{
  "runId": "mimo-2",
  "model": "xiaomi/mimo-v2.6-pro",
  "abstain": false,
  "primaryIndex": 0,
  "candidates": [
    {
      "script": "strategy \"Dual SMA 17/280\" version 1.0\nasset: BTC\ntimeframe: 4h\ncapital: 10000\n\nentry:\n    condition: SMA(17) > SMA(280)\n    size: 100%\n\nexit:\n    # let winners run\n\nparams:\n    fast: min=10 max=80 type=int default=17\n    slow: min=100 max=300 type=int default=280\n",
      "params": {"fast": 17, "slow": 280},
      "label": "exploratory",
      "researchEvidence": "Full research tape (0-6569, 6570 bars, fee 10 bps/side): pnl +267.87%, maxDD 56.67%, 38 trades, win 57.9%, return/DD 4.73 (buy-and-hold same tape: +136.43%, maxDD 77.04%, return/DD 1.77). Sub-periods: P1 2020-10..2021-09 +326.89% maxDD 26.56%; P2 2021-10..2022-09 -35.05% maxDD 43.30%; P3 2022-10..2023-09 +16.64% maxDD 23.99%. Holdout 30% (train 0-4598, holdout 4599-6569): train +243.6% maxDD 46.63%, holdout +36.66% maxDD 16.46% (best holdout qualityScore 46.576 among 69 plateau sets), beats holdout buy-and-hold +26.94% raw / +59.61% tradable on return/DD only vs full-cycle. Holdout 40% sensitivity (holdout 3942-6569): +3.38% maxDD 32.5% — only candidate of 10 still positive. Parameter plateau fast 14-19 x slow 240-300: holdout-positive cluster around slow=240-280. trainWinnerDegrades=true on both splits (train overfit)."
    },
    {
      "script": "strategy \"Dual SMA 70/250\" version 1.0\nasset: BTC\ntimeframe: 4h\ncapital: 10000\n\nentry:\n    condition: SMA(70) > SMA(250)\n    size: 100%\n\nexit:\n    # let winners run\n\nparams:\n    fast: min=10 max=80 type=int default=70\n    slow: min=100 max=300 type=int default=250\n",
      "params": {"fast": 70, "slow": 250},
      "label": "exploratory",
      "researchEvidence": "Full research tape: pnl +240.94%, maxDD 47.18%, 14 trades, win 50%, return/DD 5.11 (beats buy-and-hold 1.77). Sub-periods: P1 +289.52% maxDD 26.56% (2 trades); P2 -24.21% maxDD 31.02% (less bad than 17/280 in bear); P3 +7.47% maxDD 22.83%. Holdout 30%: holdout +23.52% maxDD 13.91% (lowest DD in plateau). Holdout 40%: -10.16% maxDD 36.54% — negative, so not robust. Lower turnover (14 trades / 3y) and lower DD than 17/280 on full tape. Same regime dependence: loses in chop, wins in trend."
    }
  ],
  "reasoning": "Methodology: only research tape ohlcv-sha256-9f8ce49cc04885b6 (BTC 4h 2020-10..2023-09) via mcp_bench.py. Built ArenaScript, validated, then run_sandbox_sweep with holdoutPercent 30 and 40, explicit paramSet plateaus, and run_sandbox_backtest sub-periods (3 x 2190 bars) plus full tape and buy-and-hold control (always-long RSI(14)>0). Fee 10 bps/side throughout. Dual SMA (fast>slow, let-run, long-only) is the classic regime filter: it compounds the 2020-21 bull, sits mostly in cash in the 2022 bear (less loss than buy-and-hold), and lags pure buy-and-hold in recoveries. Full-cycle return/DD beats buy-and-hold clearly (4.7-5.1 vs 1.77). Chosen primary 17/280 because it was #1 on 30% holdout quality and the only positive on 40% holdout among 10 neighbours — the most split-consistent point on the plateau. Secondary 70/250 is the lower-turnover, lower-DD alternative on the same thesis. Labelled exploratory (not robust) because: (1) trainWinnerDegrades=true on every split; (2) holdout 40% is near-zero/negative for the family; (3) P2 sub-period is negative — the edge is regime-dependent and may not survive a sealed tape that is pure recovery or sideways. Abstention (25 pts) was considered: expected value of exploratory with ~70% P(sealed return>0) and ~30% P(beats buy-and-hold ratio) still exceeds 25, so submit. No robust claim is made.",
  "toolCallsUsed": 46
}