{
  "state": "independent_data_and_contract_checks_passed_with_findings_training_screen_incomplete_release_blocked",
  "date": "2026-10-26",
  "audit_start_epoch": 1793020726.0973315,
  "scope": {
    "new_code": ["pallet/mask_model_improvement*.py", "scripts/*mask_model_improvement*.py", "tests/test_mask_model_improvement*.py"],
    "experiments": ["outputs/pallet_mask_model_improvement_v3", "outputs/pallet_mask_model_improvement_v4"],
    "context": ["docs/pallet_mask_model_improvement_v1.md", "docs/pallet_mask_model_training_diagnosis_v1.md", "docs/pallet_mask_model_research.md", "outputs/pallet_mask_model_training_v1"],
    "source": "outputs/pallet_mask_continuation_sources_v10",
    "repository_writes": ["outputs/pallet_mask_model_improvement_independent_v1/report.json", "outputs/pallet_mask_model_improvement_independent_v1/report.md"],
    "temporary_test_fixtures": "/tmp/opencode/improvement-independent-contracts",
    "GPU_used": false,
    "training_launched": false,
    "training_interrupted_or_modified": false,
    "delegation": false,
    "old_verifier_scripts_imported_or_executed": false,
    "process_observation": "one-shot /proc and receipts; no waiting or polling; later run changes are outside this snapshot",
    "child_CPU_threads": 1
  },
  "findings": [
    {
      "id": "RELEASE_BLOCK",
      "severity": "high_release_gate_not_training_bug",
      "location": "dataset_manifest.json and pallet/mask_model_improvement.py:83-111",
      "detail": "Dev has 10 components, but calibration has only 5 images / 2 components. Semantic trial and calibrated quality probability are absent. Acceptance remains zero/null. Research qualification is not GT, robot or automatic-training authorization.",
      "disposition": "Publication/deployment remains blocked; no threshold relaxation, 22-class blind claim or automatic A4/A5/new seeds."
    },
    {
      "id": "COMPLETION_VERIFIER_NOT_FAIL_CLOSED",
      "severity": "medium",
      "location": "pallet/mask_model_improvement_evidence.py:22-24,83-86",
      "detail": "Missing completion receipts become failed_or_not_started_not_complete, but the top-level state is still passed_completed_run_evidence. The queue waiter normally defers verification, and finalize separately distinguishes partial completion; nevertheless that evidence state alone does not certify seven completed runs.",
      "disposition": "No live changes made. Main session must use per-run completion receipts and partial/failed state rather than this unconditional top-level success label."
    },
    {
      "id": "CPU_GPU_NATIVE_THRESHOLD_REPRODUCTION_LIMIT",
      "severity": "low_open_numerical_limitation",
      "location": "v3/A3_freeze/best.pt and first dev prediction",
      "detail": "Independent CPU best-checkpoint inference differs from stored GPU-native RLE at 2 pixels in one sampled image. CPU probabilities are 0.5000855922698975 and 0.4998329281806946; stored labels are false and true respectively. Cause was not established. A0/A1/A2 sampled inference is pixel-identical.",
      "disposition": "Do not claim universal bitwise CPU/GPU inference equivalence. This is not evidence of label leakage or a demonstrated critical training bug. No live retraining/fix attempted."
    },
    {
      "id": "STORAGE_GUARD_AGGREGATION",
      "severity": "low_preventive",
      "location": "pallet/mask_model_improvement.py:48-51",
      "detail": "Runtime storage guard bounds each output directory, not combined v1-v4 storage. Independent aggregate measurement is currently well below the requested total limit.",
      "disposition": "Current budget passes; do not interpret per-directory guards as a future aggregate guarantee."
    }
  ],
  "confirmed": {
    "human_source_truth": {
      "catalog": 1520,
      "saved_pose": 482,
      "saved_gallery": 1038,
      "gallery_events": 1144,
      "baseline_raw_event_prefix": 1088,
      "baseline_prefix_byte_values_preserved": true,
      "all_1520_catalog_polygons_equal_unique_raw_mask_event_and_latest_snapshot_current": true,
      "all_1520_current_canonical_payload_hashes_verified": true,
      "new_distinct_targets": 36,
      "new_distinct_formal_train": 35,
      "new_distinct_formal_val_excluded": 1,
      "new_research_masks_selected": 34,
      "one_train_ambiguous_source_quarantined": true,
      "old_unchanged_polygons": 226,
      "old_latest_revised_polygons": 3,
      "revised_class_and_revision": [{"class": 20, "revision": 3}, {"class": 11, "revision": 2}, {"class": 12, "revision": 2}],
      "every_training_polygon_and_event_hash_equal_frozen_source": true,
      "suffix56_independent_task_manifest_target_proposal_polygon_binding": true,
      "suffix56_exact_task_proposal_canonical_sha_equals_event_proposal_sha_and_base_id": true,
      "suffix56_human_edited_true": 56,
      "suffix56_status_human_edited_draft_pending_confirmation": 56,
      "suffix_events_by_task": {"preannotated_v3": 17, "under50_v1": 10, "under50_v2": 11, "under50_v3": 7, "under50_v4": 6, "under50_v5": 5},
      "under50_v5_events": 5,
      "v6_in_source_catalog": false,
      "live_gallery_once": {"events": 1144, "frozen_prefix_preserved": true, "additional_events": 0},
      "future_human_appends_policy": "Legitimate later current revisions are allowed if frozen events remain recoverable. No dynamic training append; no assertion that current must remain immutable."
    },
    "frozen_dataset": {
      "records": 263,
      "all_formal_train": true,
      "splits": {"train": 172, "dev": 86, "calibration": 5},
      "components": {"train": 13, "dev": 10, "calibration": 2},
      "manifest_files_each_version": 800,
      "all_v3_and_v4_file_hashes_recomputed_and_valid": true,
      "v3_v4_manifest_and_model_facing_bytes_identical": true,
      "no_class_group_sensor_ID_network_input": true,
      "all_five_eligibility_flags_false": true,
      "explicit_34_mask_only_research_approval_not_GT": true,
      "no_v6_or_unsaved_machine_draft_return_to_training": true,
      "metadata_only_predeclared_split_and_dev_group_macro_selection": true,
      "formal_test_used_for_model_selection": false,
      "blind_evaluation": false
    },
    "independent_original_pixel_global_graph": {
      "method": "Independent DSU and original-pixel full-frame/context signatures, not producer global_audit import. Exact group/frame/PCD/image-pixel edges plus exhaustive 3-band dHash128 Hamming<=2 candidates, contrast>=3 and NCC>=.995; all classes/policies connected before split blocking.",
      "targets": 12344,
      "original_positive_frames_decoded": 10111,
      "components": 408,
      "all_component_labels_equal_producer": true,
      "all_formal_holdout_blocked_components_equal_producer": true,
      "all_original_pixel_crop_dhash_roi_hashes_match": true,
      "new_full_frame_near_connections": 3,
      "cross_class_near_connections": 1,
      "additional_crop_near_connections": 0,
      "selected_train_to_formal_val_test_component_connections": 0,
      "train_dev_cal_component_crossings": 0,
      "limitation": "This is a conservative defined photometric/identity audit, not exhaustive proof of physical scene identity."
    },
    "supervision_and_input_pixels": {
      "independent_pixel_center_polygon_raster_matches": 263,
      "original_bbox_context_input_crop_matches": 263,
      "known_neighbor_bbox_intersections": 0,
      "ROI_rule": "original bbox: 6% horizontal / 22% vertical context; never polygon-derived input bbox",
      "mask_labels": "frozen real saved human polygon only; no invented holes/corners/pose labels",
      "unknown_unboxed_neighbors_and_semantic_correctness": "Residual risks; source authenticity does not establish contour accuracy or semantic GT."
    },
    "implementation": {
      "runs": ["A0", "A1", "A2", "A3_freeze", "A3_photometric", "B0", "B1"],
      "seed": 42,
      "optimizer_updates_per_run_budget": 1000,
      "M1": "512x160 two-head anchor; A1 inverse-group-size weighted image sampling without class weights; A2 edge weight zero; A3_freeze first 5 epochs encoder frozen with anchor LR unchanged; A3_photometric image-only gain/gamma/noise/weak blur.",
      "M2": "B0 nearest-dense matched mask loss; B1 only matched loss changes to 2048 valid points, 75% uncertain and 25% uniform. Shared matcher/padding/FP32, architecture, 13 groups, optimizer and EMA. Fixed classification-ranked query, raw logits to native, sigmoid then global .5 threshold once.",
      "generic_cache_sha_recomputed": {"resnet18-f37072fd.pth": "f37072fd47e89c5e827621c5baffa7500819f7896bbacec160b1a16c560e07ec", "rf-detr-seg-small.pt": "6de3da31b2572cac214a1c76cce4a92a13966d56390ac2b3a3de9a8dc2b2bca3"},
      "fresh_process_and_generic_initialization_each_run": true,
      "old_task_finetune_used": false,
      "v3_B0_real_lr_drop_API_failure_preserved_not_trained": true,
      "v4_process_local_lr_scheduler_kwargs_compatibility_only": true,
      "v3_v4_live_code_and_launch_snapshot_pins_match": true,
      "EMA_last_checkpoint_semantics": "Installed RFDETR 1.11 RFDETREMACallback.on_train_end loads EMA into module before custom last.pt; this path was source-checked, not assumed base weights."
    },
    "independent_CPU_contracts": {
      "pytest_passed": 77,
      "pytest_failed": 0,
      "pytest_seconds": 4.25,
      "test_files": ["tests/test_mask_model_improvement_adapter.py", "tests/test_mask_model_improvement_retry.py", "tests/test_mask_model_training_data.py", "tests/test_mask_model_training_history_v2.py"],
      "additional_actual_13group_match_indices_dense_factorized_equal": true,
      "additional_actual_match_indices_padding_invariant": true,
      "additional_all_zero_match_main_aux_encoder_spatial_query_bias_have_connected_zero_gradient_B0_and_B1": true,
      "coverage": "noncontinuous arbitrary batch/query indices; 0/1/multiple instances; all auxiliary/encoder gradients; dense/factorized loss and gradients; FP32/BF16; valid/padding invariance; width-height/native thin/slant/concave roundtrip; classification-only query; no heldout polygon label access at eligibility; mask-head edge-gradient separation; scheduler API"
    },
    "completed_native_prediction_replay": {
      "runs": ["A0", "A1", "A2"],
      "per_run_dev": 86,
      "per_run_calibration": 5,
      "membership_exact": true,
      "independent_RLE_IoU_HD95_ASSD_boundary_F1_max_absolute_error": 0.0,
      "accepted_count_each_split": 0,
      "accepted_quality_each_split": null,
      "best_checkpoint_CPU_inference_sample_count": 4,
      "native_pixel_differences_by_sampled_run": {"A0": 0, "A1": 0, "A2": 0, "A3_freeze": 2}
    },
    "preservation": {
      "before_after_records_equal": true,
      "hashes_recomputed_counts": {"configs": 40, "old_code": 19, "old_dataset": 694, "old_models": 2, "old_experiment_artifacts": 872, "old_history_audits": 15, "old_public_objects": 200887, "formal_protected_files": 20768},
      "mismatches": 0,
      "old_training_metadata_records": 229,
      "v1_prevalidation_failure_preserved": true,
      "v2_design_rejection_frozen_not_used_for_training_preserved": true,
      "old_artifact_deletions": 0,
      "protected_inventory_scope": "All inventoried members exist and match SHA256. Not a claim that no authorized future human revision or unrelated new output may exist."
    },
    "storage_once": {"all_new_versions_v1_through_v4_bytes": 2115159131, "project_free_bytes": 55592271872, "total_below_25_GiB": true, "free_at_least_10_GiB": true},
    "queue_audit": {
      "source_checked": "subprocess.wait queues, inotify startup receipts, pidfd/select completion wait with libc/syscall fallback; no sleep/status polling",
      "observed_retry_queue_pid": 21062,
      "observed_retry_worker_pid": 21133,
      "observed_verifier_pid": 21316,
      "verifier_open_pidfd_target": 21062,
      "actual_pidfd_verified_from_proc_fdinfo": true
    }
  },
  "run_receipt_snapshot": {
    "GPU_M1": {"uuid": "GPU-3315c096-fef4-7350-483a-7c5f5ca2b619", "physical_memory_MiB": 16376, "first_step_batch": 8, "first_step_peak_bytes": 1285206016},
    "GPU_M2": {"uuid": "GPU-09ed84c8-4dee-efc0-dc1d-deda5177e4e8", "physical_memory_MiB": 24564, "first_step_batch": 2, "accumulation": 4, "first_step_peak_bytes": 2284639744, "first_step_pid": 21133},
    "hardware_claim_limit": "Actual launch physical-query receipts and running-process CUDA_VISIBLE_DEVICES checked. Receipts identify NVIDIA Graphics Device, not an assumed commercial GPU name; neither is claimed to be a 3090.",
    "A0": {"state": "completion_receipt_and_CPU_best_last_verified", "updates": 1000, "best_step": 176, "last_step": 1000, "selection_replayed": true},
    "A1": {"state": "completion_receipt_and_CPU_best_last_verified", "updates": 1000, "best_step": 110, "last_step": 1000, "selection_replayed": true},
    "A2": {"state": "completion_receipt_and_CPU_best_last_verified", "updates": 1000, "best_step": 154, "last_step": 1000, "selection_replayed": true},
    "A3_freeze": {"state": "completion_receipt_and_CPU_best_last_verified_native_full_replay_not_done", "updates": 1000, "best_step": 242, "last_step": 1000, "selection_replayed": true, "epoch_freeze_flags_match_first5": true},
    "A3_photometric": {"state": "started_pending_at_receipt_snapshot", "first_optimizer_step": 1, "pid": 22200, "live_expected_worker_and_GPU_environment_verified": true},
    "B0_v4": {"state": "started_pending_at_receipt_snapshot", "first_optimizer_step": 1, "pid": 21133, "live_expected_worker_and_GPU_environment_verified": true},
    "B1_v4": {"state": "not_started_at_receipt_snapshot"},
    "B0_v3": {"state": "failed_API_pretraining_not_complete"},
    "all_seven_runs_complete": false,
    "accuracy_or_release_certified": false,
    "receipt_policy": "Only actual completion.json plus best/last CPU step/hash checks establish completion here. No pending run is credited with its planned 1000 steps."
  },
  "open_limitations": [
    "Independent source/task binding establishes actual saved supervision, not human contour accuracy, face semantics, topology correctness, calibrated confidence or 22-class generalization.",
    "All original1088 raw events were byte-prefix checked; task/proposal triples were independently reconstructed for the56-event suffix, not independently revalidated for every baseline historical publication.",
    "Frozen files and snapshots were checked over a bounded observation window, not an atomic filesystem-wide snapshot; training may legitimately advance after receipts were read once.",
    "Only A0/A1/A2 all stored native masks were independently metrically replayed; A3_freeze has completion/selection/checkpoint checks and one inference sample. Pending runs have no completed-result certification.",
    "One CPU inference sample per four completed M1 checkpoints is not full CPU inference replay. A3_freeze has a reported two-pixel numerical mismatch.",
    "No fresh independent visual semantic review of all34 added polygons; explicit producer research-only review and quarantines were checked, and all263 mask raster labels were independently verified against actual saved polygons."
  ],
  "sha256": {
    "docs/pallet_mask_model_improvement_v1.md": "3eef0fdad8d371f58a318f2d196dd6e3b15eb662bd32ce5d07cf6facf95b3234",
    "pallet/mask_model_improvement.py": "aadf3ec7dbe5dc1f38db9b1f06e5d1b321da6e2f38147da904c6e2d5bddf2b80",
    "pallet/mask_model_improvement_adapter.py": "ccb83111359f5ece1a228dec28c1a67d7083ba08b5bfeddda89b27b048ab6f7b",
    "pallet/mask_model_improvement_data.py": "db452a284bdf40196e89f4ffbe7e49cd4858a8122f82424fa02f83d34566a62f",
    "pallet/mask_model_improvement_evidence.py": "89595f740e51610ac0355846187f4725d0452fa73460aa08b666b32ebda4438d",
    "pallet/mask_model_improvement_retry.py": "6d620f82ab10db3967d2477f92b983ce0e680d2936aabebd5613f45583bc470a",
    "scripts/run_mask_model_improvement.py": "5f8d80ec4bafb11ad159702077198379c5d3092b18aff05ed1e6ff3de3fa41ee",
    "scripts/run_mask_model_improvement_retry.py": "bb3fbc6773d6065345b61d36da80c26ae44fc966b01399c8fcb19ff29a69c656",
    "scripts/verify_mask_model_improvement.py": "f71bdbcc9dc9e15761787384609d98c5aeeda002509242930f51ee9fb8cfba00",
    "tests/test_mask_model_improvement_adapter.py": "1fa507c6da79d3479ef9a89af58045e89d9d5cb3f51b219328c8f52702f52e79",
    "tests/test_mask_model_improvement_retry.py": "a1516f52571fef98bb41d10308e3db427bb44f1eab9dbce09e1a4fe31f8c74eb",
    "outputs/pallet_mask_continuation_sources_v10/audit.json": "537b3aa097267a8e16e91a3623903f1299d557c18125f3d1fcdcf4bd9bd09a58",
    "outputs/pallet_mask_continuation_sources_v10/private/source_catalog.jsonl": "482f2c96a7957edbbfdfc48b5d650631b990ec87a358f67ca3343e62989f5e38",
    "outputs/pallet_mask_model_improvement_v3/dataset_manifest.json": "8f97f35a42b70d9e967b420d4861890bb9e9cd04c6c44198cbf5316cda51d4db",
    "outputs/pallet_mask_model_improvement_v4/dataset_manifest.json": "8f97f35a42b70d9e967b420d4861890bb9e9cd04c6c44198cbf5316cda51d4db",
    "outputs/pallet_mask_model_improvement_v3/execution_plan.json": "3c70821ac0cf1e6a4d855cd3b46bacb4c92aa151dc2e186f3d4bd4a65dcb035c",
    "outputs/pallet_mask_model_improvement_v4/execution_plan.json": "bb05de1d64d31b18d0b12a698c89d6d30e54c333bd887cabce7c30fb1e559b52",
    "outputs/pallet_mask_model_improvement_v3/private/source_binding.json": "2243c1ba7f99563aa8b8a91f4281b26784400e7d88182f080821c3d676a46bdd",
    "outputs/pallet_mask_model_improvement_v3/private/global_components.json": "80acdb6bc3e7c6cb5172eddeed641621bc75bcb96c949f77681c35d1cde9354e",
    "outputs/pallet_mask_model_improvement_v3/private/preservation_before.json": "bc0c2666b9ccd3cf2afb60cf792d4897a9eacea2f5522b46af62ac2aec9aef47",
    "outputs/pallet_mask_model_improvement_v3/contracts_execution_pin.xml": "86b9737dff7e2dac1f6ea42e922c80041960140ec424253de5d8c1227838f161",
    "outputs/pallet_mask_model_improvement_v4/contracts_execution_pin.xml": "2aac4a2d9dbf2d4a4aecc6ea35dbe279ebd4b7d73ca9cf348103bcf431480419",
    "outputs/pallet_mask_model_improvement_v3/A0/completion.json": "277dbc7ba8d5bdaae28239338f1c76b1915f4ec614fda3518e0169b9dbb22092",
    "outputs/pallet_mask_model_improvement_v3/A1/completion.json": "a62f27917860eee7f20331104583b03ae22c25cf4b622f2aaf6f3085519cf589",
    "outputs/pallet_mask_model_improvement_v3/A2/completion.json": "517f1e9894275c644669b741c69c9087d6f1a790d84d4eaf3beb62f766278e09",
    "outputs/pallet_mask_model_improvement_v3/A3_freeze/completion.json": "11f1a07cc02f781bd68b1eb3964b88e72df62ac6c7b7a33587f8147f83688613",
    "outputs/pallet_mask_model_training_v1/dataset_manifest.json": "563326cddcd97962a8f2cd8ceb5fe760cb76a0c5f0572a6e9535d01b2ccc13db"
  }
}
