{
  "paper": {
    "num": 15,
    "title": "Towards Transparent Reward Design in Clinical Reinforcement Learning",
    "authors": "Caro Strickland, Baptiste Bonin, Audrey Durand",
    "affiliation": "Universit\u00e9 Laval"
  },
  "inputs": {
    "patient": "child#001",
    "weights (gly, hypo, dose)": "0.50, 0.30, 0.20",
    "hypo limit": "4%",
    "TIR tolerance": "\u00b13 pts",
    "simulated days / rule": 2,
    "dosing rules": 48,
    "seed": 1
  },
  "tiles": [
    {
      "label": "Same TIR, different hypo",
      "value": "64.0 pts",
      "sub": "max hypo-time gap among 154 rule pairs within 3 TIR points (median 8.9)",
      "tone": "accent"
    },
    {
      "label": "Your weights pick",
      "value": "48% TIR",
      "sub": "basal 2/6 \u00b7 +1 if >180; hypo 16.4%, hyper 35.3%",
      "tone": "warn"
    },
    {
      "label": "Two fixed weightings",
      "value": "50% vs 48%",
      "sub": "TIR of the w = (1,0,0) pick vs the paper's w = (0.5,0.3,0.2) pick; hypo 21.5% vs 16.4%, insulin 16 vs 15 U/day",
      "tone": "neutral"
    },
    {
      "label": "Preferences that break your limit",
      "value": "54%",
      "sub": "of 231 weight vectors pick a rule with > 4% hypo time",
      "tone": "warn"
    }
  ],
  "figures": [
    {
      "path": "fig_tradeoff.png",
      "title": "Which risk does each dosing rule trade for its time in range?",
      "caption": "Each dot is one basal-only dosing rule simulated on the UVa/Padova model. Blue = Pareto-optimal on the paper's three objectives. The orange line joins the matched pair with the biggest hypoglycaemia gap at (almost) the same TIR."
    },
    {
      "path": "fig_pair_traces.png",
      "title": "Same score, different day",
      "caption": "Glucose traces of that matched pair on simulated day 1. Green band = target range 70\u2013180 mg/dL."
    },
    {
      "path": "fig_simplex.png",
      "title": "Who chose the trade-off? Every preference vector, mapped",
      "caption": "Each hexagon is a weight vector w over the three objectives; colour = time in range of the rule that maximizes w\u00b7r. Orange outlines mark weights whose pick exceeds your hypo limit."
    }
  ],
  "compare": [
    {
      "metric": "Matched pairs within 3 TIR pts (child)",
      "paper": "148",
      "yours": "154",
      "note": "41/48 rules kept (SD \u2264 20 pts, TIR \u2265 20%); paper: 40 preference vectors"
    },
    {
      "metric": "Median hypo gap among matched pairs (pp)",
      "paper": "6.4",
      "yours": "8.9",
      "note": ""
    },
    {
      "metric": "Max hypo gap among matched pairs (pp)",
      "paper": "39.8",
      "yours": "64.0",
      "note": ""
    },
    {
      "metric": "child#001 fixed-weight failure modes",
      "paper": "Baseline A hypo 55.2%; baseline B hypo 0.0%, hyper 56.0%",
      "yours": "A-pick hypo 21.5%; B-pick hypo 16.4%, hyper 35.3%",
      "note": "paper baselines are SAC agents; yours are the best rule in the family"
    },
    {
      "metric": "Differential (risk-decreasing) reward",
      "paper": "a policy can collect higher reward by oscillating into and out of a hazardous state",
      "yours": "rank corr. with absolute reward 0.50; its favourite rule has 15.0% hypo",
      "note": "undiscounted sum telescopes to RI(start) \u2212 RI(end)"
    }
  ],
  "runbook": {
    "name": "smash-15-transparent-reward",
    "url": ""
  }
}