{ "experiment": "m64-qwen-coding-006", "date": "2026-10-02", "status": "no_validation_improvement", "decision": "rejected_for_combined_assistant", "selected_step": null, "candidate_step": 800, "candidate_adapter_sha256": "1bd3d039e63a5dab87149af7a3762ab56099dc42b8364d732afcc8b5de56db4f", "comparison_adapter": "coding-003/checkpoint-600", "warm_start_adapter": "coding-003/checkpoint-600", "manifest": { "base_model": { "repository": "mlx-community/Qwen2.5-Coder-1.5B-Instruct-4bit", "revision": "b3252a2f97102b1fb1571fec2c9b27219a8536be", "weights_sha256": "daeab4764fb420d161721791cf2e509e2de81a7af4223646e7bed2bf82c57b58", "upstream_repository": "Qwen/Qwen2.5-Coder-1.5B-Instruct", "upstream_revision": "2e1fd397ee46e1388853d2af2c993145b0f1098a", "license": "Apache-2.0", "quantization_bits": 4, "mlx_lm_version": "0.31.3" }, "initial_adapter_sha256": "a1014c6f71b46d879c09462c4a57d17d00df8f11b8178dd53bf6d5f2ac2f778a", "source_sha256": { "training/m64-qwen/improve.py": "08b2cc6dc9fd779e509a860ae6205a0c8bcee2f6349cef249949376131bb9cdd", "training/m64-qwen/picogk.py": "3357a289b0314a4e7e24f1a0d515d1bebf55021ccacd2551650b830f352ac3c6", "training/m64-engineer/openusd.py": "eb28db525af0ed21b4f80758a39a67482029510ba26fa23fc2251a3a4e6b4bb6" }, "files_sha256": { "usd-cases.jsonl": "d18eb56c5f21679b8e37f05c1dd8442eaf600d967de063ffd358b0e678b98437", "picogk-cases.json": "dd4360574826ddb1f58671e6b0472e3f64e897ff5c1127b47f417d13248d6812", "data/test.jsonl": "e7752ee9c69737a2c6e0b872ed0a8376d32773f902fc19d887df22a9eaad7a0f", "data/train.jsonl": "89a1caa49cccc375be28461309d5341439f8faaa8492f8d313c4ab2caa849565", "data/valid.jsonl": "0714942456813dc1deb42c6abed58fc249c2ac8841b1b49b3c00eea2188c4233" }, "counts": { "train": { "usd": 336, "picogk": 704 }, "valid": { "usd": 30, "picogk": 40 }, "test": { "usd": 36, "picogk": 56 } }, "maximum_sequence_tokens": 547, "steps": [ 800 ], "learning_rate": 1e-05, "seed": 42, "warm_start_sha256": "a1014c6f71b46d879c09462c4a57d17d00df8f11b8178dd53bf6d5f2ac2f778a", "replay_weight": 4, "usd_replay_weight": 4, "effective_counts": { "train": 2240, "valid": 70, "test": 92 }, "num_layers": 4, "rank": 8, "lora_scale": 20, "batch_size": 1, "mask_prompt": true, "selection": "Higher PicoGK validation count with no per-case USD or PicoGK regression.", "test_policy": "Tests open only after selection. Only pico3-test cases are fresh in a focused run; all older tests are regressions. Shared grammar, not independent families.", "native_sha256": { "PicoGK.dll": "710c4e0428efbe9c1cd1db1663247b45604f96d167a6726782ab8ddca350565a", "picogk.26.2.dylib": "5c1cd3fc12766a85bc5c1441a95f065fb274338d1333c3ae2d3396b59d0e18e7" } }, "registration": { "registered_at_utc": "2026-10-01T22:43:54.365361+00:00", "protocol_commit": "57aa24a0c19c15fb4fc4465ee060f502ba500202", "protocol_sha256": "5bfa96664a8855f5ecfb4868a377ab7c6dbaa422d6f4e6a2cb19922e7c74cd3a", "manifest_sha256": "769290786149f5f5fde22392058956b91dc5df9d37cb31c15737afb0d600ae6f", "final_tests_opened": false }, "validation_summary": { "usd": { "total": 30, "regressions": [ "usd2-valid-013", "usd2-valid-017", "usd2-valid-021", "usd2-valid-023" ], "gains": [ "usd2-valid-010" ], "before_passed": 22, "before_native_passed": 24, "after_passed": 19, "after_native_passed": 21 }, "picogk": { "total": 40, "regressions": [ "valid-003", "valid-004", "valid-007", "pico3-valid-007" ], "gains": [ "pico2-valid-014", "pico3-valid-006" ], "before_passed": 9, "before_native_passed": 37, "before_compiled": 40, "after_passed": 7, "after_native_passed": 36, "after_compiled": 38 } }, "validation_cases": { "usd": [ { "id": "usd-valid-000", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "017bc92d9d1be753373b064d1737341216e29927b0f264c3a8f1f8d6e06e22d0", "after_passed": true, "after_native_passed": true, "after_response_sha256": "017bc92d9d1be753373b064d1737341216e29927b0f264c3a8f1f8d6e06e22d0" }, { "id": "usd-valid-001", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "31964fa538d592cf703ba661026574a2a25b970b3e1d2dfd2a84df324b542593", "after_passed": true, "after_native_passed": true, "after_response_sha256": "31964fa538d592cf703ba661026574a2a25b970b3e1d2dfd2a84df324b542593" }, { "id": "usd-valid-002", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "5d06a1edc6286362da523ac9973f3b3477c7f2c29646955522fdd48073582f4a", "after_passed": true, "after_native_passed": true, "after_response_sha256": "5d06a1edc6286362da523ac9973f3b3477c7f2c29646955522fdd48073582f4a" }, { "id": "usd-valid-003", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "734c04ed5ca2ef362599e2110d01b5d0720ec139e97724dd45e19cc49fac983c", "after_passed": true, "after_native_passed": true, "after_response_sha256": "734c04ed5ca2ef362599e2110d01b5d0720ec139e97724dd45e19cc49fac983c" }, { "id": "usd-valid-004", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "6c869ea563f35ced578c2024beb1c87715dc557be85ff19f71d86b5300ca801f", "after_passed": true, "after_native_passed": true, "after_response_sha256": "6c869ea563f35ced578c2024beb1c87715dc557be85ff19f71d86b5300ca801f" }, { "id": "usd-valid-005", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "47cd578a3a2945e316b173a02a2589627ade54ea786b090a440a4c17c9b95501", "after_passed": true, "after_native_passed": true, "after_response_sha256": "47cd578a3a2945e316b173a02a2589627ade54ea786b090a440a4c17c9b95501" }, { "id": "usd2-valid-000", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "2f2893e214c453f265045eb749a974fe285eb292a31d78229c40958eba0acf24", "after_passed": true, "after_native_passed": true, "after_response_sha256": "2f2893e214c453f265045eb749a974fe285eb292a31d78229c40958eba0acf24" }, { "id": "usd2-valid-001", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "708483c19702a670a94c9bb53ead4230b21a11223d6f186d9b6de4d755132181", "after_passed": true, "after_native_passed": true, "after_response_sha256": "708483c19702a670a94c9bb53ead4230b21a11223d6f186d9b6de4d755132181" }, { "id": "usd2-valid-002", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "6a3a1505246156f3a7627e044688f55f634cecc645acbe24c4cf101036b2f31c", "after_passed": true, "after_native_passed": true, "after_response_sha256": "6a3a1505246156f3a7627e044688f55f634cecc645acbe24c4cf101036b2f31c" }, { "id": "usd2-valid-003", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "d44e2e0e51d9c0c8463bdccc91eb987d50634722ad578fd46ffb62bd5bd7974c", "after_passed": true, "after_native_passed": true, "after_response_sha256": "d44e2e0e51d9c0c8463bdccc91eb987d50634722ad578fd46ffb62bd5bd7974c" }, { "id": "usd2-valid-004", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "bf644e3ef02c369389eb56575d6108589aaabcabc1af380b61844aa28a0fd170", "after_passed": true, "after_native_passed": true, "after_response_sha256": "bf644e3ef02c369389eb56575d6108589aaabcabc1af380b61844aa28a0fd170" }, { "id": "usd2-valid-005", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "c4a6fb96ccd7bb060215740a3911e1ea8fe18444eaa1660aa1e2c2fe37db0b22", "after_passed": true, "after_native_passed": true, "after_response_sha256": "c4a6fb96ccd7bb060215740a3911e1ea8fe18444eaa1660aa1e2c2fe37db0b22" }, { "id": "usd2-valid-006", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "11829c42f67d31406e1bf2704ad3e3e983916db0afabe56ee398b40f13065383", "after_passed": true, "after_native_passed": true, "after_response_sha256": "11829c42f67d31406e1bf2704ad3e3e983916db0afabe56ee398b40f13065383" }, { "id": "usd2-valid-007", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "6dfa2971696571bf231cb6ae0528c8b91fff7a59ca6fc2521f71b1566455d93a", "after_passed": true, "after_native_passed": true, "after_response_sha256": "6dfa2971696571bf231cb6ae0528c8b91fff7a59ca6fc2521f71b1566455d93a" }, { "id": "usd2-valid-008", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "6deec35e0384186f0a8edc3728119235ce58c5981afd3ef242de9c28e70111f1", "after_passed": true, "after_native_passed": true, "after_response_sha256": "6deec35e0384186f0a8edc3728119235ce58c5981afd3ef242de9c28e70111f1" }, { "id": "usd2-valid-009", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "8796009709d94df4c1981bc7b58b596ac319014997f2b9089549aa5225926e1a", "after_passed": true, "after_native_passed": true, "after_response_sha256": "8796009709d94df4c1981bc7b58b596ac319014997f2b9089549aa5225926e1a" }, { "id": "usd2-valid-010", "group": "parameter_holdout", "before_passed": false, "before_native_passed": true, "before_response_sha256": "163f4b7f4665932a43dd252c038c8bb845ca10610d6e327e3975f63140706693", "after_passed": true, "after_native_passed": true, "after_response_sha256": "06415f7d7db0aeb7b8fcff6b2a410222fd7c4e87bda059b3a6167b6f0b20737b" }, { "id": "usd2-valid-011", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "70e49f25875754ace785ea6bc70ccec95746f5fde62206f1304535058f9dee66", "after_passed": true, "after_native_passed": true, "after_response_sha256": "70e49f25875754ace785ea6bc70ccec95746f5fde62206f1304535058f9dee66" }, { "id": "usd2-valid-012", "group": "composition_holdout", "before_passed": false, "before_native_passed": false, "before_response_sha256": "55e1961d41725b955a8cc68bfc80b06f2514ebf1b2ecd5e4ba19b9072c0a0aa4", "after_passed": false, "after_native_passed": false, "after_response_sha256": "1407e97b727e74379ae5c2330bdca11e35e0be108203e01b7a9f739444c71da4" }, { "id": "usd2-valid-013", "group": "composition_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "e0cb11fcb25bac8596c5b72b0ab766316711518e6444abd7561b9ee3b27c783c", "after_passed": false, "after_native_passed": false, "after_response_sha256": "377817ef89b68fd7a8958ac110f1cb66699b8bc4157749ef8eae5d39fec80286" }, { "id": "usd2-valid-014", "group": "composition_holdout", "before_passed": false, "before_native_passed": false, "before_response_sha256": "07f0da1fde1651ddba6233265d51a92322cdb61077a90af986bee76c60b85bab", "after_passed": false, "after_native_passed": true, "after_response_sha256": "a0cde3bc929b0d4c44ddce5c8d451263b956af12c917c20c30388d7132709113" }, { "id": "usd2-valid-015", "group": "composition_holdout", "before_passed": false, "before_native_passed": false, "before_response_sha256": "6a796d84b696938bce1d67d66dbe92d1713a44a3efc7a8f6de1a7f2de9dc9799", "after_passed": false, "after_native_passed": false, "after_response_sha256": "6a796d84b696938bce1d67d66dbe92d1713a44a3efc7a8f6de1a7f2de9dc9799" }, { "id": "usd2-valid-016", "group": "composition_holdout", "before_passed": false, "before_native_passed": true, "before_response_sha256": "64f9d36f32cc30099ef0c46cbd20fb2e6bc0a9997e1e94ac57d57325d5c93e35", "after_passed": false, "after_native_passed": false, "after_response_sha256": "bb39147d0c13ed0f39139a306fc5afa13c1a0d7c5c04d0df139caef4541ecdc3" }, { "id": "usd2-valid-017", "group": "composition_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "dbd891f8d8db3ab662ac1065cc096922a9261643a8e87d84d62e6c2fd86620e9", "after_passed": false, "after_native_passed": false, "after_response_sha256": "e36febfcfef66529bbd6215cd9c31fe07dc1f41de99afdc62db38dd7a6661732" }, { "id": "usd2-valid-018", "group": "composition_holdout", "before_passed": false, "before_native_passed": false, "before_response_sha256": "e06360550ba1d2e5577f31eaae70cf9b74fc3f0651def5309fb008d721a6ce4a", "after_passed": false, "after_native_passed": false, "after_response_sha256": "fdf0c7205ea508d201751574ce75176ee2b0eb3ee803513ba760db3dc704893d" }, { "id": "usd2-valid-019", "group": "composition_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "38dcbb8490c1cebd44f2f80894b5adf936efe319a1d93dff1408f4b3aa32d569", "after_passed": true, "after_native_passed": true, "after_response_sha256": "38dcbb8490c1cebd44f2f80894b5adf936efe319a1d93dff1408f4b3aa32d569" }, { "id": "usd2-valid-020", "group": "composition_holdout", "before_passed": false, "before_native_passed": false, "before_response_sha256": "2b5bd65fdbda6ad1617d8bd88b78fc6fa3912ed09c7cab686419b56a440fd7a1", "after_passed": false, "after_native_passed": true, "after_response_sha256": "f4be98145e0f276a60b5a5cbd7b8aed428e221ada0cf8fadaae0e316e4eb5614" }, { "id": "usd2-valid-021", "group": "composition_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "d46004260b8849ab9e3629aa12cafb76525664415ca2a3f9e098f6c8aae909bf", "after_passed": false, "after_native_passed": false, "after_response_sha256": "71449be150fdc0685fb468c093a8be1a2ad4c92ea3e72f86caa66da577bd9849" }, { "id": "usd2-valid-022", "group": "composition_holdout", "before_passed": false, "before_native_passed": false, "before_response_sha256": "a0016abe13d17d3218f7dcd8b72e3dd5fcac1619568f06c0eb3bbce08ab788f6", "after_passed": false, "after_native_passed": false, "after_response_sha256": "a0016abe13d17d3218f7dcd8b72e3dd5fcac1619568f06c0eb3bbce08ab788f6" }, { "id": "usd2-valid-023", "group": "composition_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "258384ebbe15ed7b7a8c9ff4640f05085680f6ade803d5289be02c815d6e003d", "after_passed": false, "after_native_passed": false, "after_response_sha256": "93cae9157af5efeb8b6d1ff3ebf42103aabb64ef7621bd99dbe12c2aeeb661d3" } ], "picogk": [ { "id": "valid-000", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "2fd5912a42194064927df3fcdade78486ed3acccb36e5f5931bcaaa992dbb959", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "2fd5912a42194064927df3fcdade78486ed3acccb36e5f5931bcaaa992dbb959", "after_compiled": true }, { "id": "valid-001", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "1781906d4f89ebcddbd9dfdba60fe87898c02fb11a2f8f9790f6fa5afab82303", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "1781906d4f89ebcddbd9dfdba60fe87898c02fb11a2f8f9790f6fa5afab82303", "after_compiled": true }, { "id": "valid-002", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "f99d9e2f84b88d63d397d93aa0c3294a824540594699ea16146960f9e178581b", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "f99d9e2f84b88d63d397d93aa0c3294a824540594699ea16146960f9e178581b", "after_compiled": true }, { "id": "valid-003", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "ba7179228a7841782f0abcec93e7b4e190f40e1da3b4bc3c12ae7a2451db5685", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "17ee21b257a1d10b64c299fc9254a6eba2dc20d775720f2ab7956abe606c365e", "after_compiled": true }, { "id": "valid-004", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "bdb97bb087c9a150eb086554c8a5492d761448d81248723f3d8212f7405e19de", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "39a7b1bb8163d020a51e03c2c6e6ba4fc8777470ef233838d0c936e4052cef2d", "after_compiled": true }, { "id": "valid-005", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "4fcfadb52c7faf4188b570341485fa32d49977cab78bdbc406d6b4a8b9a7f9d7", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "4fcfadb52c7faf4188b570341485fa32d49977cab78bdbc406d6b4a8b9a7f9d7", "after_compiled": true }, { "id": "valid-006", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "e2d9ba1d1e31db29697ba721211da11ee7d799b89dff94eca6497a9b47b46ffb", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "e2d9ba1d1e31db29697ba721211da11ee7d799b89dff94eca6497a9b47b46ffb", "after_compiled": true }, { "id": "valid-007", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "d560907b8af365753d51157641f1e7cd130b16a1e828b941e3e9df3bbb8a1274", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "426715cb5503e88bacfd8c2cab8f6a58332272499dbd25c95261c0d5ff5fadcb", "after_compiled": true }, { "id": "pico2-valid-000", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "3402078e5c892513e3664391710496ae21cbf5990f1d433f181758604d855c60", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "1c5801c29cfa00aec2bb9e3ea3465fe53eac9fbb5381c936b879deb9dbbcdc45", "after_compiled": true }, { "id": "pico2-valid-001", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "259f7379383b40e50f19e703295bc980348b410b90869dfadf1094410143fecf", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "4cf105a849055016f0fbe6c6cb75487f76fd2bb2bc3e13d00c619f6242c1ebd7", "after_compiled": true }, { "id": "pico2-valid-002", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "593d70e44a292946512c56619883ef17c1b2c02e8c5e8617df374e4ab81209c4", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "593d70e44a292946512c56619883ef17c1b2c02e8c5e8617df374e4ab81209c4", "after_compiled": true }, { "id": "pico2-valid-003", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "ec3950b32d74e80a243831c010b7ab2dd31dac8edfa4d490c835138a56986bba", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "8bacd1c8e0b334c11c070e4b0ad8bfbf0e26dfcef759d937991b541c6ddbad98", "after_compiled": true }, { "id": "pico2-valid-004", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "8f41b5ca9fe2116b5e3dad400db0789defdc6b159fc3274973484c7622f101fa", "before_compiled": true, "after_passed": false, "after_native_passed": false, "after_response_sha256": "8f39f7ae60958389c281653633ee81421000c874a12333d5ec6b57daa41f8ba2", "after_compiled": false }, { "id": "pico2-valid-005", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "7c17d9dcaec1f921144071d0165265816ddd5b223fdf214d45dcfe62fbbc409e", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "37d132d893e6f6df8bd89c1926cc56c3368568a368d3b1999478ef0f5c7182e0", "after_compiled": true }, { "id": "pico2-valid-006", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "16c4db74c184e330432e3a9ad8721f6eec6dfcc6ea08e2c3039209b94c8e3543", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "e1ee4644c51b368a50ac412ebc7bd9c32be1efe72a5c074fe27a8f57e842e32b", "after_compiled": true }, { "id": "pico2-valid-007", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "c8ee67e78f97cba43bf14bef3b97efc1fafe14f387eef30c09b9c8e50547411c", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "a1db14471ee7bdda45689285a0a13e1387e2bc8b1a701474811193e55c87ea45", "after_compiled": true }, { "id": "pico2-valid-008", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "099a48b605c22656b0265f6cfff576f7a37c93b6bc253a72eac2134805b031f0", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "099a48b605c22656b0265f6cfff576f7a37c93b6bc253a72eac2134805b031f0", "after_compiled": true }, { "id": "pico2-valid-009", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "8d74f660d80ace522309a4ce99b5edf16f6a9a6516c238e6cab0bf3ac1269a21", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "50d82e4797d8a9e5ede2a41c4baa3851c230844de78c6c0fb96f6bbdce9107db", "after_compiled": true }, { "id": "pico2-valid-010", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "494f220f61abae7886df86233e97900ec2f683a750abbba6d7286d85a8a80ed2", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "da8b0513396384c348a32b0927de655ff45def34d914ad322420574af78218ed", "after_compiled": true }, { "id": "pico2-valid-011", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "bae297979e0fd17668bedbd6cd253c787e7e2f9f493066011764ec4f72c857f1", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "f3dc49ff4f3d66b962a95fbb6f0d5e45ef0389782c1575064e40472f952b2fb4", "after_compiled": true }, { "id": "pico2-valid-012", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "37a21c8acbaf009355862a84a312034502f1aed5874921c7cc88107c51d5fee4", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "8b2a1e0188f8dac2cd9867abafd0760bb30d2064df6554733bc1eb279c312f67", "after_compiled": true }, { "id": "pico2-valid-013", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "0041424962fd7300abc64a86595c129556225abe03e6d0011a5a2653681ce168", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "5d09d8703cd1bb564fa9c6b830a2adff621e9131c28536842afe412b7e509f14", "after_compiled": true }, { "id": "pico2-valid-014", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "255c34b7c25869894f9db8593585278c3d8df533a5ec9ad608fba064f056042f", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "8f64ff04c587fe8d57409fb7c2542fc600155b7abdcd120ecb795f10ffec1853", "after_compiled": true }, { "id": "pico2-valid-015", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "0c3d425f31607e52b2c255159648301a53b027eb0ebff7e49545b6f7a39ba598", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "00741c0d91c5a8852ad3aaf2fa425d9e5bb9e5561307a1fb90e499fdd7bd0ff3", "after_compiled": true }, { "id": "pico3-valid-000", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "6e12f05ab1215e4f90a81e485d1137afc386ffbb43b7a3265b7f4b41b1e404af", "before_compiled": true, "after_passed": false, "after_native_passed": false, "after_response_sha256": "4bdbb1428b35d7104ecc2b3f31dc0065ac1cad6418402ec03a5ad3a15e32796b", "after_compiled": false }, { "id": "pico3-valid-001", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "0ee373a3519fb3038a2227c6a3794ec45240501b5266ec06261bd1938c29ae7f", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "7376947e8c2bbba9e56911a4c5ba48f350f5d14c3291456a95b04ca0b7af4259", "after_compiled": true }, { "id": "pico3-valid-002", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": false, "before_response_sha256": "c885046c63b31e4c2e8324117536d54b0107fc2c33733c89a72c0351acc98da6", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "124790dc85e60e378caa55ab533d5a89cfdb5c73870da3332c791380a6f00b24", "after_compiled": true }, { "id": "pico3-valid-003", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "9291117181b3f2e016a891edf16b6571488999599ce0ee51f692e7490e86fec4", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "9291117181b3f2e016a891edf16b6571488999599ce0ee51f692e7490e86fec4", "after_compiled": true }, { "id": "pico3-valid-004", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "cf2cd09c1adad1e40a5544e5f6fa3b5e6b8cd7baaf358e8b0259a3692754db9e", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "456ee44c8c7f97051427bd52944f6698eab3d29eacd8657d3295c39430f0fda4", "after_compiled": true }, { "id": "pico3-valid-005", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "bb111f12eef4dfa028f1e3aff198b70e5e9c123a3f9df046d407c1e016c8f84c", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "ba35364a0374e2f0b2fb4fafcf766db9ea90f328046da739982a9fd078b20395", "after_compiled": true }, { "id": "pico3-valid-006", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "c34dd439f08f475dbafc8160b6ea015ef1e4ee1fc8c62ca56bc630b709310230", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "fc86db2c8d2aef4f15b1373dec85e0a31a6cecd09d5cc102a330ce0f485abadc", "after_compiled": true }, { "id": "pico3-valid-007", "group": "fresh_graph_instances", "before_passed": true, "before_native_passed": true, "before_response_sha256": "94ea776d2796260545f477e5118b7565c184a97b138fdb7e2830480b0ea8ff7b", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "a76232eb3ff0810b78def9836721b51a04feff319a8393531f66a79bfab8ee27", "after_compiled": true }, { "id": "pico3-valid-008", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": false, "before_response_sha256": "312f3af870cc2c04882c8463b053e0c6ff9a489a386f650b37feb354a30a8f3e", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "24db113f658be26b0d3941f676d7f29bbd4b626f4adb7ccf8655e3f7cf38e308", "after_compiled": true }, { "id": "pico3-valid-009", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "94d33bbefa267c76ae75fe8caa56b3af99dd483028477e33025c6a164d227bad", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "82fb5b6b3ddae47fa92eb140fcfc2f3dc2944ca432d3d4637a92e7b6aa78b7ce", "after_compiled": true }, { "id": "pico3-valid-010", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "0b5f9d6e408e8a9a88f5aee7c989196254bcfa59b070481da2ba4f59177dac8d", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "c8d2bae3ff3c4cc79cb4ea1b35b8fad1b95098e7d3404c885ff1851ecb671f70", "after_compiled": true }, { "id": "pico3-valid-011", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "eea7f7f90a3b77abfa3e560467341e172a24471d95aa97a99edac85b738a2286", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "3740bee4a00d213fa6c66d51a1508886395c702aa25a4e3b66517e644cda775c", "after_compiled": true }, { "id": "pico3-valid-012", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": false, "before_response_sha256": "91d449c3ab58d368c59ed177a8e90fa1299762aed49b138130a79418ce4282b5", "before_compiled": true, "after_passed": false, "after_native_passed": false, "after_response_sha256": "8210448976f8431a43954d0ee68872c3e3a6c9c0d21b1b070eade3331243be71", "after_compiled": true }, { "id": "pico3-valid-013", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "4ba17bc72b8bc42140a36cec06bd6c595e49ce1a87cc716e09e41c24ea865028", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "cf385213128a2e2f0b1a9f0764b1635749601a28a9afa6368651932b2bd98510", "after_compiled": true }, { "id": "pico3-valid-014", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "a44312273ddc34f8ea0418746238e4e1c0b7fab61e3d5d780b2bb88e1074a548", "before_compiled": true, "after_passed": false, "after_native_passed": false, "after_response_sha256": "4729bbae341f26020831b72908b84d247d0445ee0ab60f18a8fc1debd72e4235", "after_compiled": true }, { "id": "pico3-valid-015", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "2d6ce1108429de17adcf239598477466d57b98d7433cc3a019b1f159f7740357", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "f8d3e6850e2b8c0d2bbd4a0115bb2b7b542c9c29504cb38b29d40af2d877f07b", "after_compiled": true } ] }, "final_tests_evaluated": false, "default_inference_changed": false, "training_diagnostics": { "iterations": 800, "trained_tokens": 135924, "validation_loss": 0.024, "peak_memory_gb": 2.118 }, "verification": { "main_suite_tests": 3256, "optional_skips": 160, "main_suite_passed": true, "make_check_exit_code": 2, "make_check_blocker": "Docker daemon unavailable at the existing F37 LPBF audit", "native_picogk_reference_programs": 64, "repository_checks_exit_code": 0, "strict_documentation_broken_links": 0, "strict_documentation_files": 646, "focused_tests_after_report_check": 6, "focused_tests_passed": true }, "artifacts_sha256": { "coding-006/manifest.json": "769290786149f5f5fde22392058956b91dc5df9d37cb31c15737afb0d600ae6f", "coding-006/registration.json": "d4d95c4a199a530dc85c76a4e105d09f7c2f9c15b694f7e0be237c5bf080b637", "coding-006/selection.json": "aee338ebc057174a8e77373135fa3facddd6ec6fdc144061979bffa6643c2ad7", "coding-006/native-oracles.json": "35c1ecf7271ebbf155ce76c523b07f8ba931506840e60178ad888a8c7ab9e8e7", "coding-006/training.log": "51367b7b8b0f75ba45d92938f8abfdfcc6ee34391299ac1ecbac639c2b449421", "coding-006/training-command.json": "ab1ad33373bb9a38692a68e58ad5c7aef4aba5eae358df650acec15bf2207484", "coding-005-check.log": "68f9058ca7195c6b1e69b885456067e058b7c2f55075474e91816da1a638a4f4", "coding-006/valid-800-responses.jsonl": "aee8fd53051798a0ccda22635dbfef39ec6c491c1ad15c7e5a620b38b0cc73fc", "coding-006/valid-before-responses.jsonl": "c195c9f080738cb0ebece1cd4380ee4209876f1bddf9ac2c1359fb7af9730ce5", "coding-006/valid-800-scores.jsonl": "825509e91571470eefe333aaf56b654d5715a92594427a3b336175c63431c610", "coding-006/valid-before-scores.jsonl": "c5cea91f90755ae4c19e92f029c66da6f848a6393a4cb559b5518299bc2539ef", "coding-006/valid-800-picogk/adapter-evaluation.json": "bd536cdf0f2c35de6b7007af497d1c56c69496010e7cfc8b3adc68fe19bad684", "coding-006/valid-before-picogk/adapter-evaluation.json": "9f12e7ea1149d64a467189cf9021a2c84be396c5713662baecd9b091ce0d9a3f", "coding-006-repository-checks-staged.log": "977de4bf498dadc8a394321c6fd0fb37190527739804cf8f245b9260feca8f27", "coding-006-receipt-check.log": "e82aa5d35ae538fbed08a2d6b4d4c80e76b10a22aa9545d1d115c33e5731bf62" }, "limitations": [ "Validation-driven continuation after the rejected coding-004; one adaptive experiment family.", "Only the 24 pico3-test cases were untouched by model inference before validation selection. All older tests are development-informed regressions.", "Shared synthetic grammar and API recipes; no independent task-family benchmark.", "The installed Pixar wheel lacks shader-discovery resources; ShaderPropertyTypeConformanceChecker remains excluded.", "No OpenFOAM evaluation, cylinder-head accuracy, material, manufacturing or physical validation.", "Adapters remain local and experimental; default inference has not been changed." ], "previous_attempt": { "experiment": "m64-qwen-coding-005", "date": "2026-10-02", "status": "no_validation_improvement", "decision": "rejected_for_combined_assistant", "selected_step": null, "candidate_step": 600, "candidate_adapter_sha256": "64e3c57827d867f052c095efc84265d6bcb8509cf548aea3946826c40c5ddfb4", "comparison_adapter": "coding-003/checkpoint-600", "warm_start_adapter": "coding-004/checkpoint-800", "manifest": { "base_model": { "repository": "mlx-community/Qwen2.5-Coder-1.5B-Instruct-4bit", "revision": "b3252a2f97102b1fb1571fec2c9b27219a8536be", "weights_sha256": "daeab4764fb420d161721791cf2e509e2de81a7af4223646e7bed2bf82c57b58", "upstream_repository": "Qwen/Qwen2.5-Coder-1.5B-Instruct", "upstream_revision": "2e1fd397ee46e1388853d2af2c993145b0f1098a", "license": "Apache-2.0", "quantization_bits": 4, "mlx_lm_version": "0.31.3" }, "initial_adapter_sha256": "a1014c6f71b46d879c09462c4a57d17d00df8f11b8178dd53bf6d5f2ac2f778a", "source_sha256": { "training/m64-qwen/improve.py": "08b2cc6dc9fd779e509a860ae6205a0c8bcee2f6349cef249949376131bb9cdd", "training/m64-qwen/picogk.py": "3357a289b0314a4e7e24f1a0d515d1bebf55021ccacd2551650b830f352ac3c6", "training/m64-engineer/openusd.py": "eb28db525af0ed21b4f80758a39a67482029510ba26fa23fc2251a3a4e6b4bb6" }, "files_sha256": { "usd-cases.jsonl": "d18eb56c5f21679b8e37f05c1dd8442eaf600d967de063ffd358b0e678b98437", "picogk-cases.json": "dd4360574826ddb1f58671e6b0472e3f64e897ff5c1127b47f417d13248d6812", "data/test.jsonl": "e7752ee9c69737a2c6e0b872ed0a8376d32773f902fc19d887df22a9eaad7a0f", "data/train.jsonl": "e0fcd6da66263c9a094993f955456e5f35ec66ee14bb9c8c28e6e1caa98c49e2", "data/valid.jsonl": "0714942456813dc1deb42c6abed58fc249c2ac8841b1b49b3c00eea2188c4233" }, "counts": { "train": { "usd": 336, "picogk": 704 }, "valid": { "usd": 30, "picogk": 40 }, "test": { "usd": 36, "picogk": 56 } }, "maximum_sequence_tokens": 547, "steps": [ 600 ], "learning_rate": 5e-05, "seed": 42, "warm_start_sha256": "84323d7364f61091093b6a105b4d3371cbb71c3aa723809150814d17a98979a3", "replay_weight": 1, "usd_replay_weight": 4, "effective_counts": { "train": 2048, "valid": 70, "test": 92 }, "num_layers": 16, "rank": 8, "lora_scale": 20, "batch_size": 1, "mask_prompt": true, "selection": "Higher PicoGK validation count with no per-case USD or PicoGK regression.", "test_policy": "Tests open only after selection. Only pico3-test cases are fresh in a focused run; all older tests are regressions. Shared grammar, not independent families.", "native_sha256": { "PicoGK.dll": "710c4e0428efbe9c1cd1db1663247b45604f96d167a6726782ab8ddca350565a", "picogk.26.2.dylib": "5c1cd3fc12766a85bc5c1441a95f065fb274338d1333c3ae2d3396b59d0e18e7" } }, "registration": { "registered_at_utc": "2026-10-01T22:26:15.500807+00:00", "protocol_commit": "2e993ac6bb4d0898675d2af398bf0b12a96b8306", "protocol_sha256": "db71a248d8a6cb39437234c474db5b29be1e8d6b5785b85d8fa5ed8edd8e4ccb", "manifest_sha256": "92e22a392ec0f2c28ef9856b7f1e2c79193f0a28e4372cc29bca1f11334757a1", "final_tests_opened": false }, "validation_summary": { "usd": { "total": 30, "regressions": [ "usd-valid-004", "usd2-valid-004", "usd2-valid-013", "usd2-valid-017", "usd2-valid-019", "usd2-valid-021", "usd2-valid-023" ], "gains": [ "usd2-valid-010" ], "before_passed": 22, "before_native_passed": 24, "after_passed": 16, "after_native_passed": 30 }, "picogk": { "total": 40, "regressions": [ "valid-002", "valid-006" ], "gains": [ "pico2-valid-000", "pico2-valid-001", "pico2-valid-006", "pico2-valid-007", "pico2-valid-008", "pico2-valid-011", "pico2-valid-012", "pico2-valid-013", "pico2-valid-014", "pico3-valid-000", "pico3-valid-001", "pico3-valid-002", "pico3-valid-003", "pico3-valid-006", "pico3-valid-008", "pico3-valid-012", "pico3-valid-013", "pico3-valid-014" ], "before_passed": 9, "before_native_passed": 37, "before_compiled": 40, "after_passed": 25, "after_native_passed": 31, "after_compiled": 32 } }, "validation_cases": { "usd": [ { "id": "usd-valid-000", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "017bc92d9d1be753373b064d1737341216e29927b0f264c3a8f1f8d6e06e22d0", "after_passed": true, "after_native_passed": true, "after_response_sha256": "017bc92d9d1be753373b064d1737341216e29927b0f264c3a8f1f8d6e06e22d0" }, { "id": "usd-valid-001", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "31964fa538d592cf703ba661026574a2a25b970b3e1d2dfd2a84df324b542593", "after_passed": true, "after_native_passed": true, "after_response_sha256": "31964fa538d592cf703ba661026574a2a25b970b3e1d2dfd2a84df324b542593" }, { "id": "usd-valid-002", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "5d06a1edc6286362da523ac9973f3b3477c7f2c29646955522fdd48073582f4a", "after_passed": true, "after_native_passed": true, "after_response_sha256": "5d06a1edc6286362da523ac9973f3b3477c7f2c29646955522fdd48073582f4a" }, { "id": "usd-valid-003", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "734c04ed5ca2ef362599e2110d01b5d0720ec139e97724dd45e19cc49fac983c", "after_passed": true, "after_native_passed": true, "after_response_sha256": "734c04ed5ca2ef362599e2110d01b5d0720ec139e97724dd45e19cc49fac983c" }, { "id": "usd-valid-004", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "6c869ea563f35ced578c2024beb1c87715dc557be85ff19f71d86b5300ca801f", "after_passed": false, "after_native_passed": true, "after_response_sha256": "b0ab2b7c0409a3f10dcb392b74cc1f63444dccb76f7e30a9c2491c18505633a2" }, { "id": "usd-valid-005", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "47cd578a3a2945e316b173a02a2589627ade54ea786b090a440a4c17c9b95501", "after_passed": true, "after_native_passed": true, "after_response_sha256": "47cd578a3a2945e316b173a02a2589627ade54ea786b090a440a4c17c9b95501" }, { "id": "usd2-valid-000", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "2f2893e214c453f265045eb749a974fe285eb292a31d78229c40958eba0acf24", "after_passed": true, "after_native_passed": true, "after_response_sha256": "2f2893e214c453f265045eb749a974fe285eb292a31d78229c40958eba0acf24" }, { "id": "usd2-valid-001", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "708483c19702a670a94c9bb53ead4230b21a11223d6f186d9b6de4d755132181", "after_passed": true, "after_native_passed": true, "after_response_sha256": "708483c19702a670a94c9bb53ead4230b21a11223d6f186d9b6de4d755132181" }, { "id": "usd2-valid-002", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "6a3a1505246156f3a7627e044688f55f634cecc645acbe24c4cf101036b2f31c", "after_passed": true, "after_native_passed": true, "after_response_sha256": "6a3a1505246156f3a7627e044688f55f634cecc645acbe24c4cf101036b2f31c" }, { "id": "usd2-valid-003", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "d44e2e0e51d9c0c8463bdccc91eb987d50634722ad578fd46ffb62bd5bd7974c", "after_passed": true, "after_native_passed": true, "after_response_sha256": "d44e2e0e51d9c0c8463bdccc91eb987d50634722ad578fd46ffb62bd5bd7974c" }, { "id": "usd2-valid-004", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "bf644e3ef02c369389eb56575d6108589aaabcabc1af380b61844aa28a0fd170", "after_passed": false, "after_native_passed": true, "after_response_sha256": "461d71ec5f5191487e349a9579e83ebb6ea18a024c273b5ae7e5399ac6b31947" }, { "id": "usd2-valid-005", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "c4a6fb96ccd7bb060215740a3911e1ea8fe18444eaa1660aa1e2c2fe37db0b22", "after_passed": true, "after_native_passed": true, "after_response_sha256": "c4a6fb96ccd7bb060215740a3911e1ea8fe18444eaa1660aa1e2c2fe37db0b22" }, { "id": "usd2-valid-006", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "11829c42f67d31406e1bf2704ad3e3e983916db0afabe56ee398b40f13065383", "after_passed": true, "after_native_passed": true, "after_response_sha256": "11829c42f67d31406e1bf2704ad3e3e983916db0afabe56ee398b40f13065383" }, { "id": "usd2-valid-007", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "6dfa2971696571bf231cb6ae0528c8b91fff7a59ca6fc2521f71b1566455d93a", "after_passed": true, "after_native_passed": true, "after_response_sha256": "6dfa2971696571bf231cb6ae0528c8b91fff7a59ca6fc2521f71b1566455d93a" }, { "id": "usd2-valid-008", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "6deec35e0384186f0a8edc3728119235ce58c5981afd3ef242de9c28e70111f1", "after_passed": true, "after_native_passed": true, "after_response_sha256": "6deec35e0384186f0a8edc3728119235ce58c5981afd3ef242de9c28e70111f1" }, { "id": "usd2-valid-009", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "8796009709d94df4c1981bc7b58b596ac319014997f2b9089549aa5225926e1a", "after_passed": true, "after_native_passed": true, "after_response_sha256": "8796009709d94df4c1981bc7b58b596ac319014997f2b9089549aa5225926e1a" }, { "id": "usd2-valid-010", "group": "parameter_holdout", "before_passed": false, "before_native_passed": true, "before_response_sha256": "163f4b7f4665932a43dd252c038c8bb845ca10610d6e327e3975f63140706693", "after_passed": true, "after_native_passed": true, "after_response_sha256": "06415f7d7db0aeb7b8fcff6b2a410222fd7c4e87bda059b3a6167b6f0b20737b" }, { "id": "usd2-valid-011", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "70e49f25875754ace785ea6bc70ccec95746f5fde62206f1304535058f9dee66", "after_passed": true, "after_native_passed": true, "after_response_sha256": "70e49f25875754ace785ea6bc70ccec95746f5fde62206f1304535058f9dee66" }, { "id": "usd2-valid-012", "group": "composition_holdout", "before_passed": false, "before_native_passed": false, "before_response_sha256": "55e1961d41725b955a8cc68bfc80b06f2514ebf1b2ecd5e4ba19b9072c0a0aa4", "after_passed": false, "after_native_passed": true, "after_response_sha256": "14156e1315e831f08c4174ed33053d027d6c536090b255efd8d774b4ea434f98" }, { "id": "usd2-valid-013", "group": "composition_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "e0cb11fcb25bac8596c5b72b0ab766316711518e6444abd7561b9ee3b27c783c", "after_passed": false, "after_native_passed": true, "after_response_sha256": "64597afb3153cdaf5d966abb048200fcfb8c384c4858e9299456709bec181b55" }, { "id": "usd2-valid-014", "group": "composition_holdout", "before_passed": false, "before_native_passed": false, "before_response_sha256": "07f0da1fde1651ddba6233265d51a92322cdb61077a90af986bee76c60b85bab", "after_passed": false, "after_native_passed": true, "after_response_sha256": "6511501facffe2c1433bd54bb5139645bf966feec7ce1ab68d3ecc0e5d988afd" }, { "id": "usd2-valid-015", "group": "composition_holdout", "before_passed": false, "before_native_passed": false, "before_response_sha256": "6a796d84b696938bce1d67d66dbe92d1713a44a3efc7a8f6de1a7f2de9dc9799", "after_passed": false, "after_native_passed": true, "after_response_sha256": "adbe8360c8bd12ecb35f604dd972872e1c8d088628add1a3cd54f182afd71829" }, { "id": "usd2-valid-016", "group": "composition_holdout", "before_passed": false, "before_native_passed": true, "before_response_sha256": "64f9d36f32cc30099ef0c46cbd20fb2e6bc0a9997e1e94ac57d57325d5c93e35", "after_passed": false, "after_native_passed": true, "after_response_sha256": "743d4a070c92b9c9a1ece72ce2955a5473d6ec267d704cfb895baddb2e564041" }, { "id": "usd2-valid-017", "group": "composition_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "dbd891f8d8db3ab662ac1065cc096922a9261643a8e87d84d62e6c2fd86620e9", "after_passed": false, "after_native_passed": true, "after_response_sha256": "cd067ae4f38d7f62e8600f3a533e70d08258cb4b8413b201f03be99021dca662" }, { "id": "usd2-valid-018", "group": "composition_holdout", "before_passed": false, "before_native_passed": false, "before_response_sha256": "e06360550ba1d2e5577f31eaae70cf9b74fc3f0651def5309fb008d721a6ce4a", "after_passed": false, "after_native_passed": true, "after_response_sha256": "2917aaf7ea00e32df3cbf8b69578d8d9bc0a1c561efed92a115c811958040344" }, { "id": "usd2-valid-019", "group": "composition_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "38dcbb8490c1cebd44f2f80894b5adf936efe319a1d93dff1408f4b3aa32d569", "after_passed": false, "after_native_passed": true, "after_response_sha256": "1bc9938a877657d0b1bd499e4bae1c96665a4805ccf90b2fb01304be3833d447" }, { "id": "usd2-valid-020", "group": "composition_holdout", "before_passed": false, "before_native_passed": false, "before_response_sha256": "2b5bd65fdbda6ad1617d8bd88b78fc6fa3912ed09c7cab686419b56a440fd7a1", "after_passed": false, "after_native_passed": true, "after_response_sha256": "6945514c2320ec93e8d5f45f01407b3f8148ca8360e2b752ee5cd6a4ea6a04d8" }, { "id": "usd2-valid-021", "group": "composition_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "d46004260b8849ab9e3629aa12cafb76525664415ca2a3f9e098f6c8aae909bf", "after_passed": false, "after_native_passed": true, "after_response_sha256": "cbe76c4652d3e15b0367bb1a18f651b8819087881afaef055083bd217f3b3b06" }, { "id": "usd2-valid-022", "group": "composition_holdout", "before_passed": false, "before_native_passed": false, "before_response_sha256": "a0016abe13d17d3218f7dcd8b72e3dd5fcac1619568f06c0eb3bbce08ab788f6", "after_passed": false, "after_native_passed": true, "after_response_sha256": "467ff0ff8a4816f4cd9104786fe014550671cbbba74764e8d87756106b05e9d5" }, { "id": "usd2-valid-023", "group": "composition_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "258384ebbe15ed7b7a8c9ff4640f05085680f6ade803d5289be02c815d6e003d", "after_passed": false, "after_native_passed": true, "after_response_sha256": "6b33d77b8cd3ba10fd5a06c598d9b2491afb892095e912250f852ec53a1de64f" } ], "picogk": [ { "id": "valid-000", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "2fd5912a42194064927df3fcdade78486ed3acccb36e5f5931bcaaa992dbb959", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "2fd5912a42194064927df3fcdade78486ed3acccb36e5f5931bcaaa992dbb959", "after_compiled": true }, { "id": "valid-001", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "1781906d4f89ebcddbd9dfdba60fe87898c02fb11a2f8f9790f6fa5afab82303", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "1781906d4f89ebcddbd9dfdba60fe87898c02fb11a2f8f9790f6fa5afab82303", "after_compiled": true }, { "id": "valid-002", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "f99d9e2f84b88d63d397d93aa0c3294a824540594699ea16146960f9e178581b", "before_compiled": true, "after_passed": false, "after_native_passed": false, "after_response_sha256": "4632810a639a3c0c8667a0283371bb7359164d09dcfc13a8b62a6b5c57881608", "after_compiled": false }, { "id": "valid-003", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "ba7179228a7841782f0abcec93e7b4e190f40e1da3b4bc3c12ae7a2451db5685", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "ba7179228a7841782f0abcec93e7b4e190f40e1da3b4bc3c12ae7a2451db5685", "after_compiled": true }, { "id": "valid-004", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "bdb97bb087c9a150eb086554c8a5492d761448d81248723f3d8212f7405e19de", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "bdb97bb087c9a150eb086554c8a5492d761448d81248723f3d8212f7405e19de", "after_compiled": true }, { "id": "valid-005", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "4fcfadb52c7faf4188b570341485fa32d49977cab78bdbc406d6b4a8b9a7f9d7", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "4fcfadb52c7faf4188b570341485fa32d49977cab78bdbc406d6b4a8b9a7f9d7", "after_compiled": true }, { "id": "valid-006", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "e2d9ba1d1e31db29697ba721211da11ee7d799b89dff94eca6497a9b47b46ffb", "before_compiled": true, "after_passed": false, "after_native_passed": false, "after_response_sha256": "699c9b623b99a08b70ae9d871e6c2e7200fb77f4cc76a191c5a899011b9be294", "after_compiled": false }, { "id": "valid-007", "group": "parameter_holdout", "before_passed": true, "before_native_passed": true, "before_response_sha256": "d560907b8af365753d51157641f1e7cd130b16a1e828b941e3e9df3bbb8a1274", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "d560907b8af365753d51157641f1e7cd130b16a1e828b941e3e9df3bbb8a1274", "after_compiled": true }, { "id": "pico2-valid-000", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "3402078e5c892513e3664391710496ae21cbf5990f1d433f181758604d855c60", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "971c899aef72c8bebf9fdb3b6785b321f8146c82d51ae49c1e9f8797aba0e3d6", "after_compiled": true }, { "id": "pico2-valid-001", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "259f7379383b40e50f19e703295bc980348b410b90869dfadf1094410143fecf", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "0b98bd8ca0a562ace3723b2d5dd5b37883eb84a4991d9f04d86656bcd7d59cab", "after_compiled": true }, { "id": "pico2-valid-002", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "593d70e44a292946512c56619883ef17c1b2c02e8c5e8617df374e4ab81209c4", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "9797004bad2e6f4158770e3b3ffe943bd6c113069846d1764d02ba0a84d0b319", "after_compiled": true }, { "id": "pico2-valid-003", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "ec3950b32d74e80a243831c010b7ab2dd31dac8edfa4d490c835138a56986bba", "before_compiled": true, "after_passed": false, "after_native_passed": false, "after_response_sha256": "93dcb5e3521a30f2a7fd3fd64589eb879cc4b9ffdb2e94539ab6f0b98eed8398", "after_compiled": false }, { "id": "pico2-valid-004", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "8f41b5ca9fe2116b5e3dad400db0789defdc6b159fc3274973484c7622f101fa", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "993fa2b8d4d9174cb13e01ca6d73495997a52c2415e154c2f36e2a627ae53d3c", "after_compiled": true }, { "id": "pico2-valid-005", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "7c17d9dcaec1f921144071d0165265816ddd5b223fdf214d45dcfe62fbbc409e", "before_compiled": true, "after_passed": false, "after_native_passed": false, "after_response_sha256": "4948f7a5dee270629d4fb6e2316362e9705c7020e4e6a5a300f15a33512ec82c", "after_compiled": true }, { "id": "pico2-valid-006", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "16c4db74c184e330432e3a9ad8721f6eec6dfcc6ea08e2c3039209b94c8e3543", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "527e1a0dd0be07c035728b0d53f608ac02328832a4b5eaa98a6cabc051b3aa14", "after_compiled": true }, { "id": "pico2-valid-007", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "c8ee67e78f97cba43bf14bef3b97efc1fafe14f387eef30c09b9c8e50547411c", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "d99b08ab0343f23fdc7bec0583ab779f8fa200f379db758007a3c5b05fe0b24e", "after_compiled": true }, { "id": "pico2-valid-008", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "099a48b605c22656b0265f6cfff576f7a37c93b6bc253a72eac2134805b031f0", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "ba55add88e735d83d7ee05290e9b26a7bf1ce1e95f47db29e973a6a16f7297c1", "after_compiled": true }, { "id": "pico2-valid-009", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "8d74f660d80ace522309a4ce99b5edf16f6a9a6516c238e6cab0bf3ac1269a21", "before_compiled": true, "after_passed": false, "after_native_passed": false, "after_response_sha256": "bfec81fac137ed04a3cdf4c81bfc6d2d534577d474a47fbcd2445d4890533248", "after_compiled": false }, { "id": "pico2-valid-010", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "494f220f61abae7886df86233e97900ec2f683a750abbba6d7286d85a8a80ed2", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "b33837b86879414cef70b27d42d5436f4fbe22b1db7b391911f1088e6dad6305", "after_compiled": true }, { "id": "pico2-valid-011", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "bae297979e0fd17668bedbd6cd253c787e7e2f9f493066011764ec4f72c857f1", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "9795fbf340a8878861442d5fe3022e97651c57c5862162e8fe54f5681e5ad408", "after_compiled": true }, { "id": "pico2-valid-012", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "37a21c8acbaf009355862a84a312034502f1aed5874921c7cc88107c51d5fee4", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "4d454a22ed6e8d23d469890e83268f89e53266adfc1a6b363e50c60cfe2375d0", "after_compiled": true }, { "id": "pico2-valid-013", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "0041424962fd7300abc64a86595c129556225abe03e6d0011a5a2653681ce168", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "ae3552e9f148f5100c6b3d17e903cb83367bb74ecb955e5fc5af7871ca9b43f1", "after_compiled": true }, { "id": "pico2-valid-014", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "255c34b7c25869894f9db8593585278c3d8df533a5ec9ad608fba064f056042f", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "8f64ff04c587fe8d57409fb7c2542fc600155b7abdcd120ecb795f10ffec1853", "after_compiled": true }, { "id": "pico2-valid-015", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "0c3d425f31607e52b2c255159648301a53b027eb0ebff7e49545b6f7a39ba598", "before_compiled": true, "after_passed": false, "after_native_passed": false, "after_response_sha256": "985a9ac842d9fb1ba52744f0cff5a567ef2e28852abb660dae2a15e9badfc7e5", "after_compiled": false }, { "id": "pico3-valid-000", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "6e12f05ab1215e4f90a81e485d1137afc386ffbb43b7a3265b7f4b41b1e404af", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "bcf5b3718c7b37fa7c70307934495e49e84b8b70d01d64a631ae8fea5b9e9b7b", "after_compiled": true }, { "id": "pico3-valid-001", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "0ee373a3519fb3038a2227c6a3794ec45240501b5266ec06261bd1938c29ae7f", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "69115227e79413cda7042d7b1d843bc9af221a444ec1de811cfdd83c0ef2d52a", "after_compiled": true }, { "id": "pico3-valid-002", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": false, "before_response_sha256": "c885046c63b31e4c2e8324117536d54b0107fc2c33733c89a72c0351acc98da6", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "350698c6d7c6364f72b97d8b5f4a7d66414dce4bb5cfc87651012e35fde073c9", "after_compiled": true }, { "id": "pico3-valid-003", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "9291117181b3f2e016a891edf16b6571488999599ce0ee51f692e7490e86fec4", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "b551869b03b4438f8d1b72cd24f78c30a5affd32626a9762b8e171fcf56ba95f", "after_compiled": true }, { "id": "pico3-valid-004", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "cf2cd09c1adad1e40a5544e5f6fa3b5e6b8cd7baaf358e8b0259a3692754db9e", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "f775fb982e4c125f20718e6a5f21516f85129428b94a9ca6fbbc2a583f271311", "after_compiled": true }, { "id": "pico3-valid-005", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "bb111f12eef4dfa028f1e3aff198b70e5e9c123a3f9df046d407c1e016c8f84c", "before_compiled": true, "after_passed": false, "after_native_passed": false, "after_response_sha256": "1300e270ba46bd6835524eadab697178eb2f574f4843fcc209f0a4cc71e3cb8f", "after_compiled": false }, { "id": "pico3-valid-006", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "c34dd439f08f475dbafc8160b6ea015ef1e4ee1fc8c62ca56bc630b709310230", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "fc86db2c8d2aef4f15b1373dec85e0a31a6cecd09d5cc102a330ce0f485abadc", "after_compiled": true }, { "id": "pico3-valid-007", "group": "fresh_graph_instances", "before_passed": true, "before_native_passed": true, "before_response_sha256": "94ea776d2796260545f477e5118b7565c184a97b138fdb7e2830480b0ea8ff7b", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "81171cea890b36ab09b933e3f24dabc0c08d167e2ea1193e662400fc494eee5b", "after_compiled": true }, { "id": "pico3-valid-008", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": false, "before_response_sha256": "312f3af870cc2c04882c8463b053e0c6ff9a489a386f650b37feb354a30a8f3e", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "b838ace0e083b1cd2649f7e50163c9332d13487db7c7d26b9abd5d31ac6720dc", "after_compiled": true }, { "id": "pico3-valid-009", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "94d33bbefa267c76ae75fe8caa56b3af99dd483028477e33025c6a164d227bad", "before_compiled": true, "after_passed": false, "after_native_passed": false, "after_response_sha256": "1ccaa2679c60ef58452b85d82d688e898512d4a1abb38fbbf4ceb6136eae5186", "after_compiled": false }, { "id": "pico3-valid-010", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "0b5f9d6e408e8a9a88f5aee7c989196254bcfa59b070481da2ba4f59177dac8d", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "aa1e94de2221b5de221c230c2efc7e817995bbbf76b4ada1c7122d43372ca423", "after_compiled": true }, { "id": "pico3-valid-011", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "eea7f7f90a3b77abfa3e560467341e172a24471d95aa97a99edac85b738a2286", "before_compiled": true, "after_passed": false, "after_native_passed": true, "after_response_sha256": "21ce68a2a2e82ec82ac6b62b7078a57c3f11ad15988c9dd973bf4ec292d762c7", "after_compiled": true }, { "id": "pico3-valid-012", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": false, "before_response_sha256": "91d449c3ab58d368c59ed177a8e90fa1299762aed49b138130a79418ce4282b5", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "3d9f580878fca038fd2737c76dca9dd12418d23e5e7e4ac3c57b3e4537bfe06e", "after_compiled": true }, { "id": "pico3-valid-013", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "4ba17bc72b8bc42140a36cec06bd6c595e49ce1a87cc716e09e41c24ea865028", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "6eb45169703616500f348b19dfebd96f3c7245925409da75e83ace8023e29b6e", "after_compiled": true }, { "id": "pico3-valid-014", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "a44312273ddc34f8ea0418746238e4e1c0b7fab61e3d5d780b2bb88e1074a548", "before_compiled": true, "after_passed": true, "after_native_passed": true, "after_response_sha256": "ded341402f4acf5c8f0f3f5a4e12f28042ad00d4ef91301b59351b68e9410bf7", "after_compiled": true }, { "id": "pico3-valid-015", "group": "fresh_graph_instances", "before_passed": false, "before_native_passed": true, "before_response_sha256": "2d6ce1108429de17adcf239598477466d57b98d7433cc3a019b1f159f7740357", "before_compiled": true, "after_passed": false, "after_native_passed": false, "after_response_sha256": "2904151e14f01627f35b199d7c349f8f7da410dbe67b294c4ba52709b1ca78e2", "after_compiled": false } ] }, "final_tests_evaluated": false, "default_inference_changed": false, "training_diagnostics": { "iterations": 600, "trained_tokens": 105813, "validation_loss": 0.012, "peak_memory_gb": 3.106 }, "verification": { "main_suite_tests": 3256, "optional_skips": 160, "main_suite_passed": true, "make_check_exit_code": 2, "make_check_blocker": "Docker daemon unavailable at the existing F37 LPBF audit", "native_picogk_reference_programs": 64 }, "artifacts_sha256": { "coding-005/manifest.json": "92e22a392ec0f2c28ef9856b7f1e2c79193f0a28e4372cc29bca1f11334757a1", "coding-005/registration.json": "dffeaa1151390ef5098457fb6610268b2e49f2bf13d1cfd14ec2d90f3c3b19c2", "coding-005/selection.json": "f6bca0084cf6651f5a03f14c1abcafe2bb2b849e34b5c11fe54d8738b0739a24", "coding-005/native-oracles.json": "35c1ecf7271ebbf155ce76c523b07f8ba931506840e60178ad888a8c7ab9e8e7", "coding-005/training.log": "f8d41a31fefd30d8fff1a88887c7379afcec898823b3da5b0ad2f6223139fc73", "coding-005/training-command.json": "6c772f0aa0ae4972bf6e05d491bc9e589fd08a49b881e8ed82e168d71ceba32f", "coding-005-check.log": "68f9058ca7195c6b1e69b885456067e058b7c2f55075474e91816da1a638a4f4", "coding-005/valid-600-responses.jsonl": "e19d7e3fcb8ad51576181e0a6b0dc973edf09d6a9b622278f70205be4ec8b2a9", "coding-005/valid-before-responses.jsonl": "c195c9f080738cb0ebece1cd4380ee4209876f1bddf9ac2c1359fb7af9730ce5", "coding-005/valid-600-scores.jsonl": "e25c3286f11fda600a8adabe62e1c2b7cdfc6f2bc49f46358270256e2ec0d102", "coding-005/valid-before-scores.jsonl": "c5cea91f90755ae4c19e92f029c66da6f848a6393a4cb559b5518299bc2539ef", "coding-005/valid-600-picogk/adapter-evaluation.json": "8aa06f3e211f112a9c8d6f16e69718a3728e5f76af366a33a77c63f0568a1688", "coding-005/valid-before-picogk/adapter-evaluation.json": "b0ed8cea1477ab15130707a99fdc8509b7cb3049803016aa1d0c04fb1efa10ea" }, "limitations": [ "Validation-driven continuation after the rejected coding-004; one adaptive experiment family.", "Only the 24 pico3-test cases were untouched by model inference before validation selection. All older tests are development-informed regressions.", "Shared synthetic grammar and API recipes; no independent task-family benchmark.", "The installed Pixar wheel lacks shader-discovery resources; ShaderPropertyTypeConformanceChecker remains excluded.", "No OpenFOAM evaluation, cylinder-head accuracy, material, manufacturing or physical validation.", "Adapters remain local and experimental; default inference has not been changed." ] }, "source_commit": "8819aa3e3c472ce30d1beed0aa75659c6f39a5cb", "source_url": "https://github.com/cluster2600/porscheparts/blob/8819aa3e3c472ce30d1beed0aa75659c6f39a5cb/training/m64-qwen/retention-results.json" }