#!/usr/bin/env python3
"""Checks simple language-regression records produced by an evaluator.

Input JSONL: {"language":"es", "ok":true, "tokens":42, "text":"..."}
This is deliberately a checker, not an LLM judge. Keep raw prompts/results
outside Git if they contain private data.
"""
import json
import sys
from collections import defaultdict
from pathlib import Path


def report(path: Path) -> dict:
    rows = defaultdict(list)
    for raw in path.read_text().splitlines():
        if raw.strip():
            item = json.loads(raw)
            rows[item["language"]].append(item)
    out = {}
    for language, items in sorted(rows.items()):
        out[language] = {
            "n": len(items),
            "pass_rate": sum(bool(x["ok"]) for x in items) / len(items),
            "mean_tokens": sum(int(x.get("tokens", 0)) for x in items) / len(items),
        }
    return out


if __name__ == "__main__":
    if len(sys.argv) != 2:
        raise SystemExit("usage: language_regression.py RESULTS.jsonl")
    print(json.dumps(report(Path(sys.argv[1])), ensure_ascii=False, indent=2))
