| """Portable integrity/development admission, never a model or rights certificate.""" |
| import hashlib |
| from pathlib import Path |
|
|
| from . import _windows |
| from .protocol import decode |
|
|
| TOKENIZER_FILES = {'tokenizer.json', 'tokenizer_config.json', 'special_tokens_map.json'} |
| WINDOW_GATE_MODE = 'supervised-owned-parent-window-gated' |
| WINDOW_GATE_PROFILE = 'CLS-window-gated-linear-heads-v1' |
| SOURCE_KIND_MODE = 'supervised-owned-parent-source-kind-gated' |
| SOURCE_KIND_PROFILE = 'CLS-source-kind-gated-linear-heads-v1' |
|
|
|
|
| def sha(path, limit): |
| total = 0 |
| digest = hashlib.sha256() |
| with Path(path).open('rb') as stream: |
| while chunk := stream.read(1024 * 1024): |
| total += len(chunk) |
| if total > limit: |
| raise ValueError('artifact_byte_limit') |
| digest.update(chunk) |
| return digest.hexdigest() |
|
|
|
|
| def _file(root, name): |
| path = (root / name).resolve(strict=True) |
| if not path.is_relative_to(root) or not path.is_file(): |
| raise ValueError('artifact_path_escape') |
| return path |
|
|
|
|
| def _json(root, name, limit): |
| path = _file(root, name) |
| with path.open('rb') as stream: |
| data = stream.read(limit + 1) |
| if len(data) > limit: |
| raise ValueError('artifact_byte_limit') |
| return decode(data) |
|
|
|
|
| def _inspect_source_kind(root, report, revision): |
| recipe = report['recipe'] |
| parent = report.get('parent_checkpoint_sha256') |
| if (report.get('runtime_admitted') is not True |
| or not isinstance(parent, str) or len(parent) != 64 or parent == revision |
| or any(c not in '0123456789abcdef' for c in parent) |
| or recipe.get('gate_input') != 'encoder_CLS_per_window' |
| or type(recipe.get('gate_threshold')) is not int or recipe['gate_threshold'] != 0 |
| or recipe.get('branch_order') != ['service', 'log', 'base'] |
| or recipe.get('log_bias') != 'shared_frozen_base_head_bias' |
| or recipe.get('additional_parameters') != 1537 |
| or any(recipe.get(k) is not True for k in ('encoder_frozen', 'base_head_frozen', |
| 'service_gate_and_head_frozen', 'original_parameters_bitwise_unchanged'))): |
| raise ValueError('frozen_source_kind_parent_profile_required') |
| files = {'__init__.py', '__main__.py', '_model.py', '_window_gate.py', '_source_kind.py', |
| '_windows.py', 'artifact.py', 'bootstrap.py', 'protocol.py', 'runtime.py'} |
| package = Path(__file__).resolve().parent |
| pins = report.get('runtime_files_sha256') |
| if (not isinstance(pins, dict) or set(pins) != files |
| or {p.name for p in package.glob('*.py')} != files |
| or any(sha(package / n, 128 * 1024) != digest for n, digest in pins.items())): |
| raise ValueError('complete_source_kind_runtime_code_pins_required') |
| if (sha(package / '_window_gate.py', 128 * 1024) != report.get('window_gate_code_sha256') |
| or sha(package / '_source_kind.py', 128 * 1024) != report.get('source_kind_model_code_sha256')): |
| raise ValueError('source_kind_model_code_hash_mismatch') |
| evidence = report.get('source_kind_evidence') |
| required = {'SOURCE-KIND-TRAIN-AUDIT.json', 'SOURCE-KIND-GPU-AUDIT.json', 'SOURCE-KIND-NATIVE-AUDIT.json'} |
| if not isinstance(evidence, dict) or set(evidence) != required: |
| raise ValueError('complete_source_kind_development_evidence_required') |
| proofs = {} |
| for name, digest in evidence.items(): |
| if sha(_file(root, name), 1024 * 1024) != digest: |
| raise ValueError('source_kind_evidence_hash_mismatch') |
| proof = _json(root, name, 1024 * 1024) |
| if (not isinstance(proof, dict) or proof.get('checkpoint_sha256') != revision |
| or any(proof.get(k) is not False for k in ('qualification', 'promotion_ready', 'release_approved'))): |
| raise ValueError('unqualified_source_kind_evidence_required') |
| proofs[name] = proof |
| train = proofs['SOURCE-KIND-TRAIN-AUDIT.json'] |
| gpu = proofs['SOURCE-KIND-GPU-AUDIT.json'] |
| native = proofs['SOURCE-KIND-NATIVE-AUDIT.json'] |
| if (train.get('status') != 'independent_source_kind_TRAIN_audit_passed_not_qualification' |
| or train.get('parent_checkpoint_sha256') != parent |
| or train.get('encoder_and_original_heads_bitwise_identical') is not True |
| or train.get('gate_constraints') != 3384 |
| or type(train.get('minimum_FP32_signed_margin')) not in (int, float) |
| or not .9999 <= train['minimum_FP32_signed_margin'] <= 1e6 |
| or train.get('log_head_TRAIN_audit_passed') is not True |
| or train.get('DEV_used_in_fit') is not False or train.get('heldout_access') is not False |
| or gpu.get('status') != 'independent_source_kind_906window_GPU_audit_passed_not_qualification' |
| or gpu.get('actual_GPU_cases') != 880 or gpu.get('actual_GPU_windows') != 906 |
| or gpu.get('all871_CPU_GPU_source_parity_verified') is not True |
| or gpu.get('Java_log_windows') != 5 or gpu.get('five_family_windows') != 30 |
| or gpu.get('original_mask_bindings_verified') != 871 |
| or gpu.get('route_misses') != [] or gpu.get('all_parameters_bitwise_unchanged') is not True |
| or gpu.get('new_revision_response_identity_verified') is not True |
| or native.get('status') != 'independent_source_kind_native_DEV_audit_passed_not_qualification' |
| or native.get('all348_native_source_allocations_exact_required_budget_valid') is not True |
| or native.get('all16_Java_native_allocations_exact_required_budget_valid') is not True |
| or native.get('budgets') != [.25, .5] or native.get('known_support_losses') != [] |
| or native.get('original_masks_unchanged') is not True |
| or native.get('retained_group_identity_verified') is not True): |
| raise ValueError('source_kind_full_development_evidence_required_not_qualification') |
| if any(report.get(k) is not False for k in ('qualification', 'promotion_ready', 'release_approved', 'redistribution_cleared')): |
| raise ValueError('source_kind_development_is_not_release_clearance') |
|
|
|
|
| def inspect_artifact(directory, revision): |
| """Check bytes/declared completion only, with no imports of ML frameworks.""" |
| if (not isinstance(revision, str) or len(revision) != 64 |
| or any(c not in '0123456789abcdef' for c in revision)): |
| raise ValueError('checkpoint_SHA256_required') |
| root = Path(directory) |
| if not root.is_absolute(): |
| raise ValueError('absolute_artifact_path_required') |
| root = root.resolve(strict=True) |
| report = _json(root, 'training-report.json', 1024 * 1024) |
| recipe = report.get('recipe', {}) if isinstance(report, dict) else {} |
| mode = report.get('mode') if isinstance(report, dict) else None |
| supported_recipe = isinstance(recipe, dict) and ( |
| (mode == 'supervised-foundation-feature-distilled' |
| and recipe.get('objective_profile') == 'optional-balanced-window-hardpair-headwarm-v4') |
| or (mode == WINDOW_GATE_MODE and recipe.get('objective_profile') == WINDOW_GATE_PROFILE) |
| or (mode == SOURCE_KIND_MODE and recipe.get('objective_profile') == SOURCE_KIND_PROFILE)) |
| if (not isinstance(report, dict) or not isinstance(report.get('recipe'), dict) |
| or report.get('status') != 'completed_development_only' |
| or report.get('checkpoint_sha256') != revision |
| or not supported_recipe |
| or report.get('recipe', {}).get('precision') != 'fp32' |
| or report.get('recipe', {}).get('student_token_cap') != 8192): |
| raise ValueError('completed_clean_compact12_development_artifact_required') |
| _windows.verify_profile(report) |
| if sha(Path(__file__).with_name('_model.py'), 128 * 1024) != report.get('joint_model_code_sha256'): |
| raise ValueError('model_code_hash_mismatch') |
| if mode == SOURCE_KIND_MODE: |
| _inspect_source_kind(root, report, revision) |
| if mode == WINDOW_GATE_MODE: |
| parent = report.get('parent_checkpoint_sha256') |
| if (not isinstance(parent, str) or len(parent) != 64 or parent == revision |
| or any(c not in '0123456789abcdef' for c in parent) |
| or recipe.get('gate_input') != 'encoder_CLS_per_window' |
| or recipe.get('encoder_frozen') is not True or recipe.get('base_head_frozen') is not True): |
| raise ValueError('frozen_owned_parent_window_gate_profile_required') |
| if sha(Path(__file__).with_name('_window_gate.py'), 128 * 1024) != report.get('window_gate_code_sha256'): |
| raise ValueError('window_gate_code_hash_mismatch') |
| evidence = report.get('window_gate_evidence') |
| required = {'WINDOW-GATE-TRAIN-AUDIT.json', 'WINDOW-GATE-DEV-AUDIT.json'} |
| if not isinstance(evidence, dict) or set(evidence) != required: |
| raise ValueError('window_gate_evidence_pins_required') |
| for name, digest in evidence.items(): |
| if sha(_file(root, name), 1024 * 1024) != digest: |
| raise ValueError('window_gate_evidence_hash_mismatch') |
| train = _json(root, 'WINDOW-GATE-TRAIN-AUDIT.json', 1024 * 1024) |
| dev = _json(root, 'WINDOW-GATE-DEV-AUDIT.json', 1024 * 1024) |
| if (not isinstance(train, dict) or not isinstance(dev, dict) |
| or train.get('status') != 'independent_CLS_window_gate_TRAIN_audit_passed_not_qualification' |
| or train.get('checkpoint_sha256') != revision |
| or train.get('encoder_and_original_head_bitwise_identical') is not True |
| or dev.get('status') != 'independently_replayed_prerequisite_pass' |
| or dev.get('all871_CPU_GPU_source_parity_verified') is not True |
| or dev.get('known_support_losses') != [] |
| or any(proof.get('qualification') is not False or proof.get('promotion_ready') is not False |
| for proof in (train, dev))): |
| raise ValueError('window_gate_development_evidence_required_not_qualification') |
| config = _json(root, 'config.json', 65536) |
| if (not isinstance(config, dict) or config.get('model_type') != 'modernbert' or config.get('num_hidden_layers') != 12 |
| or config.get('hidden_size') != 768 or config.get('max_position_embeddings') != 8192 |
| or config.get('layer_types') != ['full_attention' if i % 3 == 0 else 'sliding_attention' for i in range(12)]): |
| raise ValueError('compact12_configuration_required') |
| if sha(_file(root, 'config.json'), 65536) != report.get('config_sha256'): |
| raise ValueError('config_hash_mismatch') |
| if sha(_file(root, 'checkpoint.pt'), 1024**3) != revision: |
| raise ValueError('weights_hash_mismatch') |
| pins = report.get('tokenizer_sha256') |
| if not isinstance(pins, dict) or set(pins) != TOKENIZER_FILES: |
| raise ValueError('complete_tokenizer_pins_required') |
| if {p.name for p in (root / 'tokenizer').iterdir()} != TOKENIZER_FILES: |
| raise ValueError('tokenizer_files_mismatch') |
| for name, digest in pins.items(): |
| if sha(_file(root, 'tokenizer/' + name), 32 * 1024**2) != digest: |
| raise ValueError('tokenizer_hash_mismatch') |
| decision = _json(root, 'DEVELOPMENT-DECISION.json', 1024 * 1024) |
| if (not isinstance(decision, dict) or decision.get('checkpoint_sha256') != revision |
| or any(decision.get(key) is not True for key in |
| ('training_completed', 'completion_content_audit_passed', 'development_stop_go_passed')) |
| or any(decision.get(key) is not False for key in ('qualification', 'release_approved', 'redistribution_cleared'))): |
| raise ValueError('accepted_development_decision_required_not_release') |
| return report |
|
|
