"""Validation for leakage-free SolveEdit-Harness records.""" from __future__ import annotations from typing import Any FORBIDDEN_PLANNER_KEYS = { "contract", "contract_path", "reward_questions", "penalty_questions", "manual_facts", "intended_final_state", "gt_image", "reference_image", "dependency_regime", "dependency_level", "visual_dependency_level", } REQUIRED_SOURCES = {"binding", "state", "rule", "preservation"} EVIDENCE_SOURCES = {"binding", "observed_state", "in_scene_rule", "preservation"} def find_forbidden_keys(value: Any, path: str = "$") -> list[str]: """Return forbidden key paths from an arbitrary JSON-like object.""" issues: list[str] = [] if isinstance(value, dict): for key, child in value.items(): child_path = f"{path}.{key}" if str(key).lower() in FORBIDDEN_PLANNER_KEYS: issues.append(child_path) issues.extend(find_forbidden_keys(child, child_path)) elif isinstance(value, list): for index, child in enumerate(value): issues.extend(find_forbidden_keys(child, f"{path}[{index}]")) return issues def _list(value: Any, field: str, issues: list[str]) -> list[Any]: if not isinstance(value, list): issues.append(f"{field}_must_be_list") return [] return value def _bbox(value: Any) -> bool: return ( isinstance(value, list) and len(value) == 4 and all(isinstance(item, (int, float)) for item in value) and 0 <= value[0] < value[2] <= 1 and 0 <= value[1] < value[3] <= 1 ) def validate_planner_input(record: dict[str, Any]) -> list[str]: issues = [f"forbidden_key:{path}" for path in find_forbidden_keys(record)] for field in ("uid", "case_id", "input_image", "instruction"): if not str(record.get(field) or "").strip(): issues.append(f"missing:{field}") return issues def validate_inspection(value: dict[str, Any], expected_uid: str) -> list[str]: issues = [f"forbidden_key:{path}" for path in find_forbidden_keys(value)] if value.get("uid") != expected_uid: issues.append("uid_mismatch") variables = _list(value.get("unresolved_variables"), "unresolved_variables", issues) queries = _list(value.get("visual_queries"), "visual_queries", issues) _list(value.get("preservation_hypotheses"), "preservation_hypotheses", issues) for index, item in enumerate(variables): if not isinstance(item, dict): issues.append(f"variable_{index}_not_object") continue if item.get("required_source") not in REQUIRED_SOURCES: issues.append(f"variable_{index}_bad_source") for field in ("name", "question", "why_unresolved"): if not str(item.get(field) or "").strip(): issues.append(f"variable_{index}_missing:{field}") for index, item in enumerate(queries): if not isinstance(item, dict): issues.append(f"query_{index}_not_object") continue if not _bbox(item.get("bbox")): issues.append(f"query_{index}_bad_bbox") if not str(item.get("query") or "").strip(): issues.append(f"query_{index}_missing_query") return issues def validate_plan(value: dict[str, Any], expected_uid: str) -> list[str]: issues = [f"forbidden_key:{path}" for path in find_forbidden_keys(value)] if value.get("uid") != expected_uid: issues.append("uid_mismatch") variables = _list(value.get("unresolved_variables"), "unresolved_variables", issues) evidence = _list(value.get("evidence_ledger"), "evidence_ledger", issues) candidates = _list(value.get("candidate_transitions"), "candidate_transitions", issues) _list(value.get("self_contract"), "self_contract", issues) _list(value.get("uncertainties"), "uncertainties", issues) for index, item in enumerate(variables): if not isinstance(item, dict) or item.get("status") not in {"resolved", "uncertain"}: issues.append(f"variable_{index}_bad_status") claim_ids: set[str] = set() for index, item in enumerate(evidence): if not isinstance(item, dict): issues.append(f"evidence_{index}_not_object") continue claim_id = str(item.get("claim_id") or "") if not claim_id or claim_id in claim_ids: issues.append(f"evidence_{index}_bad_claim_id") claim_ids.add(claim_id) if item.get("source") not in EVIDENCE_SOURCES: issues.append(f"evidence_{index}_bad_source") if item.get("bbox") is not None and not _bbox(item.get("bbox")): issues.append(f"evidence_{index}_bad_bbox") candidate_ids = { str(item.get("transition_id")) for item in candidates if isinstance(item, dict) and item.get("transition_id") } selected = str(value.get("selected_transition_id") or "") if not candidate_ids: issues.append("candidate_transitions_empty") if selected not in candidate_ids: issues.append("selected_transition_missing") challenge = value.get("contrastive_challenge") if not isinstance(challenge, dict): issues.append("contrastive_challenge_missing") else: if challenge.get("outcome") not in {"supported", "uncertain", "not_applicable"}: issues.append("contrastive_challenge_bad_outcome") cited = challenge.get("discriminating_claim_ids", []) if not isinstance(cited, list) or any(str(item) not in claim_ids for item in cited): issues.append("contrastive_challenge_bad_claim_ids") instruction = str(value.get("compiled_instruction") or "").strip() if not instruction: issues.append("compiled_instruction_missing") if len(instruction.split()) > 180: issues.append("compiled_instruction_over_180_words") return issues