code/verify_anima_intervention_pilot.py

from __future__ import annotations

import argparse
import hashlib
import json
from pathlib import Path
import time

import torch
from safetensors.torch import load_file


def read_json(path: Path) -> dict:
    return json.loads(path.read_text(encoding="utf-8"))


def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument("--run-root", type=Path, required=True)
    parser.add_argument("--selection", choices=("pilot", "train"), default="pilot")
    args = parser.parse_args()
    started = time.perf_counter()
    manifest_path = args.run_root / "manifest.jsonl"
    output_dir = args.run_root / f"anima_{args.selection}_features"
    expected_records = 4_096 if args.selection == "pilot" else 104_000
    expected_kind = f"factor_intervention_anima_{args.selection}"
    manifest_sha256 = hashlib.sha256(manifest_path.read_bytes()).hexdigest()
    with manifest_path.open(encoding="utf-8") as handle:
        manifest = {
            row["record_id"]: row
            for line in handle
            if line.strip()
            for row in [json.loads(line)]
            if (
                row.get("anima_pilot")
                if args.selection == "pilot"
                else row.get("split") == "train"
            )
        }
    errors: list[str] = []
    summaries = []
    for worker in range(4):
        try:
            summary = read_json(output_dir / f"anima-worker-{worker}.json")
        except (FileNotFoundError, OSError, json.JSONDecodeError) as error:
            errors.append(f"worker {worker} summary: {error}")
            continue
        summaries.append(summary)
        if summary.get("status") != "complete":
            errors.append(f"worker {worker} is incomplete")
        if summary.get("manifest_sha256") != manifest_sha256:
            errors.append(f"worker {worker} manifest hash mismatch")
    if sum(int(row.get("records", 0)) for row in summaries) != expected_records:
        errors.append("worker row total mismatch")

    json_parts = sorted(output_dir.glob("anima-w*-p*.json"))
    tensor_parts = sorted(output_dir.glob("anima-w*-p*.safetensors"))
    if {path.stem for path in json_parts} != {
        path.name.removesuffix(".safetensors") for path in tensor_parts
    }:
        errors.append("part JSON/safetensors pairing mismatch")
    seen: set[str] = set()
    tensor_bytes = 0
    for metadata_path in json_parts:
        metadata = read_json(metadata_path)
        records = metadata.get("records", [])
        expected_contract = {
            "kind": expected_kind,
            "blocks": [8, 18, 26],
            "sigma": 0.1,
            "noise_seed": 20260715,
            "preprocess_version": "square-cover-v1",
            "transform_resolution": 768,
            "manifest_sha256": manifest_sha256,
        }
        for key, expected in expected_contract.items():
            if metadata.get(key) != expected:
                errors.append(f"{metadata_path.name}: {key} contract mismatch")
        tensor_path = metadata_path.with_suffix(".safetensors")
        tensors = load_file(tensor_path, device="cpu")
        tensor_bytes += tensor_path.stat().st_size
        features = tensors.get("features")
        if set(tensors) != {"features"} or features is None:
            errors.append(f"{tensor_path.name}: feature key mismatch")
            continue
        if features.shape != (len(records), 3, 4096) or features.dtype != torch.bfloat16:
            errors.append(f"{tensor_path.name}: tensor contract mismatch")
        elif not torch.isfinite(features).all().item():
            errors.append(f"{tensor_path.name}: non-finite features")
        for row in records:
            record_id = str(row.get("record_id"))
            if record_id in seen:
                errors.append(f"duplicate record ID: {record_id}")
            seen.add(record_id)
            expected = manifest.get(record_id)
            if expected is None:
                errors.append(f"record absent from pilot manifest: {record_id}")
                continue
            for field in (
                "source_record_id",
                "style_id",
                "source",
                "split",
                "shard",
                "factor",
                "family",
                "level",
                "sign",
                "signed_intensity",
                "operation_seed",
                "transform_version",
            ):
                if row.get(field) != expected.get(field):
                    errors.append(f"{record_id}: {field} alignment mismatch")
                    break
            if row.get("source_shard") != row.get("shard"):
                errors.append(f"{record_id}: source shard mismatch")
    missing = set(manifest) - seen
    extra = seen - set(manifest)
    if missing or extra:
        errors.append(f"record coverage mismatch: missing={len(missing)}, extra={len(extra)}")
    report = {
        "status": "pass" if not errors else "fail",
        "selection": args.selection,
        "manifest_sha256": manifest_sha256,
        "records": len(seen),
        "parts": len(json_parts),
        "tensor_shape": [3, 4096],
        "tensor_dtype": "bfloat16",
        "tensor_bytes": tensor_bytes,
        "blocks": [8, 18, 26],
        "sigma": 0.1,
        "noise_seed": 20260715,
        "transform_resolution": 768,
        "worker_summaries": summaries,
        "elapsed_seconds": time.perf_counter() - started,
        "errors": errors[:100],
    }
    output = args.run_root / f"anima_{args.selection}_verification.json"
    temporary = output.with_suffix(".json.tmp")
    temporary.write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8")
    temporary.replace(output)
    print(json.dumps(report, indent=2))
    return 0 if not errors else 1


if __name__ == "__main__":
    raise SystemExit(main())