{
 "item": "combo-flag",
 "batch": "e02-flag-ceilings",
 "status": "done (replica oracle diagnostic; no contract, no submission)",
 "credit": "Replica (replica2.py), Gaussian filter, fee rule and qm tables are by Taylor (gtmcknight), github.com/gtmcknight/amm-sim @4812a6382311624e1506aca131ca35c378f930d9 research/solved. qt1-ref (QMRef), TwoComp and COMBO are IdentityMD items built on Taylor's work; this item only adds a one-line oracle override to COMBO.",
 "headline": {
  "report_10000_12047": {
   "abs_mean_edge": {
    "qt1_ref": 533.262831,
    "combo": 533.860222,
    "combo_flag": 536.430341,
    "HA_calib": 529.557434
   },
   "flag-combo": [
    2.570119,
    0.020763
   ],
   "flag-qt1": [
    3.16751,
    0.022386
   ],
   "combo-qt1": [
    0.597391,
    0.013388
   ],
   "flag-HA": [
    6.872907,
    0.043751
   ],
   "combo-HA": [
    4.302787,
    0.037598
   ],
   "qt1-HA": [
    3.705396,
    0.038086
   ],
   "format": "[mean_diff, se]"
  },
  "choose_64000_65023": {
   "abs_mean_edge": {
    "qt1_ref": 535.720495,
    "combo": 536.3301,
    "combo_flag": 538.833479
   },
   "flag-combo": [
    2.503378,
    0.02856
   ],
   "flag-qt1": [
    3.112983,
    0.030987
   ],
   "combo-qt1": [
    0.609605,
    0.019401
   ],
   "format": "[mean_diff, se]"
  },
  "headroom": {
   "target_vs_qt1": 3.5,
   "flag_minus_qt1_report": 3.16751,
   "remaining_to_plus3.5_report": 0.33249,
   "combo_minus_qt1_report": 0.597391,
   "flag_value_on_combo_report": 2.570119,
   "share_of_3.5_closed_by_combo": 0.1707,
   "share_of_3.5_reachable_with_true_flag": 0.905
  },
  "reading": "With the true arb/retail flag on each step's first own trade, and nothing else changed, COMBO gains +2.570 (se 0.021) on the report seeds and reaches +3.168 (se 0.022) over qt1-ref. That is 0.332 short of the +3.5 target. COMBO alone is +0.597 over qt1-ref. So for this frozen base a perfect first-trade classifier closes at most about 90.5% of the +3.5; the remaining ~0.33 must come from something other than the flag call (quotes, the filter between pins, tables). This is the exact flag-ablation ceiling for this base on Taylor's replica. It is not a board forecast and not a global bound: a different base could gain more from the flag."
 },
 "hashes": {
  "amm_sim_commit": "4812a6382311624e1506aca131ca35c378f930d9 (git rev-parse HEAD after checkout)",
  "artifacts_sha256_equals_url_suffix": {
   "qt1": "8a6e303b662e1a5ea9bb47f31acedd3988ef65b3ddb66d074517aa59ad41ca40",
   "TC": "d3c849db2ac31e567b57e52b80b86e8c24c842e5b4125b9d75c78e3cb07115b9",
   "COMBO": "d360bea6c8c982717b7b43a8069a87013169c89ee3e4da25622db4d880b936d4",
   "QMS1": "17d249ab0c33fc44931d9b105cf3f9e3846a560bd518b67310c2a56defacdecc",
   "result": "all 4 OK (sha256sum -c)"
  },
  "extracted_vs_published": {
   "qmref.py": {
    "sha256": "1dfc9b7a3c2237ff7a93dd9f0af54b9bdab16f2ce0a7eb00784cb4432598db92",
    "published": "1dfc9b7a3c2237ff7a93dd9f0af54b9bdab16f2ce0a7eb00784cb4432598db92 (COMBO artifact)",
    "from": "qt1 artifact, python fence after \"## Python reference\", + trailing newline"
   },
   "twocomp.py": {
    "sha256": "9d3c1833321493fb51bbde4e1b4dabc2f8978bd6d774ee2af8dc8aed13f0f0af",
    "published": "9d3c1833321493fb51bbde4e1b4dabc2f8978bd6d774ee2af8dc8aed13f0f0af (COMBO artifact)",
    "from": "TC artifact, first python fence after \"## Code\", + trailing newline"
   },
   "combo.py": {
    "sha256": "c5b711a0b4fb4783e8879a484bf87a427171f4a11aa5eb5ec3513c24ead1a796",
    "published": "c5b711a0b4fb4783e8879a484bf87a427171f4a11aa5eb5ec3513c24ead1a796 (COMBO artifact)",
    "from": "COMBO artifact, first python fence after \"## Code\", + trailing newline"
   },
   "blob.bin": {
    "sha256": "3ac7cb8811ecfb6a17f499aa4e9b70dd55f7fb6432b3aec3b904e11a3dece38e",
    "bytes": 10167,
    "published": "3ac7cb8811ecfb6a17f499aa4e9b70dd55f7fb6432b3aec3b904e11a3dece38e (QMS1 / qt1 artifacts)",
    "from": "QMS1 artifact base64 fence after \"## Data: base64\""
   },
   "qt1_Strategy.sol": {
    "sha256": "c68887cff61180e12ba2b1ceb333edf76a830f91f63b5d775a8806fb9ec08aca",
    "bytes": 42990,
    "published": "c68887cff61180e12ba2b1ceb333edf76a830f91f63b5d775a8806fb9ec08aca (qt1 artifact)"
   },
   "result": "all five match"
  },
  "frozen_before_report": {
   "comboflag.py": "794594887b95060561246070f74aa5a4141afb5520ae4b002fbd014cab429f4e",
   "comboflag.patch": "bb81d5703f5e7600ee0263a09a966e28d0b329dfac8bd05e97b914806d80d505",
   "combo.py": "c5b711a0b4fb4783e8879a484bf87a427171f4a11aa5eb5ec3513c24ead1a796",
   "twocomp.py": "9d3c1833321493fb51bbde4e1b4dabc2f8978bd6d774ee2af8dc8aed13f0f0af",
   "qmref.py": "1dfc9b7a3c2237ff7a93dd9f0af54b9bdab16f2ce0a7eb00784cb4432598db92",
   "blob.bin": "3ac7cb8811ecfb6a17f499aa4e9b70dd55f7fb6432b3aec3b904e11a3dece38e",
   "run.py": "c174cb2035799f6c771f6a431accb817fced14247451a7a8dd2a0d0232991589",
   "tests.py": "938f5afd90424861f06fce962c4423646fe4e0dc6b237b279772d0bf4288fb47"
  },
  "frozen_check_after_report": "sha256sum -c frozen.sha256: all OK"
 },
 "gates_output": {
  "command": "cd amm-sim/research/solved && python3 -I gates.py (unmodified checkout, before any change)",
  "output": "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",
  "result": "PASS: all 12 values equal the README table to 6 dp"
 },
 "qt1_ref_check": {
  "seeds_10000_10002": [
   "443.101741",
   "571.454483",
   "525.908441"
  ],
  "required": [
   "443.101741",
   "571.454483",
   "525.908441"
  ],
  "result": "PASS (QMRef = qt1's Python reference; it is not Taylor's Gaussian real.py)"
 },
 "combo_base_check": {
  "seeds_10000_10002": [
   "444.195330",
   "572.172682",
   "526.274760"
  ],
  "published_in_COMBO": [
   "444.195330",
   "572.172682",
   "526.274760"
  ],
  "report_combo-qt1_published": "+0.597391 se 0.013388, 1753/2048 wins",
  "report_combo-qt1_here": "+0.597391 se 0.013388, 1753/2048 wins",
  "result": "PASS (identical)"
 },
 "frozen_spec": {
  "name": "ComboFlag (oracle diagnostic, not a policy)",
  "base": "COMBO combo.py (sha256 c5b711a0...), unchanged except one executable line (plus header/class name)",
  "rule": "On a step's first own trade (dt = t - tlast >= 1), immediately after COMBO computes pi = la/(la+lr), set pi = float(bool(arb)), where arb is replica2's true arb flag for that callback. This is before is_arb = pi > .5, before the .005<pi<.995 two-component/MIX branch, the posterior update, the rival clip/sync, the phase and the quote. With pi in {0,1} the ambiguous branch is never taken: pi=1 takes the existing hard-arb branch (comp=[(1,lp,0)], phase 0, sd 0); pi=0 takes the existing hard-retail branch (comp=[(1,mt,vt)], phase 1, rival moved by the fill). Every other callback (t=0, later trades in a step) is unchanged and never reads arb. p, s2, j and the hidden run config (sigma, lam, size) are never read. Instantiated fresh per seed: ComboFlag('blob.bin').",
  "patch_unified_diff": "--- combo.py\t2026-11-10 03:38:40.316385676 +0000\n+++ comboflag.py\t2026-11-10 03:39:46.518800002 +0000\n@@ -1,4 +1,9 @@\n-\"\"\"COMBO = two-component filter + weighted-average offset quoting.\n+\"\"\"ComboFlag = COMBO with ONE change: on a step's first own trade (dt >= 1) pi := float(true_arb),\n+right after pi is computed and before the posterior update and every phase/rival/quote branch.\n+ORACLE DIAGNOSTIC ONLY (reads replica2's true arb flag); not a deployable policy. p, s2, j stay unused.\n+Credit: Taylor (gtmcknight) -- replica, Gaussian filter, fee rule, qm tables (github.com/gtmcknight/amm-sim\n+@4812a6382311624e1506aca131ca35c378f930d9, research/solved); COMBO/TwoComp/qt1-ref are IdentityMD items on them.\n+COMBO = two-component filter + weighted-average offset quoting.\n Credit: Taylor (gtmcknight) -- replica, Gaussian filter, fee rule and qm tables\n (github.com/gtmcknight/amm-sim @4812a638, research/solved); qt1 reference (QMRef),\n TwoComp (two-comp item) and the wavg-offsets quoting (MIX item) are IdentityMD items built on them.\n@@ -17,7 +22,7 @@\n from twocomp import TwoComp, SIG2, LD_M, LD_S, MAXF, phi, trunc\n \n \n-class Combo(TwoComp):\n+class ComboFlag(TwoComp):\n     def __init__(self, blob_path):\n         super().__init__(blob_path)\n         self.new_path_calls = 0\n@@ -66,6 +71,7 @@\n             d = max(abs(math.log(s / self.sp)), 1e-9)\n             lr = Z * .26 * phi((math.log(d) - LD_M) / LD_S) / (d * LD_S)\n             pi = la / (la + lr) if la + lr > 0 else 0.\n+            pi = float(bool(arb))   # FLAG DIAGNOSTIC (oracle, not a policy): true arb/retail label replaces pi\n             is_arb = pi > .5\n             if .005 < pi < .995:\n                 self.comp = [(pi, lp, 0.), (1 - pi, mt, vt)]\n",
  "tuned": "nothing; there is no parameter to choose. The choose window was run first, the files were frozen (hashes above), then the report window was run once."
 },
 "tests": {
  "command": "python3 -I tests.py",
  "result": "PASS",
  "counts": {
   "cb": 37257,
   "t0": 6,
   "same_step": 11452,
   "first": 25799,
   "first_dt1": 18542,
   "first_silent": 7257,
   "forced1": 14206,
   "forced0": 11593,
   "forced1_over_retail_or_amb": 12830,
   "forced0_over_arb_or_amb": 8149,
   "hard_agree_identical": 4820,
   "invariance": 37257
  },
  "trajectories": [
   "10000/flag",
   "10000/combo",
   "64000/flag",
   "64000/combo"
  ],
  "what": [
   "replays every callback of 4 full replica runs (seeds 10000 and 64000, each driven once by ComboFlag and once by Combo); at each callback the pre-state is copied into fresh Combo and ComboFlag objects",
   "dt<1 callbacks (t=0 and later trades in a step): ComboFlag output and state == Combo, and flipping arb changes nothing",
   "first own trade, true arb (forced pi=1): comp==[(1.0,lp,0.)], m==lp, v==0, cnt==0, synced False, s2h==clip(old s2h, e^lp*.997, e^lp/.997), fees == lookup(w, w2, phase 0, sd 0); new_path_calls unchanged",
   "first own trade, true retail (forced pi=0): comp==[(1.0,mt,vt)] with TwoComp's retail moments, cnt==1, rival clip then the cand sync rule, fees == lookup(..., phase 1, sd sqrt(vt)*1e4)",
   "first own trades after silent steps (dt>1, propagate with no-arb cuts) are counted separately from dt==1 and pass the same checks",
   "when the base pi is already hard (>=.995 or <=.005) and agrees with the truth, ComboFlag output and state == Combo exactly",
   "invariance on fixed callbacks: randomizing p, s2, j leaves ComboFlag and Combo outputs bit-identical; Combo output is identical with arb flipped (the base never reads truth); init fees identical"
  ]
 },
 "policy": {
  "choose": {
   "seeds": "64000..65023",
   "n": 1024,
   "mean_edge": {
    "qt1": 535.720495,
    "combo": 536.3301,
    "flag": 538.833479
   },
   "flag-combo": {
    "n": 1024,
    "mean_diff": 2.503378,
    "se": 0.02856,
    "worst5": [
     [
      64901,
      -1.316929
     ],
     [
      64003,
      -1.15932
     ],
     [
      64132,
      -0.028623
     ],
     [
      64778,
      0.009376
     ],
     [
      64150,
      0.051428
     ]
    ],
    "wins": 1021,
    "sum_d": 2563.459541,
    "sum_d2": 7271.747013
   },
   "flag-qt1": {
    "n": 1024,
    "mean_diff": 3.112983,
    "se": 0.030987,
    "worst5": [
     [
      64322,
      0.08756
     ],
     [
      64509,
      0.148183
     ],
     [
      64593,
      0.208364
     ],
     [
      64109,
      0.570966
     ],
     [
      64636,
      0.581355
     ]
    ],
    "wins": 1024,
    "sum_d": 3187.694699,
    "sum_d2": 10929.118199
   },
   "combo-qt1": {
    "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
     ]
    ],
    "wins": 875,
    "sum_d": 624.235158,
    "sum_d2": 774.814573
   },
   "vs_HA": "no stored HA rows for these seeds (calib_out_2048.json covers 10000..12047 only)"
  },
  "report": {
   "seeds": "10000..12047",
   "n": 2048,
   "mean_edge": {
    "qt1": 533.262831,
    "combo": 533.860222,
    "flag": 536.430341
   },
   "flag-combo": {
    "n": 2048,
    "mean_diff": 2.570119,
    "se": 0.020763,
    "worst5": [
     [
      11701,
      -1.138705
     ],
     [
      11519,
      -0.810231
     ],
     [
      11911,
      -0.365037
     ],
     [
      10578,
      -0.297158
     ],
     [
      10367,
      -0.244856
     ]
    ],
    "wins": 2039,
    "sum_d": 5263.604449,
    "sum_d2": 15335.322586
   },
   "flag-qt1": {
    "n": 2048,
    "mean_diff": 3.16751,
    "se": 0.022386,
    "worst5": [
     [
      11062,
      -1.300795
     ],
     [
      10146,
      -0.016155
     ],
     [
      11519,
      0.095586
     ],
     [
      11376,
      0.277364
     ],
     [
      10141,
      0.297378
     ]
    ],
    "wins": 2046,
    "sum_d": 6487.061142,
    "sum_d2": 22648.668498
   },
   "combo-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
     ]
    ],
    "wins": 1753,
    "sum_d": 1223.456693,
    "sum_d2": 1482.295158
   },
   "HA_mean_edge_calib": 529.557434,
   "flag-HA": {
    "n": 2048,
    "mean_diff": 6.872907,
    "se": 0.043751,
    "worst5": [
     [
      10146,
      1.675938
     ],
     [
      11275,
      1.794051
     ],
     [
      10618,
      2.153025
     ],
     [
      11828,
      2.200931
     ],
     [
      11523,
      2.607312
     ]
    ],
    "wins": 2048,
    "sum_d": 14075.712926,
    "sum_d2": 104765.817406
   },
   "combo-HA": {
    "n": 2048,
    "mean_diff": 4.302787,
    "se": 0.037598,
    "worst5": [
     [
      10146,
      -1.860404
     ],
     [
      10618,
      -0.141834
     ],
     [
      11031,
      0.316455
     ],
     [
      10940,
      0.415619
     ],
     [
      11828,
      0.625383
     ]
    ],
    "wins": 2046,
    "sum_d": 8812.108477,
    "sum_d2": 43842.841918
   },
   "qt1-HA": {
    "n": 2048,
    "mean_diff": 3.705396,
    "se": 0.038086,
    "worst5": [
     [
      10618,
      -1.633785
     ],
     [
      11031,
      -0.966923
     ],
     [
      11181,
      -0.198083
     ],
     [
      11275,
      -0.179574
     ],
     [
      10447,
      -0.014874
     ]
    ],
    "wins": 2042,
    "sum_d": 7588.651784,
    "sum_d2": 34200.087915
   }
  },
  "definitions": "diff = first - second per seed; se = sample_sd(diff)/sqrt(n); worst5 = [seed, diff], most negative first; wins = seeds with diff > 0; units are replica edge"
 },
 "findings": {
  "first_trade_hard_call_error_rate": {
   "choose": {
    "combo": {
     "first": 7201469,
     "wrong": 1154223,
     "rate": 0.160276
    },
    "qt1": {
     "first": 7197347,
     "wrong": 1239361,
     "rate": 0.172197
    },
    "flag": {
     "wrong": 0
    }
   },
   "report": {
    "combo": {
     "first": 14382447,
     "wrong": 2318251,
     "rate": 0.161186
    },
    "qt1": {
     "first": 14374445,
     "wrong": 2487830,
     "rate": 0.173073
    },
    "flag": {
     "wrong": 0
    }
   },
   "note": "share of first own trades whose pi>.5 call disagrees with the true flag (the label is used only for counting, after the call): COMBO 16.1%, qt1-ref 17.3%, consistent with Taylor's README ('wrong 16% of the time')"
  },
  "combo_ambiguous_first_trades_per_run": {
   "choose": 5876.74,
   "report": 5873.05
  },
  "flag_ambiguous_first_trades_total": "choose 0, report 0 (a forced pi never enters the MIX branch)",
  "comparison_with_README": "Taylor's README gives +6.98 se 0.04 vs HA for real.py with the true flag. ComboFlag is +6.873 se 0.044 vs HA on the same report seeds. With the flag known, COMBO's ambiguous-trade (two-component/MIX) path never runs, so ComboFlag is close to Taylor's own flag run; the small gap is consistent with the different filter details and the spline-compressed tables, but that attribution was not measured here.",
  "flag_gain_consistent_across_windows": "flag-combo +2.503 (choose) vs +2.570 (report); flag-qt1 +3.113 vs +3.168",
  "tail": "flag-combo is negative on 3/1024 choose and 9/2048 report seeds; flag-qt1 is negative on 0/1024 choose and 2/2048 report seeds (11062 -1.300795, 10146 -0.016155)"
 },
 "seed_coverage": {
  "choose": "64000..65023: 1024 seeds x {qt1-ref, COMBO, ComboFlag}, each seed once per policy, no gaps, no duplicates (asserted)",
  "report": "10000..12047: 2048 seeds x {qt1-ref, COMBO, ComboFlag}, each once, complete (asserted)",
  "HA": "stored HA rows from Taylor's calib_out_2048.json (10000..12047 only). There are no HA rows for the choose seeds, so vs-HA is reported on the report window only",
  "test_seeds": "10000 and 64000 were also replayed by tests.py (correctness checks only; nothing chosen)"
 },
 "data_regeneration": {
  "recipe": "commands below; replica2 is deterministic per seed, so rerunning run.py reproduces each per-seed edge",
  "checksums": {
   "qt1": {
    "rows": 3072,
    "sha256_of_sorted_lines_seed_comma_edge6dp": "21710ef3a9c200c2143ed765758570155af6a39969aa31b34f84a021cb42f0d7"
   },
   "combo": {
    "rows": 3072,
    "sha256_of_sorted_lines_seed_comma_edge6dp": "f2b0b1b47feeaf062ce23e8e24bd2d7bf0bbe0afe8d7cbe842e9a3c61ccdb47b"
   },
   "flag": {
    "rows": 3072,
    "sha256_of_sorted_lines_seed_comma_edge6dp": "05d3b6d52c57fc88f0799dad296f85739a8f300596600a3b9435adc896fe1921"
   }
  }
 },
 "runtime": {
  "machine": "4 cores, Linux x86_64, CPython 3.12.3, stdlib only",
  "choose_wall_s": {
   "qt1": 32.69,
   "combo": 54.46,
   "flag": 40.2
  },
  "report_wall_s": {
   "qt1": 62.1,
   "combo": 93.62,
   "flag": 65.23
  },
  "seconds_per_seed_4workers": {
   "choose": {
    "qt1": 0.1226,
    "combo": 0.2069,
    "flag": 0.152
   },
   "report": {
    "qt1": 0.1187,
    "combo": 0.1803,
    "flag": 0.1254
   }
  },
  "tests_wall_s": 4.7,
  "gates_wall_s": 3.4
 },
 "versions": {
  "python": "3.12.3 (python3 -I, standard library only)",
  "replica2.py": "unmodified, amm-sim @4812a638; run(seed, policy) with defaults (10000 steps, per-seed config)"
 },
 "commands": [
  "git clone https://github.com/gtmcknight/amm-sim test/scratch/amm-sim && git -C test/scratch/amm-sim checkout 4812a6382311624e1506aca131ca35c378f930d9",
  "mkdir -p test/scratch/art && cd test/scratch/art && for h in 8a6e303b...ca40 d3c849db...15b9 d360bea6...36d4 17d249ab...cacc; do curl -sSL https://api.imdnotdead.fun/artifacts/$h -o $h; echo \"$h  $h\" | sha256sum -c -; done",
  "(cd test/scratch/amm-sim/research/solved && python3 -I gates.py)   # gate P",
  "mkdir test/scratch/w && cd test/scratch/w && save bundle_zlib_b64 as bundle.b64; python3 -I -c \"import json,zlib,base64,sys;[open(k,'w').write(v) for k,v in json.loads(zlib.decompress(base64.b64decode(open(sys.argv[1]).read()))).items()]\" bundle.b64",
  "python3 -I extract.py ../art .   # writes qmref.py twocomp.py combo.py Strategy.sol blob.bin and prints sha256",
  "cp combo.py comboflag.py && patch comboflag.py comboflag.patch",
  "python3 -I tests.py",
  "for m in qt1 combo flag; do python3 -I run.py $m 64000 65024 4; done   # choose window",
  "sha256sum combo.py comboflag.py twocomp.py qmref.py blob.bin run.py tests.py comboflag.patch > frozen.sha256   # freeze",
  "for m in qt1 combo flag; do python3 -I run.py $m 10000 12048 4; done   # report window, once",
  "sha256sum -c frozen.sha256 && python3 -I agg.py"
 ],
 "bundle_files": [
  "comboflag.patch",
  "extract.py",
  "run.py",
  "tests.py",
  "agg.py"
 ],
 "bundle_format": "base64(zlib(json {filename: text}))",
 "bundle_zlib_b64": "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",
 "limitations": [
  "Oracle diagnostic: it reads the true arb flag, so it is not a deployable policy. It bounds what a perfect first-trade classifier is worth ON THIS FROZEN COMBO BASE only.",
  "Replica edge units on Taylor's replica, which uses different RNG streams from the Rust engine. The +3.5 target is a board gap vs qt1, so the 0.332 headroom is a replica measurement read against a board gap, not a board forecast.",
  "qt1 here is qt1-ref (QMRef, the qt1 contract's Python reference), not Taylor's Gaussian real.py.",
  "No contract, gas, Gate F or engine seeds (none required for this item).",
  "Not a global bound: a base that uses the flag differently (for example quotes chosen against the full posterior) could gain more.",
  "HA comparison on the report window only (no stored HA rows for 64000..65023)."
 ]
}