#!/usr/bin/env python3 """Build the sealed, Hugging Face-ready FinGovBench v1 release candidate.""" from __future__ import annotations import hashlib import json from collections import Counter, defaultdict from pathlib import Path ROOT = Path(__file__).resolve().parents[1] SOURCES = [ ROOT / "data/sec_revisions_v1.jsonl", ROOT / "data/controlled_structural_v1.jsonl", ROOT / "data/delegated_action_v1.jsonl", ] def rows(path: Path) -> list[dict]: return [json.loads(line) for line in path.read_text().splitlines() if line.strip()] def sha256(path: Path) -> str: return hashlib.sha256(path.read_bytes()).hexdigest() def write_jsonl(path: Path, records: list[dict]) -> None: path.parent.mkdir(parents=True, exist_ok=True) path.write_text("".join(json.dumps(record, sort_keys=True) + "\n" for record in records)) def public_record(record: dict, *, include_telemetry: bool) -> dict: result = {key: value for key, value in record.items() if key not in {"gold", "post_action_observation"}} # Generator bookkeeping is never a method input. In particular, controlled # fault names and topology indices would leak the answer-key construction. result.pop("controlled_metadata", None) result["workflow_context"] = result.pop("ddas") graph = result.pop("graph") # The public object is a method-neutral workflow declaration. SSEG, # flat monitors, rules engines and other methods must build their own # internal representation from the same components and dependencies. result["workflow"] = { "representation": "declared_workflow_dependency_map_v1", "components": graph["nodes"], "dependencies": graph["edges"], } if not include_telemetry: result.pop("telemetry", None) # Construction flags are labels, not method inputs. result["event"] = { key: value for key, value in result["event"].items() if key != "controlled_fault" } result["label_status"] = "sealed" result["task_mode"] = "propagation" if include_telemetry else "closed_loop" return result def label_record(record: dict) -> dict: gold = dict(record["gold"]) gold["affected_components"] = gold.pop("affected_nodes") gold["revalidation_components"] = gold.pop("revalidation_scope") return { "case_id": record["case_id"], "event_group_id": record["event_group_id"], "split": record["split"], "gold": gold, "post_action_observation": record["post_action_observation"], "evaluation_strata": { "origin": record["case_origin"], "track": record["track"], "domain": record["domain"], "scenario": ( record.get("controlled_metadata", {}).get("counterfactual_archetype", "controlled") if record["case_origin"] == "controlled" else record.get("revision_metadata", {}).get("revision_kind", "reported_value_revision") ), }, "label_basis": ( "official_sec_reported_value_pair_and_declared_workflow_contract_pending_independent_adjudication" if record["case_origin"] == "genuine_revision" else "prespecified_controlled_event_and_declared_workflow_contract" ), } def main() -> None: records = [record for path in SOURCES for record in rows(path)] if len(records) != 3000: raise SystemExit(f"expected 3000 cases, found {len(records)}") ids = [record["case_id"] for record in records] if len(ids) != len(set(ids)): raise SystemExit("duplicate case IDs") group_splits: dict[str, set[str]] = defaultdict(set) authority_splits: dict[str, set[str]] = defaultdict(set) for record in records: group_splits[record["event_group_id"]].add(record["split"]) authority_splits[record["ddas"]["authority_id"]].add(record["split"]) leaking_groups = {group: sorted(splits) for group, splits in group_splits.items() if len(splits) > 1} leaking_authorities = {group: sorted(splits) for group, splits in authority_splits.items() if len(splits) > 1} if leaking_groups or leaking_authorities: raise SystemExit(f"split leakage: groups={leaking_groups}, authorities={leaking_authorities}") records.sort(key=lambda record: record["case_id"]) release = ROOT / "data/fingovbench_v1.jsonl" write_jsonl(release, records) for split in ("development", "validation", "test"): selected = [record for record in records if record["split"] == split] write_jsonl(ROOT / f"hf/closed_loop_inputs/{split}.jsonl", [public_record(record, include_telemetry=False) for record in selected]) write_jsonl(ROOT / f"hf/propagation_inputs/{split}.jsonl", [public_record(record, include_telemetry=True) for record in selected]) label_root = ROOT / ("private/sealed_labels" if split == "test" else "hf/public_labels") write_jsonl(label_root / f"{split}.jsonl", [label_record(record) for record in selected]) artifacts = [release, *sorted((ROOT / "hf/closed_loop_inputs").glob("*.jsonl")), *sorted((ROOT / "hf/propagation_inputs").glob("*.jsonl")), *sorted((ROOT / "hf/public_labels").glob("*.jsonl")), *sorted((ROOT / "private/sealed_labels").glob("*.jsonl"))] manifest = { "benchmark": "FinGovBench", "version": "1.0.0-rc4", "cases": len(records), "origins": dict(Counter(record["case_origin"] for record in records)), "domains": dict(Counter(record["domain"] for record in records)), "splits": dict(Counter(record["split"] for record in records)), "iceberg_visibility": dict(Counter(record["gold"]["iceberg_visibility"] for record in records)), "governance_actions": dict(Counter(record["gold"]["governance_action"] for record in records)), "event_groups": len(group_splits), "authorities": len(authority_splits), "group_leakage": leaking_groups, "authority_leakage": leaking_authorities, "source_tracks": [str(path.relative_to(ROOT)) for path in SOURCES], "checksums": {str(path.relative_to(ROOT)): sha256(path) for path in artifacts}, "label_status": "algorithmically verified; external domain-expert adjudication pending", "benchmark_method_dependency": "none; SSEG is an optional evaluated method", } (ROOT / "release_manifest.json").write_text(json.dumps(manifest, indent=2, sort_keys=True) + "\n") print(json.dumps(manifest, indent=2, sort_keys=True)) if __name__ == "__main__": main()