{"item":"eight-joint","batch":"e02-residual-head","kind":"replica-only residual fee head on COMBO; no contract, no submission","credit":"Taylor (gtmcknight): replica2, Gaussian filter, fee rule, qm tables (github.com/gtmcknight/amm-sim @4812a6382311624e1506aca131ca35c378f930d9, research/solved). qt1 (QMRef), TC (TwoComp), COMBO (Combo) and QMS1 (blob) are IdentityMD items built on them and are used unmodified.","status":"done","hashes":{"amm-sim_commit":"4812a6382311624e1506aca131ca35c378f930d9","artifacts_sha256_equal_url_suffix":{"qt1":"8a6e303b662e1a5ea9bb47f31acedd3988ef65b3ddb66d074517aa59ad41ca40","TC":"d3c849db2ac31e567b57e52b80b86e8c24c842e5b4125b9d75c78e3cb07115b9","COMBO":"d360bea6c8c982717b7b43a8069a87013169c89ee3e4da25622db4d880b936d4","QMS1":"17d249ab0c33fc44931d9b105cf3f9e3846a560bd518b67310c2a56defacdecc","all_verified":true},"extracted":{"qmref.py (qt1 \"## Python reference\" fence + newline, 7739 B, published in COMBO)":"1dfc9b7a3c2237ff7a93dd9f0af54b9bdab16f2ce0a7eb00784cb4432598db92","twocomp.py (TC first python fence after \"## Code\" + newline, 11202 B)":"9d3c1833321493fb51bbde4e1b4dabc2f8978bd6d774ee2af8dc8aed13f0f0af","combo.py (COMBO first python fence after \"## Code\" + newline, 5929 B)":"c5b711a0b4fb4783e8879a484bf87a427171f4a11aa5eb5ec3513c24ead1a796","blob.bin (QMS1 base64 under \"## Data:\", 10167 B)":"3ac7cb8811ecfb6a17f499aa4e9b70dd55f7fb6432b3aec3b904e11a3dece38e","Strategy.sol (qt1 \"## Strategy.sol\" fence + newline, 42990 B)":"c68887cff61180e12ba2b1ceb333edf76a830f91f63b5d775a8806fb9ec08aca","all_match_published":true},"authored":{"head.py":"0ac8757a7f9ac517bae78948735c4b02694410da9855f1149ad8198000fdeb74","frozen_head.json":"42a7ee9a7b93a97c23f29b7e176982c4fadda293a5da7caeeb686b2c25802c3a","train.py":"57a401c85718b07bf71d1a964558abca18be87a99d2e01a0a1ee17b9fcbe30c4","score.py":"aa23053552e5b88676f8dbc8890bd482beb80d8766824eef49faa6e7cfb254e9","pairs.py":"03834967e8d5ad01e120971a7912356035eedfc7efb4276dc7eaa477af42dfca"}},"gates_output":{"P_command":"cd amm-sim/research/solved && python3 -I gates.py","P_stdout":"seed   perfect-info      gaussian+plain    gaussian+qm       lattice+qm\n10000  448.261189  443.455598  443.522608  443.663520\n10001  576.779199  571.463409  571.623551  571.977293\n10002  531.886497  525.168028  525.865859  526.411631","P_all_12_match_README_6dp":true,"P_wall_seconds":9.3,"qt1_QMRef_check_10000_10002":{"measured":[443.101741,571.454483,525.908441],"expected":[443.101741,571.454483,525.908441],"match":true,"note":"Taylor gaussian+qm gates column is 443.522608/571.623551/525.865859: qt1 (QMRef over the QMS1 spline blob) is a different policy from Taylor Gaussian."},"COMBO_check_10000_10002":{"measured":[444.19533,572.172682,526.27476],"published_in_COMBO":[444.19533,572.172682,526.27476],"match":true},"zero_head_equals_COMBO_10000_10002":true,"COMBO_minus_qt1_report_reproduces_published":{"measured":{"n":2048,"mean_diff":0.597391,"se":0.013388,"worst5":[[10146,-3.552497],[11062,-2.334849],[11808,-1.822738],[10519,-1.782691],[11376,-1.750296]],"sum_d":1223.456693,"sum_d2":1482.295158,"wins":1753,"mean_a":533.860222,"mean_b":533.262831,"missing":0},"published":{"mean_diff":0.597391,"se":0.013388,"worst_seed":[10146,-3.552497]},"match":true},"live_policy_invariance":{"test":"record all callbacks of HEAD (random theta) on seed 10005, replay to two fresh HEAD instances: one with the true (arb, p, s2, j), one with arb negated, p and s2 scaled by U(0.5,2), j random","callbacks":10128,"fee_mismatches":0,"mechanism":"Head.on_trade forwards only (is_buy, ax, ay, t, x, y) to Combo.on_trade; features read only own carried state (m, v, s2h, cnt, comp) and the callback."}},"architecture":{"base":"COMBO (TwoComp two-component filter + pi-weighted offset quoting), unmodified; HEAD adds a residual fee correction after every COMBO quote","inputs_8":["x0 = (m_prior - ln s)*1e4/50: belief mean before this callback relative to own spot after it","x1 = sd*1e4/20: posterior belief sd","x2 = 2*pi-1 on a step-first own trade (pi from COMBO comp: 2 components -> pi, point mass -> 1, truncated Gaussian -> 0); 0 inside a step","x3 = w*1e4/50: own displacement ln s - m (posterior)","x4 = R*1e4/30: inferred rival displacement ln s2h - m","x5 = phase-1, inferred phase 0/1/2 from cnt","x6 = ln(amountY/20)/1.2: size of the trade just seen","x7 = min(dt,12)/4: silent steps since own last trade (0 inside a step)","all clipped to [-3,3]"],"hidden":"8 tanh units, 9 weights each (8 inputs + bias) = 72","output":"2 linear units (bid, ask) over 8 hidden + bias = 18; d_bps = CAP*tanh(.)","n_params":90,"CAP_bps":5.0,"application":"fee' = clamp(fee_COMBO + d*1e-4, 0, 0.1) for bid and ask; stored fb/fa are the corrected fees, so the filter's no-arb interval uses the quotes actually in force","init":"hidden weights ~ N(0,0.5) from random.Random(20261110), output weights 0 => exactly COMBO at init (verified bit-identical on 10000..10002)","inference_stdlib":"head.py (zlib/base64 below): math + json only; per callback 8 tanh + 2 tanh, ~100 multiply-adds; measured mean |correction| on report seeds: bid 0.4332 bps, ask 0.4834 bps over 10964 callbacks/run"},"training":{"algorithm":"sep-CMA-ES (diagonal covariance, Ros & Hansen 2008 learning rates scaled (n+2)/3), lambda=8, mu=4, log weights, pure stdlib (no numpy on this machine)","objective":"minimize -(mean over block of edge_HEAD - edge_COMBO) + 1e-3*|theta|^2 (shrinkage); common random numbers: every candidate of a generation on the same 256 seeds, same replica2 streams","blocks":"generation g uses seeds 64000+256*(g%4) .. +255; four blocks cover 64000..65023 exactly once in 4 generations","zero_candidate":"COMBO itself = head with zero output; its block gain is 0 by construction (COMBO edges cached per seed in combo_choose.jsonl); kept as a selection candidate","rng_seed":20261110,"sigma0":0.05,"sigma_final":0.04498724596286317,"generations":4,"evaluations_replica_runs":8192,"benchmark_generation_discarded":{"note":"first attempt with sigma0=0.15: all 8 candidates lost to COMBO (mean |correction| ~2 bps); restarted with sigma0=0.05; recorded here, not used","record":{"gen":0,"block":0,"seeds":[64000,64255],"n_cand":8,"evals_gen":2048,"gain_best":-0.31886,"gain_mean":-0.62564,"gain_worst":-1.11077,"sigma":0.15,"sec":265.3}},"generation_cap":"benchmarked 265-281 s per generation (2048 replica runs on 4 cores) under an 8-minute-per-command limit; capped at 4 generations (8192 evaluations) to reserve time for full choose selection and the untouched report run","curve_per_generation":[{"gen":0,"block":0,"seeds":[64000,64255],"n_cand":8,"evals_gen":2048,"evals_total":2048,"gain_best":0.04405,"gain_mean":-0.11396,"gain_worst":-0.2485,"gain_zero":0.0,"fit_best":-0.02912,"sigma":0.05,"sigma_new":0.04837,"norm_m":3.8781,"sec":267.7},{"gen":1,"block":1,"seeds":[64256,64511],"n_cand":8,"evals_gen":2048,"evals_total":4096,"gain_best":-0.00347,"gain_mean":-0.09982,"gain_worst":-0.20711,"gain_zero":0.0,"fit_best":0.0188,"sigma":0.04837,"sigma_new":0.04702,"norm_m":3.8955,"sec":271.7},{"gen":2,"block":2,"seeds":[64512,64767],"n_cand":8,"evals_gen":2048,"evals_total":6144,"gain_best":0.06377,"gain_mean":0.01598,"gain_worst":-0.04083,"gain_zero":0.0,"fit_best":-0.04903,"sigma":0.04702,"sigma_new":0.04593,"norm_m":3.8883,"sec":280.6},{"gen":3,"block":3,"seeds":[64768,65023],"n_cand":8,"evals_gen":2048,"evals_total":8192,"gain_best":-0.00492,"gain_mean":-0.03709,"gain_worst":-0.09876,"gain_zero":0.0,"fit_best":0.02033,"sigma":0.04593,"sigma_new":0.04499,"norm_m":3.9173,"sec":271.4}],"curve_units":"gain_* = paired mean(edge_HEAD - edge_COMBO) over the 256-seed block, replica edge units (Y); fit_best includes shrinkage"},"selection_on_choose_64000_65023":{"candidates":{"cand_mean (CMA mean after gen 4)":{"n":1024,"mean_diff":0.020765,"se":0.010804,"worst5":[[64884,-1.479717],[64225,-1.352821],[64483,-1.299268],[64901,-1.233013],[64844,-1.228185]],"sum_d":21.263595,"sum_d2":122.711096,"wins":551,"mean_a":536.350865,"mean_b":536.3301,"missing":0},"cand_best (best block sample, gen 2, block gain +0.06377)":{"n":1024,"mean_diff":0.008283,"se":0.011322,"worst5":[[64901,-1.444943],[64483,-1.437043],[64024,-1.251373],[64844,-1.055303],[64507,-0.969996]],"sum_d":8.482059,"sum_d2":134.363822,"wins":528,"mean_a":536.338383,"mean_b":536.3301,"missing":0},"zero (COMBO)":{"n":1024,"mean_diff":0.0,"se":0.0}},"chosen":"cand_mean","rule":"highest full-choose paired mean over COMBO among the three; frozen as frozen_head.json before any report seed was run","cand_mean_vs_qt1_choose":{"n":1024,"mean_diff":0.63037,"se":0.019088,"worst5":[[64101,-2.104126],[64322,-1.299637],[64670,-1.256512],[64374,-1.002902],[64875,-0.945862]],"sum_d":645.498753,"sum_d2":788.567268,"wins":878,"mean_a":536.350865,"mean_b":535.720495,"missing":0},"COMBO_vs_qt1_choose":{"n":1024,"mean_diff":0.609605,"se":0.019401,"worst5":[[64101,-2.776983],[64845,-1.163868],[64374,-1.023434],[64795,-0.983171],[64555,-0.946189]],"sum_d":624.235158,"sum_d2":774.814573,"wins":875,"mean_a":536.3301,"mean_b":535.720495,"missing":0}},"frozen_spec":{"file":"frozen_head.json","sha256":"42a7ee9a7b93a97c23f29b7e176982c4fadda293a5da7caeeb686b2c25802c3a","theta_layout":"theta[0:72] = W1 rows (8 x [w_x0..w_x7, bias]); theta[72:90] = W2 rows (2 x [w_h0..w_h7, bias]), row 0 bid, row 1 ask","theta":[-0.047060289434038644,0.2819238180367228,0.518957960132667,-0.6828335654591189,-0.09652312406741784,-0.475905075829594,0.12973492143867454,0.36290228065760943,0.08346623882969981,0.3565898438574507,-0.47314790218035213,-0.23160434982550698,0.3818347705265289,-0.11672439119162469,-0.6528250525869578,-0.47386613143341744,0.13255109486477676,-0.29324012275747113,-0.5305993542277816,-0.2901129391126765,0.27054829252586055,1.2370719519062892,0.2717694137283435,0.7694089608878566,1.6016720547887238,0.7312442416451757,0.10591144629372552,-0.12868428421696454,0.38212499576416903,0.6282965645547157,0.6894255911815408,-0.11241922448067293,-0.9829536256475487,0.16502876195119046,0.46210089551615185,0.28608017000996944,-0.23211953602950056,0.06592629103189004,0.1278807443971403,-0.883272910222915,0.42550376019763836,0.04082323897972095,0.25929970557163406,0.2895962888015897,0.06886706429764905,-0.4868511379612188,0.2197955738742778,0.12050144594180053,0.6457239279601702,-0.4505110695924679,0.14422837238059497,-0.44066593323572634,0.24745200249689628,-0.3201376316965225,-0.06121598380768149,0.7369737972506111,-0.25632796192440394,0.16279288433905043,-0.11898007222596547,-0.31409063978846885,0.4473779182042242,0.03890767122514886,0.15762637362302087,-0.47232482508550416,0.16071326983754522,0.5423885351243543,-0.16938406811465279,-0.025853673308277106,0.7523884045563476,0.05916892509642309,0.28127339647410204,-0.06116132097164738,-0.05593740006653234,0.0219397076506412,0.048294416630735776,0.03367412405179379,-0.02064682334539528,-0.05390278146984731,-0.005146186684280484,0.040612218656415885,-0.03834983084757128,-0.05277895672755991,0.07995743750485894,0.0039101551171243695,-0.08446005291575624,0.01909094955995722,-0.1068404002878748,-0.021254539417622186,0.020287442911024446,-0.006756374146578623],"CAP_bps":5.0,"frozen_before_report":true},"report_10000_12047":{"HEAD_minus_COMBO":{"n":2048,"mean_diff":0.032444,"se":0.007277,"worst5":[[10256,-1.519313],[11855,-1.499502],[11672,-1.259581],[10862,-1.09196],[10971,-1.034933]],"sum_d":66.445688,"sum_d2":224.126358,"wins":1099,"mean_a":533.892666,"mean_b":533.860222,"missing":0},"HEAD_minus_qt1":{"n":2048,"mean_diff":0.629835,"se":0.013682,"worst5":[[10146,-3.612504],[11808,-1.999071],[10162,-1.439128],[10331,-1.3612],[10236,-1.347464]],"sum_d":1289.902381,"sum_d2":1597.213125,"wins":1757,"mean_a":533.892666,"mean_b":533.262831,"missing":0},"COMBO_minus_qt1":{"n":2048,"mean_diff":0.597391,"se":0.013388,"worst5":[[10146,-3.552497],[11062,-2.334849],[11808,-1.822738],[10519,-1.782691],[11376,-1.750296]],"sum_d":1223.456693,"sum_d2":1482.295158,"wins":1753,"mean_a":533.860222,"mean_b":533.262831,"missing":0},"mean_edges":{"HEAD":533.892666,"COMBO":533.860222,"qt1":533.262831},"run_once":true,"no_training_selection_or_retuning_on_report_seeds":true},"findings":["HEAD (COMBO + frozen 8-tanh residual head) scores +0.032 (se 0.007, n=2048, 1099/2048 wins) over COMBO and +0.630 (se 0.014) over qt1 on the report seeds; the choose-window estimate was +0.021 (se 0.011), so the report confirms a small positive residual rather than selection noise.","The gain is small and the corrections are small (mean |d| about 0.45 bps, CAP 5 bps never binding on average): after only 4 generations of sep-CMA-ES the head is a slight shrinkage-regularized perturbation of COMBO, not a learned classifier fix. The +2.7 arb/retail call is untouched by this item.","Corrections of ~2 bps scale (sigma0=0.15) lose 0.3 to 1.1 per candidate: COMBO quotes are locally near-optimal, so the residual must stay sub-bps; the loss surface around COMBO is steep in the correction magnitude.","COMBO - qt1 on the report seeds reproduces the COMBO artifact exactly (+0.597391, se 0.013388, worst 10146 -3.552497), so the replica, extracted sources and blob are byte-identical to what COMBO ran.","HEAD over qt1 (+0.63) is above the +0.5 contract threshold, but almost all of it is COMBO; the head adds 0.03 and would cost ~100 multiply-adds and 10 tanh per callback in Solidity. Not worth a contract item on its own.","Replica runtime on this 4-core, no-numpy machine is the binding constraint: 2048 replica runs per generation take ~270 s, so the 256-seed-block spec yields about one generation per 8-minute command."],"seed_coverage":{"choose":{"range":[64000,65023],"n":1024,"used_for":"training blocks (each seed once in 4 generations), COMBO/qt1 base cache, candidate selection"},"report":{"range":[10000,12047],"n":2048,"used_for":"one scoring run of frozen HEAD, COMBO, qt1; nothing chosen on them"},"diagnostics":[10000,10001,10002,10005],"engine_seeds_consumed":0,"replica_settings":"unmodified replica2.run(seed, policy): 10000 steps, default per-seed (size_mean, lam, sigma) from Random(10_000_000+seed), fresh policy object per seed"},"runtime_seconds":{"gates_py":9.3,"combo_choose_1024":103.2,"qt1_choose_1024":60.5,"training_per_generation":[267.7,271.7,280.6,271.4],"training_total":1091.4,"benchmark_generation_discarded":265.3,"head_choose_1024_per_candidate":[137.4,135.3],"qt1_report_2048":115.9,"combo_report_2048":189.4,"head_report_2048":268.0,"head_sec_per_seed_single_core":0.52,"machine":"4 cores, 7 GB RAM, Python 3.12.3, multiprocessing.Pool(4)"},"versions":{"python":"3.12.3 (system, no numpy/numba/cma; everything stdlib)","amm-sim":"4812a6382311624e1506aca131ca35c378f930d9 research/solved (replica2.py unmodified)","amm-challenge_setup":"not run: replica-only item, no contract"},"commands":["git clone https://github.com/gtmcknight/amm-sim test/scratch/scratch/src/amm-sim && git -C test/scratch/scratch/src/amm-sim checkout 4812a6382311624e1506aca131ca35c378f930d9","curl -LsS -o test/scratch/scratch/art/<sha> https://api.imd.fun/artifacts/<sha> for qt1, TC, COMBO, QMS1; sha256sum -c","python3 -I test/scratch/scratch/w/extract.py test/scratch/scratch/art test/scratch/scratch/w   # qmref.py twocomp.py combo.py blob.bin Strategy.sol + hash check","(cd test/scratch/scratch/src/amm-sim/research/solved && python3 -I gates.py)","cd test/scratch/scratch/w && python3 -I check0.py   # zero-head == COMBO, invariance test","python3 -I score.py combo 64000 65024 combo_choose; python3 -I score.py qt1 64000 65024 qt1_choose","python3 -I train.py 100 1   # x4 (one generation per command; resumes from ../state/cma.json)","python3 -I mkspec.py","python3 -I score.py head 64000 65024 mean_choose cand_mean.json; python3 -I score.py head 64000 65024 best_choose cand_best.json; python3 -I pairs.py mean_choose combo_choose 64000 65024","cp cand_mean.json frozen_head.json; sha256sum frozen_head.json","python3 -I score.py qt1 10000 12048 qt1_report; python3 -I score.py combo 10000 12048 combo_report; python3 -I score.py head 10000 12048 head_report frozen_head.json","python3 -I pairs.py head_report combo_report 10000 12048; python3 -I pairs.py head_report qt1_report 10000 12048; python3 -I pairs.py combo_report qt1_report 10000 12048"],"regeneration":{"layout":"test/scratch/scratch/w/{head.py,train.py,score.py,pairs.py,frozen_head.json} (below) + combo.py, twocomp.py, blob.bin extracted from the COMBO/TC/QMS1 artifacts (hashes above) + test/scratch/scratch/src/amm-sim at the commit; state files under test/scratch/scratch/state/","decode":"python3 -c \"import json,zlib,base64;d=json.load(open('src.json'));[open(k,'wb').write(zlib.decompress(base64.b64decode(v))) for k,v in d.items()]\"","determinism":"replica2 streams depend only on the seed; policies are deterministic; CMA sampling uses random.Random(20261111) with state checkpointed, so rerunning train.py 4 times reproduces frozen_head.json (sha256 above) given identical float arithmetic","source_zlib_b64":{"head.py":"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","train.py":"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","score.py":"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","pairs.py":"eNp9VN9P2zAQfs9fcepLbJaGX2MPnYrUTkjbHgYC3qLIcrFbDImT2S6sQvzvu7NDoQwtUmv77vN3d/7OHo1GF9I4rcAHGYwP5sbDQodHrS2Exw7Kcp88en+vvPOdbUCrlYalabSHzoJR2uIe2YDXWvkyW3u50hPoN+G2s8cw/gE98vuy38AM5nB1Pbu8xv/zC4DxKfy8Ov8FT7aAVksrlFkuCyQq4LFzPpzgfN0KNQxHaDbWD1g5jIvnzGuYgpdt32jhFSMWvu9/u8AsL+EbFmfCBK7lpukcsFVob+6tWd0GDk73DSZfjkajzLR95wL4DUbo8EfVYgwZbrPvZ5dnGKLDKnBZKuOsbDV7WcuFp5EJQcciBOdZlim9hKaTigW54pMM8FPI8fQcp0vMpDFWg7HQ9dpuue46YxnFKyAvyxz/4+njBHngExqjCjkfOOkzy8hV+uBMz/gEHAYiWEkJeEZO/hVU5aqcVMrrGgG4ICnzOvI4HdbOghoyJ80YHvGioL5wgYauH2LOChRymqqTvEiTBY++2AXorHys0VOBTtqVZm95KOXomoG0Kk3nKZHWeG/sCikYQWGcEuA4afCcIn8KZSmHXRNJH6MzFHBW+Rp3zXHgr8lEdArllaDGWbdMRb8oUCKERBY8MK+Otv49+AiTMsY2JByy7YONpgfpYgFIMMY89xImDRxRzKL98OXEEEptVsaWbeUfhtsLOCgPCGoTKl4IioI9qlOPY4n3ejNtZLtQErDDQ3VY82pysqPoUy7zCaCU+QLHBTUUnQDOqzeKUDo1+izasevzQQVcDbNi22305dvrigjXra1iZCngC48Btla6y2jb3ZwuN2VQYQ0JqAhXJ5mGI46w+t3e+BK80ivxEjO+EG8cR4ODngw0k4qH/ypIjajgFA74RwXKVz7cHvtpt4+iQEOg9Brt7Jj/Z8cz3jQMLgQ9JQIbcQq5EK00Vog83bPeGRtYvMhq3faexVuJD1Qp3eqhIsG2iyNcEHprOK75O8vnmuOX/QUVecYv","frozen_head.json":"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"}},"limitations":["Only 4 CMA-ES generations (8192 evaluations) fit the clock on this machine; the head is under-trained and the measured +0.032 over COMBO is a lower bound on what the architecture could give, not a converged result.","sep-CMA-ES (diagonal covariance) was used instead of full-covariance CMA-ES because numpy is unavailable; stated as such.","Choose-window gain (+0.021, se 0.011) is under 2 se on its own; the report run (+0.032, se 0.007, n=2048) is the evidence, and was run once after freezing.","Selection among three candidates on the choose window introduces a mild optimistic bias on the choose numbers only.","No engine, Solidity or gas work: replica-only item. Gas and bytecode are not estimated.","First training attempt (sigma0=0.15, one generation) was discarded and restarted; its record is included.","Features use COMBO internals (comp, cnt, s2h, m, v) that exist only in the Python reference; a contract port would need the same carried state (COMBO/TC have no contract yet)."]}