{
  "headline": "Among 154 child#001 rule pairs within 3 TIR points, the worst pair differs by 64.0 pp in hypo time; your weights pick a rule with 16.4% hypo, well above the 4% limit.",
  "idea": [
    "A scalar reward for insulin dosing silently encodes whoever designed it’s trade-off between glycaemic control, avoiding dangerous lows, and minimising insulin dose — the choice is never shown to the clinician or patient.",
    "Split the reward into three separate objectives (glycaemic risk, hypo risk, insulin exposure), simulate or train a family of policies, and report the Pareto front: now every trade-off is visible and attributable.",
    "Time-in-range (TIR) alone cannot distinguish a ‘dangerously low’ policy from a ‘dangerously high’ one — two rules can share nearly identical TIR scores while differing by tens of percentage points in time spent below 70 mg/dL."
  ],
  "findings": [
    "**Matched pairs (paper Table 1 replication, child cohort):** This run found 154 pairs of rules within 3 TIR points among 41 reliable rules; the paper reports 148 pairs for the child cohort. The median hypo-time gap is 8.9 pp (paper: 6.4 pp) and the maximum is 64.0 pp (paper: 39.8 pp) — our enumerated family produces a larger worst-case gap than the paper’s trained agents. The worst pair is ‘basal 2/6 · suspend≤80’ (TIR 35.5%, hypo 1.6%) vs ‘basal 5/6 · +1 if >180’ (TIR 33.2%, hypo 65.5%): a 2.3-point TIR difference masks a 64-point difference in blood glucose below 70 mg/dL. The glucose traces in fig_pair_traces.png show the suspend rule keeping glucose elevated above 80 mg/dL throughout the day, while the high-basal rule drives prolonged episodes of severe hypoglycaemia.",
    "**Your weights’ pick vs alternatives:** With weights (0.50, 0.30, 0.20) your pick is ‘basal 2/6 · +1 if >180’ (TIR 48.3%, hypo 16.4%, hyper 35.3%) — identical to the paper’s baseline B pick, and 16.4% hypo time is four times the 4% international consensus limit. The single-objective baseline A (w = 1,0,0) picks ‘basal 2/6 · +2 if >180’ (TIR 50.0%, hypo 21.5%) — even worse. Crucially, there is a Pareto-optimal rule ‘basal 2/6 · suspend<110 · +1 if >180’ with TIR 50.4% and hypo only 1.7% (well under the 4% limit), at the cost of higher hyperglycaemia (47.9%). Your weights assigned too little weight to hypo safety to select it; only a weight vector emphasising r_hypo would reach that Pareto rule.",
    "**Preference-simplex map:** 54% of the 231 weight vectors tested pick a rule that exceeds the 4% hypo limit, and only 9 distinct rules are reachable across the entire simplex. The simplex map (fig_simplex.png) shows that weight vectors near the glycaemic-risk corner (high r_gly) and near the centre all select rules with dangerous hypo exposure; only vectors strongly emphasising hypo safety (upper-right region) select rules within the target.",
    "**Differential vs absolute reward:** The rank correlation between undiscounted differential return and absolute return is only 0.50, confirming the paper’s warning: the differential reward’s favourite rule (‘basal 4/6 · suspend<110’) has 15.0% hypo time. The differential return telescopes to RI(start) − RI(end) and is blind to every excursion in between; a rule that oscillates into hypoglycaemia and then recovers collects the same endpoint reward as one that avoids it entirely."
  ],
  "try_next": [
    {
      "label": "Adult#001 — chase the paper’s 42.8-point gap",
      "why": "The paper reports a maximum 42.8-point matched-pair hypo gap for adult#001, making it the cohort where the TIR blind-spot is most dramatic. Running adult#001 lets you compare directly with the paper’s headline figure.",
      "params": {"patient": "adult#001", "w_glycemic": 0.5, "w_hypo": 0.3, "w_dose": 0.2, "max_hypo_pct": 4, "tir_tolerance": 3, "n_days": 2, "seed": 1}
    },
    {
      "label": "Hypo-safety-first weights — does your pick change?",
      "why": "Shifting most weight onto r_hypo should push the scalarized optimum toward the suspend-rule corner of the Pareto front and bring hypo time inside the 4% limit, demonstrating how the trade-off attribution shifts with the designer’s choice.",
      "params": {"patient": "child#001", "w_glycemic": 0.1, "w_hypo": 0.8, "w_dose": 0.1, "max_hypo_pct": 4, "tir_tolerance": 3, "n_days": 2, "seed": 1}
    },
    {
      "label": "Tighter TIR tolerance — how many pairs survive at ±1 point?",
      "why": "Reducing tir_tolerance to 1 requires rules to be within 1 TIR point to count as ‘same’, giving a stricter test of the paper’s claim. If large hypo gaps persist even at ±1 point, the finding is more alarming.",
      "params": {"patient": "child#001", "w_glycemic": 0.5, "w_hypo": 0.3, "w_dose": 0.2, "max_hypo_pct": 4, "tir_tolerance": 1, "n_days": 2, "seed": 1}
    }
  ],
  "caveats": [
    "All patients are simulated UVa/Padova virtual subjects run through the simglucose Python package — not real people. Simglucose’s meal scenarios and CGM noise model are synthetic; real physiology is more variable.",
    "This runbook replaces the paper’s trained SAC agents and preference-conditioned Q-network (envelope MORL) with an enumerated family of 48 simple basal-only rules. ‘Your pick’ is the oracle scalarized optimum within that family, not a trained policy; the matched-pair gaps reflect rule variety, not policy learning.",
    "Episodes are one simulated day with random meals and no meal boluses; real closed-loop systems also deliver correction boluses at meals, which would change TIR and hypo profiles substantially.",
    "This is an interactive companion to the SMASH 2026 paper and is not clinical guidance. Do not use these results to inform real insulin dosing decisions."
  ]
}