mean correction. Do not duplicate this wave's expected-VN action integration, invisible-rival transition or asymmetric tail-risk lookup, nor the pending continuation-edge-trained fee residual. Previous history correction was flat; raw size channel negative. Existing posterior RMS or lower misread alone is not the target: explain a causal path to higher edge. Keep it small enough for plain Python, with one isolated change and a clear zero-change control.\nOnly six afterSwap inputs and causal state online; no seed, hidden flag/price/rival/run parameters/future. Offline labels only on tuning seeds. Prove disabled change matches base callbacks; replay with hidden inputs scrambled. <=4 settings on64000..64255; freeze before confirmation64256..64511, then score ONCE on10000..12047 via unchanged run.py and pair.py against fresh unchanged base AND ref/lattice_qm.json. No holdout retuning. Budget60min; commands<470s, scratch checkpoints. Acceleration needs callback parity and pure-Python fallback; no speed gates, report actual partial n, never impute. Failed math gates return counterexamples, not unchanged-fallback performance claims.\nReturn one JSON <60000 bytes: mechanism, frozen settings, tuning/confirmation/holdout paired{n,mean_diff,se,sum,sumsq,worst5}, actual seed ranges, first-trade misread rate and price RMS for both arms, checks/limitations, incremental operations/slots and labeled unmeasured gas estimate; reproducible policy/checks/code_zb64+SHA256 of uncompressed tar, no tables/traces. Report incremental gain over base separately from inherited threshold gain. >=+0.3 incremental gets a port, not an automatic promotion on a partial holdout.", "items": [ { "key": "fresh7-1", "vars": { "copy": "1" } }, { "key": "fresh7-2", "vars": { "copy": "2" } } ], "output": { "format": "json", "maxBytes": 60000 }, "merge": { "groupSize": 25, "brief": "Preserve each item separately with its evidence, source hashes, actual n, paired sum/sumsq and limitations. Never pool different policies, duplicate seeds, invalid gas or partial/proxy measurements as a completed experiment." } } ], "read": [ "a7-confirm-m09-a/eng-12007424", "a7-confirm-m09-a/eng-12007680", "a7-confirm-m09-a/eng-12007936", "a7-confirm-m09-a/eng-12008192", "a7-confirm-m09-a/eng-12008448", "a7-confirm-m09-b/eng-12008704", "a7-confirm-m09-b/eng-12008960", "a7-confirm-m09-b/eng-12009216", "a7-confirm-m09-b/eng-12009472", "b7-posterior-values/values7-1", "b7-posterior-values/values7-2", "b7-rival-transition/rival7-1", "b7-tail-risk/tails7-1", "b7-fresh/fresh7-1", "b7-fresh/fresh7-2", "b5-fresh/fresh5-2" ], "cancel": [], "wake": { "batches": [ "a7-confirm-m09-a", "a7-confirm-m09-b", "b7-posterior-values", "b7-rival-transition", "b7-tail-risk", "b7-fresh" ], "afterMinutes": 120 }, "finish": false }