{
  "runId": "deepseek-2",
  "model": "deepseek-flash",
  "abstain": false,
  "primaryIndex": 0,
  "candidates": [
    {
      "script": "strategy \"BTC 4h SMA trend filter\" version 1.0\nasset: BTC\ntimeframe: 4h\ncapital: 10000\n# Long-only trend filter: hold while close > SMA(240) (~40 days on 4h), otherwise cash.\n# Let-run: no take-profit, no stop, no timeout; exit when the entry thesis is false.\n\nparams:\n    smaLength: min=60 max=400 type=int default=240\n\nentry:\n    condition: price > SMA(240)\n    size: 100%\n\nexit:\n",
      "params": {
        "smaLength": 240
      },
      "label": "robust",
      "researchEvidence": "Research tape BTC 4h 2020-10-01 -> 2023-09-30, 6570 bars, fee 10 bps per side (same as evaluation fee). Full window: +320.63% pnl, maxDD 37.76%, 70 trades, winRate 20%. Buy&Hold reference measured with same engine/tape: +136.43%, maxDD 77.04% -> risk-adjusted ratio 8.49 vs 1.77. Chronological 65/35 split: train +275.03% (maxDD 30.31); untouched last 35% (2022-09-27 -> 2023-09-30, 2299 bars) +24.75% (maxDD 19.99) vs B&H +20.39%; Lab flags train-winner degradation vs length-scaled train. Chronological 80/20 split: train +333.88%; last 20% (2023-02-23 -> 2023-09-30, 1314 bars) -10.14% (maxDD 19.99) vs B&H +12.56% - whipsaw range, the weakest evidence. Regime sub-periods (same script): 2020-10 -> 2021-12 bull +338.13% (maxDD 30.31); 2022 bear -33.42% (maxDD 34.20) vs B&H approx -64%; 2023-01 -> 2023-09 recovery +6.27% (maxDD 19.99). Parameter plateau (explicit CPU curve, 18 points, n=60..400): every n in 160..400 positive on the full tape; 220-280 peak (220:+239.8, 240:+320.6, 260:+290.3, 280:+289.8, 300:+226.0, 340:+240.9, 400:+146.9). Script SHA-256 9e237b54f3e6b9990a6de97895a4572bb40d03d6c32a499a448652f9d4a58237; the exact script was re-run end-to-end with run_sandbox_backtest on the pinned research dataset."
    }
  ],
  "reasoning": "One mechanism chosen a priori (the canonical long-only slow-SMA trend filter), then measured instead of score-hunted: explicit parameter curve, two chronological holdout splits, three regime sub-periods and a measured Buy&Hold, all at the evaluation fee of 10 bps. n=240 sits at the centre/top of a broad plateau (160..400 all positive; 220-280 peak) and has the best full-tape risk-adjusted ratio (8.49 vs B&H 1.77). It preserves capital in the 2022 bear (-33% vs B&H -64%), captures bull phases, and is only mildly negative in pure range chop (last-20% split -10%, flagged). Label robust (a forecast): the sealed tape starts from a post-bear base (about 27k on 2023-09-30) where a let-run trend filter has high odds of a positive compounding outcome; no leverage, no shorts, about 23 trades per year, fees immaterial. Single candidate: only the primary is scored, so extra candidates would only add selection noise.",
  "toolCallsUsed": 23
}