| """Machine-readable split, overlap, and sealed-test audit for DENIM data.""" |
|
|
| from __future__ import annotations |
|
|
| import hashlib |
| import json |
| from collections import Counter, defaultdict |
| from pathlib import Path |
|
|
| import numpy as np |
|
|
| from src.evaluate_t2_denim_sealed_test_v1 import load_trajectories as load_sealed |
| from src.train_t2_denim_capability_model_v4 import load_v4_trajectories |
| from src.train_t2_denim_conditional_v2 import load_all_trajectories |
|
|
|
|
| ROOT = Path(__file__).resolve().parents[1] |
| OUTPUT = ROOT / "artifacts" / "t2_denim_protocol_audit_v1.json" |
|
|
|
|
| def _hash(*arrays: np.ndarray) -> str: |
| digest = hashlib.sha256() |
| for value in arrays: |
| selected = np.ascontiguousarray(value) |
| digest.update(str(selected.shape).encode()) |
| digest.update(selected.view(np.uint8)) |
| return digest.hexdigest() |
|
|
|
|
| def run() -> dict[str, object]: |
| old = load_all_trajectories() |
| new, _ = load_v4_trajectories() |
| sealed, _ = load_sealed() |
| all_items = old + new + sealed |
| role_counts = Counter(item.role for item in all_items) |
| role_families: dict[str, set[str]] = defaultdict(set) |
| full_input_roles: dict[str, set[str]] = defaultdict(set) |
| input_target_roles: dict[str, set[str]] = defaultdict(set) |
| strain_roles: dict[str, set[str]] = defaultdict(set) |
| for item in all_items: |
| role_families[item.role].add(item.family) |
| descriptors = np.concatenate((item.descriptor.numpy(), np.asarray((item.young, item.poisson, item.yield_stress)))) |
| full_input_roles[_hash(item.strain.numpy(), descriptors)].add(item.role) |
| input_target_roles[_hash(item.strain.numpy(), descriptors, item.stress.numpy())].add(item.role) |
| strain_roles[_hash(item.strain.numpy())].add(item.role) |
|
|
| def collisions(values: dict[str, set[str]]) -> int: |
| return sum(len(roles) > 1 for roles in values.values()) |
|
|
| training_families = role_families["train"] |
| report = { |
| "schema": "agentfem.denim-ml-protocol-audit.v1", |
| "trajectory_count": len(all_items), |
| "role_counts": dict(sorted(role_counts.items())), |
| "exact_cross_role_collisions": { |
| "model_input_plus_target": collisions(input_target_roles), |
| "full_model_input": collisions(full_input_roles), |
| "strain_only": collisions(strain_roles), |
| }, |
| "family_overlap_with_training": { |
| role: sorted(families & training_families) for role, families in sorted(role_families.items()) |
| }, |
| "sealed_test": { |
| "trajectory_count": len(sealed), |
| "family_count": len(role_families["sealed_test"]), |
| "families": sorted(role_families["sealed_test"]), |
| "family_overlap_with_training": sorted(role_families["sealed_test"] & training_families), |
| "eligible_for_training_or_selection": False, |
| }, |
| "interpretation": { |
| "strain_only_collisions": "Expected transfer pairs can share strain histories across different material descriptors and targets; they are not direct input/target leakage.", |
| "existing_test_policy": "The 1,349 pre-existing test trajectories are development benchmarks because prior aggregate results informed later model design.", |
| "sealed_policy": "sealed_test_v1 is the first prospective frozen-model test and must never be reused for training or checkpoint selection.", |
| }, |
| } |
| report["passed"] = bool( |
| report["exact_cross_role_collisions"]["model_input_plus_target"] == 0 |
| and report["exact_cross_role_collisions"]["full_model_input"] == 0 |
| and not report["sealed_test"]["family_overlap_with_training"] |
| and len(sealed) == 128 |
| ) |
| OUTPUT.parent.mkdir(parents=True, exist_ok=True) |
| OUTPUT.write_text(json.dumps(report, indent=2, sort_keys=True) + "\n", encoding="utf-8") |
| print(json.dumps(report, indent=2, sort_keys=True)) |
| return report |
|
|
|
|
| if __name__ == "__main__": |
| run() |
|
|
