#!/usr/bin/env python3
"""Recalculate published Mac summaries, using Python 3 standard libraries only.

Download this file and both CSVs into one directory, then run:
    python3 recalculate-mac.py
Or supply their paths in Mini, Air order. No network calls or file writes occur.
This verifies arithmetic in the supplied exports, not measurement provenance,
answer quality, independent replication or performance on other tasks.
"""

import csv
import json
import math
from pathlib import Path
import statistics
import sys


def load(path):
    rows = {}
    with Path(path).open(encoding="utf-8-sig", newline="") as handle:
        for row in csv.DictReader(handle):
            # Mini's aggregate footer is not a request. Do not count it twice.
            if row.get("profile") == "TOTAL PROMPT TIME" and not row.get("prompt_id"):
                continue
            key = (row["profile"], row["prompt_id"])
            if not all(key) or key in rows:
                raise ValueError("Missing or duplicate profile/prompt identity")
            values = {}
            for field in ("output_tps", "ttft_ms", "total_request_time_s", "output_tokens"):
                value = float(row[field])
                if not math.isfinite(value) or value < 0:
                    raise ValueError("Invalid numeric measurement: " + field)
                values[field] = value
            for field in ("completed", "truncated"):
                if row[field].lower() not in ("true", "false"):
                    raise ValueError("Invalid Boolean measurement: " + field)
                values[field] = row[field].lower() == "true"
            values["prompt"] = " ".join(row["prompt"].split())
            rows[key] = values
    if len(rows) != 24:
        raise ValueError("Expected the 24 published request rows per device")
    return rows


def summarize(rows):
    return {
        "request_count": len(rows),
        "mean_output_tokens_per_second": statistics.mean(r["output_tps"] for r in rows.values()),
        "mean_time_to_first_token_ms": statistics.mean(r["ttft_ms"] for r in rows.values()),
        "summed_request_seconds": sum(r["total_request_time_s"] for r in rows.values()),
        "output_tokens": sum(int(r["output_tokens"]) for r in rows.values()),
        "completed_flags": sum(r["completed"] for r in rows.values()),
        "truncated_flags": sum(r["truncated"] for r in rows.values()),
    }


def recalculate(mini_path, air_path):
    mini, air = load(mini_path), load(air_path)
    if mini.keys() != air.keys() or any(mini[k]["prompt"] != air[k]["prompt"] for k in mini):
        raise ValueError("The two files do not contain matched prompts")
    a, b = summarize(mini), summarize(air)
    return {
        "mac_mini": a,
        "macbook_air": b,
        "ratio_of_unrounded_mean_output_rates": a["mean_output_tokens_per_second"] / b["mean_output_tokens_per_second"],
        "mini_faster_output_pairs": sum(mini[k]["output_tps"] > air[k]["output_tps"] for k in mini),
        "mini_lower_ttft_pairs": sum(mini[k]["ttft_ms"] < air[k]["ttft_ms"] for k in mini),
        "scope": "One run of each of 24 varied prompts per device; not repeated trials or independently scored task completion.",
    }


if __name__ == "__main__":
    directory = Path(__file__).resolve().parent
    paths = sys.argv[1:] or [directory / "macmini_20260929_141518_all_profiles.csv", directory / "macair_20260929_131947_Vinci-Piccolo-1.0-GGUF_Q4_K_M.csv"]
    if len(paths) != 2:
        raise SystemExit("Usage: python3 recalculate-mac.py [MINI.csv AIR.csv]")
    try:
        print(json.dumps(recalculate(*paths), indent=2, allow_nan=False))
    except (OSError, ValueError, KeyError, ZeroDivisionError) as error:
        raise SystemExit("Cannot calculate this comparison: " + str(error)) from error
