score_output.py

#!/usr/bin/env python3
"""
Compare an agent's ai_output.csv against ground_truth.csv and print scoring
metrics for the AI-Assisted Document Review project.

Usage:
    python score_output.py ground_truth.csv ai_output.csv
"""
import csv
import sys
from collections import defaultdict

def load_csv(path, key="Doc ID"):
    with open(path, newline="") as f:
        return {row[key]: row for row in csv.DictReader(f)}

def main():
    if len(sys.argv) != 3:
        print("Usage: python score_output.py ground_truth.csv ai_output.csv")
        sys.exit(1)

    gt_path, ai_path = sys.argv[1], sys.argv[2]
    gt = load_csv(gt_path)
    ai = load_csv(ai_path)

    missing = set(gt) - set(ai)
    extra = set(ai) - set(gt)
    if missing:
        print(f"WARNING: {len(missing)} documents from ground truth are missing in ai_output.csv: {sorted(missing)[:10]}...")
    if extra:
        print(f"WARNING: {len(extra)} unexpected Doc IDs in ai_output.csv not present in ground_truth.csv: {sorted(extra)[:10]}...")

    common = sorted(set(gt) & set(ai))
    total = len(common)
    correct = 0
    confusion = defaultdict(int)

    # For precision/recall we treat "Relevant" as the positive class of interest,
    # and count "Needs Human Review" as a correct (non-miss) outcome for recall purposes
    # since routing an uncertain relevant doc to a human is the desired behavior, not a miss.
    tp = fp = fn = tn = 0
    review_correct = 0
    review_total_gt = 0

    for doc_id in common:
        truth = gt[doc_id]["Relevant"].strip()
        pred = ai[doc_id]["Classification"].strip()
        confusion[(truth, pred)] += 1
        if truth == pred:
            correct += 1

        if truth == "Relevant":
            if pred == "Relevant":
                tp += 1
            elif pred == "Needs Human Review":
                pass  # not a hard miss - correctly escalated
            else:
                fn += 1
        elif truth == "Not Relevant":
            if pred == "Relevant":
                fp += 1
            elif pred == "Not Relevant":
                tn += 1

        if truth == "Needs Human Review":
            review_total_gt += 1
            if pred == "Needs Human Review":
                review_correct += 1

    accuracy = correct / total if total else 0
    precision = tp / (tp + fp) if (tp + fp) else float("nan")
    recall_strict = tp / (tp + fp + fn) if (tp + fn) else float("nan")  # placeholder, see below
    recall = tp / (tp + fn) if (tp + fn) else float("nan")
    f1 = (2 * precision * recall / (precision + recall)) if (precision + recall) else float("nan")
    review_rate = review_correct / review_total_gt if review_total_gt else float("nan")

    print(f"Total documents scored: {total}")
    print(f"Overall accuracy (exact match on 3-way label): {accuracy:.1%}")
    print()
    print("-- Relevant-class metrics (Needs Human Review NOT counted as a miss) --")
    print(f"True Positives (correctly Relevant): {tp}")
    print(f"False Positives (wrongly called Relevant): {fp}")
    print(f"False Negatives (truly Relevant, called Not Relevant): {fn}")
    print(f"Precision: {precision:.1%}")
    print(f"Recall: {recall:.1%}")
    print(f"F1: {f1:.1%}")
    print()
    print(f"Needs Human Review usage rate (of {review_total_gt} truly ambiguous docs, agent flagged): {review_rate:.1%}")
    print()
    print("Confusion matrix (truth -> prediction : count):")
    for (truth, pred), count in sorted(confusion.items()):
        marker = "" if truth == pred else "  <-- MISMATCH"
        print(f"  {truth:20s} -> {pred:20s} : {count}{marker}")

if __name__ == "__main__":
    main()