#!/usr/bin/env python3
"""Summarize the beta-independent Rain hardware-bias development screen."""

from __future__ import annotations

import json
from pathlib import Path


ROOT = Path(__file__).resolve().parents[1]
RESULT_ROOT = ROOT / "results" / "ep_bias" / "s3"
FILES = {
    "clean": "clean-e3-s1988.json",
    "raw": "raw-free2e-4-e3-s1988.json",
    "constant": "constant-free2e-4-e3-s1988.json",
    "innovation": "innovation-free2e-4-e3-s1988.json",
    "random_beta": "random-beta-free2e-4-e3-s1988.json",
    "random_beta_zero_bias": "random-beta-zero-bias-e3-s1988.json",
    "fixed_negative_zero_bias": "fixed-negative-zero-bias-e3-s1988.json",
}


def read(name: str) -> dict:
    with (RESULT_ROOT / FILES[name]).open(encoding="utf-8") as handle:
        return json.load(handle)


def main() -> None:
    records = {name: read(name) for name in FILES}
    rows = {
        name: {
            "final_validation_accuracy": record["final"]["test_accuracy"],
            "wall_seconds": record["final"]["wall_seconds"],
            "finite": all(metric["finite"] for metric in record["metrics"]),
            "final_corrector": record["final"].get("corrector", {}),
            "beta_sign_counts": record.get("beta_sign_counts", {}),
        }
        for name, record in records.items()
    }
    clean = rows["clean"]["final_validation_accuracy"]
    raw = rows["raw"]["final_validation_accuracy"]
    constant = rows["constant"]["final_validation_accuracy"]
    innovation = rows["innovation"]["final_validation_accuracy"]
    report = {
        "stage": "rain_ep_beta_independent_hardware_bias_s3",
        "status": "positive_single_seed_development_negative_beta_unavailable",
        "bias_model": {
            "normalization": "initial_free_layer_state_rms",
            "ratio": 0.0002,
            "depends_on_beta_sign": False,
            "depends_on_clean_task_difference": False,
        },
        "rows": rows,
        "effects": {
            "innovation_minus_raw_accuracy_points": 100.0 * (
                innovation - raw),
            "innovation_minus_constant_accuracy_points": 100.0 * (
                innovation - constant),
            "clean_minus_innovation_accuracy_points": 100.0 * (
                clean - innovation),
            "innovation_wall_over_clean": (
                rows["innovation"]["wall_seconds"]
                / rows["clean"]["wall_seconds"]),
            "innovation_final_residual_to_clean_state_difference_rms": (
                rows["innovation"]["final_corrector"]
                ["residual_to_clean_state_difference_rms"]),
            "constant_final_residual_to_clean_state_difference_rms": (
                rows["constant"]["final_corrector"]
                ["residual_to_clean_state_difference_rms"]),
        },
        "random_beta_boundary": {
            "random_sign_fails_with_zero_bias": (
                rows["random_beta_zero_bias"]["final_validation_accuracy"] < 0.2),
            "fixed_negative_fails_with_zero_bias": (
                rows["fixed_negative_zero_bias"]["final_validation_accuracy"] < 0.2),
            "interpretation": (
                "Negative-beta relaxation is not a valid drop-in baseline under "
                "the pinned author's positive-EP hyperparameters. Random-sign "
                "beta requires separate stability work before comparison; this "
                "screen is not an SDIL win over a functioning random-beta method."
            ),
        },
        "test_policy": "new development train holdout data seed 6200",
    }
    output = RESULT_ROOT.parent / "s3_summary.json"
    output.write_text(
        json.dumps(report, indent=2, sort_keys=True, allow_nan=False) + "\n")
    print(json.dumps(report, indent=2, sort_keys=True, allow_nan=False))


if __name__ == "__main__":
    main()