#!/usr/bin/env python3 """Export viewer tables from the recorded LHF v3 release; runs no experiments. Usage: python export_tables.py SOURCE_ROOT OUTPUT_DIR Both arguments are directories. SOURCE_ROOT contains the original data/ JSON files. Only scalar fields are exported; the original JSON remains authoritative. Uses Python's standard library and validates each CSV against its source values. """ from __future__ import annotations import argparse import csv import hashlib import json from pathlib import Path # Field name, CSV / Hugging Face type, units, meaning. SCHEMAS = { "neural_results.csv": [ ("seed", "int64", "none", "Training random seed; one of 7, 23, 61."), ("noise_norm", "float64", "latent-observation Euclidean norm", "Fixed norm of paired post-observation noise for each test sample and scenario."), ("method", "string", "none", "Transport contraction: uniform or optimized."), ("mean_scenario_accuracy", "float64", "fraction [0, 1]", "Mean test-label accuracy across eight observation scenarios."), ("worst_scenario_accuracy", "float64", "fraction [0, 1]", "Minimum test-label accuracy among the eight scenarios."), ("mean_reconstruction_RMSE", "float64", "latent-state Euclidean norm", "Mean across scenarios of sqrt(mean squared Euclidean state error across test samples); not per-coordinate RMSE."), ("maximum_state_error", "float64", "latent-state Euclidean norm", "Largest Euclidean state-reconstruction error across all tested samples and scenarios."), ("deterministic_error_radius", "float64", "latent-state Euclidean norm", "Recorded deterministic reconstruction bound noise_norm / sqrt(spectral information floor)."), ("certified_prediction_stability_fraction", "float64", "fraction [0, 1]", "Fraction of clean predictions certified unchanged under the prescribed radius; does not certify label correctness."), ("mean_prediction_stability", "float64", "fraction [0, 1]", "Mean across scenarios of the fraction of reconstructed predictions equal to clean predictions."), ], "neural_aggregates.csv": [ ("noise_norm", "float64", "latent-observation Euclidean norm", "Fixed norm of paired post-observation noise."), ("method", "string", "none", "Transport contraction: uniform or optimized."), ("mean_accuracy", "float64", "fraction [0, 1]", "Mean of mean_scenario_accuracy across the three training seeds."), ("seed_std", "float64", "accuracy fraction", "Sample standard deviation (ddof=1) of the three seeds' mean_scenario_accuracy."), ("mean_worst_scenario_accuracy", "float64", "fraction [0, 1]", "Mean across seeds of each seed's minimum scenario accuracy."), ("mean_reconstruction_RMSE", "float64", "latent-state Euclidean norm", "Mean across seeds of scenario-mean root-mean-squared Euclidean state error."), ("mean_certified_stability", "float64", "fraction [0, 1]", "Mean across seeds of certified_prediction_stability_fraction; does not certify label correctness."), ], "verification_groups.csv": [ ("name", "string", "none", "Recorded numerical-check group identifier."), ("status", "string", "none", "Status recorded in the original verification JSON; not a newly executed verification status."), ("cases", "int64", "cases", "Number of seeded cases recorded for the group."), ("max_residual", "float64", "group-dependent mathematical residual", "Largest residual recorded for this group; residual definitions differ by group and must not be compared as a common metric."), ("interpretation", "string", "none", "Original group interpretation, including empty strings when no interpretation was recorded."), ], "frontier_curve.csv": [ ("B", "float64", "dimensionless natural-log determinant budget", "Budget in |det(C)| = exp(-B) for the fixed two-dimensional noncommuting example."), ("exact", "float64", "spectral information floor in model coordinates", "Recorded float evaluation of the exact closed-form optimum; CSV contains the evaluation, not a symbolic expression."), ("computed", "float64", "spectral information floor in model coordinates", "Recorded numerical solver value for the same budget and example."), ("uniform", "float64", "spectral information floor in model coordinates", "Recorded uniform-contraction floor at the matched determinant budget."), ], "exact_checks.csv": [ ("identity", "string", "none", "Identifier from the original exact-certificate identities array."), ("recorded_status", "string", "none", "Original JSON's overall certificate-run status copied to each identity row. Individual identity statuses are not separately stored and were not independently rerun during packaging."), ("scope", "string", "none", "Scope statement copied verbatim from the original exact-certificate JSON."), ], } def load_json(path: Path) -> dict: with path.open(encoding="utf-8") as stream: value = json.load(stream) if not isinstance(value, dict): raise ValueError(f"Expected a JSON object: {path}") return value def scalar_projection(rows: list, filename: str) -> list[dict]: fields = [field[0] for field in SCHEMAS[filename]] projected = [] for index, source in enumerate(rows): row = {name: source[name] for name in fields} for name, type_name, _, _ in SCHEMAS[filename]: value = row[name] if type_name == "string": valid = isinstance(value, str) elif type_name == "int64": valid = isinstance(value, int) and not isinstance(value, bool) else: valid = isinstance(value, (float, int)) and not isinstance(value, bool) if not valid: raise TypeError(f"{filename}, row {index}, field {name}: invalid scalar type") projected.append(row) return projected def write_and_validate(path: Path, rows: list[dict]) -> dict: schema = SCHEMAS[path.name] fields = [field[0] for field in schema] with path.open("w", encoding="utf-8", newline="") as stream: writer = csv.DictWriter(stream, fieldnames=fields) writer.writeheader() writer.writerows(rows) with path.open(encoding="utf-8", newline="") as stream: reader = csv.DictReader(stream) if reader.fieldnames != fields: raise AssertionError(f"CSV header mismatch: {path.name}") parsed = list(reader) if len(parsed) != len(rows): raise AssertionError(f"CSV row-count mismatch: {path.name}") checked = 0 for index, (csv_row, source_row) in enumerate(zip(parsed, rows)): for name, type_name, _, _ in schema: value = csv_row[name] if type_name == "int64": value = int(value) elif type_name == "float64": value = float(value) if value != source_row[name]: raise AssertionError(f"CSV value mismatch: {path.name}, row {index}, {name}") checked += 1 return { "rows": len(rows), "columns": len(fields), "scalar_values_checked": checked, "validation": "PASS: parsed CSV values equal corresponding original JSON scalar values", } def dictionary_text(counts: dict) -> str: text = [ "# Viewer data dictionary", "", "These CSV files are packaging derivatives of recorded LHF v3.0.0 results. " "They add no experiments, recomputed outcomes, proof verification, or independent review. " "The original JSON, experiment code, saved arrays and manuscript remain the authoritative records.", "", "Numbers are written with Python's CSV writer without rounding or percentage conversion. " "Validation reads the CSV values back with their stated types and compares them to the " "original JSON-decoded scalar values. This preserves scalar numeric values, rather than the " "original JSON text formatting. Fractions such as 0.935 are accuracy proportions, not percentages.", "", "The neural study uses one fixed optical-digits split (1,257 training / 540 test samples), " "three training seeds, eight-dimensional frozen learned features, eight constructed observation " "scenarios, and six paired noise norms. Its transport comparison does not demonstrate " "general AI intelligence improvement or state-of-the-art superiority. Latent coordinates " "have no physical SI unit.", "", "Source mappings:", "", "| CSV | Original JSON field | Rows |", "|---|---|---:|", f"| neural_results.csv | data/neural_digits_results.json → records (scalar fields only) | {counts['neural_results.csv']} |", f"| neural_aggregates.csv | data/neural_digits_results.json → aggregates | {counts['neural_aggregates.csv']} |", f"| verification_groups.csv | data/mathematical_verification.json → groups | {counts['verification_groups.csv']} |", f"| frontier_curve.csv | data/mathematical_verification.json → frontier_curve | {counts['frontier_curve.csv']} |", f"| exact_checks.csv | data/exact_certificates.json → identities, overall status and scope | {counts['exact_checks.csv']} |", "", "The nested scenario_accuracies list is deliberately omitted from neural_results.csv; " "all original per-scenario values remain in data/neural_digits_results.json. " "The aggregate table copies the source aggregates; it does not recalculate them.", "", "The exact-check JSON records one overall PASS and twenty identity labels. " "Its recorded_status is therefore copied from the overall run to all twenty rows. " "These rows are not twenty independently rerun checks and do not supply proof-assistant verification.", "", ] for filename, schema in SCHEMAS.items(): text.extend([ f"## {filename}", "", "| Field | Type | Units | Definition |", "|---|---|---|---|", ]) for name, type_name, units, definition in schema: escaped_units = units.replace("|", "\\|") escaped_definition = definition.replace("|", "\\|") text.append(f"| {name} | {type_name} | {escaped_units} | {escaped_definition} |") text.append("") return "\n".join(text) def export_tables(source_root: Path, output_dir: Path) -> dict: paths = { "neural": source_root / "data" / "neural_digits_results.json", "verification": source_root / "data" / "mathematical_verification.json", "exact": source_root / "data" / "exact_certificates.json", } neural, verification, exact = (load_json(paths[name]) for name in ("neural", "verification", "exact")) if verification["group_count"] != len(verification["groups"]): raise AssertionError("Recorded verification group count disagrees with group list") if verification["total_cases"] != sum(group["cases"] for group in verification["groups"]): raise AssertionError("Recorded verification case total disagrees with group sum") if exact["checked_identities"] != len(exact["identities"]): raise AssertionError("Recorded exact-identity count disagrees with identity list") rows = { "neural_results.csv": scalar_projection(neural["records"], "neural_results.csv"), "neural_aggregates.csv": scalar_projection(neural["aggregates"], "neural_aggregates.csv"), "verification_groups.csv": scalar_projection(verification["groups"], "verification_groups.csv"), "frontier_curve.csv": scalar_projection(verification["frontier_curve"], "frontier_curve.csv"), "exact_checks.csv": scalar_projection([ {"identity": identity, "recorded_status": exact["status"], "scope": exact["scope"]} for identity in exact["identities"] ], "exact_checks.csv"), } output_dir.mkdir(parents=True, exist_ok=True) reports = {name: write_and_validate(output_dir / name, table) for name, table in rows.items()} counts = {name: report["rows"] for name, report in reports.items()} (output_dir / "DATA_DICTIONARY.md").write_text(dictionary_text(counts), encoding="utf-8") source_names = { "neural_results.csv": "data/neural_digits_results.json:records", "neural_aggregates.csv": "data/neural_digits_results.json:aggregates", "verification_groups.csv": "data/mathematical_verification.json:groups", "frontier_curve.csv": "data/mathematical_verification.json:frontier_curve", "exact_checks.csv": "data/exact_certificates.json:identities + overall status + scope", } for name, report in reports.items(): report["source"] = source_names[name] report["csv_sha256"] = hashlib.sha256((output_dir / name).read_bytes()).hexdigest() report = { "status": "PASS", "kind": "data-packaging derivative validation; no new experiment execution", "source_release": "3.0.0", "source_json_sha256": { str(path.relative_to(source_root)): hashlib.sha256(path.read_bytes()).hexdigest() for path in paths.values() }, "tables": reports, "recorded_numerical_case_total": verification["total_cases"], "total_exported_rows": sum(counts.values()), "total_scalar_values_checked": sum(entry["scalar_values_checked"] for entry in reports.values()), "omitted_nested_neural_record_field": "scenario_accuracies", "exact_status_note": "The source stores overall PASS, not per-identity status. recorded_status copies the overall status; identities were not independently rerun during packaging.", "numeric_preservation": "CSV writer serializes JSON-decoded numbers without rounding; parsed CSV numeric values equal originals.", } (output_dir / "table_export_validation.json").write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8") return report def main() -> None: parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("source_root", type=Path) parser.add_argument("output_dir", type=Path) args = parser.parse_args() report = export_tables(args.source_root.resolve(), args.output_dir.resolve()) print(json.dumps({"status": report["status"], "rows": {name: value["rows"] for name, value in report["tables"].items()}, "scalar_values_checked": report["total_scalar_values_checked"]}, indent=2)) if __name__ == "__main__": main()