{"approach":"A searched passive strategy quotes one tick inside the competitor's best prices only when their gap is at least three ticks. Tiny inventory skew and the competitiveness filter limit further same-side accumulation. Base quantity is 8*(1-0.75*elapsed_fraction)/price**1.5, with a net inventory cap of 2000, a synthetic-NO cap of 1000, and conservative collateral budgeting; out-of-range quotes are omitted. Every spread, size, level, inventory, skew, anchor, refill and schedule choice was parameterized; the complete winning values are in Strategy.PARAMS and the final tried entry. Discrete grids, seeded random search and mutation search evaluated parameters only on seeds 0-199; only complete 200-seed scores with zero failures could win. Screening-prefix gains that disappeared on all 200 seeds were rejected. All search processes finished before the exact winning bytes were frozen, rechecked with the unmodified CLI and its 20-seed --sandbox mode, and evaluated once on the untouched 400-seed holdout. The class uses only public StepState fields, never seeds or replayed paths. Means follow BatchResult.mean_edge applied to CLI per-simulation JSON. The local Python 3.14.4 starter score differs from the requester-provided server target, so leaderboard calibration remains unconfirmed.","blocked":null,"starter_mean_edge":-6.4826639323115,"score":{"mean_edge":39.514229904341,"simulations":200,"seed_start":0,"failures":0},"holdout":{"mean_edge":45.51548203690575,"simulations":400,"seed_start":10000,"failures":0},"code":"from orderbook_pm_challenge.strategy import BaseStrategy\nfrom orderbook_pm_challenge.types import CancelAll, CancelOrder, PlaceOrder, Side\n\n\nclass Strategy(BaseStrategy):\n    # Parameters are replaced by the search script; no seed or private state access.\n    PARAMS = {'spread': -1, 'spread_time': 0, 'size': 8, 'size_time': -0.75, 'levels': 5, 'levels_time': 2, 'cap': 2000, 'cap_time': 0, 'skew': 0.1, 'skew_time': 0, 'blend': 1, 'spacing': 1, 'decay': 0.6, 'refill': 1.01, 'keep_max': 2, 'improve': 1, 'time_power': 1, 'asymmetry': 0, 'size_power': 1.5, 'sell_ratio': 1, 'mint_cap': 1000, 'min_gap': 3, 'gap_time': 0, 'start': 0, 'stop': 1, 'spread_floor': -100, 'spread_price_power': 0}\n\n    def on_step(self, state):\n        p = self.PARAMS\n        if state.step == 0:\n            self.horizon = max(1, state.steps_remaining)\n        fraction = 1.0 - state.steps_remaining / self.horizon\n        if fraction < p['start'] or fraction >= p['stop']:\n            return [CancelAll()]\n        elapsed = fraction ** p['time_power']\n        spread = p['spread'] + p['spread_time'] * elapsed\n        size = max(0.01, p['size'] * (1.0 + p['size_time'] * elapsed))\n        levels = max(1, int(p['levels'] + p['levels_time'] * elapsed))\n        cap = max(0.0, p['cap'] * (1.0 + p['cap_time'] * elapsed))\n        skew = (p['skew'] + p['skew_time'] * elapsed) * (state.yes_inventory - state.no_inventory) / max(1.0, cap)\n        cb = state.competitor_best_bid_ticks\n        ca = state.competitor_best_ask_ticks\n        cb = 0 if cb is None else cb\n        ca = 100 if ca is None else ca\n        if ca - cb < p['min_gap'] + p['gap_time'] * elapsed:\n            return [CancelAll()]\n        mid = (cb + ca) / 2.0\n        # blend=0 anchors to midpoint; blend=1 anchors to competitor quotes.\n        bid = int(mid - spread - p['asymmetry'] + p['blend'] * (cb - mid) - skew)\n        ask = int(-(-(mid + spread - p['asymmetry'] + p['blend'] * (ca - mid) - skew) // 1))\n        # Do not clamp an out-of-range quote into a knowingly stale boundary.\n        bid = min(99, bid)\n        ask = max(bid + 1, ask)\n        # Preserve queue priority at unchanged target prices.  A configurable\n        # refill threshold controls when small residual orders are replaced.\n        targets = {}\n        for level in range(levels):\n            b = bid - level * p['spacing']\n            a = ask + level * p['spacing']\n            if b >= 1 and b >= cb + p['improve']: targets[(Side.BUY, b)] = size * p['decay'] ** level / (max(1, b) / 100.0) ** p['size_power']\n            if 1 <= a <= 99 and a <= ca - p['improve']: targets[(Side.SELL, a)] = size * p['decay'] ** level * p['sell_ratio'] / (max(1, a) / 100.0) ** p['size_power']\n        retained = {}\n        actions = []\n        for o in state.own_orders:\n            key = (o.side, o.price_ticks)\n            target = targets.get(key, 0)\n            if key not in retained and target * p['refill'] <= o.remaining_quantity <= target * p['keep_max'] and target > 0:\n                retained[key] = o\n            else:\n                actions.append(CancelOrder(o.order_id))\n        cash = max(0.0, state.cash - 0.000001)\n        yes = state.yes_inventory\n        net = state.yes_inventory - state.no_inventory\n        buy_room = max(0.0, cap - net)\n        sell_room = min(max(0.0, cap + net), yes + max(0.0, p['mint_cap'] - state.no_inventory))\n        for o in retained.values():\n            if o.side == Side.BUY:\n                cash = max(0.0, cash - o.price_ticks / 100.0 * o.remaining_quantity)\n                buy_room = max(0.0, buy_room - o.remaining_quantity)\n            else:\n                covered = min(yes, o.remaining_quantity)\n                yes -= covered\n                cash = max(0.0, cash - o.remaining_quantity * (1.0 - o.price_ticks / 100.0))\n                sell_room = max(0.0, sell_room - o.remaining_quantity)\n        for level in range(levels):\n            b = bid - level * p['spacing']\n            a = ask + level * p['spacing']\n            q0 = size * p['decay'] ** level\n            if b >= 1 and b >= cb + p['improve'] and (Side.BUY, b) not in retained:\n                qbuy = q0 / (max(1, b) / 100.0) ** p['size_power']\n                q = int(min(qbuy, buy_room, cash / (b / 100.0)) * 100) / 100.0\n                if q >= 0.01:\n                    actions.append(PlaceOrder(Side.BUY, b, q))\n                    cash = max(0.0, cash - b / 100.0 * q)\n                    buy_room -= q\n            if 1 <= a <= 99 and a <= ca - p['improve'] and (Side.SELL, a) not in retained:\n                qsell = q0 * p['sell_ratio'] / (max(1, a) / 100.0) ** p['size_power']\n                mint_room = max(0.0, p['mint_cap'] - state.no_inventory)\n                q = int(min(qsell, sell_room, yes + mint_room, yes + cash / (1.0 - a / 100.0)) * 100) / 100.0\n                if q >= 0.01:\n                    actions.append(PlaceOrder(Side.SELL, a, q))\n                    covered = min(yes, q)\n                    yes -= covered\n                    cash = max(0.0, cash - (q - covered) * (1.0 - a / 100.0))\n                    sell_room -= q\n        return actions\n","code_sha256":"d5b1dd9f3b4e18d2a99bb9a019bad0055db485208ac297f6afedfe0c98c9dd0f","tried":["Starter, unmodified public simulator b244530ee2a6247b5f930830eeca9dd950da0e7e, Python 3.14.4, seeds 0-199: mean_edge=-6.4826639323115, failures=0; does NOT match requester-provided server target -6.504278. Calibration discrepancy remains unexplained; leaderboard equivalence is unconfirmed.","No-trade sanity control: mean_edge=0.0, n=200, seed_start=0, failures=0.","Initial cancel/requote midpoint-versus-competitor spread/size grid (boundary-clamped): 42 evaluations, 0 rejected for failures; best g_0_10_2 mean_edge=-1.359150447975, n=12, seed_start=0. Partial-seed results were screening only.","Queue-preserving refill, spread, size and level-count grid (boundary-clamped): 56 evaluations, 0 rejected for failures; best p_3_2_1 mean_edge=-2.6653563948550003, n=20, seed_start=0. Partial-seed results were screening only.","Deep ladders of 10/30/99 levels with small share sizes (boundary-clamped): 60 evaluations, 1 rejected for failures; best d_10_5_0.1 mean_edge=-0.3274369568666667, n=12, seed_start=0. Partial-seed results were screening only.","Omit out-of-range quotes instead of repeatedly clamping them to 1/99: 56 evaluations, 7 rejected for failures; best b_0_5_2 mean_edge=-0.852668079885, n=20, seed_start=0. Partial-seed results were screening only.","Broad seeded random search of spread, size, levels, cap, inventory skew and time slopes (RNG 812367): 400 evaluations, 0 rejected for failures; best r_279 mean_edge=3.4447589791750004, n=24, seed_start=0. Partial-seed results were screening only.","Grid requiring quotes to match or improve competitor prices: 48 evaluations, 0 rejected for failures; best i_1.5_8_1 mean_edge=0.8367607574275, n=40, seed_start=0. Partial-seed results were screening only.","Mutation search over the core parameters (RNG 991624): 300 evaluations, 0 rejected for failures; best e_200 mean_edge=7.048039767500001, n=40, seed_start=0. Partial-seed results were screening only.","Add asymmetric sizes/prices, price scaling, mint limits and nonlinear time schedules (RNG 776432): 300 evaluations, 0 rejected for failures; best x_278 mean_edge=24.2819873898525, n=40, seed_start=0. Partial-seed results were screening only.","Minimum competitor-gap and anchor/spread/size grid: 90 evaluations, 0 rejected for failures; best gap_1_-1_5_3 mean_edge=0.565528535405, n=40, seed_start=0. Partial-seed results were screening only.","Trading-window grid scored directly on all 200 training seeds: 42 evaluations, 0 rejected for failures; best w_0.7_2_10 mean_edge=0.37838887544149996, n=200, seed_start=0. Partial-seed results were screening only.","Core mutations scored directly on all 200 training seeds (RNG 471983): 120 evaluations, 0 rejected for failures; best f_93 mean_edge=13.101871912756001, n=200, seed_start=0. Partial-seed results were screening only.","Price scaling, size and cap grid scored directly on all 200 training seeds: 36 evaluations, 0 rejected for failures; best scale_1_8_2000 mean_edge=10.882449555628, n=200, seed_start=0. Partial-seed results were screening only.","Small-cap, size and spread grid scored directly on all 200 training seeds: 75 evaluations, 0 rejected for failures; best cap_0.75_2_8 mean_edge=2.0647205375295, n=200, seed_start=0. Partial-seed results were screening only.","Deduplicated full-training mutation search (RNG 1276543): 90 evaluations, 0 rejected for failures; best h_76 mean_edge=13.8087430073875, n=200, seed_start=0. Partial-seed results were screening only.","Adaptive mutations of the best complete training candidates across grids (RNG 9215523): 100 evaluations, 0 rejected for failures; best a_99 mean_edge=39.514229904341, n=200, seed_start=0. Partial-seed results were screening only.","Greedy coordinate sweeps around the best complete training candidates: 44 evaluations, 0 rejected for failures; best z_16 mean_edge=18.423045429792, n=200, seed_start=0. Partial-seed results were screening only.","Expanded cap, skew and price-scaling grid on all 200 training seeds: 32 evaluations, 0 rejected for failures; best l_9 mean_edge=17.41805213596, n=200, seed_start=0. Partial-seed results were screening only.","Coupled queue-refill, competitiveness, spread and size grid on all 200 training seeds: 54 evaluations, 0 rejected for failures; best q_34 mean_edge=20.661268492441, n=200, seed_start=0. Partial-seed results were screening only.","Synthetic-short collateral cap, size and price-scaling grid on all 200 training seeds: 30 evaluations, 0 rejected for failures; best m_27 mean_edge=18.423045429792, n=200, seed_start=0. Partial-seed results were screening only.","Competitor-anchor negative-offset and minimum-gap grid with price-scaled sizes: 32 evaluations, 0 rejected for failures; best n_8 mean_edge=31.576750948761, n=200, seed_start=0. Partial-seed results were screening only.","Price-dependent spread power/floor and base-size grid on all 200 training seeds: 32 evaluations, 0 rejected for failures; best v_12 mean_edge=27.684546571327, n=200, seed_start=0. Partial-seed results were screening only.","Coupled size, price scaling, inventory cap and mint cap around the competitive-anchor winner: 36 evaluations, 0 rejected for failures; best t_23 mean_edge=38.3205789581765, n=200, seed_start=0. Partial-seed results were screening only.","Intermediate synthetic-short caps around the best competitive-anchor sizing candidate: 6 evaluations, 0 rejected for failures; best u_5 mean_edge=38.418734917031, n=200, seed_start=0. Partial-seed results were screening only.","Coupled time-schedule power and queue-refill grid around the best sizing candidate: 12 evaluations, 0 rejected for failures; best j_11 mean_edge=38.439992243574004, n=200, seed_start=0. Partial-seed results were screening only.","Noise check r_11: screening mean=0.26795915193333336 on 24 seeds versus full training mean=-0.1695644763415 on seeds 0-199, failures=0; not selected.","Noise check i_1.5_8_1: screening mean=0.8367607574275 on 40 seeds versus full training mean=-1.6950159173810002 on seeds 0-199, failures=0; not selected.","Noise check x_244: screening mean=22.901810370187498 on 40 seeds versus full training mean=13.1920314160105 on seeds 0-199, failures=0; not selected.","Performed 64 full 200-seed rescoring evaluations of screening candidates; all selection used seeds 0-199. Failed candidates were excluded regardless of partial edge.","Final exact-code --sandbox check: n=20, seeds 0-19, mean_edge=44.721551580124995, failures=0, per-seed edges identical to unsandboxed run; Python-level sandbox, nsjail unavailable.","Untouched holdout, opened only after freezing code: mean_edge=45.51548203690575, n=400, seeds 10000-10399, failures=0; holdout-minus-training=6.001252132564751; per-simulation standard errors training=6.599906475962597, holdout=5.349510966533766. No post-holdout tuning or code changes.","Winner a_99: final exact-code mean_edge=39.514229904341, n=200, seeds 0-199, failures=0; retail_edge=72.5119998623615, arb_edge=-32.9977699580205; parameters={\"asymmetry\":0,\"blend\":1,\"cap\":2000,\"cap_time\":0,\"decay\":0.6,\"gap_time\":0,\"improve\":1,\"keep_max\":2,\"levels\":5,\"levels_time\":2,\"min_gap\":3,\"mint_cap\":1000,\"refill\":1.01,\"sell_ratio\":1,\"size\":8,\"size_power\":1.5,\"size_time\":-0.75,\"skew\":0.1,\"skew_time\":0,\"spacing\":1,\"spread\":-1,\"spread_floor\":-100,\"spread_price_power\":0,\"spread_time\":0,\"start\":0,\"stop\":1,\"time_power\":1}"],"next":"With two more hours, investigate the starter/server calibration discrepancy, then search a finer spread/size schedule and inventory-cap grid using nested validation within seeds 0-199. Reserve a fresh external holdout before any further tuning; the reported 10000-10399 set has now been observed."}