{"item":"sixteen-output","batch":"e02-residual-head","credit":"Replica (replica2.py), Gaussian filter, fee rule, qm tables: Taylor (gtmcknight), github.com/gtmcknight/amm-sim @4812a6382311624e1506aca131ca35c378f930d9 research/solved. qt1/QMRef, TwoComp, COMBO: IdentityMD items built on Taylor's work.","status":"completed_negative: the trained head did not beat its zero candidate; frozen policy = zero head = COMBO unchanged. No gain over COMBO is claimed.","hashes":{"amm_sim_commit":"4812a6382311624e1506aca131ca35c378f930d9 (git checkout confirmed)","artifacts_sha256_equal_url_suffix":{"qt1":"8a6e303b662e1a5ea9bb47f31acedd3988ef65b3ddb66d074517aa59ad41ca40","TC":"d3c849db2ac31e567b57e52b80b86e8c24c842e5b4125b9d75c78e3cb07115b9","COMBO":"d360bea6c8c982717b7b43a8069a87013169c89ee3e4da25622db4d880b936d4","QMS1":"17d249ab0c33fc44931d9b105cf3f9e3846a560bd518b67310c2a56defacdecc","PR_qt1_report_edges":"d03074104960a361564cfb6d7db752763e9c9b651fa9ed212c10d7acc2738610"},"extracted":{"blob.bin":"3ac7cb8811ecfb6a17f499aa4e9b70dd55f7fb6432b3aec3b904e11a3dece38e (10,167 B, matches published)","qmref.py":"1dfc9b7a3c2237ff7a93dd9f0af54b9bdab16f2ce0a7eb00784cb4432598db92 (matches)","twocomp.py":"9d3c1833321493fb51bbde4e1b4dabc2f8978bd6d774ee2af8dc8aed13f0f0af (matches COMBO's stated hash)","combo.py":"c5b711a0b4fb4783e8879a484bf87a427171f4a11aa5eb5ec3513c24ead1a796 (matches)","qt1 Strategy.sol":"c68887cff61180e12ba2b1ceb333edf76a830f91f63b5d775a8806fb9ec08aca, 42,990 B (matches)"},"blob_and_qmref_source":"raw.githubusercontent.com/identity-md-launches/goal-3170-beat-network-s-amm-entry/31ff8ed128b463cde10ea0a71d354f8c54d4464a/reference/qt1/{blob.bin,qmref.py} (URL as given in COMBO's Commands)","authored_files_sha256":{"head.py":"5d1afb7e9ec4cfcf86634106ee2ff79202d37dc7a32f41ea7c779b861985d79a","ev.py":"9f3c7f694f19d70994263a268a4aebe618c541386a80abe26a9f5b896aed84b8","cand.py":"b7dbe0f74d8e9ac94f49323e57ffebf28fb492c1ab43068b427840406e73eadc","train.py":"c689389e4f2f2cdbd20b4d550e698fc32900fc00afd7b5b32016bc71d4c3f647","rep.py":"9ac434858b956e58360dcfcb0a03abfbd76e15dbbe7c7a78481df03d29cdf0d3","inv.py":"5a5bf65d41cf80ac29d25c8f0c9235f1ec83afeb1d2112aec5bbc73f9a43fc5a","agg.py":"d4ad6bcf53333f594383656c1b14b93b5ee63dad9429d0ce94be03d6a3e5afc4"}},"gates_output":{"gate_P":"PASS, unmodified python3 -I gates.py, 7.3 s, all 12 values equal README to 6dp","table":["10000  448.261189  443.455598  443.522608  443.663520","10001  576.779199  571.463409  571.623551  571.977293","10002  531.886497  525.168028  525.865859  526.411631"],"qmref_edges_10000_10002":[443.101741,571.454483,525.908441],"combo_edges_10000_10002":[444.19533,572.172682,526.27476]},"architecture":{"base":"COMBO (combo.Combo, unmodified) ; head added after COMBO's fee output","inputs_10":["w/30 (w=(ln s - m) bps)","R/20 (R=(ln s2h - m) bps)","sd/15-1","2*pi-1 (pi of the step's first own trade; 1/0 when outside the ambiguous band)","phase0 +-1","phase1 +-1","ln(1+size_bps)/3-1, size_bps=amountY/reserveY*1e4","min(silent,12)/4-0.5, silent=dt-1 on first trade else 0","first-trade +-1","mrel/30, mrel=(m-ln s) bps = belief mean relative to own spot (= -w, redundant by construction)"],"hidden":"16 tanh units, W1[16x10]~N(0,1)*1.5/sqrt(10), b1~N(0,0.5), random.Random(20261110) gauss order: W1 row-major then b1; fixed","output":"theta[34]: ask weights 16, bid weights 16, ask bias, bid bias; dask=8*tanh(za), dbid=8*tanh(zb) bps (cap 8 bps/side); fee=clamp(fee+d*1e-4,0,0.1), stored fees = returned fees","zero_init":"theta=0 bypasses the head exactly: fees bit-identical to COMBO"},"training":{"optimizer":"sep-CMA-ES (diagonal), mirrored sampling, sigma0=0.05, RNG seed 7101 (generation g uses random.Random(7101*100003+g))","benchmark":"machine cgroup cpu.max=100000/100000 (1 CPU though nproc=16). Measured: COMBO 0.42 s/run, head 0.79-0.85 s/run, qt1 ~0.2-0.4 s/run. A 256-seed block costs ~210 s per candidate.","cap":"Generations capped to 1 with lambda=2 (one mirrored pair) to reserve the full zero-candidate choose run and the full report run, per the brief's 'benchmark then cap generations'.","curve":[{"gen":0,"block":[64000,64255],"evaluations":512,"candidate_plus_vs_zero":{"n":256,"mean_diff":-0.725521,"se":0.038399,"worst5":[[64034,-2.262943],[64124,-2.175668],[64143,-2.046409],[64028,-2.016042],[64012,-2.012249]],"sum_d":-185.733445,"sum_d2":231.007503},"candidate_minus_vs_zero":{"n":256,"mean_diff":-0.516533,"se":0.036154,"worst5":[[64105,-2.581743],[64003,-2.580636],[64116,-2.198044],[64132,-2.060736],[64044,-1.938788]],"sum_d":-132.232478,"sum_d2":153.632865},"best_gain":-0.516533}],"zero_candidate_choose":"COMBO scored on all 1024 seeds 64000..65023 (mean edge 536.330100)"},"selection":"Choose-window rule: keep the best of {zero, trained candidates}. Both candidates lose to zero by >13 se on block 64000..64255 (part of the choose window), so zero is selected. The losing candidates were not extended to the other 768 choose seeds (time); the zero candidate needs no extension. No report seed was used for selection.","frozen_spec":{"policy":"Head16(blob, theta=[0]*34) == COMBO","head.py_sha256":"5d1afb7e9ec4cfcf86634106ee2ff79202d37dc7a32f41ea7c779b861985d79a","frozen_before_report":true},"policy_scores":{"choose_64000_65023":{"frozen_vs_COMBO":"identical policy: diff 0 on every seed by construction (zero bypass)","frozen_vs_qt1":"not measured on choose window (time; 1 CPU)"},"report_10000_12047":{"frozen_vs_COMBO":{"n":2048,"mean_diff":0.0,"se":0.0,"note":"identical code path; not separately rerun"},"frozen_vs_qt1":{"n":2048,"mean_diff":0.597391,"se":0.013388,"worst5":[[10146,-3.552497],[11062,-2.334849],[11808,-1.822738],[10519,-1.782692],[11376,-1.750296]],"sum_d":1223.456704,"sum_d2":1482.295137,"wins":1753,"qt1_base":"published per-seed qt1 edges from PR artifact (6dp; verified on 10000..10002 here)"},"frozen_mean_edge":533.860222,"qt1_mean_edge":533.262831,"note":"Reproduces COMBO's published report figure (+0.597391, se 0.013388) exactly."}},"findings":["Random 34-dim output perturbations of only ~1 bps scale (sigma 0.05) cost 0.52-0.73 edge per run on 256 seeds: COMBO sits in a sharp optimum w.r.t. a uniform fee residual, so any learned head needs many CMA generations at small sigma (or a gradient-like signal) to find gain; one generation cannot.","Live-policy invariance: 1926 recorded callbacks (seed 64000, 2000 steps) replayed with junk arb flag/p/s2/j gave 0 fee mismatches for a nonzero head.","Compute needed for the brief's full plan (8 candidates x 256 seeds x ~20 generations) is ~40k head runs ~ 9 core-hours; this machine had 1 core."],"seed_coverage":{"choose_used":"64000..65023 zero candidate (1024); 64000..64255 candidates (2x256)","report_used":"10000..12047 frozen policy (2048), once","engine_seeds":"none"},"runtime_s":{"gate_P":7.3,"zero_choose":432,"cand0":202,"cand1":218,"report":850},"commands":["git clone https://github.com/gtmcknight/amm-sim test/scratch/src/amm-sim && git checkout 4812a6382311624e1506aca131ca35c378f930d9","curl the 4 artifacts + PR into test/scratch/art; sha256sum == URL suffix","extract first ```python fence after '## Code' of TC -> w/twocomp.py and of COMBO -> w/combo.py; curl blob.bin, qmref.py into w/","(cd src/amm-sim/research/solved && python3 -I gates.py)","cd w; python3 -I -c 'import train' (zero_choose cache: train.zero_edges via main is not needed; zero edges = [ev.job((\"h\",None,s)) for s in 64000..65023])","python3 -I cand.py 0 64000 64256; python3 -I cand.py 1 64000 64256; python3 -I agg.py","python3 -I inv.py; python3 -I rep.py 10000 11024; python3 -I rep.py 11024 12048"],"versions":{"python":"CPython 3.12 (system), stdlib only","replica":"replica2.py unmodified, 10000 steps, default configs, fresh policy per seed"},"source_package":{"format":"base64(zlib(tar)) of head.py,ev.py,cand.py,train.py,rep.py,inv.py,agg.py; place in a dir with blob.bin, qmref.py, twocomp.py, combo.py; ev.py expects the amm-sim clone at ../src/amm-sim/research/solved","data":"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"},"limitations":["Only 1 CMA generation (lambda=2) ran: 1-CPU machine. This is not a test of whether a residual head can help; it only shows these two candidates hurt.","qt1 comparison on report uses published per-seed qt1 edges (PR artifact) rather than a fresh rerun; 3 seeds re-verified here.","train.py (8-candidate loop) is included but was not run to completion; cand.py is what produced the curve.","No contract, no engine run, no gas (out of scope for this item)."]}