"""Generate the fake learner responses in data/responses.csv.

Nobody real took this quiz. Each simulated learner gets an ability score,
and each item gets a difficulty and a discrimination value, using a standard
three-parameter logistic (3PL) model. A few items have problems planted on
purpose so the analysis has something to find:

  Q01  far too easy (almost everyone gets it right)
  Q05  ambiguous: stronger learners who know about mid-conversation system
       messages are drawn to option A, so the item barely separates learners
  Q08  miskeyed: the answer key in quiz.json says C, but B is the real answer
  Q12  option D is a distractor almost nobody picks

Run it again with the same seed and you get the same file.
"""

import csv
import json
import math
import random
from pathlib import Path

SEED = 2026
N_LEARNERS = 80
ROOT = Path(__file__).resolve().parent.parent

# (true answer, difficulty b, discrimination a, distractor weights)
# Distractor weights say how attractive each wrong option is.
ITEM_PARAMS = {
    "Q01": ("B", -2.4, 1.0, {"A": 3, "C": 2, "D": 2}),
    "Q02": ("B", -0.6, 1.3, {"A": 3, "C": 3, "D": 3}),
    "Q03": ("B", 0.2, 1.4, {"A": 5, "C": 1, "D": 2}),
    "Q04": ("B", 0.4, 1.1, {"A": 3, "C": 3, "D": 1}),
    "Q05": None,  # simulated separately below
    "Q06": ("B", 0.0, 1.5, {"A": 4, "C": 3, "D": 2}),
    "Q07": ("A", 0.8, 1.2, {"B": 5, "C": 2, "D": 1}),
    "Q08": ("B", 0.1, 1.3, {"A": 2, "C": 5, "D": 2}),
    "Q09": ("A", 1.0, 1.0, {"B": 5, "C": 2, "D": 3}),
    "Q10": ("B", -0.8, 1.2, {"A": 3, "C": 3, "D": 2}),
    "Q11": ("D", 0.3, 1.3, {"A": 2, "B": 4, "C": 3}),
    "Q12": ("B", -0.3, 1.2, {"A": 5, "C": 3, "D": 0.05}),
}


def p_correct(theta, a, b, guess=0.2):
    """3PL probability that a learner with ability theta answers correctly."""
    return guess + (1 - guess) / (1 + math.exp(-1.7 * a * (theta - b)))


def pick_weighted(rng, weights):
    options = list(weights)
    return rng.choices(options, weights=[weights[o] for o in options])[0]


def answer_q05(rng, theta):
    # Learners who know the API well hesitate between A and B, because
    # role "system" messages are allowed later in the messages array.
    # Weaker learners mostly guess.
    pull_to_a = 1 / (1 + math.exp(-2.0 * (theta - 0.2)))
    if rng.random() < 0.55 * pull_to_a:
        return "A"
    if rng.random() < 0.5:
        return "B"
    return pick_weighted(rng, {"A": 2, "C": 2, "D": 2})


def main():
    rng = random.Random(SEED)
    quiz = json.loads((ROOT / "data" / "quiz.json").read_text())
    item_ids = [item["id"] for item in quiz["items"]]

    rows = []
    for n in range(1, N_LEARNERS + 1):
        theta = rng.gauss(0, 1)
        row = {"learner_id": f"L{n:03d}"}
        for item_id in item_ids:
            if item_id == "Q05":
                row[item_id] = answer_q05(rng, theta)
                continue
            truth, b, a, distractors = ITEM_PARAMS[item_id]
            if rng.random() < p_correct(theta, a, b):
                row[item_id] = truth
            else:
                row[item_id] = pick_weighted(rng, distractors)
        rows.append(row)

    out = ROOT / "data" / "responses.csv"
    with out.open("w", newline="") as f:
        writer = csv.DictWriter(f, fieldnames=["learner_id", *item_ids])
        writer.writeheader()
        writer.writerows(rows)
    print(f"Wrote {len(rows)} learners to {out.relative_to(ROOT)}")


if __name__ == "__main__":
    main()