{
 "credit": "Taylor (gtmcknight), github.com/gtmcknight/amm-sim@4812a6382311624e1506aca131ca35c378f930d9 research/solved: the solved policy, Gaussian filter, qm tables, exact-arithmetic arb/retail tell and U1 tables that LEAN and QLX5 are built on. QLX5 = LEAN plus the drift-tolerant split test, decoded from code_zb64.QLX5 of A/1ef80bfb; knob taken from the c-map map (A/f921f213); block driver: driver_zb64 of A/764c9461, unchanged; kit: identity-md-launches/goal-3178-build-shared-evaluation-kit@45e7563d1a29ffca9d608fdf60c90bf2624d2d01. This item only changes one constant of that source and measures it; all credit for the policy is Taylor's.",
 "n": 28,
 "id": "level_cap",
 "family": "table levels",
 "find": "while(k<6&&",
 "lo": "while(k<4&&",
 "hi": "while(k<5&&",
 "seeds": {
  "first": 10206584,
  "last": 10206711,
  "n": 128,
  "unused": []
 },
 "arms": {
  "QLX5": {
   "mean_edge": 546.0684020453724,
   "peak_gas": 207834,
   "out_of_gas": 0,
   "runtime_bytes": 22440,
   "source_sha256": "e183e5bc3fb794b7067703b10a7da9920267aa5f4caa968e16ca101b02338f40"
  },
  "LO": {
   "mean_edge": 546.0738622776324,
   "peak_gas": 207834,
   "out_of_gas": 0,
   "runtime_bytes": 22440,
   "source_sha256": "e54b30b015deffa4340bc477c965faaedca07edef005621426d05211c4e21e6c"
  },
  "HI": {
   "mean_edge": 546.0718868713873,
   "peak_gas": 207834,
   "out_of_gas": 0,
   "runtime_bytes": 22440,
   "source_sha256": "0592b87f94b518e42df5ed5acf920c3f423d14ed3cf32175beff90d975aa4553"
  }
 },
 "diffs": {
  "LO-QLX5": {
   "n": 128,
   "mean_diff": 0.005460232259991926,
   "se": 0.004933277425606423,
   "sum_d": 0.6989097292789666,
   "sum_d2": 0.3994421578750457,
   "seeds_positive": 55,
   "min": -0.16332208960784556,
   "max": 0.32908203876240805
  },
  "HI-QLX5": {
   "n": 128,
   "mean_diff": 0.0034848260148394594,
   "se": 0.0027783738074849363,
   "sum_d": 0.4460577298994508,
   "sum_d2": 0.12704036622678166,
   "seeds_positive": 24,
   "min": -0.14068542488928415,
   "max": 0.21362203861201579
  },
  "HI-LO": {
   "n": 128,
   "mean_diff": -0.0019754062451524668,
   "se": 0.00478011232821626,
   "sum_d": -0.25285199937951575,
   "sum_d2": 0.3719405326553283,
   "seeds_positive": 53,
   "min": -0.3322924043216062,
   "max": 0.15651543789135758
  }
 },
 "pass_peak": {
  "LO": true,
  "HI": true
 },
 "reading": "On 128 fresh engine seeds, capping the sd level index at 4 (LO, top level 20 bps) gives LO - QLX5 = +0.0055 se 0.0049 and capping it at 5 (HI, top level 26 bps) gives HI - QLX5 = +0.0035 se 0.0028 (1.1 and 1.3 standard errors; HI - LO = -0.0020 se 0.0048; the edge is exactly QLX5's on 25 of the 128 seeds for LO and on 85 for HI), so neither cap is told apart from QLX5's k<6, both point estimates are under 0.006 against the +0.04 the goal asks for, and both variants stay inside every limit (peak 207,834, 0 out of gas, 22,440 runtime bytes).",
 "not_done": [
  "No gate (P, Q, T, F, G, E) was run, no replica run was made and nothing was submitted or prepared for the leaderboard; the arms were paired with QLX5 only, not with qt1.",
  "128 seeds only: the standard errors (0.003 to 0.005) are the same size as the measured differences, so the sign of either effect is not established; no confirmation run on further seeds was made (none were assigned to this item).",
  "Peaks and the out-of-gas count come from these 128 seeds only (4 complete blocks of 32; about 295 s of engine time against the 35 min allowed, so no block was skipped and no seed was burnt, failed or left partial).",
  "LO and HI were run alone, not stacked with any other knob; the tables were not refitted for the lower cap.",
  "Why the caps move the edge so little was not measured directly: how often the belief sd reaches the 20 or 26 bps level was not counted; the only evidence is that the per-seed edge equals QLX5's exactly on 25 seeds (LO) and 85 seeds (HI), with 55 seeds up and 48 down for LO and 24 up and 19 down for HI.",
  "No seed outside 10206584..10206711 was simulated, and each of those was run once per arm."
 ]
}