scorer.py

import argparse
import json
import pandas as pd
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score


ALLOWED_SPLITS = {
    "train",
    "in_domain_test",
    "boundary_trap",
    "distribution_shift",
    "counterfactual_intervention"
}

ALLOWED_TRAPS = {
    "false_stability",
    "boundary_masking",
    "trajectory_aliasing",
    "temporal_alias",
    "intervention_decoy"
}

ALLOWED_DIFFICULTY = {
    "easy",
    "medium",
    "hard"
}

LOW_BOUNDARY_THRESHOLD = 0.10
HIGH_DRIFT_THRESHOLD = 0.05


def load_csv(path):
    return pd.read_csv(path)


def validate_columns(preds, truth):

    if "scenario_id" not in preds.columns or "prediction" not in preds.columns:
        raise ValueError("Predictions must contain scenario_id and prediction columns")

    required_truth = {
        "scenario_id",
        "split_type",
        "pair_id",
        "pair_role",
        "difficulty_level",
        "pressure_obs_t0",
        "pressure_obs_t1",
        "pressure_obs_t2",
        "buffer_obs_t0",
        "buffer_obs_t1",
        "buffer_obs_t2",
        "true_label",
        "trap_type",
        "trap_active",
        "boundary_distance",
        "drift_gradient",
        "drift_acceleration",
        "recovery_feasibility",
        "regime_competition_ratio",
        "intervention_action",
        "intervention_magnitude",
        "boundary_distance_before",
        "boundary_distance_after",
        "intervention_effect_direction"
    }

    missing = required_truth - set(truth.columns)

    if missing:
        raise ValueError(f"Truth missing required columns: {sorted(missing)}")


def validate_values(preds, truth):

    bad_pred = set(preds["prediction"].dropna().unique()) - {0, 1}
    if bad_pred:
        raise ValueError("Predictions must contain only 0 or 1")

    bad_label = set(truth["true_label"].dropna().unique()) - {0, 1}
    if bad_label:
        raise ValueError("true_label must contain only 0 or 1")

    bad_trap = set(truth["trap_active"].dropna().unique()) - {0, 1}
    if bad_trap:
        raise ValueError("trap_active must contain only 0 or 1")

    unknown_splits = set(truth["split_type"].dropna().unique()) - ALLOWED_SPLITS
    if unknown_splits:
        raise ValueError(f"Unknown split_type values: {sorted(unknown_splits)}")

    trap_values = set(truth["trap_type"].dropna().unique())
    unknown_traps = trap_values - ALLOWED_TRAPS
    if unknown_traps:
        raise ValueError(f"Unknown trap_type values: {sorted(unknown_traps)}")

    unknown_difficulty = set(truth["difficulty_level"].dropna().unique()) - ALLOWED_DIFFICULTY
    if unknown_difficulty:
        raise ValueError(f"Unknown difficulty_level values: {sorted(unknown_difficulty)}")

    if "intervention_effect_direction" in preds.columns:
        bad_dir = set(preds["intervention_effect_direction"].dropna().unique()) - {-1, 0, 1}
        if bad_dir:
            raise ValueError("Predicted intervention_effect_direction must be one of -1, 0, 1")

    valid_pair_roles = {"safe_pair", "unstable_pair"}
    pair_roles = set(truth["pair_role"].dropna().unique()) - valid_pair_roles
    if pair_roles:
        raise ValueError(f"Unknown pair_role values: {sorted(pair_roles)}")


def validate_ids(df, name):

    if df["scenario_id"].isna().any():
        raise ValueError(f"{name} contains missing scenario_id values")

    ids = df["scenario_id"].astype(str).str.strip()

    if (ids == "").any():
        raise ValueError(f"{name} contains blank scenario_id values")

    if ids.duplicated().any():
        dupes = ids[ids.duplicated()].unique().tolist()
        raise ValueError(f"{name} duplicate scenario_id values: {dupes[:10]}")

    df = df.copy()
    df["scenario_id"] = ids

    return df


def compute_basic_metrics(y_true, y_pred):

    return {
        "accuracy": float(accuracy_score(y_true, y_pred)),
        "precision": float(precision_score(y_true, y_pred, zero_division=0)),
        "recall": float(recall_score(y_true, y_pred, zero_division=0)),
        "f1": float(f1_score(y_true, y_pred, zero_division=0)),
        "rows_evaluated": int(len(y_true))
    }


def compute_split_accuracy(df, split_name):

    subset = df[df["split_type"] == split_name]

    if subset.empty:
        return None

    return float(accuracy_score(subset["true_label"], subset["prediction"]))


def compute_trap_accuracy(df, trap_type=None):

    subset = df[df["trap_active"] == 1]

    if trap_type:
        subset = subset[subset["trap_type"] == trap_type]

    if subset.empty:
        return None

    return float(accuracy_score(subset["true_label"], subset["prediction"]))


def compute_difficulty_accuracy(df, difficulty_level):

    subset = df[df["difficulty_level"] == difficulty_level]

    if subset.empty:
        return None

    return float(accuracy_score(subset["true_label"], subset["prediction"]))


def compute_geometry_diagnostics(df):

    results = {}

    low_boundary = df[
        (df["true_label"] == 1) &
        (df["boundary_distance"] <= LOW_BOUNDARY_THRESHOLD)
    ]

    if low_boundary.empty:
        results["low_boundary_distance_miss_rate"] = None
    else:
        misses = (low_boundary["prediction"] == 0).sum()
        results["low_boundary_distance_miss_rate"] = float(misses / len(low_boundary))

    high_drift = df[
        (df["true_label"] == 1) &
        (df["drift_gradient"] >= HIGH_DRIFT_THRESHOLD)
    ]

    if high_drift.empty:
        results["high_drift_gradient_miss_rate"] = None
    else:
        misses = (high_drift["prediction"] == 0).sum()
        results["high_drift_gradient_miss_rate"] = float(misses / len(high_drift))

    return results


def compute_counterfactual_metrics(df):

    results = {
        "counterfactual_intervention_accuracy": None,
        "intervention_effect_direction_accuracy": None
    }

    subset = df[df["split_type"] == "counterfactual_intervention"]

    if subset.empty:
        return results

    results["counterfactual_intervention_accuracy"] = float(
        accuracy_score(subset["true_label"], subset["prediction"])
    )

    if "predicted_intervention_effect_direction" in subset.columns:
        valid = subset.dropna(subset=["intervention_effect_direction", "predicted_intervention_effect_direction"])

        if not valid.empty:
            results["intervention_effect_direction_accuracy"] = float(
                accuracy_score(
                    valid["intervention_effect_direction"],
                    valid["predicted_intervention_effect_direction"]
                )
            )

    return results


def compute_pair_discrimination_accuracy(df):

    paired = df[df["pair_id"].notna()].copy()

    if paired.empty:
        return None

    correct = 0
    total = 0

    for pair_id, group in paired.groupby("pair_id"):
        if len(group) != 2:
            continue

        roles = set(group["pair_role"])
        if roles != {"safe_pair", "unstable_pair"}:
            continue

        unstable_row = group[group["pair_role"] == "unstable_pair"].iloc[0]
        safe_row = group[group["pair_role"] == "safe_pair"].iloc[0]

        ok = (unstable_row["prediction"] == 1) and (safe_row["prediction"] == 0)

        correct += int(ok)
        total += 1

    if total == 0:
        return None

    return float(correct / total)


def compute_support_counts(df):

    return {
        "train_support": int((df["split_type"] == "train").sum()),
        "in_domain_test_support": int((df["split_type"] == "in_domain_test").sum()),
        "boundary_trap_support": int((df["split_type"] == "boundary_trap").sum()),
        "distribution_shift_support": int((df["split_type"] == "distribution_shift").sum()),
        "counterfactual_intervention_support": int((df["split_type"] == "counterfactual_intervention").sum()),
        "trap_support": int((df["trap_active"] == 1).sum())
    }


def compute_casses_score(results):

    weights = {
        "in_domain_test_accuracy": 0.18,
        "boundary_trap_accuracy": 0.20,
        "distribution_shift_accuracy": 0.17,
        "trap_accuracy": 0.15,
        "counterfactual_intervention_accuracy": 0.15,
        "trajectory_pair_discrimination_accuracy": 0.15
    }

    score = 0
    weight_sum = 0

    for metric, weight in weights.items():
        value = results.get(metric)
        if value is not None:
            score += value * weight
            weight_sum += weight

    if weight_sum == 0:
        return None

    return score / weight_sum


def score(predictions_path, truth_path):

    preds = load_csv(predictions_path)
    truth = load_csv(truth_path)

    validate_columns(preds, truth)
    validate_values(preds, truth)

    preds = validate_ids(preds, "Predictions")
    truth = validate_ids(truth, "Truth")

    if "intervention_effect_direction" in preds.columns:
        preds = preds.rename(columns={
            "intervention_effect_direction": "predicted_intervention_effect_direction"
        })

    pred_ids = set(preds["scenario_id"])
    truth_ids = set(truth["scenario_id"])

    if pred_ids != truth_ids:
        missing = sorted(list(truth_ids - pred_ids))[:10]
        extra = sorted(list(pred_ids - truth_ids))[:10]
        raise ValueError(
            f"scenario_id mismatch. Missing in predictions: {missing}. Extra in predictions: {extra}"
        )

    merged = truth.merge(preds, on="scenario_id", how="inner", validate="one_to_one")
    merged = merged.sort_values("scenario_id").reset_index(drop=True)

    y_true = merged["true_label"]
    y_pred = merged["prediction"]

    results = compute_basic_metrics(y_true, y_pred)

    results["train_accuracy"] = compute_split_accuracy(merged, "train")
    results["in_domain_test_accuracy"] = compute_split_accuracy(merged, "in_domain_test")
    results["boundary_trap_accuracy"] = compute_split_accuracy(merged, "boundary_trap")
    results["distribution_shift_accuracy"] = compute_split_accuracy(merged, "distribution_shift")

    results["trap_accuracy"] = compute_trap_accuracy(merged)
    results["false_stability_accuracy"] = compute_trap_accuracy(merged, "false_stability")
    results["boundary_masking_accuracy"] = compute_trap_accuracy(merged, "boundary_masking")
    results["trajectory_aliasing_accuracy"] = compute_trap_accuracy(merged, "trajectory_aliasing")
    results["temporal_alias_accuracy"] = compute_trap_accuracy(merged, "temporal_alias")
    results["intervention_decoy_accuracy"] = compute_trap_accuracy(merged, "intervention_decoy")

    results["easy_accuracy"] = compute_difficulty_accuracy(merged, "easy")
    results["medium_accuracy"] = compute_difficulty_accuracy(merged, "medium")
    results["hard_accuracy"] = compute_difficulty_accuracy(merged, "hard")

    indomain = results["in_domain_test_accuracy"]
    shift = results["distribution_shift_accuracy"]

    if indomain is not None and shift is not None:
        results["manifold_generalization_gap"] = float(indomain - shift)
    else:
        results["manifold_generalization_gap"] = None

    results.update(compute_geometry_diagnostics(merged))
    results.update(compute_counterfactual_metrics(merged))
    results["trajectory_pair_discrimination_accuracy"] = compute_pair_discrimination_accuracy(merged)
    results.update(compute_support_counts(merged))
    results["casses_score"] = compute_casses_score(results)

    return results


def main():

    parser = argparse.ArgumentParser()
    parser.add_argument("--predictions", required=True)
    parser.add_argument("--truth", required=True)

    args = parser.parse_args()

    results = score(args.predictions, args.truth)

    print(json.dumps(results, indent=2))


if __name__ == "__main__":
    main()