app/training/reward_functions.py

"""Standalone reward functions with strict [0.001, 0.999] output."""

from __future__ import annotations

from app.common.normalization import clamp_reward


def format_compliance_score(valid: bool) -> float:
    """Schema validity: valid->0.999, invalid->0.001."""
    return clamp_reward(0.999 if valid else 0.001)


def candidate_alignment_score(aligned: bool) -> float:
    """Whether selected action references legal candidate set."""
    return clamp_reward(0.999 if aligned else 0.001)


def legality_score(legal: bool) -> float:
    """Hard constraint satisfaction score."""
    return clamp_reward(0.999 if legal else 0.001)


def safety_delta_score(delta: float) -> float:
    """Risk-delta mapping where positive delta means lower safety risk."""
    return clamp_reward(0.5 + delta * 0.4)


def burden_improvement_score(delta: float) -> float:
    """Burden reduction score; positive delta indicates lower burden."""
    return clamp_reward(0.5 + delta * 0.4)


def disease_stability_score(stability: float) -> float:
    """Stability proxy in [0,1], default caller-side imputation when missing."""
    return clamp_reward(stability)


def dosing_quality_score(quality: float) -> float:
    """Dose quality proxy in [0,1], neutral caller default for non-dose scenarios."""
    return clamp_reward(quality)


def abstention_quality_score(good_abstain: bool) -> float:
    """Judges abstention quality; not merely abstaining."""
    return clamp_reward(0.8 if good_abstain else 0.3)


def efficiency_score(step_fraction: float) -> float:
    """Shorter successful trajectories receive higher score."""
    return clamp_reward(1.0 - step_fraction)


def process_fidelity_score(fidelity: float) -> float:
    """Process-supervision score for valid clinical decision sequence."""
    return clamp_reward(fidelity)


def explanation_grounding_score(grounded: float) -> float:
    """Grounded explanation support score."""
    return clamp_reward(grounded)


def anti_cheat_score(exploit: bool) -> float:
    """Exploit-like behavior gets floor score."""
    return clamp_reward(0.001 if exploit else 0.999)


def uncertainty_calibration_score(calibration: float) -> float:
    """Confidence calibration score."""
    return clamp_reward(calibration)