"""
Metabolic Forensics — N-of-1 biosignal evidence engine.

Skeleton on synthetic data. The deterministic evidence pipeline is real;
the LLM narration layer is stubbed (plug in a local llama.cpp model where
marked). No personal data ships in this repo — privacy by construction.
"""
import numpy as np
import pandas as pd
import gradio as gr

# ---------------------------------------------------------------------------
# Synthetic demo data — stands in for ~/Health/data/*.jsonl (CGM/HRV/recovery)
# Real version: load the user's Ultrahuman OAuth metrics locally, never in repo.
# ---------------------------------------------------------------------------
def make_demo_data(days: int = 46, seed: int = 7) -> pd.DataFrame:
    rng = np.random.default_rng(seed)
    dates = pd.date_range("2026-04-19", periods=days, freq="D")
    late_meal = rng.random(days) < 0.35          # hidden "cause"
    glucose_spike = 110 + late_meal * 45 + rng.normal(0, 8, days)   # mg/dL peak
    # next-day HRV suffers after a spike (with noise + counterexamples)
    hrv = 55 - late_meal * 9 + rng.normal(0, 6, days)
    recovery = np.clip(70 - late_meal * 18 + rng.normal(0, 10, days), 1, 100)
    alertness = np.clip(75 - late_meal * 15 + rng.normal(0, 12, days), 1, 100)
    return pd.DataFrame({
        "date": dates,
        "late_meal": late_meal.astype(int),
        "glucose_peak": glucose_spike.round(0),
        "hrv": hrv.round(0),
        "recovery": recovery.round(0),
        "morning_alertness": alertness.round(0),
    })

DF = make_demo_data()

# ---------------------------------------------------------------------------
# THE EVIDENCE ENGINE (deterministic — this is the moat, not the model)
# Detect an event, split days into event vs non-event, report the contrast
# AND the counterexamples (days where the pattern breaks).
# ---------------------------------------------------------------------------
QUESTIONS = {
    "What precedes my low-recovery mornings?": ("recovery", "low", "late_meal"),
    "What precedes my glucose spikes?":        ("glucose_peak", "high", "late_meal"),
    "What precedes my low-alertness mornings?":("morning_alertness", "low", "late_meal"),
}

def forensics(question: str):
    metric, direction, candidate = QUESTIONS[question]
    df = DF.copy()
    thr = df[metric].quantile(0.25 if direction == "low" else 0.75)
    event = df[metric] <= thr if direction == "low" else df[metric] >= thr

    with_cand = event & (df[candidate] == 1)
    base_rate = df[candidate].mean()
    event_rate = df.loc[event, candidate].mean()
    lift = (event_rate / base_rate) if base_rate else float("nan")

    # counterexamples: event days where the candidate cause was ABSENT
    counterex = df[event & (df[candidate] == 0)]

    evidence = {
        "question": question,
        "n_days": len(df),
        "n_event_days": int(event.sum()),
        "candidate_signal": candidate,
        "candidate_present_on_event_days_pct": round(100 * event_rate, 0),
        "candidate_baseline_pct": round(100 * base_rate, 0),
        "lift": round(lift, 2),
        "n_counterexamples": len(counterex),
    }

    # ---- LLM NARRATION LAYER (stub) -------------------------------------
    # Real version: hand `evidence` to a local llama.cpp ≤32B model with a
    # strict prompt: narrate ONLY this evidence, surface the counterexamples,
    # propose ONE testable experiment. Cannot invent associations.
    narration = (
        f"**Observed association** — on your {evidence['n_event_days']} "
        f"'{question.split('my ')[-1].rstrip('?')}' days, "
        f"`{candidate}` was present {evidence['candidate_present_on_event_days_pct']:.0f}% "
        f"of the time vs a {evidence['candidate_baseline_pct']:.0f}% baseline "
        f"(**{evidence['lift']}× lift**).\n\n"
        f"**Counterexamples** — but {evidence['n_counterexamples']} of those days had "
        f"NO `{candidate}`, so this is a tendency, not a law. Don't overfit.\n\n"
        f"**Next experiment** — deliberately vary `{candidate}` for 7 days and watch "
        f"whether '{metric}' separates. That turns correlation into something testable."
    )
    plot_df = df[["date", metric]].rename(columns={metric: "value"})
    return narration, plot_df, evidence

with gr.Blocks(title="Metabolic Forensics", theme=gr.themes.Soft()) as demo:
    gr.Markdown(
        "# 🩸 Metabolic Forensics\n"
        "*N-of-1 biosignal **evidence engine** — forensics, not coaching. "
        "Running on synthetic demo data.*"
    )
    q = gr.Dropdown(list(QUESTIONS), value=list(QUESTIONS)[0], label="Forensic question")
    btn = gr.Button("Investigate", variant="primary")
    out_md = gr.Markdown()
    out_plot = gr.LinePlot(x="date", y="value", label="Metric over your history")
    out_json = gr.JSON(label="Raw evidence (what the model is allowed to narrate)")
    btn.click(forensics, q, [out_md, out_plot, out_json])

if __name__ == "__main__":
    demo.launch()