tasks/task_hard.py

"""
tasks/task_hard.py -- Hard task: 4 variables, high noise, random domain.

The agent must discover a complex causal graph with high noise and a
tight budget. This is designed to challenge frontier models.

Grader returns 0.0-1.0 based on:
  - Hypothesis accuracy (40%)
  - Precision bonus (15%)
  - Efficiency bonus (15%)
  - Calibration (15%)
  - Avoiding contradiction penalty (15%)
"""

from __future__ import annotations

from typing import Any

TASK_HARD = {
    "id": "hard",
    "name": "Hard -- Complex Graph Under Noise",
    "description": (
        "Discover causal relationships among four variables under "
        "high noise (sigma=0.50) with a tight budget (8 steps). "
        "Requires strategic experiment design and careful reasoning."
    ),
    "difficulty": "hard",
    "reset_kwargs": {
        "noise_level": "high",
        "seed": 999,
    },
}


def grade_hard(episode_result: dict[str, Any]) -> float:
    """
    Grade a hard-task episode. Returns a score in [0.0, 1.0].
    """
    accuracy = episode_result.get("accuracy_score", 0.0)
    precision = episode_result.get("precision_bonus", 0.0)
    efficiency = episode_result.get("efficiency_bonus", 0.0)
    calibration = episode_result.get("calibration_score", 0.0)
    contradiction = episode_result.get("contradiction_penalty", 0.0)

    no_contradiction = 1.0 if contradiction >= 0.0 else 0.0

    raw = (
        0.40 * min(accuracy, 1.0)
        + 0.15 * min(precision / 0.10, 1.0)
        + 0.15 * min(efficiency / 0.15, 1.0)
        + 0.15 * min(calibration / 0.20, 1.0)
        + 0.15 * no_contradiction
    )

    return round(max(0.0, min(1.0, raw)), 4)