Replace JSONLite with native source-schema JSON model and verified SFT results

Update selected weights to the native JSON continuation; ship the matching raw-output runner, tokenizer/configuration, complete training lineage, SFT data/settings, held-out and playground results, examples, verification records and checksums. One whole-object generation with no fieldwise assembly or repair.

CITATION.cff

@@ -1,12 +1,12 @@
 cff-version: 1.2.0
-message: "Please cite this source repository and the selected checkpoint/adapter identity when reporting OpenSML-JSONLite-175M."
-title: "OpenSML-JSONLite-175M: Field Extraction with Schema-Guided JSON Assembly"
+message: "Please cite this source repository and the selected checkpoint and native inference identity when reporting OpenSML-JSONLite-175M."
+title: "OpenSML-JSONLite-175M: Native JSON Generation from Source Text and Schema"
 authors:
   - family-names: Zebrowski
     given-names: William
-version: 0.1.0
+version: 0.2.0
 date-released: 2026-10-14
 url: "https://github.com/williamzebrowskI/OpenSML-JSONLite-175M"
 repository-code: "https://github.com/williamzebrowskI/OpenSML-JSONLite-175M"
 license: Apache-2.0
-abstract: "Source, tokenizer, training provenance and observed evaluation records for a 175.6M-parameter field extractor with JSON assembly. The Hugging Face package includes the selected weights and native JSON adapter; the source license does not assign a model-weight license."
+abstract: "Source, tokenizer, training provenance and observed evaluation records for a 175.6M-parameter native text-to-JSON model. Source license only; weights are not included."

INFERENCE_VERIFICATION.json

@@ -1,571 +1,454 @@
 {
-  "complete": true,
-  "requests": 40,
-  "exact_correct": 29,
-  "all_values_and_stops_match_recorded_playground": true,
-  "weight_sha256": "d552c1380c71c7c0569e9d692ec9c46c5c94bb87a7b024e56df0f0e36df4d798",
-  "versions": {
-    "mlx": "0.32.3",
-    "numpy": "2.4.2",
-    "tokenizers": "0.22.2",
-    "jsonschema": "4.26.0"
-  },
-  "environment": "clean temporary venv with published PyPI dependencies, independent of workspace PYTHONPATH",
-  "elapsed_seconds": 3.6218272920232266,
+  "constrained_decoding": false,
+  "correct_against_reference": 42,
+  "json_assembly": false,
+  "model": "OpenSML-JSONLite-175M-Instruct",
+  "model_calls_per_request": 1,
+  "output_mode": "native-json",
+  "output_repair": false,
   "outputs": [
     {
+      "correct": false,
       "id": "FIELD-01",
-      "value": {
-        "playlist_owner": [
-          "my"
-        ],
-        "playlist": [
-          "Rock Gaming"
-        ],
-        "entity_name": []
-      },
-      "exact": false,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 1,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"playlist_owner\":[],\"playlist\":[\"Rock Gaming\"],\"entity_name\":[\"Mariem Hassan\"]}"
     },
     {
+      "correct": false,
       "id": "FIELD-02",
-      "value": {
-        "artist": [
-          "Mariem Hassan"
-        ],
-        "music_item": []
-      },
-      "exact": false,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 2,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"music_item\":[],\"artist\":[\"Mariem Hassan\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-03",
-      "value": {
-        "playlist_owner": [
-          "my"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 1
+      "index": 3,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"playlist_owner\":[\"my\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-04",
-      "value": {
-        "artist": [],
-        "entity_name": [],
-        "playlist": [
-          "Digster Future Hits"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 4,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"artist\":[],\"entity_name\":[],\"playlist\":[\"Digster Future Hits\"]}"
     },
     {
+      "correct": false,
       "id": "FIELD-05",
-      "value": {
-        "restaurant_type": [
-          "brasserie"
-        ],
-        "party_size_number": [
-          "nine"
-        ],
-        "state": [
-          "AR"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 5,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"state\":[\"AR\"],\"party_size_number\":[\"nine\"],\"state\":[\"AR\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-06",
-      "value": {
-        "party_size_number": [
-          "nine"
-        ],
-        "timeRange": [
-          "nine"
-        ],
-        "sort": []
-      },
-      "exact": false,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 6,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"party_size_number\":[\"nine\"],\"timeRange\":[],\"sort\":[]}"
     },
     {
+      "correct": true,
       "id": "FIELD-07",
-      "value": {
-        "sort": [
-          "best"
-        ],
-        "city": [
-          "Wrigley"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 7,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"sort\":[\"best\"],\"city\":[\"Wrigley\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-08",
-      "value": {
-        "timeRange": [],
-        "poi": []
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 8,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"timeRange\":[],\"poi\":[]}"
     },
     {
+      "correct": true,
       "id": "FIELD-09",
-      "value": {
-        "condition_temperature": [
-          "colder"
-        ],
-        "state": [
-          "Louisiana"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 9,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"condition_temperature\":[\"colder\"],\"state\":[\"Louisiana\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-10",
-      "value": {
-        "current_location": [],
-        "geographic_poi": [],
-        "condition_temperature": [
-          "colder"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 10,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"current_location\":[],\"geographic_poi\":[],\"condition_temperature\":[\"colder\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-11",
-      "value": {
-        "timeRange": [
-          "August"
-        ],
-        "geographic_poi": [
-          "Stunt Ranch Santa Monica Mountains Reserve"
-        ],
-        "condition_temperature": [
-          "colder"
-        ]
-      },
-      "exact": false,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 11,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"timeRange\":[\"last August\"],\"geographic_poi\":[\"Stunt Ranch Santa Monica Mountains Reserve\"],\"condition_temperature\":[\"colder\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-12",
-      "value": {
-        "current_location": [],
-        "geographic_poi": [
-          "Stunt Ranch Santa Monica Mountains Reserve"
-        ],
-        "state": []
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 12,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"current_location\":[],\"geographic_poi\":[\"Stunt Ranch Santa Monica Mountains Reserve\"],\"state\":[]}"
     },
     {
+      "correct": true,
       "id": "FIELD-13",
-      "value": {
-        "artist": [
-          "Behzad Mirkhani"
-        ],
-        "music_item": [
-          "theme"
-        ],
-        "sort": [
-          "last"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 13,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"artist\":[\"Behzad Mirkhani\"],\"music_item\":[\"theme\"],\"sort\":[\"last\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-14",
-      "value": {
-        "service": [],
-        "music_item": [
-          "theme"
-        ],
-        "playlist": []
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 14,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"playlist\":[],\"music_item\":[\"theme\"],\"service\":[]}"
     },
     {
+      "correct": true,
       "id": "FIELD-15",
-      "value": {
-        "artist": [
-          "Yui"
-        ],
-        "music_item": [
-          "track"
-        ],
-        "service": [
-          "Vimeo"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 15,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"artist\":[\"Yui\"],\"music_item\":[\"track\"],\"service\":[\"Vimeo\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-16",
-      "value": {
-        "playlist": [],
-        "sort": []
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 16,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"sort\":[],\"playlist\":[]}"
     },
     {
+      "correct": true,
       "id": "FIELD-17",
-      "value": {
-        "rating_value": [
-          "3"
-        ],
-        "rating_unit": [
-          "stars"
-        ],
-        "object_name": [
-          "Things Fall Apart"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 17,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"rating_value\":[\"3\"],\"rating_unit\":[\"stars\"],\"object_name\":[\"Things Fall Apart\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-18",
-      "value": {
-        "object_name": [
-          "Things Fall Apart"
-        ],
-        "object_select": [],
-        "object_type": []
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 18,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"object_name\":[\"Things Fall Apart\"],\"object_select\":[],\"object_type\":[]}"
     },
     {
+      "correct": true,
       "id": "FIELD-19",
-      "value": {
-        "rating_value": [
-          "1"
-        ],
-        "object_name": []
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 19,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"rating_value\":[\"1\"],\"object_name\":[]}"
     },
     {
+      "correct": true,
       "id": "FIELD-20",
-      "value": {
-        "object_part_of_series_type": [],
-        "object_name": []
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 20,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"object_part_of_series_type\":[],\"object_name\":[]}"
     },
     {
+      "correct": true,
       "id": "FIELD-21",
-      "value": {
-        "object_name": [
-          "Happiness Was Free"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 1
+      "index": 21,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"object_name\":[\"Happiness Was Free\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-22",
-      "value": {
-        "object_type": [],
-        "object_name": [
-          "Happiness Was Free"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 22,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"object_type\":[],\"object_name\":[\"Happiness Was Free\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-23",
-      "value": {
-        "object_name": [
-          "Martin Morning"
-        ],
-        "object_type": [
-          "saga"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 23,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"object_name\":[\"Martin Morning\"],\"object_type\":[\"saga\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-24",
-      "value": {
-        "object_name": [
-          "Martin Morning"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 1
+      "index": 24,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"object_name\":[\"Martin Morning\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-25",
-      "value": {
-        "movie_type": [
-          "animated movies"
-        ],
-        "spatial_relation": [
-          "in the neighborhood"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 25,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"movie_type\":[\"animated movies\"],\"spatial_relation\":[\"in the neighborhood\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-26",
-      "value": {
-        "movie_type": [
-          "animated movies"
-        ],
-        "timeRange": [],
-        "movie_name": []
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 26,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"movie_name\":[],\"timeRange\":[],\"movie_type\":[\"animated movies\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-27",
-      "value": {
-        "movie_name": [
-          "Mabel's Dramatic Career"
-        ],
-        "timeRange": [
-          "twenty two o'clock"
-        ]
-      },
-      "exact": false,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 27,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"timeRange\":[\"twenty two o'clock\"],\"movie_name\":[\"Mabel’s Dramatic Career\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-28",
-      "value": {
-        "timeRange": [
-          "twenty two o'clock"
-        ],
-        "spatial_relation": [],
-        "object_location_type": []
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 28,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"timeRange\":[\"twenty two o'clock\"],\"spatial_relation\":[],\"object_location_type\":[]}"
     },
     {
+      "correct": true,
       "id": "FIELD-29",
-      "value": {
-        "track": [
-          "Blue Lantern"
-        ],
-        "artist": [
-          "Mira Sol"
-        ],
-        "album": []
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 29,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"album\":[],\"artist\":[\"Mira Sol\"],\"track\":[\"Blue Lantern\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-30",
-      "value": {
-        "restaurant_name": [
-          "River Moon"
-        ],
-        "city": [
-          "Denver"
-        ],
-        "party_size_number": [
-          "four"
-        ],
-        "timeRange": [
-          "tomorrow"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 4
+      "index": 30,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"restaurant_name\":[\"River Moon\"],\"city\":[\"Denver\"],\"party_size_number\":[\"four\"],\"timeRange\":[\"tomorrow\"]}"
     },
     {
+      "correct": false,
       "id": "FIELD-31",
-      "value": {
-        "city": [
-          "Oslo"
-        ],
-        "timeRange": [
-          "Friday"
-        ],
-        "country": [
-          "Norway"
-        ]
-      },
-      "exact": false,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 31,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"country\":[],\"timeRange\":[\"Friday\"],\"city\":[\"Odense\"]}"
     },
     {
+      "correct": false,
       "id": "FIELD-32",
-      "value": {
-        "object_name": [
-          "Glass Harbor"
-        ],
-        "rating_value": [
-          "four"
-        ],
-        "best_rating": [],
-        "rating_unit": [
-          "stars"
-        ]
-      },
-      "exact": false,
-      "service_parity": true,
-      "model_calls": 4
+      "index": 32,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"object_name\":[\"Glass Harbor\"],\"rating_value\":[\"four\"],\"best_rating\":[],\"rating_unit\":[\"stars\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-33",
-      "value": {
-        "movie_name": [
-          "Silver Orchard"
-        ],
-        "location_name": [
-          "Maple Cinema"
-        ],
-        "timeRange": []
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 33,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"timeRange\":[],\"location_name\":[\"Maple Cinema\"],\"movie_name\":[\"Silver Orchard\"]}"
     },
     {
+      "correct": true,
       "id": "FIELD-34",
-      "value": {
-        "entity_name": [
-          "Night Steps"
-        ],
-        "playlist": [
-          "Evening Walks"
-        ]
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 34,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"entity_name\":[\"Night Steps\"],\"playlist\":[\"Evening Walks\"]}"
     },
     {
+      "correct": false,
       "id": "FIELD-35",
-      "value": {
-        "artist": [
-          "Mira Sol"
-        ],
-        "album": []
-      },
-      "exact": false,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 35,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"artist\":[\"Mira Sol and Theo Lane\"],\"album\":[]}"
     },
     {
+      "correct": false,
       "id": "FIELD-36",
-      "value": {
-        "city": [
-          "Bergen"
-        ],
-        "country": [
-          "Norway"
-        ]
-      },
-      "exact": false,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 36,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"city\":[\"Bergen\"],\"country\":[]}"
     },
     {
+      "correct": false,
       "id": "FIELD-37",
-      "value": {
-        "cuisine": [],
-        "city": [
-          "Boston"
-        ],
-        "restaurant_name": []
-      },
-      "exact": false,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 37,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"cuisine\":[],\"city\":[\"Boston\"],\"restaurant_name\":[]}"
     },
     {
+      "correct": true,
       "id": "FIELD-38",
-      "value": {
-        "album": [
-          "Quiet Signals"
-        ],
-        "artist": [
-          "Mira Sol"
-        ],
-        "track": [
-          "Quiet Signals"
-        ]
-      },
-      "exact": false,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 38,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"album\":[\"Quiet Signals\"],\"artist\":[\"Mira Sol\"],\"track\":[]}"
     },
     {
+      "correct": true,
       "id": "FIELD-39",
-      "value": {
-        "object_name": [
-          "Paper Bridge"
-        ],
-        "rating_value": []
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 2
+      "index": 39,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"object_name\":[\"Paper Bridge\"],\"rating_value\":[]}"
     },
     {
+      "correct": true,
       "id": "FIELD-40",
-      "value": {
-        "country": [],
-        "timeRange": [],
-        "artist": []
-      },
-      "exact": true,
-      "service_parity": true,
-      "model_calls": 3
+      "index": 40,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"country\":[],\"timeRange\":[],\"artist\":[]}"
+    },
+    {
+      "correct": true,
+      "id": "native:snips:AddToPlaylist:train:80:view1:native:alias",
+      "index": 41,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"selected_playlist\":[\"night vibes\"],\"selected_artist\":[],\"selected_playlist_owner\":[]}"
+    },
+    {
+      "correct": false,
+      "id": "native:snips:AddToPlaylist:train:80:view1:native:nullable",
+      "index": 42,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"playlist_owner\":null,\"artist\":\"gn\",\"playlist\":night vibes\"}"
+    },
+    {
+      "correct": false,
+      "id": "native:snips:BookRestaurant:train:684:view1:native:alias",
+      "index": 43,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"selected_restaurant_name\":[\"mughlai restaurant\"],\"selected_timeRange\":[\"Nov. the 1st\"],\"selected_sort\":[]}"
+    },
+    {
+      "correct": false,
+      "id": "native:snips:BookRestaurant:train:684:view1:native:nullable",
+      "index": 44,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"restaurant_name\":\"Mughlai Restaurant\",\"timeRange\":\"Nov. the 1st\",\"sort\":null}"
+    },
+    {
+      "correct": true,
+      "id": "native:snips:GetWeather:train:109:view0:native:alias",
+      "index": 45,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"selected_condition_temperature\":[\"colder\"],\"selected_state\":[\"Louisiana\"]}"
+    },
+    {
+      "correct": true,
+      "id": "native:snips:GetWeather:train:109:view0:native:nullable",
+      "index": 46,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"state\":\"Louisiana\",\"condition_temperature\":\"colder\"}"
+    },
+    {
+      "correct": true,
+      "id": "native:snips:PlayMusic:train:801:view1:native:alias",
+      "index": 47,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"selected_playlist\":[],\"selected_music_item\":[\"theme\"],\"selected_service\":[]}"
+    },
+    {
+      "correct": true,
+      "id": "native:snips:PlayMusic:train:801:view1:native:nullable",
+      "index": 48,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"playlist\":null,\"music_item\":\"theme\",\"service\":null}"
+    },
+    {
+      "correct": true,
+      "id": "native:snips:RateBook:train:293:view0:native:alias",
+      "index": 49,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"selected_rating_value\":[\"2\"],\"selected_object_select\":[\"current\"]}"
+    },
+    {
+      "correct": true,
+      "id": "native:snips:RateBook:train:293:view0:native:nullable",
+      "index": 50,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"object_select\":\"current\",\"rating_value\":\"2\"}"
+    },
+    {
+      "correct": true,
+      "id": "native:snips:SearchCreativeWork:train:602:view1:native:alias",
+      "index": 51,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"selected_object_name\":[\"Playlist: The Very Best of Dolly Parton\"]}"
+    },
+    {
+      "correct": true,
+      "id": "native:snips:SearchCreativeWork:train:602:view1:native:nullable",
+      "index": 52,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"object_name\":\"Playlist: The Very Best of Dolly Parton\"}"
+    },
+    {
+      "correct": true,
+      "id": "native:snips:SearchScreeningEvent:train:897:view1:native:alias",
+      "index": 53,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"selected_movie_type\":[\"animated movies\"],\"selected_movie_name\":[],\"selected_location_name\":[]}"
+    },
+    {
+      "correct": false,
+      "id": "native:snips:SearchScreeningEvent:train:897:view1:native:nullable",
+      "index": 54,
+      "raw_parity": true,
+      "stop_reason": "turn_end",
+      "text": "{\"movie_name\":null,\"movie_type\":\"movies\",\"location_name\":null}"
     }
   ],
-  "scope": "40 observed fixtures, same greedy native MLX path; no fresh blind benchmark or full-suite stock-runtime parity claim"
+  "raw_text_and_stop_matches": 54,
+  "requests": 54,
+  "scope": "Clean stock-MLX packaging parity on observed playground fixtures; not full-heldout runtime parity or fresh blind accuracy.",
+  "step": 469,
+  "versions": {
+    "jsonschema": "4.26.0",
+    "mlx": "0.32.3",
+    "numpy": "2.4.2",
+    "tokenizers": "0.22.2"
+  },
+  "weight_sha256": "656f45015aab548465d93bed0b57e687497163741885da7364872851fa17e9a4"
 }

JSONLITE_RESULTS.json

@@ -1,1873 +1,1162 @@
 {
-  "model": "OpenSML-JSONLite-175M",
-  "step": 160,
-  "weight_sha256": "d552c1380c71c7c0569e9d692ec9c46c5c94bb87a7b024e56df0f0e36df4d798",
-  "parent_same_route": {
-    "all": {
-      "count": 1388,
-      "counts": {
-        "exact": 953,
-        "format_ok": 1384,
-        "json_parse": 1384,
-        "schema_valid": 1384,
-        "field_correct": 2796,
-        "fields": 3347,
-        "tp": 1759,
-        "fp": 524,
-        "fn": 120,
-        "present_correct": 1759,
-        "present_total": 1879,
-        "absent_correct": 1037,
-        "absent_total": 1468,
-        "false_abstention": 18,
-        "missed_abstention": 431,
-        "unsupported_values": 43,
-        "predicted_values": 2283,
-        "natural_stop": 1388,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 1388,
-        "requested_order": 1384
-      },
-      "exact": 0.6865994236311239,
-      "field_accuracy": 0.8353749626531222,
-      "precision": 0.7704774419623303,
-      "recall": 0.9361362426822778,
-      "f1": 0.8452666987025469,
-      "present_accuracy": 0.9361362426822778,
-      "absent_accuracy": 0.7064032697547684,
-      "schema_valid_rate": 0.9971181556195965
-    },
-    "AddToPlaylist": {
-      "count": 198,
-      "counts": {
-        "exact": 148,
-        "format_ok": 198,
-        "json_parse": 198,
-        "schema_valid": 198,
-        "field_correct": 444,
-        "fields": 508,
-        "tp": 259,
-        "fp": 56,
-        "fn": 19,
-        "present_correct": 259,
-        "present_total": 278,
-        "absent_correct": 185,
-        "absent_total": 230,
-        "false_abstention": 8,
-        "missed_abstention": 45,
-        "unsupported_values": 4,
-        "predicted_values": 315,
-        "natural_stop": 198,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 198,
-        "requested_order": 198
-      },
-      "exact": 0.7474747474747475,
-      "field_accuracy": 0.8740157480314961,
-      "precision": 0.8222222222222222,
-      "recall": 0.9316546762589928,
-      "f1": 0.8735244519392917,
-      "present_accuracy": 0.9316546762589928,
-      "absent_accuracy": 0.8043478260869565,
-      "schema_valid_rate": 1.0
-    },
-    "BookRestaurant": {
-      "count": 200,
-      "counts": {
-        "exact": 122,
-        "format_ok": 200,
-        "json_parse": 200,
-        "schema_valid": 200,
-        "field_correct": 446,
-        "fields": 543,
-        "tp": 273,
-        "fp": 96,
-        "fn": 26,
-        "present_correct": 273,
-        "present_total": 299,
-        "absent_correct": 173,
-        "absent_total": 244,
-        "false_abstention": 1,
-        "missed_abstention": 71,
-        "unsupported_values": 5,
-        "predicted_values": 369,
-        "natural_stop": 200,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 200,
-        "requested_order": 200
-      },
-      "exact": 0.61,
-      "field_accuracy": 0.8213627992633518,
-      "precision": 0.7398373983739838,
-      "recall": 0.9130434782608695,
-      "f1": 0.8173652694610778,
-      "present_accuracy": 0.9130434782608695,
-      "absent_accuracy": 0.7090163934426229,
-      "schema_valid_rate": 1.0
-    },
-    "GetWeather": {
-      "count": 198,
-      "counts": {
-        "exact": 101,
-        "format_ok": 198,
-        "json_parse": 198,
-        "schema_valid": 198,
-        "field_correct": 373,
-        "fields": 501,
-        "tp": 245,
-        "fp": 126,
-        "fn": 18,
-        "present_correct": 245,
-        "present_total": 263,
-        "absent_correct": 128,
-        "absent_total": 238,
-        "false_abstention": 2,
-        "missed_abstention": 110,
-        "unsupported_values": 22,
-        "predicted_values": 371,
-        "natural_stop": 198,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 198,
-        "requested_order": 198
-      },
-      "exact": 0.51010101010101,
-      "field_accuracy": 0.7445109780439122,
-      "precision": 0.660377358490566,
-      "recall": 0.9315589353612167,
-      "f1": 0.7728706624605678,
-      "present_accuracy": 0.9315589353612167,
-      "absent_accuracy": 0.5378151260504201,
-      "schema_valid_rate": 1.0
-    },
-    "PlayMusic": {
-      "count": 200,
-      "counts": {
-        "exact": 97,
-        "format_ok": 198,
-        "json_parse": 198,
-        "schema_valid": 198,
-        "field_correct": 356,
-        "fields": 499,
-        "tp": 230,
-        "fp": 135,
-        "fn": 23,
-        "present_correct": 230,
-        "present_total": 253,
-        "absent_correct": 126,
-        "absent_total": 246,
-        "false_abstention": 3,
-        "missed_abstention": 120,
-        "unsupported_values": 2,
-        "predicted_values": 365,
-        "natural_stop": 200,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 200,
-        "requested_order": 198
-      },
-      "exact": 0.485,
-      "field_accuracy": 0.7134268537074149,
-      "precision": 0.6301369863013698,
-      "recall": 0.9090909090909091,
-      "f1": 0.744336569579288,
-      "present_accuracy": 0.9090909090909091,
-      "absent_accuracy": 0.5121951219512195,
-      "schema_valid_rate": 0.99
-    },
-    "RateBook": {
-      "count": 196,
-      "counts": {
-        "exact": 176,
-        "format_ok": 196,
-        "json_parse": 196,
-        "schema_valid": 196,
-        "field_correct": 521,
-        "fields": 541,
-        "tp": 300,
-        "fp": 20,
-        "fn": 5,
-        "present_correct": 300,
-        "present_total": 305,
-        "absent_correct": 221,
-        "absent_total": 236,
-        "false_abstention": 0,
-        "missed_abstention": 15,
-        "unsupported_values": 1,
-        "predicted_values": 320,
-        "natural_stop": 196,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 196,
-        "requested_order": 196
-      },
-      "exact": 0.8979591836734694,
-      "field_accuracy": 0.9630314232902033,
-      "precision": 0.9375,
-      "recall": 0.9836065573770492,
-      "f1": 0.96,
-      "present_accuracy": 0.9836065573770492,
-      "absent_accuracy": 0.9364406779661016,
-      "schema_valid_rate": 1.0
-    },
-    "SearchCreativeWork": {
-      "count": 200,
-      "counts": {
-        "exact": 189,
-        "format_ok": 200,
-        "json_parse": 200,
-        "schema_valid": 200,
-        "field_correct": 263,
-        "fields": 274,
-        "tp": 226,
-        "fp": 11,
-        "fn": 11,
-        "present_correct": 226,
-        "present_total": 237,
-        "absent_correct": 37,
-        "absent_total": 37,
-        "false_abstention": 0,
-        "missed_abstention": 0,
-        "unsupported_values": 0,
-        "predicted_values": 237,
-        "natural_stop": 200,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 200,
-        "requested_order": 200
-      },
-      "exact": 0.945,
-      "field_accuracy": 0.9598540145985401,
-      "precision": 0.9535864978902954,
-      "recall": 0.9535864978902954,
-      "f1": 0.9535864978902954,
-      "present_accuracy": 0.9535864978902954,
-      "absent_accuracy": 1.0,
-      "schema_valid_rate": 1.0
-    },
-    "SearchScreeningEvent": {
-      "count": 196,
-      "counts": {
-        "exact": 120,
-        "format_ok": 194,
-        "json_parse": 194,
-        "schema_valid": 194,
-        "field_correct": 393,
-        "fields": 481,
-        "tp": 226,
-        "fp": 80,
-        "fn": 18,
-        "present_correct": 226,
-        "present_total": 244,
-        "absent_correct": 167,
-        "absent_total": 237,
-        "false_abstention": 4,
-        "missed_abstention": 70,
-        "unsupported_values": 9,
-        "predicted_values": 306,
-        "natural_stop": 196,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 196,
-        "requested_order": 194
-      },
-      "exact": 0.6122448979591837,
-      "field_accuracy": 0.817047817047817,
-      "precision": 0.738562091503268,
-      "recall": 0.9262295081967213,
-      "f1": 0.8218181818181819,
-      "present_accuracy": 0.9262295081967213,
-      "absent_accuracy": 0.7046413502109705,
-      "schema_valid_rate": 0.9897959183673469
-    },
-    "paired_requests": {
-      "count": 694,
-      "both_correct": 360
-    },
-    "macro_exact": 0.6868256913154873
-  },
-  "candidate_same_route": {
-    "all": {
-      "count": 1388,
-      "counts": {
-        "exact": 1132,
-        "format_ok": 1385,
-        "json_parse": 1385,
-        "schema_valid": 1385,
-        "field_correct": 3047,
-        "fields": 3347,
-        "tp": 1740,
-        "fp": 223,
-        "fn": 139,
-        "present_correct": 1740,
-        "present_total": 1879,
-        "absent_correct": 1307,
-        "absent_total": 1468,
-        "false_abstention": 69,
-        "missed_abstention": 161,
-        "unsupported_values": 5,
-        "predicted_values": 1963,
-        "natural_stop": 1388,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 1388,
-        "requested_order": 1385
-      },
-      "exact": 0.8155619596541787,
-      "field_accuracy": 0.910367493277562,
-      "precision": 0.8863983698420784,
-      "recall": 0.9260244811069718,
-      "f1": 0.9057782404997397,
-      "present_accuracy": 0.9260244811069718,
-      "absent_accuracy": 0.8903269754768393,
-      "schema_valid_rate": 0.9978386167146974
-    },
-    "AddToPlaylist": {
-      "count": 198,
-      "counts": {
-        "exact": 164,
-        "format_ok": 197,
-        "json_parse": 197,
-        "schema_valid": 197,
-        "field_correct": 465,
-        "fields": 508,
-        "tp": 257,
-        "fp": 31,
-        "fn": 21,
-        "present_correct": 257,
-        "present_total": 278,
-        "absent_correct": 208,
-        "absent_total": 230,
-        "false_abstention": 10,
-        "missed_abstention": 22,
-        "unsupported_values": 0,
-        "predicted_values": 288,
-        "natural_stop": 198,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 198,
-        "requested_order": 197
-      },
-      "exact": 0.8282828282828283,
-      "field_accuracy": 0.9153543307086615,
-      "precision": 0.8923611111111112,
-      "recall": 0.9244604316546763,
-      "f1": 0.9081272084805655,
-      "present_accuracy": 0.9244604316546763,
-      "absent_accuracy": 0.9043478260869565,
-      "schema_valid_rate": 0.9949494949494949
-    },
-    "BookRestaurant": {
-      "count": 200,
-      "counts": {
-        "exact": 151,
-        "format_ok": 200,
-        "json_parse": 200,
-        "schema_valid": 200,
-        "field_correct": 488,
-        "fields": 543,
-        "tp": 269,
-        "fp": 41,
-        "fn": 30,
-        "present_correct": 269,
-        "present_total": 299,
-        "absent_correct": 219,
-        "absent_total": 244,
-        "false_abstention": 14,
-        "missed_abstention": 25,
-        "unsupported_values": 1,
-        "predicted_values": 310,
-        "natural_stop": 200,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 200,
-        "requested_order": 200
-      },
-      "exact": 0.755,
-      "field_accuracy": 0.8987108655616943,
-      "precision": 0.867741935483871,
-      "recall": 0.8996655518394648,
-      "f1": 0.8834154351395731,
-      "present_accuracy": 0.8996655518394648,
-      "absent_accuracy": 0.8975409836065574,
-      "schema_valid_rate": 1.0
-    },
-    "GetWeather": {
-      "count": 198,
-      "counts": {
-        "exact": 143,
-        "format_ok": 198,
-        "json_parse": 198,
-        "schema_valid": 198,
-        "field_correct": 438,
-        "fields": 501,
-        "tp": 246,
-        "fp": 57,
-        "fn": 17,
-        "present_correct": 246,
-        "present_total": 263,
-        "absent_correct": 192,
-        "absent_total": 238,
-        "false_abstention": 6,
-        "missed_abstention": 46,
-        "unsupported_values": 2,
-        "predicted_values": 303,
-        "natural_stop": 198,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 198,
-        "requested_order": 198
-      },
-      "exact": 0.7222222222222222,
-      "field_accuracy": 0.874251497005988,
-      "precision": 0.8118811881188119,
-      "recall": 0.935361216730038,
-      "f1": 0.8692579505300353,
-      "present_accuracy": 0.935361216730038,
-      "absent_accuracy": 0.8067226890756303,
-      "schema_valid_rate": 1.0
-    },
-    "PlayMusic": {
-      "count": 200,
-      "counts": {
-        "exact": 139,
-        "format_ok": 198,
-        "json_parse": 198,
-        "schema_valid": 198,
-        "field_correct": 420,
-        "fields": 499,
-        "tp": 215,
-        "fp": 48,
-        "fn": 38,
-        "present_correct": 215,
-        "present_total": 253,
-        "absent_correct": 205,
-        "absent_total": 246,
-        "false_abstention": 25,
-        "missed_abstention": 41,
-        "unsupported_values": 1,
-        "predicted_values": 263,
-        "natural_stop": 200,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 200,
-        "requested_order": 198
-      },
-      "exact": 0.695,
-      "field_accuracy": 0.8416833667334669,
-      "precision": 0.8174904942965779,
-      "recall": 0.849802371541502,
-      "f1": 0.8333333333333334,
-      "present_accuracy": 0.849802371541502,
-      "absent_accuracy": 0.8333333333333334,
-      "schema_valid_rate": 0.99
-    },
-    "RateBook": {
-      "count": 196,
-      "counts": {
-        "exact": 183,
-        "format_ok": 196,
-        "json_parse": 196,
-        "schema_valid": 196,
-        "field_correct": 527,
-        "fields": 541,
-        "tp": 298,
-        "fp": 10,
-        "fn": 7,
-        "present_correct": 298,
-        "present_total": 305,
-        "absent_correct": 229,
-        "absent_total": 236,
-        "false_abstention": 4,
-        "missed_abstention": 7,
-        "unsupported_values": 1,
-        "predicted_values": 308,
-        "natural_stop": 196,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 196,
-        "requested_order": 196
-      },
-      "exact": 0.9336734693877551,
-      "field_accuracy": 0.9741219963031423,
-      "precision": 0.9675324675324676,
-      "recall": 0.9770491803278688,
-      "f1": 0.9722675367047309,
-      "present_accuracy": 0.9770491803278688,
-      "absent_accuracy": 0.9703389830508474,
-      "schema_valid_rate": 1.0
-    },
-    "SearchCreativeWork": {
-      "count": 200,
-      "counts": {
-        "exact": 192,
-        "format_ok": 200,
-        "json_parse": 200,
-        "schema_valid": 200,
-        "field_correct": 266,
-        "fields": 274,
-        "tp": 229,
-        "fp": 8,
-        "fn": 8,
-        "present_correct": 229,
-        "present_total": 237,
-        "absent_correct": 37,
-        "absent_total": 37,
-        "false_abstention": 0,
-        "missed_abstention": 0,
-        "unsupported_values": 0,
-        "predicted_values": 237,
-        "natural_stop": 200,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 200,
-        "requested_order": 200
-      },
-      "exact": 0.96,
-      "field_accuracy": 0.9708029197080292,
-      "precision": 0.9662447257383966,
-      "recall": 0.9662447257383966,
-      "f1": 0.9662447257383966,
-      "present_accuracy": 0.9662447257383966,
-      "absent_accuracy": 1.0,
-      "schema_valid_rate": 1.0
-    },
-    "SearchScreeningEvent": {
-      "count": 196,
-      "counts": {
-        "exact": 160,
-        "format_ok": 196,
-        "json_parse": 196,
-        "schema_valid": 196,
-        "field_correct": 443,
-        "fields": 481,
-        "tp": 226,
-        "fp": 28,
-        "fn": 18,
-        "present_correct": 226,
-        "present_total": 244,
-        "absent_correct": 217,
-        "absent_total": 237,
-        "false_abstention": 10,
-        "missed_abstention": 20,
-        "unsupported_values": 0,
-        "predicted_values": 254,
-        "natural_stop": 196,
-        "token_limited": 0,
-        "nonrepetitive_nonempty": 196,
-        "requested_order": 196
-      },
-      "exact": 0.8163265306122449,
-      "field_accuracy": 0.920997920997921,
-      "precision": 0.889763779527559,
-      "recall": 0.9262295081967213,
-      "f1": 0.9076305220883534,
-      "present_accuracy": 0.9262295081967213,
-      "absent_accuracy": 0.9156118143459916,
-      "schema_valid_rate": 1.0
-    },
-    "paired_requests": {
-      "count": 694,
-      "both_correct": 498
-    },
-    "macro_exact": 0.8157864357864357
-  },
+  "model": "OpenSML-JSONLite-175M-Instruct",
+  "selected_stage": "sft-100",
+  "step": 469,
+  "model_sha256": "656f45015aab548465d93bed0b57e687497163741885da7364872851fa17e9a4",
+  "output_mode": "native-json",
   "development": {
-    "fit": {
-      "AddToPlaylist": {
-        "absent_accuracy": 0.9523809523809523,
-        "count": 32,
+    "metrics": {
+      "all": {
+        "absent_accuracy": 0.9294117647058824,
+        "count": 252,
         "counts": {
-          "absent_correct": 20,
-          "absent_total": 21,
-          "exact": 31,
-          "false_abstention": 0,
-          "field_correct": 79,
-          "fields": 80,
-          "fn": 0,
-          "format_ok": 32,
-          "fp": 1,
-          "json_parse": 32,
-          "missed_abstention": 1,
-          "natural_stop": 32,
-          "nonrepetitive_nonempty": 32,
-          "predicted_values": 60,
-          "present_correct": 59,
-          "present_total": 59,
-          "requested_order": 32,
-          "schema_valid": 32,
+          "absent_correct": 237,
+          "absent_total": 255,
+          "exact": 200,
+          "false_abstention": 25,
+          "field_correct": 546,
+          "fields": 604,
+          "fn": 40,
+          "fp": 29,
+          "json_parse_ok": 251,
+          "missed_abstention": 18,
+          "natural_stop": 252,
+          "nonrepetitive_nonempty": 252,
+          "predicted_values": 339,
+          "present_correct": 309,
+          "present_total": 349,
+          "schema_ok": 251,
           "token_limited": 0,
-          "tp": 59,
-          "unsupported_values": 0
-        },
-        "exact": 0.96875,
-        "f1": 0.9915966386554621,
-        "field_accuracy": 0.9875,
-        "json_parse_rate": 1.0,
-        "output_nll": 0.008540522438109594,
-        "precision": 0.9833333333333333,
-        "present_accuracy": 1.0,
-        "recall": 1.0,
-        "schema_valid_rate": 1.0
+          "tp": 310,
+          "unsupported_values": 4
+        },
+        "exact": 0.7936507936507936,
+        "f1": 0.899854862119013,
+        "false_abstention_rate": 0.07163323782234957,
+        "field_accuracy": 0.9039735099337748,
+        "json_parse_rate": 0.996031746031746,
+        "missed_abstention_rate": 0.07058823529411765,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.03873514435736791,
+        "precision": 0.9144542772861357,
+        "present_accuracy": 0.8853868194842407,
+        "recall": 0.8857142857142857,
+        "schema_valid_rate": 0.996031746031746,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.011799410029498525
       },
-      "all": {
-        "absent_accuracy": 0.9523809523809523,
-        "count": 32,
+      "domain/AddToPlaylist": {
+        "absent_accuracy": 0.9024390243902439,
+        "count": 36,
         "counts": {
-          "absent_correct": 20,
-          "absent_total": 21,
-          "exact": 31,
-          "false_abstention": 0,
-          "field_correct": 79,
-          "fields": 80,
-          "fn": 0,
-          "format_ok": 32,
-          "fp": 1,
-          "json_parse": 32,
-          "missed_abstention": 1,
-          "natural_stop": 32,
-          "nonrepetitive_nonempty": 32,
-          "predicted_values": 60,
-          "present_correct": 59,
-          "present_total": 59,
-          "requested_order": 32,
-          "schema_valid": 32,
+          "absent_correct": 37,
+          "absent_total": 41,
+          "exact": 20,
+          "false_abstention": 9,
+          "field_correct": 72,
+          "fields": 91,
+          "fn": 15,
+          "fp": 7,
+          "json_parse_ok": 35,
+          "missed_abstention": 4,
+          "natural_stop": 36,
+          "nonrepetitive_nonempty": 36,
+          "predicted_values": 42,
+          "present_correct": 35,
+          "present_total": 50,
+          "schema_ok": 35,
           "token_limited": 0,
-          "tp": 59,
-          "unsupported_values": 0
-        },
-        "exact": 0.96875,
-        "f1": 0.9915966386554621,
-        "field_accuracy": 0.9875,
-        "json_parse_rate": 1.0,
-        "output_nll": 0.008540522438109594,
-        "precision": 0.9833333333333333,
-        "present_accuracy": 1.0,
-        "recall": 1.0,
-        "schema_valid_rate": 1.0
+          "tp": 35,
+          "unsupported_values": 2
+        },
+        "exact": 0.5555555555555556,
+        "f1": 0.7608695652173915,
+        "false_abstention_rate": 0.18,
+        "field_accuracy": 0.7912087912087912,
+        "json_parse_rate": 0.9722222222222222,
+        "missed_abstention_rate": 0.0975609756097561,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.08232390080795288,
+        "precision": 0.8333333333333334,
+        "present_accuracy": 0.7,
+        "recall": 0.7,
+        "schema_valid_rate": 0.9722222222222222,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.047619047619047616
       },
-      "macro_exact": 0.96875,
-      "paired_requests": {
-        "both_correct": 0,
-        "count": 0
-      }
-    },
-    "json": {
-      "AddToPlaylist": {
-        "absent_accuracy": 0.8709677419354839,
-        "count": 28,
+      "domain/BookRestaurant": {
+        "absent_accuracy": 0.8604651162790697,
+        "count": 36,
         "counts": {
-          "absent_correct": 27,
-          "absent_total": 31,
-          "exact": 21,
+          "absent_correct": 37,
+          "absent_total": 43,
+          "exact": 24,
           "false_abstention": 4,
-          "field_correct": 63,
-          "fields": 71,
-          "fn": 4,
-          "format_ok": 28,
-          "fp": 4,
-          "json_parse": 28,
-          "missed_abstention": 4,
-          "natural_stop": 28,
-          "nonrepetitive_nonempty": 28,
-          "predicted_values": 40,
-          "present_correct": 36,
-          "present_total": 40,
-          "requested_order": 28,
-          "schema_valid": 28,
+          "field_correct": 82,
+          "fields": 95,
+          "fn": 7,
+          "fp": 9,
+          "json_parse_ok": 36,
+          "missed_abstention": 6,
+          "natural_stop": 36,
+          "nonrepetitive_nonempty": 36,
+          "predicted_values": 54,
+          "present_correct": 45,
+          "present_total": 52,
+          "schema_ok": 36,
           "token_limited": 0,
-          "tp": 36,
-          "unsupported_values": 0
+          "tp": 45,
+          "unsupported_values": 2
         },
-        "exact": 0.75,
-        "f1": 0.9,
-        "field_accuracy": 0.8873239436619719,
+        "exact": 0.6666666666666666,
+        "f1": 0.8490566037735849,
+        "false_abstention_rate": 0.07692307692307693,
+        "field_accuracy": 0.8631578947368421,
         "json_parse_rate": 1.0,
-        "output_nll": 0.06179538822800083,
-        "precision": 0.9,
-        "present_accuracy": 0.9,
-        "recall": 0.9,
-        "schema_valid_rate": 1.0
+        "missed_abstention_rate": 0.13953488372093023,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.0473247706718717,
+        "precision": 0.8333333333333334,
+        "present_accuracy": 0.8653846153846154,
+        "recall": 0.8653846153846154,
+        "schema_valid_rate": 1.0,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.037037037037037035
       },
-      "BookRestaurant": {
-        "absent_accuracy": 0.8064516129032258,
-        "count": 28,
+      "domain/GetWeather": {
+        "absent_accuracy": 0.925,
+        "count": 36,
         "counts": {
-          "absent_correct": 25,
-          "absent_total": 31,
-          "exact": 21,
-          "false_abstention": 0,
-          "field_correct": 65,
-          "fields": 73,
-          "fn": 2,
-          "format_ok": 28,
-          "fp": 8,
-          "json_parse": 28,
-          "missed_abstention": 6,
-          "natural_stop": 28,
-          "nonrepetitive_nonempty": 28,
-          "predicted_values": 48,
-          "present_correct": 40,
-          "present_total": 42,
-          "requested_order": 28,
-          "schema_valid": 28,
+          "absent_correct": 37,
+          "absent_total": 40,
+          "exact": 28,
+          "false_abstention": 5,
+          "field_correct": 85,
+          "fields": 94,
+          "fn": 6,
+          "fp": 3,
+          "json_parse_ok": 36,
+          "missed_abstention": 3,
+          "natural_stop": 36,
+          "nonrepetitive_nonempty": 36,
+          "predicted_values": 52,
+          "present_correct": 48,
+          "present_total": 54,
+          "schema_ok": 36,
           "token_limited": 0,
-          "tp": 40,
+          "tp": 49,
           "unsupported_values": 0
         },
-        "exact": 0.75,
-        "f1": 0.888888888888889,
-        "field_accuracy": 0.8904109589041096,
+        "exact": 0.7777777777777778,
+        "f1": 0.9158878504672897,
+        "false_abstention_rate": 0.09259259259259259,
+        "field_accuracy": 0.9042553191489362,
         "json_parse_rate": 1.0,
-        "output_nll": 0.055805700505931914,
-        "precision": 0.8333333333333334,
-        "present_accuracy": 0.9523809523809523,
-        "recall": 0.9523809523809523,
-        "schema_valid_rate": 1.0
+        "missed_abstention_rate": 0.075,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.044021505151047476,
+        "precision": 0.9423076923076923,
+        "present_accuracy": 0.8888888888888888,
+        "recall": 0.8909090909090909,
+        "schema_valid_rate": 1.0,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.0
       },
-      "GetWeather": {
-        "absent_accuracy": 0.8666666666666667,
-        "count": 28,
+      "domain/PlayMusic": {
+        "absent_accuracy": 0.9767441860465116,
+        "count": 36,
         "counts": {
-          "absent_correct": 26,
-          "absent_total": 30,
-          "exact": 22,
-          "false_abstention": 0,
-          "field_correct": 63,
-          "fields": 70,
-          "fn": 3,
-          "format_ok": 28,
-          "fp": 6,
-          "json_parse": 28,
-          "missed_abstention": 4,
-          "natural_stop": 28,
-          "nonrepetitive_nonempty": 28,
-          "predicted_values": 44,
-          "present_correct": 37,
-          "present_total": 40,
-          "requested_order": 28,
-          "schema_valid": 28,
+          "absent_correct": 42,
+          "absent_total": 43,
+          "exact": 30,
+          "false_abstention": 6,
+          "field_correct": 81,
+          "fields": 88,
+          "fn": 6,
+          "fp": 1,
+          "json_parse_ok": 36,
+          "missed_abstention": 1,
+          "natural_stop": 36,
+          "nonrepetitive_nonempty": 36,
+          "predicted_values": 40,
+          "present_correct": 39,
+          "present_total": 45,
+          "schema_ok": 36,
           "token_limited": 0,
-          "tp": 38,
+          "tp": 39,
           "unsupported_values": 0
         },
-        "exact": 0.7857142857142857,
-        "f1": 0.8941176470588236,
-        "field_accuracy": 0.9,
+        "exact": 0.8333333333333334,
+        "f1": 0.9176470588235294,
+        "false_abstention_rate": 0.13333333333333333,
+        "field_accuracy": 0.9204545454545454,
         "json_parse_rate": 1.0,
-        "output_nll": 0.041639312657675225,
-        "precision": 0.8636363636363636,
-        "present_accuracy": 0.925,
-        "recall": 0.926829268292683,
-        "schema_valid_rate": 1.0
+        "missed_abstention_rate": 0.023255813953488372,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.04131003013268734,
+        "precision": 0.975,
+        "present_accuracy": 0.8666666666666667,
+        "recall": 0.8666666666666667,
+        "schema_valid_rate": 1.0,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.0
       },
-      "PlayMusic": {
-        "absent_accuracy": 0.9696969696969697,
-        "count": 28,
+      "domain/RateBook": {
+        "absent_accuracy": 0.9767441860465116,
+        "count": 36,
         "counts": {
-          "absent_correct": 32,
-          "absent_total": 33,
-          "exact": 27,
+          "absent_correct": 42,
+          "absent_total": 43,
+          "exact": 34,
           "false_abstention": 0,
-          "field_correct": 69,
-          "fields": 70,
-          "fn": 0,
-          "format_ok": 28,
-          "fp": 1,
-          "json_parse": 28,
+          "field_correct": 94,
+          "fields": 96,
+          "fn": 1,
+          "fp": 2,
+          "json_parse_ok": 36,
           "missed_abstention": 1,
-          "natural_stop": 28,
-          "nonrepetitive_nonempty": 28,
-          "predicted_values": 38,
-          "present_correct": 37,
-          "present_total": 37,
-          "requested_order": 28,
-          "schema_valid": 28,
+          "natural_stop": 36,
+          "nonrepetitive_nonempty": 36,
+          "predicted_values": 54,
+          "present_correct": 52,
+          "present_total": 53,
+          "schema_ok": 36,
           "token_limited": 0,
-          "tp": 37,
+          "tp": 52,
           "unsupported_values": 0
         },
-        "exact": 0.9642857142857143,
-        "f1": 0.9866666666666666,
-        "field_accuracy": 0.9857142857142858,
+        "exact": 0.9444444444444444,
+        "f1": 0.9719626168224299,
+        "false_abstention_rate": 0.0,
+        "field_accuracy": 0.9791666666666666,
         "json_parse_rate": 1.0,
-        "output_nll": 0.011643990869006736,
-        "precision": 0.9736842105263158,
-        "present_accuracy": 1.0,
-        "recall": 1.0,
-        "schema_valid_rate": 1.0
+        "missed_abstention_rate": 0.023255813953488372,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.01588757585947557,
+        "precision": 0.9629629629629629,
+        "present_accuracy": 0.9811320754716981,
+        "recall": 0.9811320754716981,
+        "schema_valid_rate": 1.0,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.0
       },
-      "RateBook": {
+      "domain/SearchCreativeWork": {
         "absent_accuracy": 1.0,
-        "count": 28,
+        "count": 36,
         "counts": {
-          "absent_correct": 31,
-          "absent_total": 31,
-          "exact": 25,
-          "false_abstention": 0,
-          "field_correct": 71,
-          "fields": 74,
+          "absent_correct": 4,
+          "absent_total": 4,
+          "exact": 33,
+          "false_abstention": 1,
+          "field_correct": 44,
+          "fields": 47,
           "fn": 3,
-          "format_ok": 28,
-          "fp": 3,
-          "json_parse": 28,
+          "fp": 2,
+          "json_parse_ok": 36,
           "missed_abstention": 0,
-          "natural_stop": 28,
-          "nonrepetitive_nonempty": 28,
-          "predicted_values": 43,
+          "natural_stop": 36,
+          "nonrepetitive_nonempty": 36,
+          "predicted_values": 42,
           "present_correct": 40,
           "present_total": 43,
-          "requested_order": 28,
-          "schema_valid": 28,
+          "schema_ok": 36,
           "token_limited": 0,
           "tp": 40,
           "unsupported_values": 0
         },
-        "exact": 0.8928571428571429,
-        "f1": 0.9302325581395349,
-        "field_accuracy": 0.9594594594594594,
+        "exact": 0.9166666666666666,
+        "f1": 0.9411764705882352,
+        "false_abstention_rate": 0.023255813953488372,
+        "field_accuracy": 0.9361702127659575,
         "json_parse_rate": 1.0,
-        "output_nll": 0.01846137226675637,
-        "precision": 0.9302325581395349,
+        "missed_abstention_rate": 0.0,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.017151051737730287,
+        "precision": 0.9523809523809523,
         "present_accuracy": 0.9302325581395349,
         "recall": 0.9302325581395349,
-        "schema_valid_rate": 1.0
+        "schema_valid_rate": 1.0,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.0
       },
-      "SearchCreativeWork": {
-        "absent_accuracy": 1.0,
-        "count": 28,
+      "domain/SearchScreeningEvent": {
+        "absent_accuracy": 0.926829268292683,
+        "count": 36,
         "counts": {
-          "absent_correct": 4,
-          "absent_total": 4,
-          "exact": 28,
+          "absent_correct": 38,
+          "absent_total": 41,
+          "exact": 31,
           "false_abstention": 0,
-          "field_correct": 39,
-          "fields": 39,
-          "fn": 0,
-          "format_ok": 28,
-          "fp": 0,
-          "json_parse": 28,
-          "missed_abstention": 0,
-          "natural_stop": 28,
-          "nonrepetitive_nonempty": 28,
-          "predicted_values": 35,
-          "present_correct": 35,
-          "present_total": 35,
-          "requested_order": 28,
-          "schema_valid": 28,
+          "field_correct": 88,
+          "fields": 93,
+          "fn": 2,
+          "fp": 5,
+          "json_parse_ok": 36,
+          "missed_abstention": 3,
+          "natural_stop": 36,
+          "nonrepetitive_nonempty": 36,
+          "predicted_values": 55,
+          "present_correct": 50,
+          "present_total": 52,
+          "schema_ok": 36,
           "token_limited": 0,
-          "tp": 35,
+          "tp": 50,
           "unsupported_values": 0
         },
-        "exact": 1.0,
-        "f1": 1.0,
-        "field_accuracy": 1.0,
+        "exact": 0.8611111111111112,
+        "f1": 0.9345794392523366,
+        "false_abstention_rate": 0.0,
+        "field_accuracy": 0.946236559139785,
         "json_parse_rate": 1.0,
-        "output_nll": 0.0027189763324193855,
-        "precision": 1.0,
-        "present_accuracy": 1.0,
-        "recall": 1.0,
-        "schema_valid_rate": 1.0
+        "missed_abstention_rate": 0.07317073170731707,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.023127176140810154,
+        "precision": 0.9090909090909091,
+        "present_accuracy": 0.9615384615384616,
+        "recall": 0.9615384615384616,
+        "schema_valid_rate": 1.0,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.0
+      },
+      "macro_exact": 0.7936507936507936,
+      "paired_requests": {
+        "accuracy": 0.7040816326530612,
+        "both_correct": 69,
+        "count": 98
       },
-      "SearchScreeningEvent": {
-        "absent_accuracy": 0.9354838709677419,
+      "variant/alias": {
+        "absent_accuracy": 0.96875,
         "count": 28,
         "counts": {
-          "absent_correct": 29,
-          "absent_total": 31,
+          "absent_correct": 31,
+          "absent_total": 32,
           "exact": 24,
-          "false_abstention": 0,
-          "field_correct": 67,
-          "fields": 71,
-          "fn": 2,
-          "format_ok": 28,
-          "fp": 4,
-          "json_parse": 28,
-          "missed_abstention": 2,
+          "false_abstention": 3,
+          "field_correct": 64,
+          "fields": 68,
+          "fn": 3,
+          "fp": 1,
+          "json_parse_ok": 28,
+          "missed_abstention": 1,
           "natural_stop": 28,
           "nonrepetitive_nonempty": 28,
-          "predicted_values": 42,
-          "present_correct": 38,
-          "present_total": 40,
-          "requested_order": 28,
-          "schema_valid": 28,
+          "predicted_values": 34,
+          "present_correct": 33,
+          "present_total": 36,
+          "schema_ok": 28,
           "token_limited": 0,
-          "tp": 38,
-          "unsupported_values": 1
+          "tp": 33,
+          "unsupported_values": 0
         },
         "exact": 0.8571428571428571,
-        "f1": 0.9268292682926829,
-        "field_accuracy": 0.9436619718309859,
+        "f1": 0.9428571428571428,
+        "false_abstention_rate": 0.08333333333333333,
+        "field_accuracy": 0.9411764705882353,
         "json_parse_rate": 1.0,
-        "output_nll": 0.011078711751820047,
-        "precision": 0.9047619047619048,
-        "present_accuracy": 0.95,
-        "recall": 0.95,
-        "schema_valid_rate": 1.0
+        "missed_abstention_rate": 0.03125,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.058719543459509235,
+        "precision": 0.9705882352941176,
+        "present_accuracy": 0.9166666666666666,
+        "recall": 0.9166666666666666,
+        "schema_valid_rate": 1.0,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.0
       },
-      "all": {
-        "absent_accuracy": 0.9109947643979057,
+      "variant/canonical": {
+        "absent_accuracy": 0.9319371727748691,
         "count": 196,
         "counts": {
-          "absent_correct": 174,
+          "absent_correct": 178,
           "absent_total": 191,
-          "exact": 168,
-          "false_abstention": 4,
-          "field_correct": 437,
+          "exact": 157,
+          "false_abstention": 17,
+          "field_correct": 426,
           "fields": 468,
-          "fn": 14,
-          "format_ok": 196,
-          "fp": 26,
-          "json_parse": 196,
-          "missed_abstention": 17,
+          "fn": 29,
+          "fp": 24,
+          "json_parse_ok": 196,
+          "missed_abstention": 13,
           "natural_stop": 196,
           "nonrepetitive_nonempty": 196,
-          "predicted_values": 290,
-          "present_correct": 263,
+          "predicted_values": 273,
+          "present_correct": 248,
           "present_total": 277,
-          "requested_order": 196,
-          "schema_valid": 196,
+          "schema_ok": 196,
           "token_limited": 0,
-          "tp": 264,
-          "unsupported_values": 1
+          "tp": 249,
+          "unsupported_values": 3
         },
-        "exact": 0.8571428571428571,
-        "f1": 0.9295774647887325,
-        "field_accuracy": 0.9337606837606838,
+        "exact": 0.8010204081632653,
+        "f1": 0.9038112522686025,
+        "false_abstention_rate": 0.061371841155234655,
+        "field_accuracy": 0.9102564102564102,
         "json_parse_rate": 1.0,
-        "output_nll": 0.02902049323023008,
-        "precision": 0.9103448275862069,
-        "present_accuracy": 0.9494584837545126,
-        "recall": 0.9496402877697842,
-        "schema_valid_rate": 1.0
+        "missed_abstention_rate": 0.06806282722513089,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.02958408025570323,
+        "precision": 0.9120879120879121,
+        "present_accuracy": 0.8953068592057761,
+        "recall": 0.89568345323741,
+        "schema_valid_rate": 1.0,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.01098901098901099
       },
-      "macro_exact": 0.8571428571428571,
-      "paired_requests": {
-        "both_correct": 73,
-        "count": 98
-      }
-    },
-    "labeled": {
-      "AddToPlaylist": {
-        "absent_accuracy": 0.8709677419354839,
+      "variant/nullable": {
+        "absent_accuracy": 0.875,
         "count": 28,
         "counts": {
-          "absent_correct": 27,
-          "absent_total": 31,
+          "absent_correct": 28,
+          "absent_total": 32,
           "exact": 19,
-          "false_abstention": 4,
-          "field_correct": 61,
-          "fields": 71,
-          "fn": 6,
-          "format_ok": 28,
-          "fp": 6,
+          "false_abstention": 5,
+          "field_correct": 56,
+          "fields": 68,
+          "fn": 8,
+          "fp": 4,
+          "json_parse_ok": 27,
           "missed_abstention": 4,
           "natural_stop": 28,
           "nonrepetitive_nonempty": 28,
-          "predicted_values": 40,
-          "present_correct": 34,
-          "present_total": 40,
-          "requested_order": 28,
+          "predicted_values": 32,
+          "present_correct": 28,
+          "present_total": 36,
+          "schema_ok": 27,
           "token_limited": 0,
-          "tp": 34,
-          "unsupported_values": 0
+          "tp": 28,
+          "unsupported_values": 1
         },
         "exact": 0.6785714285714286,
-        "f1": 0.85,
-        "field_accuracy": 0.8591549295774648,
-        "output_nll": 0.05353730236882776,
-        "precision": 0.85,
-        "present_accuracy": 0.85,
-        "recall": 0.85
-      },
-      "BookRestaurant": {
-        "absent_accuracy": 0.8387096774193549,
-        "count": 28,
+        "f1": 0.823529411764706,
+        "false_abstention_rate": 0.1388888888888889,
+        "field_accuracy": 0.8235294117647058,
+        "json_parse_rate": 0.9642857142857143,
+        "missed_abstention_rate": 0.125,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.08280819396687937,
+        "precision": 0.875,
+        "present_accuracy": 0.7777777777777778,
+        "recall": 0.7777777777777778,
+        "schema_valid_rate": 0.9642857142857143,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.03125
+      }
+    },
+    "protocol": {
+      "constrained_decoding": false,
+      "json_assembly": false,
+      "max_tokens": 256,
+      "model_calls_per_request": 1,
+      "output_repair": false,
+      "prompt_truncation": false,
+      "repetition_penalty": 1,
+      "scorer": "strict-raw-json100-v1",
+      "serialization": "native-sft-chat-v1",
+      "temperature": 0
+    },
+    "scope": "Raw model-generated JSON, source-span annotation agreement. Duplicate keys, nonfinite JSON, fences, trailing prose and schema violations fail exact correctness. Flat string arrays/nullable strings in seven SNIPS domains; observed research panels, not a new blind or external benchmark."
+  },
+  "test": {
+    "metrics": {
+      "all": {
+        "absent_accuracy": 0.9246395806028833,
+        "count": 1444,
         "counts": {
-          "absent_correct": 26,
-          "absent_total": 31,
-          "exact": 22,
-          "false_abstention": 2,
-          "field_correct": 65,
-          "fields": 73,
-          "fn": 3,
-          "format_ok": 28,
-          "fp": 6,
-          "missed_abstention": 5,
-          "natural_stop": 28,
-          "nonrepetitive_nonempty": 28,
-          "predicted_values": 45,
-          "present_correct": 39,
-          "present_total": 42,
-          "requested_order": 28,
+          "absent_correct": 1411,
+          "absent_total": 1526,
+          "exact": 1191,
+          "false_abstention": 111,
+          "field_correct": 3193,
+          "fields": 3497,
+          "fn": 189,
+          "fp": 186,
+          "json_parse_ok": 1442,
+          "missed_abstention": 115,
+          "natural_stop": 1444,
+          "nonrepetitive_nonempty": 1444,
+          "predicted_values": 1968,
+          "present_correct": 1782,
+          "present_total": 1971,
+          "schema_ok": 1441,
           "token_limited": 0,
-          "tp": 39,
-          "unsupported_values": 0
-        },
-        "exact": 0.7857142857142857,
-        "f1": 0.896551724137931,
-        "field_accuracy": 0.8904109589041096,
-        "output_nll": 0.04426888480416632,
-        "precision": 0.8666666666666667,
-        "present_accuracy": 0.9285714285714286,
-        "recall": 0.9285714285714286
+          "tp": 1782,
+          "unsupported_values": 30
+        },
+        "exact": 0.824792243767313,
+        "f1": 0.9047981721249048,
+        "false_abstention_rate": 0.0563165905631659,
+        "field_accuracy": 0.9130683442951101,
+        "json_parse_rate": 0.9986149584487535,
+        "missed_abstention_rate": 0.07536041939711664,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.028334847962383555,
+        "precision": 0.9054878048780488,
+        "present_accuracy": 0.9041095890410958,
+        "recall": 0.9041095890410958,
+        "schema_valid_rate": 0.9979224376731302,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.01524390243902439
       },
-      "GetWeather": {
-        "absent_accuracy": 0.9,
-        "count": 28,
+      "domain/AddToPlaylist": {
+        "absent_accuracy": 0.9110169491525424,
+        "count": 206,
         "counts": {
-          "absent_correct": 27,
-          "absent_total": 30,
-          "exact": 23,
-          "false_abstention": 1,
-          "field_correct": 64,
-          "fields": 70,
-          "fn": 3,
-          "format_ok": 28,
-          "fp": 4,
-          "missed_abstention": 3,
-          "natural_stop": 28,
-          "nonrepetitive_nonempty": 28,
-          "predicted_values": 42,
-          "present_correct": 37,
-          "present_total": 40,
-          "requested_order": 28,
+          "absent_correct": 215,
+          "absent_total": 236,
+          "exact": 168,
+          "false_abstention": 13,
+          "field_correct": 475,
+          "fields": 526,
+          "fn": 30,
+          "fp": 38,
+          "json_parse_ok": 206,
+          "missed_abstention": 21,
+          "natural_stop": 206,
+          "nonrepetitive_nonempty": 206,
+          "predicted_values": 298,
+          "present_correct": 260,
+          "present_total": 290,
+          "schema_ok": 206,
           "token_limited": 0,
-          "tp": 38,
-          "unsupported_values": 0
+          "tp": 260,
+          "unsupported_values": 13
         },
-        "exact": 0.8214285714285714,
-        "f1": 0.9156626506024096,
-        "field_accuracy": 0.9142857142857143,
-        "output_nll": 0.053932554667491264,
-        "precision": 0.9047619047619048,
-        "present_accuracy": 0.925,
-        "recall": 0.926829268292683
+        "exact": 0.8155339805825242,
+        "f1": 0.8843537414965987,
+        "false_abstention_rate": 0.04482758620689655,
+        "field_accuracy": 0.903041825095057,
+        "json_parse_rate": 1.0,
+        "missed_abstention_rate": 0.08898305084745763,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.04326067907575776,
+        "precision": 0.87248322147651,
+        "present_accuracy": 0.896551724137931,
+        "recall": 0.896551724137931,
+        "schema_valid_rate": 1.0,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.0436241610738255
       },
-      "PlayMusic": {
-        "absent_accuracy": 0.9696969696969697,
-        "count": 28,
+      "domain/BookRestaurant": {
+        "absent_accuracy": 0.908,
+        "count": 208,
         "counts": {
-          "absent_correct": 32,
-          "absent_total": 33,
-          "exact": 24,
-          "false_abstention": 4,
-          "field_correct": 65,
-          "fields": 70,
-          "fn": 4,
-          "format_ok": 28,
-          "fp": 1,
-          "missed_abstention": 1,
-          "natural_stop": 28,
-          "nonrepetitive_nonempty": 28,
-          "predicted_values": 34,
-          "present_correct": 33,
-          "present_total": 37,
-          "requested_order": 28,
+          "absent_correct": 227,
+          "absent_total": 250,
+          "exact": 156,
+          "false_abstention": 21,
+          "field_correct": 506,
+          "fields": 569,
+          "fn": 40,
+          "fp": 42,
+          "json_parse_ok": 208,
+          "missed_abstention": 23,
+          "natural_stop": 208,
+          "nonrepetitive_nonempty": 208,
+          "predicted_values": 321,
+          "present_correct": 279,
+          "present_total": 319,
+          "schema_ok": 208,
           "token_limited": 0,
-          "tp": 33,
-          "unsupported_values": 0
+          "tp": 279,
+          "unsupported_values": 4
         },
-        "exact": 0.8571428571428571,
-        "f1": 0.9295774647887325,
-        "field_accuracy": 0.9285714285714286,
-        "output_nll": 0.020421282849773497,
-        "precision": 0.9705882352941176,
-        "present_accuracy": 0.8918918918918919,
-        "recall": 0.8918918918918919
+        "exact": 0.75,
+        "f1": 0.8718750000000001,
+        "false_abstention_rate": 0.06583072100313479,
+        "field_accuracy": 0.8892794376098418,
+        "json_parse_rate": 1.0,
+        "missed_abstention_rate": 0.092,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.032639438736133834,
+        "precision": 0.8691588785046729,
+        "present_accuracy": 0.8746081504702194,
+        "recall": 0.8746081504702194,
+        "schema_valid_rate": 1.0,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.012461059190031152
       },
-      "RateBook": {
-        "absent_accuracy": 1.0,
-        "count": 28,
+      "domain/GetWeather": {
+        "absent_accuracy": 0.9349593495934959,
+        "count": 206,
         "counts": {
-          "absent_correct": 31,
-          "absent_total": 31,
-          "exact": 25,
-          "false_abstention": 0,
-          "field_correct": 71,
-          "fields": 74,
-          "fn": 3,
-          "format_ok": 28,
-          "fp": 3,
-          "missed_abstention": 0,
-          "natural_stop": 28,
-          "nonrepetitive_nonempty": 28,
-          "predicted_values": 43,
-          "present_correct": 40,
-          "present_total": 43,
-          "requested_order": 28,
+          "absent_correct": 230,
+          "absent_total": 246,
+          "exact": 168,
+          "false_abstention": 15,
+          "field_correct": 479,
+          "fields": 525,
+          "fn": 30,
+          "fp": 27,
+          "json_parse_ok": 205,
+          "missed_abstention": 16,
+          "natural_stop": 206,
+          "nonrepetitive_nonempty": 206,
+          "predicted_values": 276,
+          "present_correct": 249,
+          "present_total": 279,
+          "schema_ok": 204,
           "token_limited": 0,
-          "tp": 40,
-          "unsupported_values": 0
-        },
-        "exact": 0.8928571428571429,
-        "f1": 0.9302325581395349,
-        "field_accuracy": 0.9594594594594594,
-        "output_nll": 0.017240766815250805,
-        "precision": 0.9302325581395349,
-        "present_accuracy": 0.9302325581395349,
-        "recall": 0.9302325581395349
+          "tp": 249,
+          "unsupported_values": 2
+        },
+        "exact": 0.8155339805825242,
+        "f1": 0.8972972972972973,
+        "false_abstention_rate": 0.053763440860215055,
+        "field_accuracy": 0.9123809523809524,
+        "json_parse_rate": 0.9951456310679612,
+        "missed_abstention_rate": 0.06504065040650407,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.028848352069651157,
+        "precision": 0.9021739130434783,
+        "present_accuracy": 0.8924731182795699,
+        "recall": 0.8924731182795699,
+        "schema_valid_rate": 0.9902912621359223,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.007246376811594203
       },
-      "SearchCreativeWork": {
-        "absent_accuracy": 1.0,
-        "count": 28,
+      "domain/PlayMusic": {
+        "absent_accuracy": 0.94140625,
+        "count": 208,
         "counts": {
-          "absent_correct": 4,
-          "absent_total": 4,
-          "exact": 28,
-          "false_abstention": 0,
-          "field_correct": 39,
-          "fields": 39,
-          "fn": 0,
-          "format_ok": 28,
-          "fp": 0,
-          "missed_abstention": 0,
-          "natural_stop": 28,
-          "nonrepetitive_nonempty": 28,
-          "predicted_values": 35,
-          "present_correct": 35,
-          "present_total": 35,
-          "requested_order": 28,
+          "absent_correct": 241,
+          "absent_total": 256,
+          "exact": 149,
+          "false_abstention": 44,
+          "field_correct": 451,
+          "fields": 523,
+          "fn": 57,
+          "fp": 28,
+          "json_parse_ok": 208,
+          "missed_abstention": 15,
+          "natural_stop": 208,
+          "nonrepetitive_nonempty": 208,
+          "predicted_values": 238,
+          "present_correct": 210,
+          "present_total": 267,
+          "schema_ok": 208,
           "token_limited": 0,
-          "tp": 35,
-          "unsupported_values": 0
+          "tp": 210,
+          "unsupported_values": 2
         },
-        "exact": 1.0,
-        "f1": 1.0,
-        "field_accuracy": 1.0,
-        "output_nll": 0.0035197215885481065,
-        "precision": 1.0,
-        "present_accuracy": 1.0,
-        "recall": 1.0
+        "exact": 0.7163461538461539,
+        "f1": 0.8316831683168316,
+        "false_abstention_rate": 0.1647940074906367,
+        "field_accuracy": 0.8623326959847036,
+        "json_parse_rate": 1.0,
+        "missed_abstention_rate": 0.05859375,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.04051304714644525,
+        "precision": 0.8823529411764706,
+        "present_accuracy": 0.7865168539325843,
+        "recall": 0.7865168539325843,
+        "schema_valid_rate": 1.0,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.008403361344537815
       },
-      "SearchScreeningEvent": {
-        "absent_accuracy": 0.8709677419354839,
-        "count": 28,
+      "domain/RateBook": {
+        "absent_accuracy": 0.9552845528455285,
+        "count": 204,
         "counts": {
-          "absent_correct": 27,
-          "absent_total": 31,
-          "exact": 23,
-          "false_abstention": 0,
-          "field_correct": 66,
-          "fields": 71,
-          "fn": 1,
-          "format_ok": 28,
-          "fp": 5,
-          "missed_abstention": 4,
-          "natural_stop": 28,
-          "nonrepetitive_nonempty": 28,
-          "predicted_values": 44,
-          "present_correct": 39,
-          "present_total": 40,
-          "requested_order": 28,
+          "absent_correct": 235,
+          "absent_total": 246,
+          "exact": 188,
+          "false_abstention": 2,
+          "field_correct": 548,
+          "fields": 567,
+          "fn": 8,
+          "fp": 14,
+          "json_parse_ok": 203,
+          "missed_abstention": 11,
+          "natural_stop": 204,
+          "nonrepetitive_nonempty": 204,
+          "predicted_values": 327,
+          "present_correct": 313,
+          "present_total": 321,
+          "schema_ok": 203,
           "token_limited": 0,
-          "tp": 39,
-          "unsupported_values": 1
-        },
-        "exact": 0.8214285714285714,
-        "f1": 0.9285714285714285,
-        "field_accuracy": 0.9295774647887324,
-        "output_nll": 0.011952276432566578,
-        "precision": 0.8863636363636364,
-        "present_accuracy": 0.975,
-        "recall": 0.975
+          "tp": 313,
+          "unsupported_values": 2
+        },
+        "exact": 0.9215686274509803,
+        "f1": 0.9660493827160493,
+        "false_abstention_rate": 0.006230529595015576,
+        "field_accuracy": 0.9664902998236331,
+        "json_parse_rate": 0.9950980392156863,
+        "missed_abstention_rate": 0.044715447154471545,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.011863740803485545,
+        "precision": 0.9571865443425076,
+        "present_accuracy": 0.9750778816199377,
+        "recall": 0.9750778816199377,
+        "schema_valid_rate": 0.9950980392156863,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.0061162079510703364
       },
-      "all": {
-        "absent_accuracy": 0.9109947643979057,
-        "count": 196,
+      "domain/SearchCreativeWork": {
+        "absent_accuracy": 0.9230769230769231,
+        "count": 208,
         "counts": {
-          "absent_correct": 174,
-          "absent_total": 191,
-          "exact": 164,
-          "false_abstention": 11,
-          "field_correct": 431,
-          "fields": 468,
-          "fn": 20,
-          "format_ok": 196,
-          "fp": 25,
-          "missed_abstention": 17,
-          "natural_stop": 196,
-          "nonrepetitive_nonempty": 196,
-          "predicted_values": 283,
-          "present_correct": 257,
-          "present_total": 277,
-          "requested_order": 196,
+          "absent_correct": 36,
+          "absent_total": 39,
+          "exact": 197,
+          "false_abstention": 2,
+          "field_correct": 275,
+          "fields": 286,
+          "fn": 8,
+          "fp": 9,
+          "json_parse_ok": 208,
+          "missed_abstention": 3,
+          "natural_stop": 208,
+          "nonrepetitive_nonempty": 208,
+          "predicted_values": 248,
+          "present_correct": 239,
+          "present_total": 247,
+          "schema_ok": 208,
           "token_limited": 0,
-          "tp": 258,
-          "unsupported_values": 1
+          "tp": 239,
+          "unsupported_values": 0
         },
-        "exact": 0.8367346938775511,
-        "f1": 0.9197860962566845,
-        "field_accuracy": 0.9209401709401709,
-        "output_nll": 0.029267541360946333,
-        "precision": 0.911660777385159,
-        "present_accuracy": 0.927797833935018,
-        "recall": 0.9280575539568345
+        "exact": 0.9471153846153846,
+        "f1": 0.9656565656565657,
+        "false_abstention_rate": 0.008097165991902834,
+        "field_accuracy": 0.9615384615384616,
+        "json_parse_rate": 1.0,
+        "missed_abstention_rate": 0.07692307692307693,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.019352024208483972,
+        "precision": 0.9637096774193549,
+        "present_accuracy": 0.9676113360323887,
+        "recall": 0.9676113360323887,
+        "schema_valid_rate": 1.0,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.0
       },
-      "macro_exact": 0.836734693877551,
-      "paired_requests": {
-        "both_correct": 71,
-        "count": 98
-      }
-    },
-    "qa": {
-      "all": {
-        "absent": {
-          "correct": 1,
-          "total": 24
-        },
-        "count": 136,
+      "domain/SearchScreeningEvent": {
+        "absent_accuracy": 0.8972332015810277,
+        "count": 204,
         "counts": {
-          "abstained": 4,
-          "all_requested_values_present": 8,
-          "exact": 34,
-          "false_abstention": 3,
-          "missed_abstention": 23,
-          "natural_stop": 136,
-          "nonrepetitive_nonempty": 136,
-          "relation_correct": 34,
-          "token_f1": 65.24622858028253,
+          "absent_correct": 227,
+          "absent_total": 253,
+          "exact": 165,
+          "false_abstention": 14,
+          "field_correct": 459,
+          "fields": 501,
+          "fn": 16,
+          "fp": 28,
+          "json_parse_ok": 204,
+          "missed_abstention": 26,
+          "natural_stop": 204,
+          "nonrepetitive_nonempty": 204,
+          "predicted_values": 260,
+          "present_correct": 232,
+          "present_total": 248,
+          "schema_ok": 204,
           "token_limited": 0,
-          "wrong_value_mentioned": 4
-        },
-        "exact": 0.25,
-        "output_nll": 0.5983618814429706,
-        "present": {
-          "correct": 33,
-          "total": 112
+          "tp": 232,
+          "unsupported_values": 7
         },
-        "token_f1": 0.47975168073737157
+        "exact": 0.8088235294117647,
+        "f1": 0.9133858267716536,
+        "false_abstention_rate": 0.056451612903225805,
+        "field_accuracy": 0.9161676646706587,
+        "json_parse_rate": 1.0,
+        "missed_abstention_rate": 0.10276679841897234,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.021568229504291415,
+        "precision": 0.8923076923076924,
+        "present_accuracy": 0.9354838709677419,
+        "recall": 0.9354838709677419,
+        "schema_valid_rate": 1.0,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.026923076923076925
       },
-      "balanced_squad_exact": 0.2734375,
-      "calibration": {
-        "absent": {
-          "correct": 0,
-          "total": 8
-        },
-        "count": 16,
-        "counts": {
-          "abstained": 0,
-          "all_requested_values_present": 2,
-          "exact": 0,
-          "false_abstention": 0,
-          "missed_abstention": 8,
-          "natural_stop": 16,
-          "nonrepetitive_nonempty": 16,
-          "relation_correct": 0,
-          "token_f1": 4.1313519813519815,
-          "token_limited": 0,
-          "wrong_value_mentioned": 0
-        },
-        "exact": 0.0,
-        "output_nll": 0.606968455016613,
-        "present": {
-          "correct": 0,
-          "total": 8
-        },
-        "relation_accuracy": 0.0,
-        "token_f1": 0.25820949883449884
+      "macro_exact": 0.8249888080699045,
+      "paired_requests": {
+        "accuracy": 0.7219020172910663,
+        "both_correct": 501,
+        "count": 694
       },
-      "corrections": {
-        "absent": {
-          "correct": 0,
-          "total": 0
-        },
-        "count": 16,
+      "variant/alias": {
+        "absent_accuracy": 0.8275862068965517,
+        "count": 28,
         "counts": {
-          "abstained": 2,
-          "all_requested_values_present": 4,
-          "exact": 0,
+          "absent_correct": 24,
+          "absent_total": 29,
+          "exact": 19,
           "false_abstention": 2,
-          "missed_abstention": 0,
-          "natural_stop": 16,
-          "nonrepetitive_nonempty": 16,
-          "relation_correct": 0,
-          "token_f1": 8.903747069923542,
-          "token_limited": 0,
-          "wrong_value_mentioned": 4
-        },
-        "exact": 0.0,
-        "output_nll": 0.49557842686772346,
-        "present": {
-          "correct": 0,
-          "total": 16
-        },
-        "relation_accuracy": 0.0,
-        "token_f1": 0.5564841918702214
-      },
-      "dolly_closed_qa": {
-        "absent": {
-          "correct": 0,
-          "total": 0
-        },
-        "count": 8,
-        "counts": {
-          "abstained": 1,
-          "all_requested_values_present": 0,
-          "exact": 0,
-          "false_abstention": 1,
-          "missed_abstention": 0,
-          "natural_stop": 8,
-          "nonrepetitive_nonempty": 8,
-          "relation_correct": 0,
-          "token_f1": 2.238761490082245,
+          "field_correct": 62,
+          "fields": 75,
+          "fn": 8,
+          "fp": 7,
+          "json_parse_ok": 27,
+          "missed_abstention": 5,
+          "natural_stop": 28,
+          "nonrepetitive_nonempty": 28,
+          "predicted_values": 45,
+          "present_correct": 38,
+          "present_total": 46,
+          "schema_ok": 26,
           "token_limited": 0,
-          "wrong_value_mentioned": 0
-        },
-        "exact": 0.0,
-        "output_nll": 0.9192326962947845,
-        "present": {
-          "correct": 0,
-          "total": 8
+          "tp": 38,
+          "unsupported_values": 2
         },
-        "token_f1": 0.2798451862602806
-      },
-      "legacy_summary": {
-        "count": 88,
-        "exact": 0.36363636363636365,
-        "false_abstention": 1
+        "exact": 0.6785714285714286,
+        "f1": 0.8351648351648352,
+        "false_abstention_rate": 0.043478260869565216,
+        "field_accuracy": 0.8266666666666667,
+        "json_parse_rate": 0.9642857142857143,
+        "missed_abstention_rate": 0.1724137931034483,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.05925792032836138,
+        "precision": 0.8444444444444444,
+        "present_accuracy": 0.8260869565217391,
+        "recall": 0.8260869565217391,
+        "schema_valid_rate": 0.9285714285714286,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.044444444444444446
       },
-      "multipart": {
-        "absent": {
-          "correct": 0,
-          "total": 0
-        },
-        "count": 16,
+      "variant/canonical": {
+        "absent_accuracy": 0.9305177111716622,
+        "count": 1388,
         "counts": {
-          "abstained": 0,
-          "all_requested_values_present": 2,
-          "exact": 2,
-          "false_abstention": 0,
-          "missed_abstention": 0,
-          "natural_stop": 16,
-          "nonrepetitive_nonempty": 16,
-          "relation_correct": 2,
-          "token_f1": 8.500593013124984,
+          "absent_correct": 1366,
+          "absent_total": 1468,
+          "exact": 1156,
+          "false_abstention": 103,
+          "field_correct": 3074,
+          "fields": 3347,
+          "fn": 171,
+          "fp": 170,
+          "json_parse_ok": 1388,
+          "missed_abstention": 102,
+          "natural_stop": 1388,
+          "nonrepetitive_nonempty": 1388,
+          "predicted_values": 1878,
+          "present_correct": 1708,
+          "present_total": 1879,
+          "schema_ok": 1388,
           "token_limited": 0,
-          "wrong_value_mentioned": 0
-        },
-        "exact": 0.125,
-        "output_nll": 0.44682544376701117,
-        "present": {
-          "correct": 2,
-          "total": 16
-        },
-        "relation_accuracy": 0.125,
-        "token_f1": 0.5312870633203115
-      },
-      "pairs_by_source": {
-        "calibration": {
-          "both_exact": 0,
-          "count": 8,
-          "same_answer": 2
-        },
-        "corrections": {
-          "both_exact": 0,
-          "count": 8,
-          "same_answer": 1
+          "tp": 1708,
+          "unsupported_values": 24
         },
-        "squad2": {
-          "both_exact": 10,
-          "count": 32,
-          "same_answer": 4
-        }
-      },
-      "question_pairs": {
-        "both_exact": 10,
-        "count": 48,
-        "same_answer": 7
+        "exact": 0.8328530259365994,
+        "f1": 0.909236092627096,
+        "false_abstention_rate": 0.05481639169771155,
+        "field_accuracy": 0.9184344188825814,
+        "json_parse_rate": 1.0,
+        "missed_abstention_rate": 0.06948228882833787,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.02666785328274954,
+        "precision": 0.9094781682641108,
+        "present_accuracy": 0.9089941458222459,
+        "recall": 0.9089941458222459,
+        "schema_valid_rate": 1.0,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.012779552715654952
       },
-      "selection_score": 0.15755208333333334,
-      "squad2": {
-        "absent": {
-          "correct": 1,
-          "total": 16
-        },
-        "count": 80,
+      "variant/nullable": {
+        "absent_accuracy": 0.7241379310344828,
+        "count": 28,
         "counts": {
-          "abstained": 1,
-          "all_requested_values_present": 0,
-          "exact": 32,
-          "false_abstention": 0,
-          "missed_abstention": 15,
-          "natural_stop": 80,
-          "nonrepetitive_nonempty": 80,
-          "relation_correct": 32,
-          "token_f1": 41.47177502579979,
+          "absent_correct": 21,
+          "absent_total": 29,
+          "exact": 16,
+          "false_abstention": 6,
+          "field_correct": 57,
+          "fields": 75,
+          "fn": 10,
+          "fp": 9,
+          "json_parse_ok": 27,
+          "missed_abstention": 8,
+          "natural_stop": 28,
+          "nonrepetitive_nonempty": 28,
+          "predicted_values": 45,
+          "present_correct": 36,
+          "present_total": 46,
+          "schema_ok": 27,
           "token_limited": 0,
-          "wrong_value_mentioned": 0
-        },
-        "exact": 0.4,
-        "output_nll": 0.6154174636933021,
-        "present": {
-          "correct": 31,
-          "total": 64
-        },
-        "token_f1": 0.5183971878224974
+          "tp": 36,
+          "unsupported_values": 4
+        },
+        "exact": 0.5714285714285714,
+        "f1": 0.7912087912087912,
+        "false_abstention_rate": 0.13043478260869565,
+        "field_accuracy": 0.76,
+        "json_parse_rate": 0.9642857142857143,
+        "missed_abstention_rate": 0.27586206896551724,
+        "natural_stop_rate": 1.0,
+        "nonrepetitive_nonempty_rate": 1.0,
+        "output_nll": 0.08004708328683462,
+        "precision": 0.8,
+        "present_accuracy": 0.782608695652174,
+        "recall": 0.782608695652174,
+        "schema_valid_rate": 0.9642857142857143,
+        "token_limited_rate": 0.0,
+        "unsupported_value_rate": 0.08888888888888889
       }
-    }
+    },
+    "protocol": {
+      "constrained_decoding": false,
+      "json_assembly": false,
+      "max_tokens": 256,
+      "model_calls_per_request": 1,
+      "output_repair": false,
+      "prompt_truncation": false,
+      "repetition_penalty": 1,
+      "scorer": "strict-raw-json100-v1",
+      "serialization": "native-sft-chat-v1",
+      "temperature": 0
+    },
+    "scope": "Raw model-generated JSON, source-span annotation agreement. Duplicate keys, nonfinite JSON, fences, trailing prose and schema violations fail exact correctness. Flat string arrays/nullable strings in seven SNIPS domains; observed research panels, not a new blind or external benchmark."
   },
-  "development_readiness": {
-    "failed_gates": [
-      "json_content_floor",
-      "json_vs_start",
-      "labeled_retention",
-      "qa_retention"
-    ],
-    "json_content_floor": {
-      "paired_delta": 3,
-      "passed": false,
-      "per_source_deltas": {
-        "AddToPlaylist": {
-          "absent_correct": 0,
-          "exact": 0,
+  "playground": {
+    "metrics": {
+      "alias": {
+        "absent_accuracy": 0.875,
+        "count": 7,
+        "counts": {
+          "absent_correct": 7,
+          "absent_total": 8,
+          "exact": 6,
           "false_abstention": 0,
-          "field_correct": 0,
+          "field_correct": 16,
+          "fields": 17,
           "fn": 0,
-          "format_ok": 0,
-          "fp": 0,
-          "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 0,
-          "tp": 0,
-          "unsupported_values": 0
-        },
-        "BookRestaurant": {
-          "absent_correct": 1,
-          "exact": 0,
-          "false_abstention": -2,
-          "field_correct": 2,
-          "fn": -1,
-          "format_ok": 0,
-          "fp": 0,
-          "missed_abstention": -1,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 1,
-          "tp": 1,
-          "unsupported_values": 0
-        },
-        "GetWeather": {
-          "absent_correct": 0,
-          "exact": 1,
-          "false_abstention": -4,
-          "field_correct": 3,
-          "fn": -3,
-          "format_ok": 0,
-          "fp": 1,
-          "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 3,
-          "tp": 3,
-          "unsupported_values": 0
-        },
-        "PlayMusic": {
-          "absent_correct": 1,
-          "exact": 2,
-          "false_abstention": -2,
-          "field_correct": 3,
-          "fn": -2,
-          "format_ok": 0,
-          "fp": -1,
-          "missed_abstention": -1,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 2,
-          "tp": 2,
-          "unsupported_values": 0
-        },
-        "RateBook": {
-          "absent_correct": 0,
-          "exact": 2,
-          "false_abstention": -2,
-          "field_correct": 2,
-          "fn": -2,
-          "format_ok": 0,
-          "fp": 0,
-          "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 2,
-          "tp": 2,
-          "unsupported_values": 0
-        },
-        "SearchCreativeWork": {
-          "absent_correct": 0,
-          "exact": 2,
-          "false_abstention": 0,
-          "field_correct": 2,
-          "fn": -2,
-          "format_ok": 0,
-          "fp": -2,
-          "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 2,
-          "tp": 2,
-          "unsupported_values": 0
-        },
-        "SearchScreeningEvent": {
-          "absent_correct": 0,
-          "exact": -1,
-          "false_abstention": 0,
-          "field_correct": -1,
-          "fn": 1,
-          "format_ok": 0,
           "fp": 1,
-          "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": -1,
-          "tp": -1,
-          "unsupported_values": 0
-        },
-        "all": {
-          "absent_correct": 2,
-          "exact": 6,
-          "false_abstention": -10,
-          "field_correct": 11,
-          "fn": -9,
-          "format_ok": 0,
-          "fp": -1,
-          "missed_abstention": -2,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
+          "json_parse_ok": 7,
+          "missed_abstention": 1,
+          "natural_stop": 7,
+          "nonrepetitive_nonempty": 7,
+          "predicted_values": 10,
           "present_correct": 9,
+          "present_total": 9,
+          "schema_ok": 7,
+          "token_limited": 0,
           "tp": 9,
           "unsupported_values": 0
-        }
-      },
-      "schema_valid": true
-    },
-    "json_vs_start": {
-      "paired_delta": 23,
-      "passed": false,
-      "per_source_deltas": {
-        "AddToPlaylist": {
-          "absent_correct": 6,
-          "exact": 4,
-          "false_abstention": 0,
-          "field_correct": 6,
-          "fn": 0,
-          "format_ok": 0,
-          "fp": -6,
-          "missed_abstention": -6,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 0,
-          "tp": 0,
-          "unsupported_values": 0
-        },
-        "BookRestaurant": {
-          "absent_correct": 5,
-          "exact": 4,
-          "false_abstention": 0,
-          "field_correct": 5,
-          "fn": 0,
-          "format_ok": 0,
-          "fp": -5,
-          "missed_abstention": -5,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 0,
-          "tp": 0,
-          "unsupported_values": -1
-        },
-        "GetWeather": {
-          "absent_correct": 8,
-          "exact": 8,
-          "false_abstention": 0,
-          "field_correct": 10,
-          "fn": -2,
-          "format_ok": 0,
-          "fp": -10,
-          "missed_abstention": -8,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 2,
-          "tp": 2,
-          "unsupported_values": -5
         },
-        "PlayMusic": {
-          "absent_correct": 8,
+        "exact": 0.8571428571428571,
+        "f1": 0.9473684210526316,
+        "field_accuracy": 0.9411764705882353,
+        "json_parse": 1.0,
+        "natural_stop_rate": 1.0,
+        "precision": 0.9,
+        "present_accuracy": 1.0,
+        "recall": 1.0,
+        "schema_valid": 1.0,
+        "unsupported_value_rate": 0.0
+      },
+      "all": {
+        "absent_accuracy": 0.9056603773584906,
+        "count": 54,
+        "counts": {
+          "absent_correct": 48,
+          "absent_total": 53,
+          "exact": 42,
+          "false_abstention": 3,
+          "field_correct": 118,
+          "fields": 135,
+          "fn": 13,
+          "fp": 7,
+          "json_parse_ok": 52,
+          "missed_abstention": 5,
+          "natural_stop": 54,
+          "nonrepetitive_nonempty": 54,
+          "predicted_values": 78,
+          "present_correct": 70,
+          "present_total": 82,
+          "schema_ok": 52,
+          "token_limited": 0,
+          "tp": 71,
+          "unsupported_values": 2
+        },
+        "exact": 0.7777777777777778,
+        "f1": 0.8765432098765431,
+        "field_accuracy": 0.8740740740740741,
+        "json_parse": 0.9629629629629629,
+        "natural_stop_rate": 1.0,
+        "precision": 0.9102564102564102,
+        "present_accuracy": 0.8536585365853658,
+        "recall": 0.8452380952380952,
+        "schema_valid": 0.9629629629629629,
+        "unsupported_value_rate": 0.02564102564102564
+      },
+      "authored_diagnostics": {
+        "absent_accuracy": 1.0,
+        "count": 12,
+        "counts": {
+          "absent_correct": 11,
+          "absent_total": 11,
           "exact": 7,
-          "false_abstention": 0,
-          "field_correct": 8,
-          "fn": 0,
-          "format_ok": 0,
-          "fp": -8,
-          "missed_abstention": -8,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 0,
-          "tp": 0,
-          "unsupported_values": 0
-        },
-        "RateBook": {
-          "absent_correct": 0,
-          "exact": 0,
-          "false_abstention": 0,
-          "field_correct": 0,
-          "fn": 0,
-          "format_ok": 0,
-          "fp": 0,
-          "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 0,
-          "tp": 0,
-          "unsupported_values": 0
-        },
-        "SearchCreativeWork": {
-          "absent_correct": 1,
-          "exact": 3,
-          "false_abstention": 0,
-          "field_correct": 3,
-          "fn": -2,
-          "format_ok": 0,
-          "fp": -3,
-          "missed_abstention": -1,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 2,
-          "tp": 2,
-          "unsupported_values": -1
-        },
-        "SearchScreeningEvent": {
-          "absent_correct": 4,
-          "exact": 3,
-          "false_abstention": 0,
-          "field_correct": 3,
-          "fn": 1,
-          "format_ok": 0,
-          "fp": -3,
-          "missed_abstention": -4,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": -1,
-          "tp": -1,
-          "unsupported_values": -1
-        },
-        "all": {
-          "absent_correct": 32,
-          "exact": 29,
-          "false_abstention": 0,
-          "field_correct": 35,
-          "fn": -3,
-          "format_ok": 0,
-          "fp": -35,
-          "missed_abstention": -32,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 3,
-          "tp": 3,
-          "unsupported_values": -8
-        }
-      }
-    },
-    "labeled_retention": {
-      "paired_delta": 1,
-      "passed": false,
-      "per_source_deltas": {
-        "AddToPlaylist": {
-          "absent_correct": 0,
-          "exact": -2,
-          "false_abstention": 0,
-          "field_correct": -2,
-          "fn": 2,
-          "format_ok": 0,
+          "false_abstention": 2,
+          "field_correct": 29,
+          "fields": 34,
+          "fn": 6,
           "fp": 2,
+          "json_parse_ok": 12,
           "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": -2,
-          "tp": -2,
-          "unsupported_values": 0
+          "natural_stop": 12,
+          "nonrepetitive_nonempty": 12,
+          "predicted_values": 21,
+          "present_correct": 18,
+          "present_total": 23,
+          "schema_ok": 12,
+          "token_limited": 0,
+          "tp": 19,
+          "unsupported_values": 1
         },
-        "BookRestaurant": {
-          "absent_correct": 2,
-          "exact": 1,
+        "exact": 0.5833333333333334,
+        "f1": 0.8260869565217391,
+        "field_accuracy": 0.8529411764705882,
+        "json_parse": 1.0,
+        "natural_stop_rate": 1.0,
+        "precision": 0.9047619047619048,
+        "present_accuracy": 0.782608695652174,
+        "recall": 0.76,
+        "schema_valid": 1.0,
+        "unsupported_value_rate": 0.047619047619047616
+      },
+      "nullable": {
+        "absent_accuracy": 0.625,
+        "count": 7,
+        "counts": {
+          "absent_correct": 5,
+          "absent_total": 8,
+          "exact": 4,
           "false_abstention": 0,
-          "field_correct": 2,
-          "fn": 0,
-          "format_ok": 0,
-          "fp": -2,
-          "missed_abstention": -2,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 0,
-          "tp": 0,
-          "unsupported_values": 0
-        },
-        "GetWeather": {
-          "absent_correct": 1,
-          "exact": 2,
-          "false_abstention": -3,
-          "field_correct": 4,
-          "fn": -3,
-          "format_ok": 0,
-          "fp": -1,
-          "missed_abstention": -1,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 3,
-          "tp": 3,
-          "unsupported_values": 0
-        },
-        "PlayMusic": {
-          "absent_correct": 1,
-          "exact": -1,
-          "false_abstention": 2,
-          "field_correct": -1,
+          "field_correct": 12,
+          "fields": 17,
           "fn": 2,
-          "format_ok": 0,
-          "fp": -1,
-          "missed_abstention": -1,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": -2,
-          "tp": -2,
-          "unsupported_values": 0
-        },
-        "RateBook": {
-          "absent_correct": 0,
-          "exact": 2,
-          "false_abstention": -2,
-          "field_correct": 2,
-          "fn": -2,
-          "format_ok": 0,
-          "fp": 0,
-          "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 2,
-          "tp": 2,
-          "unsupported_values": 0
-        },
-        "SearchCreativeWork": {
-          "absent_correct": 0,
-          "exact": 2,
-          "false_abstention": 0,
-          "field_correct": 2,
-          "fn": -2,
-          "format_ok": 0,
-          "fp": -2,
-          "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 2,
-          "tp": 2,
-          "unsupported_values": 0
-        },
-        "SearchScreeningEvent": {
-          "absent_correct": -2,
-          "exact": -2,
-          "false_abstention": 0,
-          "field_correct": -2,
-          "fn": 0,
-          "format_ok": 0,
           "fp": 2,
-          "missed_abstention": 2,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 0,
-          "tp": 0,
-          "unsupported_values": 0
+          "json_parse_ok": 6,
+          "missed_abstention": 3,
+          "natural_stop": 7,
+          "nonrepetitive_nonempty": 7,
+          "predicted_values": 9,
+          "present_correct": 7,
+          "present_total": 9,
+          "schema_ok": 6,
+          "token_limited": 0,
+          "tp": 7,
+          "unsupported_values": 1
         },
-        "all": {
-          "absent_correct": 2,
-          "exact": 2,
-          "false_abstention": -3,
-          "field_correct": 5,
-          "fn": -3,
-          "format_ok": 0,
-          "fp": -2,
-          "missed_abstention": -2,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "present_correct": 3,
-          "tp": 3,
+        "exact": 0.5714285714285714,
+        "f1": 0.7777777777777778,
+        "field_accuracy": 0.7058823529411765,
+        "json_parse": 0.8571428571428571,
+        "natural_stop_rate": 1.0,
+        "precision": 0.7777777777777778,
+        "present_accuracy": 0.7777777777777778,
+        "recall": 0.7777777777777778,
+        "schema_valid": 0.8571428571428571,
+        "unsupported_value_rate": 0.1111111111111111
+      },
+      "observed_originals": {
+        "absent_accuracy": 0.9615384615384616,
+        "count": 28,
+        "counts": {
+          "absent_correct": 25,
+          "absent_total": 26,
+          "exact": 25,
+          "false_abstention": 1,
+          "field_correct": 61,
+          "fields": 67,
+          "fn": 5,
+          "fp": 2,
+          "json_parse_ok": 27,
+          "missed_abstention": 1,
+          "natural_stop": 28,
+          "nonrepetitive_nonempty": 28,
+          "predicted_values": 38,
+          "present_correct": 36,
+          "present_total": 41,
+          "schema_ok": 27,
+          "token_limited": 0,
+          "tp": 36,
           "unsupported_values": 0
-        }
-      }
-    },
-    "passed": false,
-    "qa_retention": {
-      "absent_correct_delta": 0,
-      "answerable_correct_delta": 2,
-      "calibration_relation_delta": 0.0,
-      "corrections_relation_delta": 0.0,
-      "false_abstention_delta": 0,
-      "legacy_exact_delta": 0.022727272727272763,
-      "multipart_relation_delta": -0.0625,
-      "paired_correct_delta": 2,
-      "passed": false,
-      "per_source_deltas": {
-        "all": {
-          "all_requested_values_present": -2,
-          "exact": 1,
-          "false_abstention": 0,
-          "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "wrong_value_mentioned": 0
-        },
-        "calibration": {
-          "all_requested_values_present": 0,
-          "exact": 0,
-          "false_abstention": 0,
-          "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "wrong_value_mentioned": 0
-        },
-        "corrections": {
-          "all_requested_values_present": -1,
-          "exact": 0,
-          "false_abstention": 0,
-          "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "wrong_value_mentioned": 0
-        },
-        "dolly_closed_qa": {
-          "all_requested_values_present": 0,
-          "exact": 0,
-          "false_abstention": 0,
-          "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "wrong_value_mentioned": 0
         },
-        "multipart": {
-          "all_requested_values_present": -1,
-          "exact": -1,
+        "exact": 0.8928571428571429,
+        "f1": 0.9113924050632912,
+        "field_accuracy": 0.9104477611940298,
+        "json_parse": 0.9642857142857143,
+        "natural_stop_rate": 1.0,
+        "precision": 0.9473684210526315,
+        "present_accuracy": 0.8780487804878049,
+        "recall": 0.8780487804878049,
+        "schema_valid": 0.9642857142857143,
+        "unsupported_value_rate": 0.0
+      },
+      "schema_stress": {
+        "absent_accuracy": 0.75,
+        "count": 14,
+        "counts": {
+          "absent_correct": 12,
+          "absent_total": 16,
+          "exact": 10,
           "false_abstention": 0,
-          "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "wrong_value_mentioned": 0
+          "field_correct": 28,
+          "fields": 34,
+          "fn": 2,
+          "fp": 3,
+          "json_parse_ok": 13,
+          "missed_abstention": 4,
+          "natural_stop": 14,
+          "nonrepetitive_nonempty": 14,
+          "predicted_values": 19,
+          "present_correct": 16,
+          "present_total": 18,
+          "schema_ok": 13,
+          "token_limited": 0,
+          "tp": 16,
+          "unsupported_values": 1
         },
-        "squad2": {
-          "all_requested_values_present": 0,
-          "exact": 2,
-          "false_abstention": 0,
-          "missed_abstention": 0,
-          "natural_stop": 0,
-          "nonrepetitive_nonempty": 0,
-          "wrong_value_mentioned": 0
-        }
+        "exact": 0.7142857142857143,
+        "f1": 0.8648648648648649,
+        "field_accuracy": 0.8235294117647058,
+        "json_parse": 0.9285714285714286,
+        "natural_stop_rate": 1.0,
+        "precision": 0.8421052631578947,
+        "present_accuracy": 0.8888888888888888,
+        "recall": 0.8888888888888888,
+        "schema_valid": 0.9285714285714286,
+        "unsupported_value_rate": 0.05263157894736842
       }
     },
-    "train_fit": {
-      "correct": 31,
-      "count": 32,
-      "passed": true,
-      "required": 29
-    }
-  },
-  "playground": {
-    "correct": 29,
-    "count": 40,
-    "observed_correct": 23,
-    "observed_count": 28,
-    "authored_correct": 6,
-    "authored_count": 12
-  },
-  "evidence_files": {
-    "full_parent": "../reports/model-design/JSON-090_099_approaches_20261014/release-fulltest-extract-068.json",
-    "full_candidate": "../reports/model-design/JSON-090_099_approaches_20261014/release-fulltest-sft-093.json",
-    "playground": "../reports/model-design/JSON-090_099_approaches_20261014/playground-093-40-results.json"
+    "protocol": {
+      "constrained": false,
+      "json_assembly": false,
+      "max_tokens": 256,
+      "model_calls_per_request": 1,
+      "output_repair": false,
+      "raw_text_preserved": true,
+      "repetition_penalty": 1,
+      "route": "/api/json/generate",
+      "seed": 2026,
+      "temperature": 0,
+      "top_k": 40
+    },
+    "scope": "All actual playground source/schema fixtures, with their displayed schema property ordering and instructions unchanged. Observed original DEV cases, previously authored diagnostics, and observed schema-stress views; local diagnostic, not a fresh blind benchmark. NLL not measured through generation API."
   },
-  "scope": "Observed adapted SNIPS extraction panels and local diagnostic. Not a fresh blind benchmark or official SNIPS intent score. Model plus adapter; current checkpoint fails original retention gates. No unrelated benchmark scores claimed."
+  "historical_protocol_note": "SFT-093 results used labeled fieldwise generation plus JSON assembly; they do not describe the current raw JSON model."
 }

PACKAGING_VERIFICATION.json

@@ -1,36 +1,485 @@
 {
   "model": "OpenSML-JSONLite-175M-Instruct",
   "date": "2026-10-14",
+  "selected_stage": "sft-100",
+  "step": 469,
   "weight_file": "model.safetensors",
   "weight_bytes": 351242342,
-  "weight_sha256": "d552c1380c71c7c0569e9d692ec9c46c5c94bb87a7b024e56df0f0e36df4d798",
-  "runtime": "stock MLX 0.32.3, Apple Silicon GPU, Python 3.11",
+  "weight_sha256": "656f45015aab548465d93bed0b57e687497163741885da7364872851fa17e9a4",
+  "native_bundle_sha256": "88e6a957ec915d23a4be199b8107cd4812577f8f86d446b0e890fef4a3d22da8",
+  "runtime": "stock MLX0.32.3, Apple Silicon, Python3.11",
+  "native_generation_verified": {
+    "constrained_decoding": false,
+    "correct_against_reference": 42,
+    "json_assembly": false,
+    "model": "OpenSML-JSONLite-175M-Instruct",
+    "model_calls_per_request": 1,
+    "output_mode": "native-json",
+    "output_repair": false,
+    "outputs": [
+      {
+        "correct": false,
+        "id": "FIELD-01",
+        "index": 1,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"playlist_owner\":[],\"playlist\":[\"Rock Gaming\"],\"entity_name\":[\"Mariem Hassan\"]}"
+      },
+      {
+        "correct": false,
+        "id": "FIELD-02",
+        "index": 2,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"music_item\":[],\"artist\":[\"Mariem Hassan\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-03",
+        "index": 3,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"playlist_owner\":[\"my\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-04",
+        "index": 4,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"artist\":[],\"entity_name\":[],\"playlist\":[\"Digster Future Hits\"]}"
+      },
+      {
+        "correct": false,
+        "id": "FIELD-05",
+        "index": 5,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"state\":[\"AR\"],\"party_size_number\":[\"nine\"],\"state\":[\"AR\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-06",
+        "index": 6,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"party_size_number\":[\"nine\"],\"timeRange\":[],\"sort\":[]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-07",
+        "index": 7,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"sort\":[\"best\"],\"city\":[\"Wrigley\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-08",
+        "index": 8,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"timeRange\":[],\"poi\":[]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-09",
+        "index": 9,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"condition_temperature\":[\"colder\"],\"state\":[\"Louisiana\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-10",
+        "index": 10,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"current_location\":[],\"geographic_poi\":[],\"condition_temperature\":[\"colder\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-11",
+        "index": 11,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"timeRange\":[\"last August\"],\"geographic_poi\":[\"Stunt Ranch Santa Monica Mountains Reserve\"],\"condition_temperature\":[\"colder\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-12",
+        "index": 12,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"current_location\":[],\"geographic_poi\":[\"Stunt Ranch Santa Monica Mountains Reserve\"],\"state\":[]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-13",
+        "index": 13,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"artist\":[\"Behzad Mirkhani\"],\"music_item\":[\"theme\"],\"sort\":[\"last\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-14",
+        "index": 14,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"playlist\":[],\"music_item\":[\"theme\"],\"service\":[]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-15",
+        "index": 15,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"artist\":[\"Yui\"],\"music_item\":[\"track\"],\"service\":[\"Vimeo\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-16",
+        "index": 16,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"sort\":[],\"playlist\":[]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-17",
+        "index": 17,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"rating_value\":[\"3\"],\"rating_unit\":[\"stars\"],\"object_name\":[\"Things Fall Apart\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-18",
+        "index": 18,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"object_name\":[\"Things Fall Apart\"],\"object_select\":[],\"object_type\":[]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-19",
+        "index": 19,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"rating_value\":[\"1\"],\"object_name\":[]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-20",
+        "index": 20,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"object_part_of_series_type\":[],\"object_name\":[]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-21",
+        "index": 21,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"object_name\":[\"Happiness Was Free\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-22",
+        "index": 22,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"object_type\":[],\"object_name\":[\"Happiness Was Free\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-23",
+        "index": 23,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"object_name\":[\"Martin Morning\"],\"object_type\":[\"saga\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-24",
+        "index": 24,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"object_name\":[\"Martin Morning\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-25",
+        "index": 25,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"movie_type\":[\"animated movies\"],\"spatial_relation\":[\"in the neighborhood\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-26",
+        "index": 26,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"movie_name\":[],\"timeRange\":[],\"movie_type\":[\"animated movies\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-27",
+        "index": 27,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"timeRange\":[\"twenty two o'clock\"],\"movie_name\":[\"Mabel\u2019s Dramatic Career\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-28",
+        "index": 28,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"timeRange\":[\"twenty two o'clock\"],\"spatial_relation\":[],\"object_location_type\":[]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-29",
+        "index": 29,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"album\":[],\"artist\":[\"Mira Sol\"],\"track\":[\"Blue Lantern\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-30",
+        "index": 30,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"restaurant_name\":[\"River Moon\"],\"city\":[\"Denver\"],\"party_size_number\":[\"four\"],\"timeRange\":[\"tomorrow\"]}"
+      },
+      {
+        "correct": false,
+        "id": "FIELD-31",
+        "index": 31,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"country\":[],\"timeRange\":[\"Friday\"],\"city\":[\"Odense\"]}"
+      },
+      {
+        "correct": false,
+        "id": "FIELD-32",
+        "index": 32,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"object_name\":[\"Glass Harbor\"],\"rating_value\":[\"four\"],\"best_rating\":[],\"rating_unit\":[\"stars\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-33",
+        "index": 33,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"timeRange\":[],\"location_name\":[\"Maple Cinema\"],\"movie_name\":[\"Silver Orchard\"]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-34",
+        "index": 34,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"entity_name\":[\"Night Steps\"],\"playlist\":[\"Evening Walks\"]}"
+      },
+      {
+        "correct": false,
+        "id": "FIELD-35",
+        "index": 35,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"artist\":[\"Mira Sol and Theo Lane\"],\"album\":[]}"
+      },
+      {
+        "correct": false,
+        "id": "FIELD-36",
+        "index": 36,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"city\":[\"Bergen\"],\"country\":[]}"
+      },
+      {
+        "correct": false,
+        "id": "FIELD-37",
+        "index": 37,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"cuisine\":[],\"city\":[\"Boston\"],\"restaurant_name\":[]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-38",
+        "index": 38,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"album\":[\"Quiet Signals\"],\"artist\":[\"Mira Sol\"],\"track\":[]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-39",
+        "index": 39,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"object_name\":[\"Paper Bridge\"],\"rating_value\":[]}"
+      },
+      {
+        "correct": true,
+        "id": "FIELD-40",
+        "index": 40,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"country\":[],\"timeRange\":[],\"artist\":[]}"
+      },
+      {
+        "correct": true,
+        "id": "native:snips:AddToPlaylist:train:80:view1:native:alias",
+        "index": 41,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"selected_playlist\":[\"night vibes\"],\"selected_artist\":[],\"selected_playlist_owner\":[]}"
+      },
+      {
+        "correct": false,
+        "id": "native:snips:AddToPlaylist:train:80:view1:native:nullable",
+        "index": 42,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"playlist_owner\":null,\"artist\":\"gn\",\"playlist\":night vibes\"}"
+      },
+      {
+        "correct": false,
+        "id": "native:snips:BookRestaurant:train:684:view1:native:alias",
+        "index": 43,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"selected_restaurant_name\":[\"mughlai restaurant\"],\"selected_timeRange\":[\"Nov. the 1st\"],\"selected_sort\":[]}"
+      },
+      {
+        "correct": false,
+        "id": "native:snips:BookRestaurant:train:684:view1:native:nullable",
+        "index": 44,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"restaurant_name\":\"Mughlai Restaurant\",\"timeRange\":\"Nov. the 1st\",\"sort\":null}"
+      },
+      {
+        "correct": true,
+        "id": "native:snips:GetWeather:train:109:view0:native:alias",
+        "index": 45,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"selected_condition_temperature\":[\"colder\"],\"selected_state\":[\"Louisiana\"]}"
+      },
+      {
+        "correct": true,
+        "id": "native:snips:GetWeather:train:109:view0:native:nullable",
+        "index": 46,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"state\":\"Louisiana\",\"condition_temperature\":\"colder\"}"
+      },
+      {
+        "correct": true,
+        "id": "native:snips:PlayMusic:train:801:view1:native:alias",
+        "index": 47,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"selected_playlist\":[],\"selected_music_item\":[\"theme\"],\"selected_service\":[]}"
+      },
+      {
+        "correct": true,
+        "id": "native:snips:PlayMusic:train:801:view1:native:nullable",
+        "index": 48,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"playlist\":null,\"music_item\":\"theme\",\"service\":null}"
+      },
+      {
+        "correct": true,
+        "id": "native:snips:RateBook:train:293:view0:native:alias",
+        "index": 49,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"selected_rating_value\":[\"2\"],\"selected_object_select\":[\"current\"]}"
+      },
+      {
+        "correct": true,
+        "id": "native:snips:RateBook:train:293:view0:native:nullable",
+        "index": 50,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"object_select\":\"current\",\"rating_value\":\"2\"}"
+      },
+      {
+        "correct": true,
+        "id": "native:snips:SearchCreativeWork:train:602:view1:native:alias",
+        "index": 51,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"selected_object_name\":[\"Playlist: The Very Best of Dolly Parton\"]}"
+      },
+      {
+        "correct": true,
+        "id": "native:snips:SearchCreativeWork:train:602:view1:native:nullable",
+        "index": 52,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"object_name\":\"Playlist: The Very Best of Dolly Parton\"}"
+      },
+      {
+        "correct": true,
+        "id": "native:snips:SearchScreeningEvent:train:897:view1:native:alias",
+        "index": 53,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"selected_movie_type\":[\"animated movies\"],\"selected_movie_name\":[],\"selected_location_name\":[]}"
+      },
+      {
+        "correct": false,
+        "id": "native:snips:SearchScreeningEvent:train:897:view1:native:nullable",
+        "index": 54,
+        "raw_parity": true,
+        "stop_reason": "turn_end",
+        "text": "{\"movie_name\":null,\"movie_type\":\"movies\",\"location_name\":null}"
+      }
+    ],
+    "raw_text_and_stop_matches": 54,
+    "requests": 54,
+    "scope": "Clean stock-MLX packaging parity on observed playground fixtures; not full-heldout runtime parity or fresh blind accuracy.",
+    "step": 469,
+    "versions": {
+      "jsonschema": "4.26.0",
+      "mlx": "0.32.3",
+      "numpy": "2.4.2",
+      "tokenizers": "0.22.2"
+    },
+    "weight_sha256": "656f45015aab548465d93bed0b57e687497163741885da7364872851fa17e9a4"
+  },
   "root_runner_verification": {
-    "command": "python inference.py --example 29",
+    "command": "python inference.py --example 29 --json",
     "default_weights": true,
     "source_bundle_integrity": true,
     "frozen_tokenizer_integrity": true,
-    "schema_valid": true,
-    "natural_stop": true,
     "exact": true,
+    "raw_text": "{\"album\":[],\"artist\":[\"Mira Sol\"],\"track\":[\"Blue Lantern\"]}",
     "value": {
-      "track": [
-        "Blue Lantern"
-      ],
+      "album": [],
       "artist": [
         "Mira Sol"
       ],
-      "album": []
-    }
+      "track": [
+        "Blue Lantern"
+      ]
+    },
+    "model_calls": 1
   },
   "tokenizer_root_matches_bundle": true,
-  "prior_native_fixture_parity": {
-    "requests": 40,
-    "value_and_stop_matches": 40,
-    "correct_against_reference": 29
-  },
-  "scope": "Package entry-point validation; no new blind benchmark or full-test cross-runtime parity claim.",
-  "documentation_update": "Owner-provided pretraining loss/validation image added; native model, tokenizer, adapter and weight bytes are unchanged.",
-  "native_bundle_sha256": "5df3ada361b6fe7226170a73cd8d9cf08a70dd3cdee3ea7c78cb751e7922c94c",
-  "source_commit": "c5198ca0fbbb6f1ac10b3685846602226facb21e"
+  "scope": "54-case raw-output clean-runtime parity; not fresh blind or full-heldout cross-runtime validation."
 }

README.md

@@ -25,239 +25,121 @@ datasets:
 
 # OpenSML-JSONLite-175M-Instruct
 
-**Author:** William Zebrowski · **Model:** OpenSML-JSONLite-175M-Instruct
+**Author:** William Zebrowski · **175,611,648 parameters**
 
-[Run locally](RUN_GUIDE.md) · [Technical report](TECHNICAL_REPORT.md) · [Results](JSONLITE_RESULTS.json) · [Checkpoint provenance](checkpoint_provenance.json) · [Native inference verification](INFERENCE_VERIFICATION.json)
+[Run locally](RUN_GUIDE.md) · [Technical report](TECHNICAL_REPORT.md) · [Results](JSONLITE_RESULTS.json) · [Provenance](checkpoint_provenance.json) · [Inference verification](INFERENCE_VERIFICATION.json)
 
-**A 175.6M-parameter English field extractor with a schema-guided JSON adapter.**
+**A small language model that generates JSON directly from source text and a supplied JSON Schema.** It produces the entire object in one native autoregressive request. There is no fieldwise JSON assembler, output repair or constrained decoder in this release. Schema validation happens after generation and cannot establish that extracted values are correct.
 
-JSONLite takes short source text and a supported JSON Schema, extracts the requested values, and returns a structured object. The neural model generates labeled field values; software parses those values, assembles JSON and validates its structure. A request for multiple fields uses one model call per field.
+This update replaces the earlier extraction-plus-assembly weights and runner with a native source/schema → JSON fine-tune. The pretrained ancestor is [OpenSML-Base-175M](https://huggingface.co/wzebrowski/OpenSML-Base-175M); the model adds evidence QA, exact field extraction, focused extraction and a final directly supervised JSON stage. Full lineage and exact artifact hashes are in the report.
 
-The weights start from [OpenSML-Base-175M](https://huggingface.co/wzebrowski/OpenSML-Base-175M), trained from scratch through OpenSML Stage A pretraining, then adapted with grounded question answering, requested-field extraction and focused single-field extraction. **The inherited SFT lineage does not directly supervise raw JSON Schema → JSON targets.** The evaluated text-to-JSON capability is the checkpoint plus the supplied extraction/assembly adapter.
+## Run locally
 
-## Availability and inference
-
-This repository includes the selected **BF16 `model.safetensors`**, frozen tokenizer, configuration, native inference runner, source bundle, examples and verified extraction results. Download the complete repository to run the model locally. The verified platform is Apple Silicon macOS with Python 3.11 and stock MLX 0.32.3.
-
-This is a custom native MLX architecture. Use the supplied runner; it is not a Transformers AutoModel or mlx-lm auto-loading export. The source bundle contains the exact model implementation and schema-guided extraction/JSON adapter. Optimizer state is unnecessary for inference and is not included.
+The repository contains BF16 `model.safetensors`, tokenizer/configuration, native MLX source archive, runnable entry point, examples and verification/results records. This is a custom architecture; use the supplied runner rather than Transformers AutoModel or mlx-lm auto-loading. Verified platform: Apple Silicon macOS, Python 3.11 and stock MLX 0.32.3.
 
 ```bash
 python3 -m venv .venv
 source .venv/bin/activate
 python -m pip install huggingface_hub
-# If access requires authentication: hf auth login
-hf download wzebrowski/OpenSML-JSONLite-175M --local-dir ./OpenSML-JSONLite-175M
+# Authenticate if access requires it: hf auth login
+hf download wzebrowski/OpenSML-JSONLite-175M --local-dir OpenSML-JSONLite-175M
 cd OpenSML-JSONLite-175M
 python -m pip install -r requirements.txt
 python inference.py --example 29
-# Your own source and schema:
 python inference.py --text-file request.txt --schema-file schema.json
+# Optional: raw text plus validation/stop metadata and example correctness
+python inference.py --example 29 --json
 ```
 
-The root runner defaults to `model.safetensors`, verifies and temporarily extracts `native_bundle.zip`, then runs the frozen native implementation. After downloading the artifacts and installing dependencies, inference runs offline. The root `tokenizer.json` and `tokenizer_config.json` mirror the frozen tokenizer in the bundle, which includes its additional integrity metadata. See [RUN_GUIDE.md](RUN_GUIDE.md) for complete instructions and an example schema.
+Default stdout is the model's unmodified generated text. The fixed extraction instruction and native prompt template are supplied by the runner. Example answers are checked afterward and never passed to generation. After downloading and installing dependencies, inference runs offline. The root runner verifies the source archive and selected weight SHA-256. Tokenizer companions are in the bundle; optimizer state is not needed for inference.
 
-The uploaded weight SHA-256 is:
+## Model and inputs
 
-`d552c1380c71c7c0569e9d692ec9c46c5c94bb87a7b024e56df0f0e36df4d798`
+| Setting | Value |
+| --- | --- |
+| Parameters | 175,611,648 |
+| Layers / hidden / FFN width | 24 / 768 / 2,048 |
+| Attention | GQA, 12 query heads / 4 KV heads |
+| Normalization / MLP / positions | RMSNorm + Q/K norm / SwiGLU / RoPE 10,000 |
+| Vocabulary / context | Frozen 32,000-token byte-BPE / 2,048 shared tokens |
+| Embeddings / biases | Tied input-output / no linear biases |
+| Weights | BF16, unquantized safetensors |
+| Output | One native JSON object per request |
+| Default decoding | Temperature 0, repetition penalty 1, max 256 output tokens |
 
-The CLI verifies this weight identity and the frozen tokenizer, then generates using greedy BF16 native MLX. It returns JSON, intermediate labeled outputs, natural-stop status and per-example exact equality if an example reference is available. References are scored after inference and never passed to the model. Custom requests use `--text-file request.txt --schema-file schema.json`. Source plus generation budget must fit 2,048 tokens per field call; overflowing input is rejected.
+Supply source text containing the facts and a closed flat object schema with 1–8 required properties and `additionalProperties: false`. Supported property types are strings, nullable strings (`["string","null"]`) and arrays of strings. Clear field descriptions help specify semantics. Missing arrays target `[]`; missing nullable strings target `null`. Strings preserve source case/spelling and arrays preserve source order; number/date normalization is not trained. Overlong inputs reject without truncation.
 
-The portable runner was tested in a clean temporary environment with **stock MLX 0.32.3**, NumPy 2.4.2, Tokenizers 0.22.2 and jsonschema 4.26.0. All **40 saved playground values and stop reasons matched**; correctness against references remains **29/40**. This is packaging parity on those fixtures, not a full-suite runtime-parity or new blind accuracy result. Other platforms, framework conversions and quantized formats remain unverified.
+The run is trained primarily on seven SNIPS domains and 39 field definitions. New names/descriptions are passed to the model directly, but reliable unfamiliar-domain extraction has not been established. Nested schemas, numeric/boolean outputs, enums, references and advanced constraints are outside the verified runner contract.
 
-## Model specification
+Example source: `Play the track Blue Lantern by Mira Sol.` With a schema requesting track, artist and album as string arrays, actual native output is:
 
-| Field | Value |
-| --- | --- |
-| Parameters | **175,611,648 (175.61M)** |
-| Layers / hidden width / FFN width | 24 / 768 / 2,048 |
-| Attention | Grouped-query attention: 12 query heads / 4 KV heads, head dimension 64 |
-| Position encoding | RoPE, base 10,000 |
-| Normalization / MLP | RMSNorm, Q/K normalization, SwiGLU |
-| Vocabulary / context | 32,000 / 2,048 per field call |
-| Embeddings | Tied input and output |
-| Linear biases | Disabled |
-| Compute weights | BF16; no quantization |
-| Training optimizer precision | FP32 master weights, moments and gradient accumulation |
-| SFT serialization | Native turn-start / role / turn-end template |
-| Adapter output | Closed JSON object; required fields contain arrays of strings |
-| Default decoding | Temperature 0, repetition penalty 1, at most 160 generated tokens per field |
-
-The frozen byte-level BPE tokenizer was fitted for the earlier OpenSML project and reused unchanged. Earlier OpenSML-150M model weights were **not** used to initialize this model. Tokenizer artifacts, structural IDs and their integrity checks are in the bundle and report.
-
-## Training and model lineage
-
-```text
-Stage A pretrained/132005 → SFT-065/469 → SFT-068/379 → SFT-093/160
+```json
+{"track":["Blue Lantern"],"artist":["Mira Sol"],"album":[]}
 ```
 
-Only those selected stages are inherited. The many intervening research trials and failed RL/JSON experiments are not additional ancestors of the final weights.
-
-### 1. General pretraining
+## Training
 
-The inherited Stage A checkpoint had processed **17,300,111,360 lifetime token positions**. It was trained from random initialization with document next-token prediction on five Apple Silicon Macs using Thunderbolt/RDMA. The nominal continuation target was higher; the selected parent is a **17.30B-position artifact**, not a completed 30B model.
+The 175.6M-parameter base was pretrained from random initialization on **17,300,111,360 lifetime token positions**, using a five-Mac Apple Silicon cluster with Thunderbolt/RDMA.
 
-| Source | Configured sampling share |
-| --- | --- |
+| Pretraining source | Intended sampling share |
+| --- | ---: |
 | FineWeb-Edu | 50% |
 | DCLM-Edu | 18% |
 | English FineWiki | 13% |
 | Cosmopedia v2 | 9% |
-| Educational Python text mirror | 10% |
-
-These are intended processed-token sampling shares, not unique-document counts. Selected-base document validation loss was **2.5563359552**, perplexity **12.8885**, on the recorded fixed validation panel. This is document-prediction loss, not extraction accuracy.
+| Educational Python | 10% |
 
-Nominal pretraining peak LR was 3e-4, but the adaptive controller capped the effective LR at **3.221225472e-5** at the selected checkpoint. AdamW used betas (0.9, 0.95), epsilon 1e-8, weight decay 0.1 and gradient clipping 1.0. Exact warmup, batch-transition history, source revisions, filters and controller state are recorded in the technical report.
+These are configured token-sampling weights, not unique-document counts. Selected-base validation loss was 2.5563359552. Pretraining architecture, tokenizer, source revisions, filtering, scheduler and precision are documented in the report.
 
 ### Training run
 
-The 175.6M-parameter base was trained from random initialization using next-token prediction across the educational web, Wikipedia, Cosmopedia and Python mixture. Five Apple Silicon Macs carried out distributed pretraining over Thunderbolt/RDMA. Validation was measured every 50M tokens. The curves show a steep early loss reduction followed by slower improvement; the published base was selected using held-out validation.
-
-![Stage A pretraining training loss, EMA and held-out validation loss](pretraining-training-loss.png)
+![Training loss and held-out validation](pretraining-training-loss.png)
 
 *Training loss and held-out validation during pretraining. Left: raw training loss and exponential moving average. Right: held-out validation loss.*
 
-### 2. SFT-065: answering from supplied evidence
-
-This stage starts directly from Stage A/132005. The prepared corpus has **25,000 grounded QA examples**: 12,000 natural QA, 5,000 contrasting questions, 3,000 answerability/calibration examples, 3,000 multipart questions and 2,000 corrections. Published inputs/answers come from selected SQuAD 2.0 and Dolly closed QA; controlled fictional examples use programmatically checked fact oracles.
-
-The instruction asks the model to answer using only the supplied passage and return exactly `Not stated.` when unsupported. The complete run reached 782 updates, but **only selected update 469 is inherited**: **15,008 presentations / 3,097,331 serialized token positions**. Later updates are not part of JSONLite.
-
-### 3. SFT-068: requested-field extraction
-
-This stage continues from SFT-065/469 with **12,115 distinct curated SNIPS training queries** across playlist additions, restaurant bookings, weather, music playback, book ratings, creative-work searches and screening-event searches.
-
-Each request supplies text, requested field names and definitions. Targets copy original annotated source spans into labeled lines, preserving source spelling and order. An absent field uses `field: NOT_STATED`. Date and number normalization are not performed.
-
-It completes **379 updates**, **12,115 presentations** and **1,842,593 serialized tokens**. The selected checkpoint becomes the immediate parent for SFT-093.
-
-### 4. SFT-093: focused extraction and replay
-
-The final stage uses **2,560 original SNIPS TRAIN queries**: 1,920 accepted by a TRAIN-only verified parent-output bank plus 640 additional training queries. Each contributes an original replay view and a focused single-field view, producing **5,120 presentations**.
-
-Every batch has 16 replay and 16 focused examples. The run completes **160 updates** and **706,399 serialized tokens**. No development, held-out benchmark or authored playground answers are added to training. The single-field target pool contains 1,900 present-field and 660 missing-field views; it is not exactly balanced.
-
-All three SFT stages update the full model on Mac1 with assistant-output-only mean loss and a fresh AdamW optimizer. No RL, LoRA, added extraction head, frozen-layer adaptation or weight merging occurs in this inherited path.
-
-| Setting | SFT-065 | SFT-068 | SFT-093 |
-| --- | --- | --- | --- |
-| Batch / microbatch | 32 / 2 | 32 / 2 | 32 / 2 |
-| Selected update | 469 of 782 | 379 | 160 |
-| Peak / planned final LR | 1e-5 / 2e-6 | 5e-6 / 1e-6 | 5e-6 / 1e-6 |
-| Warmup updates | 20 | 20 | 12 |
-| Weight decay / gradient clip | 0.01 / 1.0 | 0.01 / 1.0 | 0.01 / 1.0 |
-| Decay | Cosine | Cosine | Cosine |
-
-The inherited post-training totals are **1,008 updates**, **32,243 presentations** and **5,646,323 serialized tokens**. These counts include repeated source views, and serialized token accounting includes prompt positions masked out of the loss.
-
-All scores in this card describe the fixed model identified by the weight hash below. Detailed checkpoint history and reproducibility notes are recorded in the technical report and provenance file.
-
-## Supported schema contract
-
-- Root is a closed object with 1–8 properties; all properties must be required.
-- Every property is an array with `items: {"type": "string"}`; absence is `[]`.
-- Canonical fields use the frozen 39-field glossary. Supplied field descriptions must match their trained definitions exactly.
-- An output-key alias is accepted only when its description uniquely matches a known definition.
-- Duplicate semantic field requests, arbitrary new definitions, nested records, scalars, null values, enums, advanced constraints and references are rejected.
-- If `$schema` is supplied, this adapter accepts the standard draft 2020-12 URI.
-- The instruction fixes exact copying, spelling/case and source order. No silent truncation or fuzzy repair is performed.
+### Supervised adaptation
 
-The model receives the field definition and source text, rather than the full JSON Schema. Software owns schema compilation and JSON assembly. A schema-valid object can still omit an entity, attach a value to the wrong field or invent unsupported information.
-
-## Input, schema and actual output
-
-Input:
-
-```text
-Play the track Blue Lantern by Mira Sol.
-```
-
-Schema:
-
-```json
-{
-  "type": "object",
-  "properties": {
-    "track": {"type": "array", "items": {"type": "string"}, "description": "track name"},
-    "artist": {"type": "array", "items": {"type": "string"}, "description": "performer or musical artist"},
-    "album": {"type": "array", "items": {"type": "string"}, "description": "album name"}
-  },
-  "required": ["track", "artist", "album"],
-  "additionalProperties": false
-}
-```
-
-Actual result:
-
-```json
-{"track":["Blue Lantern"],"artist":["Mira Sol"],"album":[]}
-```
-
-The model produced separate labeled `track`, `artist` and `album` answers; the adapter converted `album: NOT_STATED` to `album: []`. This is example 29 in the included fixtures.
-
-## Evaluation
-
-### Complete-request correctness
-
-| Panel | Parent SFT-068/379, same adapter | JSONLite-175M-Instruct |
-| --- | --- | --- |
-| Observed development | 139/196 — 70.92% | **168/196 — 85.71%** |
-| Observed adapted SNIPS held-out | 953/1,388 — 68.66% | **1,132/1,388 — 81.56%** |
-| Local playground diagnostic | — | **29/40 — 72.50%** |
-| Observed development portion of playground | — | 23/28 — 82.14% |
-| Newly authored portion of playground | — | **6/12 — 50.00%** |
-
-![Complete-request exact through the same fieldwise JSON adapter](complete-request-exact.svg)
-
-### Full held-out extraction metrics
-
-| Metric | Parent, same adapter | JSONLite |
+| Stage | What it teaches | Inherited training |
 | --- | --- | --- |
-| Correct fields / 3,347 | 2,796 — 83.54% | **3,047 — 91.04%** |
-| Value precision | 77.05% | **88.64%** |
-| Value recall | **93.61%** | 92.60% |
-| Value F1 | 84.53% | **90.58%** |
-| Correct present fields / 1,879 | **1,759 — 93.61%** | 1,740 — 92.60% |
-| Correct missing fields / 1,468 | 1,037 — 70.64% | **1,307 — 89.03%** |
-| False abstentions | **18** | 69 |
-| Missed abstentions | 431 | **161** |
-| Source-unsupported predicted values | 43 | **5** |
-| Both views correct / 694 pairs | 360 | **498** |
-| Strict parse/schema completion / 1,388 | 1,384 | 1,385 |
-| Natural stops / token-limit failures | 1,388 / 0 | 1,388 / 0 |
-
-The held-out panel has 694 original source texts with two different requested-field views each, across seven SNIPS categories. Whole-request exact improves in every category under the same route, while present-field recall and false abstentions regress. The parent’s older **83.50% joint-labeled** test uses a different protocol and is not this table’s fieldwise JSON baseline.
-
-The model fails the original strict per-category retention gates (`json_content_floor`, `json_vs_start`, `labeled_retention`, `qa_retention`). It is preserved at the owner's request as an experimental candidate; its automated release-ready flag remains false. Aggregate improvements do not erase these failures.
-
-<details>
-<summary>Protocol, metrics and evaluation independence</summary>
+| Grounded QA | Answers from evidence, absence, contrasting questions, multipart/corrections | 15,008 presentations through selected update 469 |
+| Requested-field extraction | Exact SNIPS annotated spans, field definitions, missing values | 12,115 requests / 379 updates |
+| Focused extraction | Single-field supervision plus labeled replay | 5,120 presentations / 160 updates |
+| **Native JSON SFT** | **Source text + full schema → JSON assistant targets** | **15,000 presentations / 469 updates** |
 
-Generation is greedy with temperature 0, repetition penalty 1 and a 160-token cap per field using native SFT serialization. Prompt overflow is rejected; no truncation, gold-derived filling, source filtering or JSON repair occurs. Expected answers remain outside generation.
+The final JSON stage has **12,115 distinct source requests**: 12,115 canonical-schema examples, 1,443 renamed-key views and 1,442 nullable-string views. Additional views reuse TRAIN-only sources. Original SNIPS chunk annotations are reconstructed, checked against source offsets and schema-validated; upstream annotation fidelity is not a guarantee of perfect semantic golds.
 
-Complete exact requires natural completion, schema-valid output and equality of every expected field list. Object-key order is ignored; array order, duplicate values, spelling and case matter. Value F1 counts `(field,value)` multisets. Missing-field accuracy counts exactly empty annotated fields. Source-unsupported values use a literal substring check; it does not detect every wrong association or boundary error.
+The final stage runs one full-weight pass on Mac1 with batch 32, microbatch 2, fresh AdamW, peak LR 5e-6/final 1e-6, 20-update warmup then cosine decay, betas (0.9,0.95), epsilon 1e-8, weight decay 0.01 and clipping 1. Only assistant JSON and stop tokens contribute to the loss. It processes 4,634,126 serialized positions; there is no RL or adapter-only training in this chain.
 
-Development and the adapted SNIPS held-out panel have prior research exposure. They are not a fresh blind benchmark or the official SNIPS intent-classification task. The 12 authored requests are a small diagnostic. Structural validity and teacher-forced NLL do not establish correct extraction. General ARC/PIQA/HellaSwag, IFEval or coding scores from other OpenSML checkpoints are not attributed to these weights.
+DEV selection happens at updates 0/100/200/300/469 on 252 fixed requests. The final artifact is selected before the 1,444-request test. TRAIN/DEV/TEST source groups are separate, but these held-out panels have prior research exposure and are not a new blind benchmark.
 
-</details>
+## Native JSON evaluation
 
-## Intended use and limitations
+| Metric | Adapted SNIPS held-out, 1,444 requests |
+| --- | ---: |
+| Entire request correct | **1,191/1,444 — 82.48%** |
+| Field accuracy | **91.31%** |
+| Extracted-value precision / recall / F1 | **90.55% / 90.41% / 90.48%** |
+| Missing-field accuracy | **92.46%** |
+| Schema-valid JSON | **99.79%** |
+| False abstentions / unsupported values | 111 / 30 |
+| Token-limited outputs | 0 |
 
-Intended use is controlled extraction experimentation on short English requests with the documented fields. The current system is a specialist, not an arbitrary-schema or general web-document extractor.
+Canonical string-array schemas alone score **1,156/1,388 (83.29%)**, with 100% schema validity. Extra renamed-key and nullable-string views score **19/28** and **16/28**. The full 54-case local playground scores **42/54 (77.78%)**, including **7/12** previously authored natural examples. These are different panels; do not pool or present them as an external leaderboard.
 
-Disclosed failures include omission of a second performer or city, inferred countries not stated in the source, a missing maximum-rating value, missed cuisine and album/track confusion. Additional weaknesses include exact punctuation changes, corrections, current-value reasoning, associations, long paragraphs and unfamiliar field definitions. All 40 fixture golds and actual outputs are documented in the report/bundle.
+All outputs are raw, greedy native model generations: one whole-object request, no constrained decoder, assembly, repair, prompt truncation or reference answer in inputs. Complete correctness requires natural completion, strict JSON/schema validity and exact values. Structural validity is not factual correctness. [Results](JSONLITE_RESULTS.json) and the report contain counts, per-domain performance, protocols and failures.
 
-General dialogue, coding, multilingual quality, calibrated confidence, prompt-injection robustness and a comprehensive safety/privacy audit remain unestablished. `success: true` means natural completion and structural validity; it is not a factual confidence score. Production suitability requires evaluation on the intended inputs and field semantics.
+![Native JSON correctness on separate research panels](complete-request-exact.svg)
 
-## Verification and provenance
+## Limitations
 
-The selected BF16 weight file has **351,242,342 bytes** and the SHA-256 above. Original metadata and sanitized published metadata have separate identities in `checkpoint_provenance.json`. The included snapshot verifier checks shipped code, tokenizer, docs and evidence; its native loader separately verifies the supplied weights.
+The model remains a focused extraction research candidate. Failures include inventing a city, omitting second values, merging two names, missing rating/cuisine information, confusing entity types, incomplete spans, duplicate keys and malformed nullable strings. Broader natural language, corrections, associations, unfamiliar definitions, long documents and complex schemas need further validation. General dialogue, coding, multilingual quality, confidence calibration and prompt-injection robustness are not established. `success` means natural completion and structure validity, not verified factuality or release readiness.
 
-The model card, metrics and technical report describe one fixed checkpoint. They do not include subsequent training. Clean-runtime 40-case parity is documented separately from the historical full-panel results. No automatic model updates, model promotion or hosted deployment is enabled.
+## Identity and packaging
 
-## Licensing and source notices
+Selected weight SHA-256:
 
-Author-controlled source contributions use the included Apache-2.0 license. **The model-weight release license has not yet been assigned in this prepared repository**, and this card does not make a weight-license grant. The YAML intentionally omits a model-weight license identifier until that decision is made.
+`656f45015aab548465d93bed0b57e687497163741885da7364872851fa17e9a4`
 
-Training sources retain their own terms, including SQuAD 2.0 (CC-BY-SA-4.0), Dolly (CC-BY-SA-3.0) and SNIPS (CC0-1.0). Pretraining declarations and the Python mirror audit limitation are detailed in `THIRD_PARTY_NOTICES.md` and the report. A project license does not relicense underlying dataset content or establish blanket redistribution rights.
+Weights have 351,242,342 bytes. The frozen tokenizer is unchanged. Clean stock-MLX packaging verification checks raw output text and stop reasons against all 54 actual playground predictions; its 42/54 correctness is the same observed result, not an independent blind test or full-heldout runtime-parity claim. CUDA, CPU-only, quantized and converted exports remain unverified.
 
-See `SOURCE_NOTICE.md`, `THIRD_PARTY_NOTICES.md`, `CITATION.cff` and the detailed technical report for attribution and exact source revisions.
+Source code uses Apache-2.0; upstream data/dependencies retain their terms. The source license does not automatically assign a model-weight license. See the included notices and provenance.

RUN_GUIDE.md

@@ -1,96 +1,67 @@
-# Run OpenSML-JSONLite-175M-Instruct locally
+# Run OpenSML-JSONLite-175M-Instruct: native JSON
 
-The repository includes the selected BF16 `model.safetensors`, frozen tokenizer,
-configuration, native MLX implementation and schema-to-extraction/JSON adapter.
-No training workspace, optimizer state or external inference service is needed.
-
-The verified runtime is **Apple Silicon macOS with Python 3.11 and MLX 0.32.3**.
-This is a custom native architecture: use the supplied runner, rather than
-`transformers.AutoModel` or `mlx_lm.load`. CUDA, CPU-only and quantized exports
-have not been verified.
-
-## Download and run
+The current weights generate the entire JSON object from source text and the
+supplied schema. No fieldwise assembler, constrained decoder or repair is used.
+The runnable package includes model.safetensors, tokenizer/config, native MLX
+model source, a source archive, 54 examples, results and verification records.
+Verified platform: Apple Silicon macOS, Python 3.11, stock MLX 0.32.3.
 
 ```bash
 python3 -m venv .venv
 source .venv/bin/activate
 python -m pip install huggingface_hub
-# If access requires authentication: hf auth login
+# Authenticate if access requires it: hf auth login
 hf download wzebrowski/OpenSML-JSONLite-175M --local-dir OpenSML-JSONLite-175M
 cd OpenSML-JSONLite-175M
 python -m pip install -r requirements.txt
 python inference.py --example 29
+python inference.py --example 29 --json
+python inference.py --text-file request.txt --schema-file schema.json
 ```
 
-Once downloaded and dependencies installed, generation runs offline. The runner
-defaults to `model.safetensors` beside it, verifies its SHA-256, verifies and
-unpacks the native source bundle into a temporary directory, and loads the
-frozen tokenizer. It cleans up the temporary code after execution.
+Default stdout is the unmodified generated text. `--json` prints a result record
+with raw `text`, parsed `value`, structural success, stop reason, token count and
+example exact correctness when a reference exists. The reference is compared
+only after generation and never enters the model input. Schema validity does
+not certify extracted values. Some included examples intentionally expose
+remaining errors. The current local diagnostic is 42/54, not 100% correctness.
 
-Example 29 source: `Play the track Blue Lantern by Mira Sol.` Its result is:
+Example source: `Play the track Blue Lantern by Mira Sol.`
 
 ```json
-{"track": ["Blue Lantern"], "artist": ["Mira Sol"], "album": []}
-```
-
-The command prints a result record including `value`, `schema_valid`,
-`success`, intermediate extractions and stop reasons. Only example mode adds
-`exact` by comparing the generated value to a caller-side reference. Reference
-answers are never given to the model.
-
-## Your own text and schema
-
-```bash
-python inference.py --text-file request.txt --schema-file schema.json
+{"type":"object","properties":{"track":{"type":"array","items":{"type":"string"},"description":"track name"},"artist":{"type":"array","items":{"type":"string"},"description":"performer or musical artist"},"album":{"type":"array","items":{"type":"string"},"description":"album name"}},"required":["track","artist","album"],"additionalProperties":false}
 ```
 
-`request.txt` contains your source text. `schema.json` requests a closed object
-with one to eight required properties, each an array of strings. For the example:
+Actual native output:
 
 ```json
-{
-  "type": "object",
-  "properties": {
-    "track": {"type": "array", "items": {"type": "string"}, "description": "track name"},
-    "artist": {"type": "array", "items": {"type": "string"}, "description": "performer or musical artist"},
-    "album": {"type": "array", "items": {"type": "string"}, "description": "album name"}
-  },
-  "required": ["track", "artist", "album"],
-  "additionalProperties": false
-}
+{"track":["Blue Lantern"],"artist":["Mira Sol"],"album":[]}
 ```
 
-Only the 39 documented field definitions are supported. Missing values are `[]`.
-The schema adapter calls the model once per field, parses labeled values,
-assembles JSON and validates it. This is not unrestricted arbitrary-schema
-JSON generation. See the model card for supported fields, limits and measured
-accuracy. Structural success does not establish content correctness.
-
-The included source archive contains 40 examples in `examples/cases.json` and
-`inference_native/adapter.py` contains the complete field glossary.
+Schemas must be flat closed objects with 1–8 required properties: strings,
+nullable strings or arrays of strings. The training scope is primarily SNIPS
+fields; unfamiliar field descriptions are accepted but are not reliable by
+construction. Missing arrays use []; missing nullable strings use null. Clear
+descriptions are important. Default decoding is greedy, repetition penalty 1,
+max 256 output tokens. `--max-tokens` supports 1..256. Source + full schema +
+fixed instruction + output budget must fit 2,048 tokens; overflow rejects.
 
-## Integrity and lower-level use
-
-On macOS, verify all downloadable artifacts with:
+The runner supplies the exact fixed instruction and native SFT serialization.
+It checks model and source-archive SHA-256 before loading. Root tokenizer files
+mirror the tokenizer inside the archive, including its integrity companions.
+After downloading and installing dependencies, generation runs offline.
+This custom architecture is not a Transformers or MLX-LM auto-loader export.
 
 ```bash
 shasum -a 256 -c SHA256SUMS
-```
-
-The chosen weight identity is
-`d552c1380c71c7c0569e9d692ec9c46c5c94bb87a7b024e56df0f0e36df4d798`.
-Root tokenizer files are byte-identical to the native bundle's frozen tokenizer.
-`tokenizer_manifest.json` records its native metadata; the bundle also contains
-all companion tokenizer files needed for its integrity checks.
-
-For direct access to the implementation and evaluator:
-
-```bash
 python -m zipfile -e native_bundle.zip native_bundle
 python native_bundle/scripts/verify_snapshot.py
-python native_bundle/inference.py --weights "$PWD/model.safetensors" --example 29
+python native_bundle/inference.py --weights "$PWD/model.safetensors" --example 29 --json
 ```
 
-Generation uses greedy BF16 native MLX, repetition penalty 1, maximum 160 tokens
-per field and a 2,048-token context. Overlong input or unsupported schemas are
-rejected. No silent truncation, fuzzy correction or online API calls occur.
+Selected weights:
+`656f45015aab548465d93bed0b57e687497163741885da7364872851fa17e9a4`.
+The current source module is inference_native/native_json.py. The old labeled
+adapter, if present in the historical source archive, is not used by this
+entry point. Optimizer state and the original training workspace are not needed
+for inference. See the technical report for lineage, metrics and limitations.

SHA256SUMS

@@ -1,19 +1,20 @@
-d0b49f29bcb68dbb1d9967dc026b972ef9fa4a22eb67c9fea88539a90a786d03  CITATION.cff
-82b9908c9fd21fd8c9d10ad809e7c48446485ee98a56bfd9d64bf9fb353761c2  INFERENCE_VERIFICATION.json
-d29cf5952a82db444a6df56446806b1f4e2ccbee1b4aa6b93cb7eeeabcab8536  JSONLITE_RESULTS.json
+983da8cdd3053471db6ffe7abefe1d7ca57403be8369d3f610a18d39be9ac9ec  CITATION.cff
+a84124bd31b3191d5b1b974285361b65ac47659f3ae29adf52694c3e5ccb9a61  INFERENCE_VERIFICATION.json
+d051b0268853461b88a3ba688987e59cfa8daedab167514f39286f7c61076bac  JSONLITE_RESULTS.json
 c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  LICENSE
-d93ce1dcf8bf09dbb15b92b3da659a89a335122f0f151b547534ef05590fb2fd  PACKAGING_VERIFICATION.json
-ba3fab6986fee521149d0ec02e1d37b7a26ceba9b8ff2bd93554d1c0ca8c1ed9  README.md
-994a1798d7b961e34b581bf5337dd5dd14f533e9ebfb4fe1be10cd7b17edf0ec  RUN_GUIDE.md
-5bc419e245a6d8f074186210d9cba59df3e2cc82dfd04452c625268108a6796d  SOURCE_NOTICE.md
-478e1b7ec484be803d3bbfb1fa3b202d89b47889f412bed5c839202b7238ba79  TECHNICAL_REPORT.md
-423f0d1e0e2690505c24e6d31308db728fba60657c5166e177bb269d6a7198d0  THIRD_PARTY_NOTICES.md
-6f8de0b591c8a07885c8adcab359b8ca1cdc81be5f90638cc0ce286f3899c6c5  checkpoint_provenance.json
-fcc46a0f69fc4afb8d6e25361a013e7656ac6758ae5a49cc6b629f95bf728c1a  complete-request-exact.svg
-2f918a15b7b06a792e48d476fa9884dfc7fb46d5ccd0f756bd3890a8ad259d96  config.json
-01d820f55a4dd08b6c9cc940ef84b45d2f80a733ae919876ac97e90ffec0c360  inference.py
-d552c1380c71c7c0569e9d692ec9c46c5c94bb87a7b024e56df0f0e36df4d798  model.safetensors
-5df3ada361b6fe7226170a73cd8d9cf08a70dd3cdee3ea7c78cb751e7922c94c  native_bundle.zip
+5ea861fe99a04e28f7b15838f555382af41fa4ea85a9405b90d58209fe5a73d1  PACKAGING_VERIFICATION.json
+0b7ab57097503cd0b45edd3833bf8c5d0e085c302dcaec24357e64bab0784b37  README.md
+566da4bfc6c074313dfccafd4e0d5670546865d63fffc7b9e25f494624842103  RUN_GUIDE.md
+51746abc1e814b667fb69ee6853add348c88cdcfcd16cc94c110426759694962  SOURCE_NOTICE.md
+fa833ca4380ff47d177578067e00a186c8ce4c2bb17b0a352122ccc5437e16f5  TECHNICAL_REPORT.md
+4e9b202c1a70b476d74a4fe9b9a55e9756195a8b5bd3fc63558875a89487a91d  THIRD_PARTY_NOTICES.md
+5ecf9a8850aef6cc75add6bacda880648e72de1b38089d585e237bbce828c2c3  checkpoint_metadata.json
+99d25e8a8ff34edf9a002e13afe743754cf6aa71d318d86a68239a0efc010f76  checkpoint_provenance.json
+83bd7d84e67ff1177057577de68fb0d3daa9afa6d0eda444b28f60ad3e2c0853  complete-request-exact.svg
+23684c5658428e3b05e3c5cec4c71163e39fee67cba8bd7b4fbde100397b1ebb  config.json
+1f78d86c084f24bbb4233282cf0015a646055103fcbc550ca3daae7fdb23f06d  inference.py
+656f45015aab548465d93bed0b57e687497163741885da7364872851fa17e9a4  model.safetensors
+88e6a957ec915d23a4be199b8107cd4812577f8f86d446b0e890fef4a3d22da8  native_bundle.zip
 641fe860d6fd24547fe4d15bd852190f00ddec31bd74eb7ca9eeae74031c2f0c  pretraining-training-loss.png
 c7b1a3c0efb09ef3b5416a24072e48b25e2cc1f122fb5ed638c0f95b72594af9  requirements.txt
 e095f0413a11fa91fbb58f37d00e4c6bcda59beb89b28f6f680db1436a7e8891  tokenizer.json

SOURCE_NOTICE.md

@@ -1,15 +1,24 @@
 # Source and data notice
 
-OpenSML-JSONLite-175M is extracted from William Zebrowski's OpenSML research workspace. Its inherited weights follow Stage A/132005 → SFT-065/469 → SFT-068/379 → recreated SFT-093/160. Supporting code dependencies and unused representations are not additional weight ancestors.
+Current inherited weights: Stage A/132005 → SFT-065/469 → SFT-068/379 →
+SFT-093/160 → SFT-100/469. The final stage directly supervises native JSON.
+Other research trials and unused helper code are not weight ancestors.
+Author-controlled source uses Apache-2.0. Upstream datasets/dependencies retain
+their notices; no trained-weight license is assigned by this source update.
+GitHub excludes weight and optimizer binaries; the HF companion has the weights.
 
-Author-controlled source contributions use the root Apache-2.0 license. Upstream dependencies, datasets and their underlying content retain their own notices. The Hugging Face package includes the selected BF16 weights. This source notice does not assign a model-weight license.
+Pretraining uses the pinned educational web/wiki/Cosmopedia/Python mixture.
+Grounded QA uses selected SQuAD2/Dolly and checked authored fictional facts.
+SNIPS original CC0 chunk annotations supply exact extraction and native JSON
+labels. The 15,000 curated JSON records are source/schema/target views of12,115
+TRAIN requests, not15,000distinct documents. Development/test are source-disjoint
+from TRAIN within this curated dataset, but have previous research exposure.
+Pretraining decontamination and perfectly correct semantic golds are not claimed.
 
-Pretraining uses pinned FineWeb-Edu, DCLM-Edu, English FineWiki, Cosmopedia v2 and the Python-Edu text mirror. Grounded QA uses original selected SQuAD 2.0/Dolly answers plus programmatically checked fictional examples. Extraction uses original SNIPS chunk labels. See the technical report for exact revisions, sampling, exclusions and selected exposure.
-
-Raw training pools, packed arrays, optimizer states, credentials, SSH keys, private launch/network receipts and unrelated model benchmark scores are excluded. Public checkpoint JSON replaces private workspace paths with `[workspace]`; provenance records both the original and published hashes. Source and tokenizer identifiers remain unchanged for compatibility.
-
-Actual SNIPS predictions and the documented 40 fixtures are included as evaluation evidence under their source notices. The recorded corpus declarations do not establish blanket rights to redistribute every underlying pretraining document or public weight licensing clearance.
-
-See [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md), [technical report](docs/TECHNICAL_REPORT.md), [training records](TRAINING_AND_REPRODUCIBILITY.md) and [CITATION.cff](CITATION.cff).
-
-The adapted 1,388-request SNIPS held-out evaluation panel is included under its CC0 notice in `evaluation/snips-heldout.jsonl`. It remains outside training; requested-field transformations and prior evaluation exposure are disclosed.
+Raw pretraining pools, packed arrays, optimizer state, credentials, SSH keys and
+private host/network receipts are excluded. Published metadata replaces private
+workspace paths with [workspace]; original and published hashes are separated.
+54actual local fixtures, raw predictions and held-out evidence are retained.
+The canonical held-out1,388and schema-stress56requests are included as JSONL in
+evaluation/native-json-heldout.jsonl and the prepared native data directory.
+See THIRD_PARTY_NOTICES.md, docs/TECHNICAL_REPORT.md and provenance/LINEAGE.json.

TECHNICAL_REPORT.md

@@ -1,100 +1,33 @@
-# OpenSML-JSONLite-175M: SFT-093/update 160
+# OpenSML-JSONLite-175M-Instruct: native JSON technical report
 
-Hugging Face companion edition. Archive-relative links in this report resolve inside `native_bundle.zip`. The Hugging Face repository additionally includes the selected `model.safetensors`, root tokenizer files and a runnable entry point; see `RUN_GUIDE.md`.
+**Author:** William Zebrowski · **Parameters:** 175,611,648 · **Date:** 2026-10-14
 
-## Technical report, training provenance, evaluation and public release specification
+## 1. Current artifact and behavior
 
-**Author/project:** William Zebrowski / OpenSML
-**Report date:** October 14, 2026 (America/New_York)
-**Artifact:** recreated SFT-093/update 160, retained at the owner's request
-The Hugging Face package includes the selected BF16 weights and offline native runner. No hosted inference endpoint is deployed.
+The current weights directly generate JSON from source text and a supplied JSON Schema. A single native autoregressive request emits the entire object. There is no fieldwise extraction-to-JSON assembler, answer repair, constrained decoder or gold-conditioned filling in the current inference path. Parsing and schema validation happen only after generation; the raw generated string is retained even when incorrect or invalid.
 
-This report describes one exact model and its accompanying inference pipeline. The model is a 175,611,648-parameter decoder pretrained from random initialization, then adapted through grounded QA, requested-field extraction and focused single-field extraction. Its learned output is a labeled list. Software accepts a supported JSON Schema, turns each requested property into an extraction question, strictly parses generated field lines, assembles JSON and validates its structure.
+Current artifact: SFT-100/update 469, BF16 `model.safetensors`, 351,242,342 bytes.
+SHA-256: `656f45015aab548465d93bed0b57e687497163741885da7364872851fa17e9a4`.
+This is a custom native MLX architecture, not a Transformers/MLX-LM auto-loader export.
 
-The selected model was **not supervised on raw JSON Schema → JSON answer pairs in its inherited SFT lineage**. The user-visible text-to-JSON functionality belongs to the model **plus** the schema compiler, fieldwise prompt renderer, parser and serializer. Shipping or evaluating only the weights does not reproduce the advertised service.
+This report supersedes the earlier fieldwise SFT-093 release. Its historical report is retained as `docs/HISTORICAL_SFT093_REPORT.md` inside the source archive; those adapter-assisted scores are not current native JSON scores. No release-readiness or arbitrary-schema reliability claim is made.
 
-The recreated checkpoint scores **1,132/1,388 (81.56%)** on the complete observed SNIPS held-out serving panel, versus **953/1,388 (68.66%)** for its parent using the same adapter. It scores **168/196 (85.71%)** on the observed development JSON-extraction panel and **29/40 (72.50%)** on the local playground diagnostic. The latter includes **23/28 observed development examples** and **6/12 newly authored requests**. It still fails the project's strict per-category retention gates. This is a focused extraction candidate, not a demonstrated arbitrary-schema, general web-extraction or production-reliability solution.
+## 2. Exact inherited lineage
 
-All numerical claims below are tied to saved metadata or actual generated outputs. Prepared corpus sizes, the examples actually inherited at selected checkpoints, document-token budgets and inference-only transformations are distinguished explicitly.
+`Stage A/132005 → SFT-065/469 → SFT-068/379 → SFT-093/160 → SFT-100/469`
 
-## Contents
+Only those selected checkpoints are inherited. Other trials and failed RL/JSON experiments are research history, not hidden additional ancestors. The architecture and frozen tokenizer remain unchanged. Each SFT stage resets its local counters and AdamW optimizer while continuing the inherited neural weights.
 
-1. Artifact identity and release description
-2. Exact inherited training lineage
-3. Architecture and parameterization
-4. Tokenizer and serialization
-5. Pretraining corpus and curation
-6. Pretraining execution, optimization and validation
-7. Common supervised objective and precision
-8. SFT-065: grounded QA
-9. SFT-068: original SNIPS extraction
-10. SFT-093: focused extraction and replay
-11. Schema-to-JSON inference contract
-12. Worked request and API example
-13. Evaluation definitions and panels
-14. Development, full held-out and playground results
-15. Retention failures, limitations and intended use
-16. Runtime measurements and dependencies
-17. Reproduction and release packaging
-18. Dataset notices and release status
-19. Artifact/evidence index
-20. Appendices: complete field glossary, playground results and metric audit
+| Stage | Inherited update | Presentations through selected artifact | Serialized positions |
+| --- | ---: | ---: | ---: |
+| Stage A pretrained | 132005 | Document stream | 17,300,111,360 lifetime pretraining positions |
+| SFT-065 grounded QA | 469 of 782 | 15,008 | 3,097,331 |
+| SFT-068 requested-field extraction | 379 | 12,115 | 1,842,593 |
+| SFT-093 focused extraction/replay | 160 | 5,120 | 706,399 |
+| SFT-100 native JSON | 469 | 15,000 | 4,634,126 |
+| Inherited post-training total | 1,477 updates | 47,243 | 10,280,449 |
 
-## 1. Artifact identity and release description
-
-Public source repository: **OpenSML-JSONLite-175M**, https://github.com/williamzebrowskI/OpenSML-JSONLite-175M. The preserved artifact is recreated SFT-093/update 160. This artifact is distinct from OpenSML-150M, the earlier base-only OpenSML-175M documentation and the deleted original SFT-093 trial.
-
-A suitable public description is:
-
-> An English-focused 175.6M-parameter MLX extractor adapted from the OpenSML Stage A base. It extracts exact values for documented field definitions. The included adapter produces schema-validated JSON objects containing arrays of strings, with empty arrays for missing values. It is an experimental extraction pipeline with published failure cases and limited natural-language transfer.
-
-The code's JSON model label selects the playground form; it does not establish direct native JSON supervision. The model retains a `native-sft-chat-v1` user/assistant serialization, despite the deployed task being independent extraction requests rather than conversational chat.
-
-| Identity | Verified value |
-| --- | --- |
-| Checkpoint bundle | step_0000160_82ffab052730 |
-| Local run | pretraining/runs/sft_093_json_approaches_20261014 |
-| Local update | 160 |
-| Parameters | 175,611,648 |
-| Weight representation | BF16 safetensors; no quantization in this artifact |
-| Weight file bytes | 351242342 |
-| Weight SHA-256 | d552c1380c71c7c0569e9d692ec9c46c5c94bb87a7b024e56df0f0e36df4d798 |
-| Training metadata SHA-256 | 32a380899c455449db066ef317de9dd8e80f6977bfb42a6fe03b67c5f23963f8 |
-| User-selected preservation | true |
-| Automated retention qualification | false |
-| Release-ready flag in receipt | false |
-| Original deleted diagnostic weight SHA-256 | a001a44290989d65430496617f66abc7f5f3886f26230d7928496f43802ce3a1 |
-| Recreation byte-identical to original | false |
-| Original diagnostic JSON dev exact | 167/196 |
-| Recreated artifact JSON dev exact | 168/196 |
-
-The original trial's weights were deleted under the failed-gate cleanup policy. The owner subsequently requested recreation and preservation. The recreation used a byte-matching training contract, source rows, order and hyperparameters, but produced different weight bytes and slightly different answers. Exact GPU numerical reproducibility is therefore **not established**. The cause of the difference has not been isolated; this report does not assert that a particular kernel caused it.
-
-A public weight release must pin the recreated hash above and use its own scores, rather than attribute the original checkpoint's metrics to it. Optimizer states exist for local continuation but are not required for inference.
-
-## 2. Exact inherited training lineage
-
-```mermaid
-flowchart LR
-    A[Stage A base: update 132005 / 17.30B tokens] --> B[SFT-065: selected update 469 / grounded QA]
-    B --> C[SFT-068: update 379 / requested-field extraction]
-    C --> D[SFT-093: recreated update 160 / focused extraction]
-```
-
-Plain-text lineage: `Stage A/132005 → SFT-065/469 → SFT-068/379 → recreated SFT-093/160`.
-
-Only these three SFT stages are inherited. SFT-064 provided a predecessor **data corpus**, not the model weights used to initialize SFT-065. The many failed JSON, QA, extraction and RL experiments are research history, not additional stages embedded in this checkpoint. In particular, SFT-074/077 and trials 080–092 do not initialize SFT-093. SFT-093 starts directly from SFT-068/379.
-
-| Stage | Selected update | Inherited presentations / positions | Inherited serialized tokens | Weight SHA-256 |
-| --- | --- | --- | --- | --- |
-| Stage A base | 132005 | 17,300,111,360 processed token positions | 17,300,111,360 | 5151567e266244c9e32fd6e01775192633abe227320269ecdefe28ce43e9ccc3 |
-| SFT-065 | 469 | 15,008 QA presentations | 3,097,331 | 0991b805a878c15733c909ef7f4127baa2b8866f8ca7d7db232ad39434f80bc4 |
-| SFT-068 | 379 | 12,115 extraction presentations | 1,842,593 | 18f0f2b8f57905ac758f392d0eb83e51e508da6a955b9f7991338e690205df86 |
-| SFT-093 recreation | 160 | 5,120 presentations / 2,560 source queries | 706,399 | d552c1380c71c7c0569e9d692ec9c46c5c94bb87a7b024e56df0f0e36df4d798 |
-
-The inherited post-training path contains **1,008 optimizer updates**, **32,243 training presentations** and **5,646,323 serialized token positions**. These are not counts of distinct passages: QA contrasts can share a passage, and SFT-093 reuses source queries previously seen by SFT-068. Post-training token accounting includes prompt/header/context positions even though the loss supervises assistant outputs only.
-
-Each SFT stage resets its optimizer and local update/token counters. Model weights continue through the lineage. A public consumer must not mistake `step:160` or `tokens:706399` in the final metadata for the entire lifetime training budget.
+Presentation counts include repeated source views. Serialized positions include masked prompt tokens; they are not supervised-output-token counts. The earlier 150M model weights did not initialize this model; the earlier frozen OpenSML tokenizer was reused.
 
 ## 3. Architecture and parameterization
 
@@ -226,844 +159,156 @@ The 175.6M-parameter base was trained from random initialization using next-toke
 
 *Training loss and held-out validation during pretraining. Left: raw training loss and exponential moving average. Right: held-out validation loss.*
 
-## 7. Common supervised objective and precision
 
-All inherited SFT stages run locally on Mac1 and update all 175,611,648 parameters. No LoRA/adapters, added span head, frozen-layer adaptation, reward model, DPO, PPO or GRPO is part of this lineage.
-
-The model computes in BF16. `MasterAdamW` keeps FP32 master parameters and optimizer moments; small optimizer changes are applied to the master copy before recasting compute weights. Parameters with fewer than two dimensions do not receive decoupled weight decay. Gradient contributions are accumulated in FP32, averaged by actual conversation count, checked for finiteness and globally clipped.
-
-For record `i`, the supervised response loss is the mean CE over its assistant/EOS positions:
-
-`L_i = sum_t(mask_it × CE(logits_it, target_it)) / max(1, sum_t mask_it)`
-
-The optimizer minimizes the batch mean of record losses. This gives equal record weight, not equal weight per output token across the whole batch. Loss does not supervise user prompt tokens. Two-record microbatches use padding to multiples of 128 under the recorded policy. Padding labels are ignored.
-
-Each stage uses a fresh optimizer. The low LR and replay reduce adaptation pressure but do not guarantee nonregression. All quality decisions depend on generated outputs, not just teacher-forced loss.
-
-| Setting | SFT-065 | SFT-068 | SFT-093 |
-| --- | --- | --- | --- |
-| Batch / microbatch | 32 / 2 | 32 / 2 | 32 / 2 |
-| Planned updates | 782 | 379 | 160 |
-| Inherited selected update | 469 | 379 | 160 |
-| Peak LR | 1e-5 | 5e-6 | 5e-6 |
-| Final LR at planned end | 2e-6 | 1e-6 | 1e-6 |
-| Warmup updates | 20 | 20 | 12 |
-| Decay | Cosine | Cosine | Cosine |
-| Betas | (0.9,0.95) | (0.9,0.95) | (0.9,0.95) |
-| Epsilon | 1e-8 | 1e-8 | 1e-8 |
-| Weight decay | 0.01 | 0.01 | 0.01 |
-| Gradient clip | 1.0 | 1.0 | 1.0 |
-| Seed | 20261013 | 20261013 | 20261014 |
-| DEV checkpoints | 0,156,313,469,625,782 | 0,100,200,300,379 | 0,80,160 |
-
-The LR shown for SFT-065's planned end applies at 782; selecting 469 inherits a mid-decay checkpoint, not the final 2e-6 endpoint. Script source and metadata encode the exact schedule.
-
-## 8. SFT-065: grounded QA
-
-Parent: Stage A/132005. The prepared 25,000-record corpus contains 17,000 original published QA rows and 8,000 controlled fictional rows. Source totals: 16,624 SQuAD 2.0, 376 Dolly closed QA, 3,000 calibration, 3,000 multipart and 2,000 corrections. Natural and contrasting buckets overlap source type, not record membership.
-
-SQuAD 2.0 supplies passage-grounded answerable/unanswerable questions; Dolly supplies selected closed QA. See the original [SQuAD source](https://rajpurkar.github.io/SQuAD-explorer/) and [Dolly dataset card](https://huggingface.co/datasets/databricks/databricks-dolly-15k). The local SQuAD training snapshot SHA is `68dcfbb971bd3e96d5b46c7177b16c1a4e7d4bdef19fb204502738552dede002`. Dolly revision is `bdd27f4d94b9c1f951818a7da7fd7aeea5dbff1a`, JSONL SHA `2df9083338b4abd6bceb5635764dab5d833b393b55759dffb0959b6fcbf794ec`.
-
-Public answers are retained rather than rewritten by a teacher model. SQuAD offsets are checked against passages. Eligibility limits include 30–450 passage words, packed sequence ≤ 1,024 and answer ≤ 150 tokens. Article/group/document partitioning excludes known cross-split duplicates. Controlled facts and target logic are programmatically reconstructed; this is not independent human validation of all generated phrasing. Controlled frame IDs and source groups are separated across splits, while structural task families remain shared.
-
-The prompt asks for a direct answer using only the passage and exactly `Not stated.` when unsupported. Contrast records ask different questions about the same passage. Calibration, multipart and correction records teach answerability, completeness and revision of facts. No JSON output targets are added.
-
-| Bucket | Prepared records | Actually inherited by 469 |
-| --- | --- | --- |
-| Natural QA | 12000 | 7226 |
-| Contrasting QA | 5000 | 2997 |
-| Calibration / missing evidence | 3000 | 1797 |
-| Multipart | 3000 | 1812 |
-| Corrections | 2000 | 1176 |
-| Total | 25000 | 15008 |
+## 7. Inherited supervised stages and objective
 
-Inherited source counts are 10,008 SQuAD 2.0, 215 Dolly, 1,797 calibration, 1,812 multipart and 1,176 corrections, reconstructed from the saved permutation seed and first 469 batches. The full run finished 782 updates / 25,000 records / 5,161,605 tokens. The selected 469 bundle contains 15,008 records / 3,097,331 tokens; later updates are not inherited.
+SFT-065 starts directly from Stage A/132005. Its prepared 25,000-record grounded-QA corpus contains 12,000 natural SQuAD 2.0/Dolly QA records, 5,000 contrasting questions over shared evidence, 3,000 answerability/missing-information cases, 3,000 multipart questions and 2,000 corrections. Unsupported answers target `Not stated.`. The whole run reached 782 updates; only update 469 and its 15,008 presentations are inherited.
 
-Development: 136 requests; local test: 272. Selected SFT-065/469 test exact 89/272 (32.72%), token F1 55.27%, answerable 71/224 and absent 18/48. These are local curated panels, not official full SQuAD or conversational leaderboards. This stage is a working QA parent, not a claim of strong general chat quality.
+SFT-068 uses 12,115 curated original SNIPS TRAIN requests across seven intent domains. Each supplies requested field names/definitions and source text. Assistant targets are labeled exact source spans, with `NOT_STATED` for absent fields. It completes 379 updates. Original joint-labeled accuracy is a different protocol from native JSON and is not attributed to the current weights.
 
-## 9. SFT-068: original SNIPS extraction
-
-Parent: SFT-065/469. Source: [SNIPS NLU benchmark, June 2017 custom-intent data](https://github.com/sonos/nlu-benchmark/tree/b86ac7f1577868c42158d0dec77db50956046696/2017-06-custom-intent-engines), pinned revision `b86ac7f1577868c42158d0dec77db50956046696`.
-
-Original chunk annotations are reconstructed into exact text spans and requested-field mappings. Raw source counts are 13,784 training and 700 validation queries. Recorded exclusions include 197 duplicate training entries, 5 duplicate test entries, 3 conflicting-text groups, 8 official-test overlaps removed from training and 147 unsupported-text entries. The PlayMusic training source uses Latin-1 decoding; other selected sources use UTF-8, with no silent replacement.
-
-After grouping case/whitespace-equivalent text and excluding conflicting annotations, 10% of remaining full-training groups are reserved. 14 groups/category form the compact 196-request DEV panel, each with two genuinely different requested-field views. Unused reserved groups stay excluded. The original validate files yield 694 test sources × 2 views = 1,388 requests. Final splits have 12,115 distinct TRAIN texts, 98 DEV texts and 694 test texts.
-
-Seven categories: AddToPlaylist, BookRestaurant, GetWeather, PlayMusic, RateBook, SearchCreativeWork, SearchScreeningEvent. Gold values are exact chunks, trimmed only at outer whitespace. Date/number normalization is not performed. Each source gets a varied field subset, field order and missing-field conditions. Multiple values are repeated field lines in source order; absence is `field: NOT_STATED`.
-
-The model sees field names and definitions, not JSON Schema. This is annotation agreement on short requests, not proof that all SNIPS categories/playlist/type conventions perfectly match production semantics. Unannotated slots are treated as absent; missing-annotation agreement is not independent semantic certification.
-
-The single pass completes 379 updates / 12,115 records / 1,842,593 tokens. DEV selection prioritizes category-macro exact extraction, then field accuracy, F1, pairs, unsupported values and NLL. The selected update 379 parent's original **joint labeled protocol** test scores 1,159/1,388 (83.50%) whole-request exact, 3,071/3,347 (91.75%) field accuracy, value F1 90.84%, present 1,705/1,879 (90.74%) and missing 1,366/1,468 (93.05%). This original labeled test result must not be conflated with its different fieldwise JSON-serving baseline later in this report.
-
-## 10. SFT-093: focused extraction and replay
-
-Parent: SFT-068/379. The bank builder checked 2,165 original TRAIN requests with that parent and accepted 1,920 where its generated labeled answer matched the requested gold. Selection is approximately balanced across seven categories. This bank filter uses only TRAIN predictions; it is an easy-example bias, not proof of general correctness. Original source rows are recovered by ID rather than retaining report-added source labels.
-
-An additional 640 SNIPS TRAIN queries outside the bank are selected by a deterministic hash order. These queries are new to the SFT-093 pool, not unseen by SFT-068, which already trained on the full 12,115-query set. No authored playground cases, DEV answers, test answers or earlier JSON outputs are included as training targets.
-
-Each of 2,560 sources provides an unchanged original labeled replay and one focused single-field view, totaling 5,120 presentations. For the focused view, the deterministic selector alternates preference for an absent field when one exists, otherwise choosing a present field; it never invents absence outside the original requested-field mapping. The resulting focused pool contains 1,900 present-field and 660 absent-field requests. There is no claim of an exactly 50/50 presence balance.
-
-Each 160-update batch contains 16 replay + 16 focused records; each presentation is visited once, under a fixed permutation. The same source can consequently appear twice in different views. The final packed corpus maximum sequence length is 230 tokens. Values and span boundaries are checked back against pinned original chunks; target decoding round-trips to exact expected mappings. Input/output masks and source-disjoint TRAIN/DEV/test identities are checked before training.
-
-The data manifest records demonstration IDs for other representation trials, but SFT-093 **does not use demonstrations or few-shot prompts**. Its training contains labeled outputs only. The 160-update recreation processes 5,120 records and 706,399 serialized tokens. Numeric finite-gradient checks and a disposable miniature-model update/checkpoint roundtrip passed before the original loop; these validate mechanics, not quality or cross-run GPU bit reproducibility.
-
-The original SFT-093/160 reached 167/196 but failed retention and was deleted. The user requested this variant again. The recreation matched the original contract but now scores 168/196 and is retained as a user-selected diagnostic. Its eligibility flags remain false; retention criteria were not changed to manufacture a pass.
-
-### 10.1 Exact examples of the supervised representations
-
-The following examples are actual retained training rows. They illustrate learned tasks, not model predictions. The QA example is among the 15,008 presentations inherited through SFT-065/469. The SNIPS replay is an original requested-field view inherited by SFT-068 and replayed in SFT-093; the focused view uses the same source with only one requested field.
-
-**Grounded QA, SFT-065** — row `qa065:controlled:train:calibration:1485:present`:
-
-```text
-USER
-Answer the question using only the passage. Give a direct answer without adding unsupported information. If the passage does not contain the answer, reply exactly: Not stated.
-
-Passage:
-Fictional exhibition entry notes, reference T-C-01485. The exhibition entry for Luca Vale lists artwork: Red Garden; gallery: Ash Gallery. The exhibition entry for Boris Turner lists artwork: Golden Field; gallery: Cedar Gallery. These notes describe separate entries; do not transfer details between people.
-
-Question: What is the gallery for Luca Vale?
-
-ASSISTANT TARGET
-The gallery for Luca Vale is Ash Gallery.
-```
-
-**SNIPS replay, SFT-068 and SFT-093** — row `snips:PlayMusic:train:1295:view0`:
-
-```text
-USER
-Read the request below and copy the values for the listed fields. Copy each value exactly, preserving spelling and case. Do not normalize numbers or dates. Output one field: value line per value, in requested-field order and source order within a field. If the text does not supply a requested field, output field: NOT_STATED. Do not add explanations.
-
-Requested fields:
-- track: track name
-- music_item: word specifying the requested music category, such as artist, track, song, or album
-
-Text:
-play the track Asleep In The Deep
-
-ASSISTANT TARGET
-track: Asleep In The Deep
-music_item: track
-```
+SFT-093 uses 2,560 original SNIPS TRAIN queries (1,920 accepted by a TRAIN-only parent-output verification bank and 640 additional queries). Each contributes a labeled replay view and a focused single-field view, yielding 5,120 presentations. Every batch has 16 replay and 16 focused examples. This stage completes 160 updates. It was not a native JSON target stage; its later software-assisted JSON route is historical.
 
-**Focused extraction, SFT-093** — same row, target view:
+All four SFT stages update full weights on Mac1 with assistant-output-only, conversation-mean cross entropy. User/schema/source tokens and role headers are masked; assistant content and turn/document stops are supervised. BF16 model computation uses FP32 master AdamW weights/moments and FP32 gradient accumulation. No LoRA, frozen-layer training, RL, added entity head or weight merge is part of this lineage.
 
-```text
-USER
-Extract only the requested fields from the text. Copy each value exactly, preserving spelling and case. Do not normalize numbers or dates. Output one field: value line per value, in requested-field order and source order within a field. If the text does not supply a requested field, output field: NOT_STATED. Do not add explanations.
-
-Requested fields:
-- track: track name
-
-Text:
-play the track Asleep In The Deep
-
-ASSISTANT TARGET
-track: Asleep In The Deep
-```
-
-For an absent field, the extraction target uses `field: NOT_STATED`; the adapter later turns that into an empty JSON array. Neither of these extraction targets contains JSON syntax or a JSON Schema. The source span `[15,33)` for the example track is indexed into the original text during corpus verification; those offsets are not supplied as this model's training answer.
-
-## 11. Schema-to-JSON inference contract
-
-```mermaid
-flowchart LR
-    S[Source text + supported JSON Schema] --> C[Validate schema and map properties to known fields]
-    C --> P[One native SFT prompt per field]
-    P --> M[SFT-093 model generates labeled values]
-    M --> R[Strict labeled-output parser]
-    R --> J[Deterministic JSON serializer]
-    J --> V[JSON Schema structural validation]
-```
+| Setting | SFT-065 | SFT-068 | SFT-093 | SFT-100 |
+| --- | --- | --- | --- | --- |
+| Batch / microbatch | 32 / 2 | 32 / 2 | 32 / 2 | 32 / 2 |
+| Planned updates | 782 | 379 | 160 | 469 |
+| Inherited selected update | 469 | 379 | 160 | 469 |
+| Peak / final LR | 1e-5 / 2e-6 | 5e-6 / 1e-6 | 5e-6 / 1e-6 | 5e-6 / 1e-6 |
+| Warmup updates | 20 | 20 | 12 | 20 |
+| Weight decay / clip | 0.01 / 1 | 0.01 / 1 | 0.01 / 1 | 0.01 / 1 |
+| Adam betas / epsilon | (0.9, 0.95) / 1e-8 | Same | Same | Same |
+| LR decay | Cosine | Cosine | Cosine | Cosine |
 
-The serving adapter is [`json_fieldwise_playground.py`](../provenance/LINEAGE.json). The model remains a causal generator; a separate decoder cache is created for each independent field call. No previous field output or chat history is supplied as evidence to the next field. A request for `k` properties makes `k` model calls. The model never receives the fixture reference answer.
+## 8. SFT-100: source text + schema → JSON targets
 
-Supported schema boundary:
+The final one-pass continuation starts from preserved SFT-093/update 160 with a fresh optimizer. Every assistant target is an actual JSON object. No labeled-line replay target is mixed into this stage.
 
-- Object root, 1–8 properties; every property required; `additionalProperties:false`.
-- Each value is an array with exactly `items:{"type":"string"}`. Missing values are `[]`.
-- Known canonical field names use the frozen glossary. An optional supplied description must match that trained definition exactly.
-- A custom output-key alias is supported only when its description uniquely matches one glossary entry. Software maps it to a canonical extraction field and remaps the output key. This is not learned arbitrary-schema interpretation.
-- Different properties cannot request the same semantic definition under multiple keys.
-- Draft 2020-12 is the accepted explicit `$schema`; it may also be omitted. Titles/descriptions are annotations, not factual source evidence.
-- Unsupported scalar/null/nested types, enums, format constraints, min/max constraints, conditional schemas, references and arbitrary new definitions are rejected.
-- The instruction is fixed to the exact-span policy shown in the playground. Changed instructions are rejected rather than silently ignored.
-- Source text must fit each native prompt plus requested output budget in 2,048 tokens. No truncation. API character ceilings do not replace the token-budget check.
+| JSON training view | Records |
+| --- | ---: |
+| Canonical string-array schemas | 12,115 |
+| Renamed property keys with field descriptions | 1,443 |
+| Nullable-string schema views | 1,442 |
+| Total presentations | 15,000 |
+| Distinct source requests | 12,115 |
 
-The canonical instruction is:
+The additional views reuse TRAIN-only sources. Original SNIPS annotation chunks at revision `b86ac7f1577868c42158d0dec77db50956046696` provide exact values and offsets. All targets are reconstructed from those annotations, compared with the source text and validated against the supplied schema. This is annotation fidelity, not a claim that every upstream annotation is independently perfect semantic truth.
 
-```text
-Copy exact values for each requested field, preserving case and source order. Use [] when a field is not stated.
-```
+Schemas are closed objects with all properties required. Arrays contain all annotated source spans in source order, or `[]` when absent. Nullable strings are used only when the gold has at most one value, with `null` for absence. Multi-value fields remain arrays. Spelling and case are copied exactly; numbers/dates are not normalized. TRAIN alias keys use `out_`; DEV/TEST use `selected_`. Gold key maps are curation/evaluation metadata only and never enter inference.
 
-The native field prompt additionally specifies no number/date normalization, one `field: value` line per value, source order and `field: NOT_STATED` for absence. The model receives those names/definitions and source text, not the full JSON Schema document.
+The seven domains are AddToPlaylist, BookRestaurant, GetWeather, PlayMusic, RateBook, SearchCreativeWork and SearchScreeningEvent; there are 39 trained canonical field definitions. TRAIN/DEV/TEST source-text groups do not overlap within this prepared dataset. The frozen development/test panels had previous research exposure; pretraining decontamination and perfect semantic golds are not claimed.
 
-Strict parsing requires a recognized requested field and nonempty value. Missing required fields, extra fields, duplicate empty declarations or an empty/value contradiction fail the representation. There is no fuzzy correction, automatic field filling or dropping of unsupported values. A syntactically valid array can still contain a hallucinated country, wrong entity or incomplete list. Schema validation checks structure, not whether values are grounded.
+Packed TRAIN has 4,634,126 serialized positions, max prompt length 425 and max target length 75. No record is truncated. Scheduled DEV checks at 0, 100, 200, 300 and 469 use 252 requests: 196 canonical, 28 alias and 28 nullable views of 98 source requests. TEST contains 1,444 requests: 1,388 canonical plus 28 alias and 28 nullable views of 694 distinct sources. Packed arrays are not in the source repository; JSONL records and the preparation/masking code are included. The original data/checkpoint manifests record workspace files (including excluded packed arrays/optimizer binaries); use provenance/SNAPSHOT_MANIFEST.json for the actual distributed file inventory.
 
-The literal `NOT_STATED` is reserved by the labeled protocol; it is not a lossless representation of an entity genuinely named that string. Multiline values and unrestricted complex paragraphs are outside the validated target format. Unicode/case/spelling are preserved by the tokenizer and parser, but multilingual extraction has not been evaluated.
+The local preflight checked target/schema/annotation agreement, evidence offsets, tokenization and assistant masks, split separation and frozen hashes. Eleven adversarial scorer/retention tests passed. GPU forward/backward accumulation was checked without optimizer updates or new checkpoints. Batch 32 uses actual microbatch 2, padding each microbatch to a multiple of 128 with no unused 64-slot fake samples. Training completed 469 updates, 15,000 presentations and 4,634,126 positions.
 
-## 12. Worked request and API example
+## 9. Native inference contract
 
-Actual newly authored playground case FIELD-29:
+Input is source text plus the supplied JSON Schema, serialized in a fixed extraction instruction followed by compact `{"source_text": ..., "schema": ...}`. Native role IDs construct a user turn and the assistant-generation header. The entire source/schema prompt reaches the model together. There is no canonical-field compiler that extracts one field at a time.
 
-```text
-Play the track Blue Lantern by Mira Sol.
-```
+The portable runner accepts one to eight required properties in a closed flat object: strings, nullable strings (`["string","null"]`) or arrays with `items: {"type":"string"}`. `$schema`, when present, is draft 2020-12. Nested records, numbers, booleans, enums, references and advanced constraints are outside this verified runner contract. Unknown names/descriptions are passed directly to the neural model, but reliable unfamiliar-domain semantics are not established.
 
-```json
-{
-  "type": "object",
-  "properties": {
-    "track": {"type": "array", "items": {"type": "string"}, "description": "track name"},
-    "artist": {"type": "array", "items": {"type": "string"}, "description": "performer or musical artist"},
-    "album": {"type": "array", "items": {"type": "string"}, "description": "album name"}
-  },
-  "required": ["track", "artist", "album"],
-  "additionalProperties": false
-}
-```
-
-Three separate model outputs are parsed as `track: Blue Lantern`, `artist: Mira Sol`, and `album: NOT_STATED`. The actual final result matches:
+Default generation is greedy BF16 native MLX, temperature 0, repetition penalty 1, max 256 output tokens. Prompt plus output budget must fit 2,048 tokens; overflow rejects without truncation. Stop IDs are turn-end 3 or document-end 1; unexpected structural tokens or token exhaustion fail complete correctness. JSON object key order is irrelevant; array order and exact source strings matter.
 
-```json
-{"track":["Blue Lantern"],"artist":["Mira Sol"],"album":[]}
-```
-
-Historical workspace server command, run from the original research workspace root (not a launch command for this dedicated repository):
+The default CLI writes raw model text. `--json` returns a result record with raw `text`, parsed `value`, schema/stop status and diagnostics. Parsing is strictly post-generation; duplicate keys and nonfinite JSON numbers reject. No generated value is filled, normalized, repaired or reserialized to create the model's raw output. The example reference stays caller-side and is compared only after inference.
 
 ```bash
-PYTHONPATH=pretraining:pretraining/scripts:pretraining/data/json_sft_validator_20261010 \
-  .venv/bin/python -u pretraining/scripts/start_json_playground.py --port 8767
-```
-
-This entry point currently reads local registration/evaluation receipts. It is a verified workspace command, not yet a standalone portable public-release loader. Public packaging must replace workspace-specific receipt resolution with a relative release manifest and package the required implementation.
-
-Example HTTP client against that running local service:
-
-```python
-import json
-import urllib.request
-
-schema = {
-    "type": "object",
-    "properties": {
-        "artist": {"type": "array", "items": {"type": "string"},
-                   "description": "performer or musical artist"},
-        "album": {"type": "array", "items": {"type": "string"},
-                  "description": "album name"},
-    },
-    "required": ["artist", "album"],
-    "additionalProperties": False,
-}
-body = {
-    "model_id": "sft-093",
-    "document": "Play songs by Mira Sol.",
-    "schema_text": json.dumps(schema),
-    "instruction": "Copy exact values for each requested field, preserving case and source order. Use [] when a field is not stated.",
-    "constrained": True,
-    "max_tokens": 160,
-    "temperature": 0,
-    "repetition_penalty": 1,
-    "top_k": 40,
-    "seed": 2026,
-}
-request = urllib.request.Request(
-    "http://127.0.0.1:8767/api/json/generate",
-    data=json.dumps(body).encode(),
-    headers={"Content-Type": "application/json"},
-)
-with urllib.request.urlopen(request, timeout=60) as response:
-    for line in response:
-        if line.startswith(b"data: "):
-            event = json.loads(line[6:])
-            if event["type"] == "error":
-                raise RuntimeError(event["message"])
-            if event["type"] == "done":
-                if not event["success"]:
-                    raise RuntimeError("Extraction did not complete validly")
-                print(event["value"])
+python inference.py --weights model.safetensors --example 29
+python inference.py --weights model.safetensors --text-file request.txt --schema-file schema.json
+python inference.py --weights model.safetensors --example 29 --json
 ```
 
-The endpoint streams server-sent events, not one ordinary JSON HTTP response. `success:true` means schema-valid structure and natural completion; it does not certify extraction correctness. The fieldwise route assembles JSON regardless of the legacy `constrained` flag; it reports `constrained:false` and `output_mode:fieldwise-assembly` because it does not apply a JSON grammar to model tokens. The UI disables that legacy switch for this route and labels assembly accurately.
-
-Sampling defaults match evaluation: temperature 0, repetition penalty 1, maximum 160 tokens per field. Nondefault temperature/penalty/top-k/seed options are honored by native-token sampling; scores in this report do not cover those nondefault settings. Total generated tokens can exceed 160 across multiple field calls.
-
-Pydantic is optional caller-side schema tooling; no Pydantic class is given to the neural network. `model_json_schema()` output works only if it falls within the supported contract. For example, use `extra="forbid"`, required `list[str]` fields and exact glossary descriptions; scalar defaults or additional constraints are unsupported.
-
-## 13. Evaluation definitions and panels
-
-All main extraction scores use generated answers with temperature 0, repetition penalty 1, 160 tokens/call, native serialization and no prompt truncation. Object-key order is ignored for exact correctness; array order, case, spelling, duplicates and required keys are significant. Every call must terminate naturally. No model sees the expected answer at inference.
-
-Metrics:
-
-- **Complete-request exact:** final parsed/schema-valid object equals the entire expected mapping and all field calls finish naturally.
-- **Field accuracy:** count of exactly correct field-value lists divided by requested fields, including missing fields.
-- **Value precision/recall/F1:** multiset counts of `(field,value)` pairs; duplicates affect counts. `precision=TP/(TP+FP)`, `recall=TP/(TP+FN)` and their harmonic mean, F1. Expected-value count can differ from present-field count when a field has multiple values.
-- **Present accuracy:** exactly correct nonempty expected field lists divided by present fields.
-- **Missing accuracy:** exactly correct empty expected field lists divided by annotated absent fields.
-- **False abstention:** a present field predicted as an empty list. **Missed abstention:** an absent field not predicted empty.
-- **Unsupported values:** predicted strings absent anywhere in the source text. This substring test does not detect every boundary error or wrong association and is not a general hallucination-rate measure.
-- **Paired correctness:** both requested-field views of the same source are completely correct. Individual request observations within a pair are correlated.
-- **Format/parse/schema validity, natural stop, token limits, nonempty/nonrepetitive output, requested order:** separate diagnostics. Valid JSON does not establish correct extraction.
-- **Output NLL:** teacher-forced mean response CE. Development NLL is averaged across the fieldwise subrequests. It is representation-dependent and cannot be compared casually with another output format or used as correctness.
-
-Compact DEV coverage is 196 JSON requests from 98 sources/7 categories, 196 original labeled requests, 136 QA controls and 32 in-sample FIT requests. The 32 FIT requests are chosen lexicographically from training IDs and are **all AddToPlaylist**, so they are not a balanced full-training acquisition test. FIT passes at ≥ 29/32 exact and 32/32 schema validity.
-
-The full serving test below contains 1,388 requests / 694 source pairs, using the original SNIPS held-out panel with a schema compiled from each request's fields. These requests have been evaluated in earlier research; neither this nor DEV is a fresh blind benchmark. Results measure annotation agreement in this modified requested-field task, not the official SNIPS intent-classification protocol.
-
-The 40 playground cases contain 28 observed DEV views plus 12 newly authored requests. These are a small diagnostic, not an external certified benchmark. Golds remain in the browser/scorer; API requests include only text, schema, instruction and generation options. All 28 observed live outputs match the recreated saved DEV outputs exactly. Nonempty/nonrepetitive checks are heuristics, not human fluency ratings.
-
-## 14. Verified results
+Example 29 source: `Play the track Blue Lantern by Mira Sol.` The actual current native output is `{"track":["Blue Lantern"],"artist":["Mira Sol"],"album":[]}`. The schema names/describes those fields as arrays of strings. The model generates all keys and values in one request.
 
-### 14.1 Same-route development comparison
+## 10. Evaluation and checkpoint selection
 
-| Metric | Parent SFT-068/379 fieldwise baseline | Recreated SFT-093/160 |
-| --- | --- | --- |
-| Whole request exact | 70.92% | 85.71% |
-| Field accuracy | 85.90% | 93.38% |
-| Value precision | 81.06% | 91.03% |
-| Value recall | 93.88% | 94.96% |
-| Value F1 | 87.00% | 92.96% |
-| Present-field accuracy | 93.86% | 94.95% |
-| Missing-field accuracy | 74.35% | 91.10% |
-| Schema-valid rate | 100.00% | 100.00% |
-| Completely correct requests | 139 | 168 |
-| Correct fields | 402 | 437 |
-| Correct present fields | 260 | 263 |
-| Correct missing fields | 142 | 174 |
-| TP values | 261 | 264 |
-| FP values | 61 | 26 |
-| FN values | 17 | 14 |
-| False abstentions | 4 | 4 |
-| Missed abstentions | 49 | 17 |
-| Unsupported values | 9 | 1 |
-| Predicted values | 322 | 290 |
-| Natural completion | 196 | 196 |
-| Token-limit failures | 0 | 0 |
-| Nonempty/nonrepetitive | 196 | 196 |
-| Requested-field order | 196 | 196 |
-| Output NLL | 0.044543960131912703 | 0.02902049323023008 |
-| Both views correct /98 | 50/98 | 73/98 |
-
-The gain is 29 additional correct requests, or 14.80 percentage points. Much of it is improved missing-field handling: 174/191 correct versus 142/191. DEV errors still include 28 incomplete/incorrect requests, 26 false-positive values, 14 missed values and one source-unsupported value. The original deleted SFT-093 had 167 correct requests, 436 correct fields and 173 correct missing fields; its full metric record is separate from this recreation.
-
-### 14.2 Recreated development category results
-
-| Category | Requests correct | Field accuracy | Precision | Recall | F1 | Present | Missing | NLL |
-| --- | --- | --- | --- | --- | --- | --- | --- | --- |
-| AddToPlaylist | 21/28 | 88.73% | 90.00% | 90.00% | 90.00% | 90.00% | 87.10% | 0.061795 |
-| BookRestaurant | 21/28 | 89.04% | 83.33% | 95.24% | 88.89% | 95.24% | 80.65% | 0.055806 |
-| GetWeather | 22/28 | 90.00% | 86.36% | 92.68% | 89.41% | 92.50% | 86.67% | 0.041639 |
-| PlayMusic | 27/28 | 98.57% | 97.37% | 100.00% | 98.67% | 100.00% | 96.97% | 0.011644 |
-| RateBook | 25/28 | 95.95% | 93.02% | 93.02% | 93.02% | 93.02% | 100.00% | 0.018461 |
-| SearchCreativeWork | 28/28 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 0.002719 |
-| SearchScreeningEvent | 24/28 | 94.37% | 90.48% | 95.00% | 92.68% | 95.00% | 93.55% | 0.011079 |
-
-### 14.3 Full observed held-out serving comparison
-
-| Metric | Parent SFT-068/379, same fieldwise adapter | Recreated SFT-093/160, same fieldwise adapter |
-| --- | --- | --- |
-| Whole request exact | 68.66% | 81.56% |
-| Field accuracy | 83.54% | 91.04% |
-| Value precision | 77.05% | 88.64% |
-| Value recall | 93.61% | 92.60% |
-| Value F1 | 84.53% | 90.58% |
-| Present-field accuracy | 93.61% | 92.60% |
-| Missing-field accuracy | 70.64% | 89.03% |
-| Schema-valid rate | 99.71% | 99.78% |
-| Completely correct requests | 953 | 1132 |
-| Correct fields | 2796 | 3047 |
-| Correct present fields | 1759 | 1740 |
-| Correct missing fields | 1037 | 1307 |
-| TP values | 1759 | 1740 |
-| FP values | 524 | 223 |
-| FN values | 120 | 139 |
-| False abstentions | 18 | 69 |
-| Missed abstentions | 431 | 161 |
-| Unsupported values | 43 | 5 |
-| Predicted values | 2283 | 1963 |
-| Natural completion | 1388 | 1388 |
-| Token-limit failures | 0 | 0 |
-| Nonempty/nonrepetitive | 1388 | 1388 |
-| Requested-field order | 1384 | 1385 |
-| Both views correct / 694 | 360/694 | 498/694 |
-
-| Category | Requests | Parent exact | Recreated exact | Recreated field accuracy | Recreated F1 | Recreated missing accuracy |
-| --- | --- | --- | --- | --- | --- | --- |
-| AddToPlaylist | 198 | 148/198 (74.75%) | 164/198 (82.83%) | 91.54% | 90.81% | 90.43% |
-| BookRestaurant | 200 | 122/200 (61.00%) | 151/200 (75.50%) | 89.87% | 88.34% | 89.75% |
-| GetWeather | 198 | 101/198 (51.01%) | 143/198 (72.22%) | 87.43% | 86.93% | 80.67% |
-| PlayMusic | 200 | 97/200 (48.50%) | 139/200 (69.50%) | 84.17% | 83.33% | 83.33% |
-| RateBook | 196 | 176/196 (89.80%) | 183/196 (93.37%) | 97.41% | 97.23% | 97.03% |
-| SearchCreativeWork | 200 | 189/200 (94.50%) | 192/200 (96.00%) | 97.08% | 96.62% | 100.00% |
-| SearchScreeningEvent | 196 | 120/196 (61.22%) | 160/196 (81.63%) | 92.10% | 90.76% | 91.56% |
-
-The recreated model gains **179 completely correct requests**, or **12.90 percentage points**, through the same fieldwise adapter. Whole-request exact improves in all seven categories. Missing-field correctness rises from 1,037/1,468 to 1,307/1,468, and unsupported predicted values fall from 43 to 5. There is also a measurable tradeoff: present-field correctness falls from 1,759/1,879 to 1,740/1,879; value recall declines from 93.61% to 92.60%, and false abstentions rise from 18 to 69. Aggregate improvement is therefore not equivalent to improvement on every metric.
-
-The comparison ran through the live local SSE endpoint, model by model, with 1,388 outputs each and identical inputs/decoding. Three candidate requests and four parent requests failed strict parsing/schema completion; both models otherwise terminated naturally without token-limit failures. Output NLL was **not measured** in this serving test. The parent's original 83.50% joint-labeled score uses a different protocol. This comparison does not select an update, alter earlier retention gates or remove the failures on newly authored examples.
-
-### 14.4 In-sample fit, original labeled extraction and QA controls
-
-| Panel | Exact | Other metrics |
-| --- | --- | --- |
-| TRAIN FIT 32 | 31/32 | Field 98.75%; F1 99.16%; NLL 0.008541; all 32 schema valid |
-| Original labeled 196 | 164/196 | Field 92.09%; F1 91.98%; present 92.78%; missing 91.10% |
-| QA 136 | 34/136 | Token F1 47.98%; NLL 0.598362; answerable 33/112; missing 1/24 |
-| Paired original labeled | 71/98 | Both requested views correct |
-| Paired QA | 10/48 | Both answers exact; same answer on 7/48 pairs |
-
-| QA category | Exact | Token F1 |
-| --- | --- | --- |
-| calibration | 0/16 | 25.82% |
-| corrections | 0/16 | 55.65% |
-| dolly_closed_qa | 0/8 | 27.98% |
-| multipart | 2/16 | 53.13% |
-| squad2 | 32/80 | 51.84% |
-
-The QA panel has 3 false abstentions, 23 missed abstentions, 4 wrong-value mentions and 8 complete-requested-values passes, versus 10 at baseline. Exact-normalized QA and conservative controlled-equivalence checks are not human semantic evaluations; an unrecognized paraphrase does not prove that every part of an answer is wrong. These panels show weak general QA and do not support marketing this model as a strong chatbot. ARC, PIQA, HellaSwag or IFEval scores from unrelated SFT checkpoints are not attributed to this artifact.
-
-### 14.5 Local playground diagnostic
+Selection was frozen from development before TEST generation. Rank prioritizes domain-macro complete correctness, then overall complete correctness, field accuracy, F1, schema validity, paired correctness, unsupported-value count and NLL. Eligibility requires no loss of complete-request or field correctness on any fixed domain/schema panel relative to the same-protocol raw parent at update 0. This is a limited observed-panel gate, not protection against every regression and not a comparison with the parent's old assembler.
 
-| Subset | Correct | Accuracy |
-| --- | --- | --- |
-| Observed 28 DEV examples | 23/28 | 82.14% |
-| New 12 authored requests | 6/12 | 50.00% |
-| Total | 29/40 | 72.50% |
+Complete correctness requires natural completion, strict JSON parsing, schema validity and exact agreement with every gold value. Fences, trailing prose, duplicate keys, nonfinite numbers, missing/extra keys, wrong types or values, wrong array ordering and exhausted budgets fail. Field accuracy is exact per requested field. Value precision/recall/F1 use `(field,value)` multisets. Unsupported values use literal source-substring checks and do not detect every association or boundary error.
 
-The new cases expose practical failures despite valid JSON:
+| Development update | Complete correct / 252 | Field accuracy | Value F1 | Schema-valid rate |
+| --- | --- | --- | --- | --- |
+| 0 | 0 (0%) | 0% | 0% | 0% |
+| 100 | 88 (34.92%) | 59.11% | 63.64% | 97.22% |
+| 200 | 162 (64.29%) | 80.30% | 81.03% | 99.60% |
+| 300 | 177 (70.24%) | 85.76% | 84.71% | 99.21% |
+| 469 selected | 200 (79.37%) | 90.40% | 89.99% | 99.60% |
 
-- `What is the weather in Oslo on Friday?` produced `country:["Norway"]`, although no country was supplied. The model inferred plausible world knowledge rather than copied evidence.
-- `Play songs by Mira Sol and Theo Lane.` produced only Mira Sol, omitting the second performer.
-- `Give me the weather in Bergen and Oslo.` omitted Oslo and inferred Norway.
-- A book rating of four out of five stars missed the maximum-rating field.
-- A Thai restaurant request missed the cuisine.
-- An album request also filled a missing track with the album title.
+### Selected candidate held-out results: native generation
 
-Reference answers and actual outputs for all 40 cases are listed in Appendix B. Most newly authored requests are still short; these failures do not require long-document complexity. No claim of reliable paragraph extraction, record binding or arbitrary schema inference follows from 85.71% DEV exact.
+| Metric | All 1,444 requests | Canonical 1,388 requests |
+| --- | ---: | ---: |
+| Entire request correct | 1,191 (82.48%) | 1,156 (83.29%) |
+| Field accuracy | 3,193/3,497 (91.31%) | 3,074/3,347 (91.84%) |
+| Value precision | 90.55% | 90.95% |
+| Value recall | 90.41% | 90.90% |
+| Value F1 | 90.48% | 90.92% |
+| Present-field accuracy | 1,782/1,971 (90.41%) | 1,708/1,879 (90.90%) |
+| Missing-field accuracy | 1,411/1,526 (92.46%) | 1,366/1,468 (93.05%) |
+| False abstentions | 111 | 103 |
+| Missed abstentions | 115 | 102 |
+| Unsupported values | 30 | 24 |
+| Strict JSON parse | 1,442 (99.86%) | 1,388 (100%) |
+| Schema-valid output | 1,441 (99.79%) | 1,388 (100%) |
+| Natural stops / token limits | 1,444 / 0 | 1,388 / 0 |
+| Output NLL | 0.028335 | 0.026668 |
 
-## 15. Retention failures, limitations and intended use
+Both canonical requested-field views are correct for 501/694 source pairs (72.19%). Alias stress is 19/28 (67.86%); nullable stress is 16/28 (57.14%). These extra schemas reuse held-out source texts, not additional independent documents.
 
-Predeclared gates compare every category's counts, pair correctness and error ceilings with the parent, requiring improvements not to conceal local regressions. Recreated SFT-093 fails `json_content_floor`, `json_vs_start`, `labeled_retention` and `qa_retention` ; TRAIN FIT passes. The criteria are strict, finite-panel checks, not a statistical guarantee of unseen nonregression.
+| Domain, including stress views | Complete-request exact |
+| --- | ---: |
+| AddToPlaylist | 168/206 (81.55%) |
+| BookRestaurant | 156/208 (75.00%) |
+| GetWeather | 168/206 (81.55%) |
+| PlayMusic | 149/208 (71.63%) |
+| RateBook | 188/204 (92.16%) |
+| SearchCreativeWork | 197/208 (94.71%) |
+| SearchScreeningEvent | 165/204 (80.88%) |
 
-Concrete recreated DEV regressions include a lost present-field/screening-event result, additional weather false-positive value relative to the original labeled parent, playlist/music labeled-extraction losses and a lost multipart QA answer. Complete requested values in QA decline 10→8. Aggregate gains coexist with those category losses. The owner explicitly chose preservation/deployment exploration despite these flags; the report does not rename that decision an automated acceptance.
+The automatic same-protocol raw-parent comparison completes with parent 0/1,444 exact versus candidate 1,191/1,444. The parent had not been trained to generate native JSON from this combined schema/source prompt. Its historical fieldwise-plus-assembler score is a separate protocol: 1,132/1,388 (81.56%), compared with current native canonical 1,156/1,388 (83.29%). Current canonical false abstentions and unsupported values are higher (103 and 24 versus 69 and 5), despite improved complete correctness. Do not claim universal retention or equate these different inference protocols.
 
-Appropriate exploratory use is constrained-domain extraction of named values from short English requests with documented fields, source-preserving strings and an explicit empty-list policy. Consumers should evaluate their own target domain and treat `success:true` as structural completion rather than factual certification.
+### All 54 actual playground examples
 
-Unsupported or unestablished capabilities include arbitrary new field definitions, nested records, typed scalar conversion, date/number normalization, complete multi-entity lists, correction/current-value reasoning, owner/entity associations, long web pages, robust prompt-injection resistance, multilingual quality, perfect source grounding and general conversation. The current instruction renderer is not a demonstrated security boundary. No general human accuracy, safety or privacy audit is claimed.
+The complete live SSE test keeps the displayed schema ordering, instruction, source and settings unchanged. Model outputs are stored and rescored verbatim; reference answers never enter a generation POST.
 
-There is no measured calibration score for extraction confidence. A softmax probability, low NLL or valid JSON must not be presented as a calibrated confidence in correct extraction. Structured success is particularly easy to achieve because software owns the JSON syntax.
+| Metric / group | Result |
+| --- | ---: |
+| Whole request exact | 42/54 (77.78%) |
+| Strict parse / schema validity | 52/54 (96.30%) |
+| Field accuracy | 118/135 (87.41%) |
+| Value precision / recall / F1 | 91.03% / 84.52% / 87.65% |
+| Present / missing-field correctness | 70/82 (85.37%) / 48/53 (90.57%) |
+| False / missed abstentions | 3 / 5 |
+| Unsupported values | 2 |
+| Natural stop / token limits | 54 / 0 |
+| Original observed SNIPS views | 25/28 (89.29%) |
+| Previously authored natural diagnostics | 7/12 (58.33%) |
+| Additional schema stress views | 10/14 (71.43%) |
 
-## 16. Runtime measurements and dependencies
+These are observed adapted SNIPS and local diagnostic panels, not official SNIPS intent classification or a fresh blind general text-to-JSON benchmark. No Paraloq, SOB, JSONSchemaBench, ARC, IFEval or coding score is attributed to these exact weights. General language quality is not established by JSON schema validity. Playground output NLL was not measured through the API.
 
-Training/recreation and the local test used Mac1: Apple M5 Ultra, 96 GB unified memory. The measured 40-case playground diagnostic made 101 model calls. Defaults were greedy decoding, repetition penalty 1 and 160 tokens per field; cases used 1–4 fields. The service's internal generation timer starts after validation and model loading and ends after field generation/assembly. It excludes full HTTP client latency, cold weight loading and separate tokenizer/schema preparation.
+## 11. Failure evidence and intended scope
 
-| Small diagnostic measurement | Value |
-| --- | --- |
-| Minimum generation timer | 33.51 ms |
-| Median generation timer | 83.05 ms |
-| 95th percentile generation timer | 130.49 ms |
-| Maximum generation timer | 147.48 ms |
-| Largest reported MLX peak allocation | 0.4270 GiB |
-| Model calls across 40 requests | 101 |
+The retained live outputs include Oslo changed to unsupported Odense, two performers combined into one string, a second city omitted, a maximum rating omitted, missing Thai cuisine, restaurant type confused with restaurant name, incomplete spans, a duplicate-key object and malformed nullable-string JSON. The model is a focused small-model extraction research candidate; it is not established as a reliable arbitrary-schema or general web-document extractor.
 
-These are small, observed measurements with a warm service. They do not establish a deployment SLA, concurrency throughput, CPU performance, total process RSS or performance on other GPUs. The MLX peak allocation is cumulative since the model loader resets its peak counter; it does not measure all Python, process or system memory. Fieldwise latency grows with the number of requested fields and differs from one joint generation.
+Schema validity does not verify values. Important gaps remain in exact copying, multiple values, boundaries/associations, correction/current-value reasoning, nullable fields, unfamiliar descriptions, long passages and schema complexity. General dialogue, coding, multilingual transfer, calibrated confidence, prompt-injection robustness and comprehensive safety/privacy behavior are unverified. Release readiness remains false in the research record.
 
-Recorded training dependencies:
+## 12. Runtime, packaging and verification
 
-| Package | Recorded version |
-| --- | --- |
-| jsonschema | 4.26.0 |
-| mlx | 0.32.3.dev20260920+7241f12 |
-| numpy | 2.4.2 |
-| tokenizers | 0.22.2 |
-| datasets | 4.5.0 |
-| huggingface-hub | 1.4.0 |
-| fastapi | 0.128.2 |
-| pydantic | 2.12.5 |
-| uvicorn | 0.40.0 |
-
-The MLX version is a custom development build; the initial cluster report records upstream commit `7241f12e631440387a2156ec1018d1c9fe8e56b9` with a native ring correction. This exact wheel is not assumed available on PyPI. The workspace resolves jsonschema 4.26.0 through `pretraining/data/json_sft_validator_20261010` on PYTHONPATH; it is not installed directly in the examined venv site-packages. A portable release must package or pin that validator explicitly.
-
-Custom Metal attention/FFN/CE paths are part of the recorded source. No CUDA, Transformers AutoModel, MLX-LM conversion, ONNX, CPU-only or quantized artifact has been verified for this checkpoint. The dedicated portable native runner has now been tested with stock MLX 0.32.3 in a clean temporary environment: all 40 saved fixture values and stop reasons agree with the original playground. See `../inference_native/INFERENCE_VERIFICATION.json`; full-suite stock-runtime parity is not claimed. Publish those formats only after conversion/parity tests and new resource/correctness measurements.
-
-## 17. Reproduction and release packaging
-
-### 17.1 Reproduction status
-
-Prepared SFT data and model, code and source hashes are preserved in manifests. SFT-093 uses 2,560 fixed query IDs, 5,120 view IDs, seed 20261014, 160 batches of 32 and the recorded optimizer. Its ancestor SNIPS and QA sources, splits and loss masks are separately recorded. Verification compares gold mappings with original source chunks, checks target encode/decode round trips, masks, lengths, split separation, finite gradients and checkpoint mechanics. Serving checks cover strict parsing, rejection of incorrect content as an exact match, and exclusion of reference answers from model inputs.
-
-The one-off workspace recreation entry point is `pretraining/scripts/recreate_sft_093_playground.py`. It requires its output/data directories not to exist and the exact previous contract. **Do not delete the preserved checkpoint to rerun this command.** It is evidence of the executed recipe, not a public one-command training product. Reproduction on a clean machine needs preserved source artifacts, compatible MLX kernels and new isolated output paths. A recreated weight hash may differ, as this report explicitly demonstrates.
-
-Public release packaging should include a frozen training recipe/data-provenance snapshot and audit/output artifacts, while distinguishing inference necessities from optional training data. Local absolute paths and report-backed model registration are not portable defaults.
-
-### 17.2 Inference artifact layout to finalize
-
-```text
-release/
-  README.md                         # public description + quick start
-  TECHNICAL_REPORT.md                # this report
-  model.safetensors                  # recreated BF16 weights, exact SHA
-  config.json                       # architecture + declared serialization
-  tokenizer/
-    tokenizer.json
-    tokenizer_config.json
-    manifest.json
-    corpus.json
-  serving/
-    model implementation and required kernels
-    native serializer/tokenizer loader
-    documented field glossary
-    supported-schema validator/compiler
-    labeled parser + JSON serializer
-    portable inference entry point
-  examples.json                      # schemas + references, never sent as gold
-  evaluation/
-    protocols, complete metrics and prediction records
-  provenance/
-    lineage, source revisions, hashes and training recipes
-  release-manifest.json              # file sizes/hashes/version
-  LICENSE                            # owner-selected license; not assigned here
-  NOTICES                            # source/dependency/dataset notices
-```
-
-This is a packaging specification, not a claim that this standalone directory has already been produced. The public configuration must record tied embeddings, grouped-query attention, RoPE, feed-forward width, normalization, context size, tokenizer compatibility and the fieldwise adapter version. Describe it as an extractor with an adapter; do not promise native arbitrary-schema JSON generation. Sanitized public metadata receives its own hash, while the original training metadata identity remains separately recorded.
+The current portable inference package is tested independently of the training workspace on Apple Silicon macOS, Python 3.11 with stock MLX 0.32.3, NumPy 2.4.2, Tokenizers 0.22.2 and jsonschema 4.26.0. Greedy raw output text and stop reasons are compared with all 54 live playground outputs. The verification artifact records coverage and measured correctness; this is packaging parity, not a fresh blind accuracy result or full 1,444-case cross-runtime parity claim.
 
-For inference, package the 351,242,342-byte BF16 weight file, tokenizer and tested model/serving implementation. The 2,107,399,512-byte optimizer file and training cursor state are unnecessary. Exclude optimizer states, private cluster identities, SSH/network receipts, credentials, host serial numbers and machine-specific paths from public inference packages. Redistribution of training data requires its own source-specific review.
+Hugging Face provides canonical `model.safetensors`, config/tokenizer files, a root runner and `native_bundle.zip`; the archive includes the native model implementation, prompt/validator, tokenizer companions, examples and reproducibility records. GitHub contains source/documentation and provenance, with weights hosted separately. Optimizer binaries, packed NumPy arrays, host network logs and private workspace paths are excluded. Original/published metadata hashes are separated in lineage provenance.
 
-The historical workspace FastAPI service binds to 127.0.0.1:8767. It is a local service. A public endpoint requires an implementation of hosting, resource and concurrency limits, authentication if appropriate, and a policy for logging and retaining submitted data. This report does not change the bind address, publish repositories or expose user inputs.
+## 13. Reproduction and evidence
 
-### 17.3 API semantics to preserve
+The frozen source/schema/JSON records, preparation/masking code, manifests, selected training metadata, DEV evaluations, parent/candidate TEST outputs and all 54 local predictions are included. Training uses the original workspace modules and the preserved SFT-093 parent; reproducing training requires reconstructing the corresponding workspace inputs/checkpoint paths and regenerating a local preflight receipt. The downloadable inference package does not require that workspace or optimizer state. Publishing training source is not a claim that all historic pretraining raw data or discarded optimizer checkpoints are included.
 
-Preserve unsupported-schema rejection, checks that prevent input truncation, the fixed exact-span instruction, the strict labeled parser and completion checks for every field. Keep structural validity separate from value correctness. Never include example `expected` or `target` fields in generation options. Keep the 40-case green/red scorer independent, and score only unchanged fixtures against their reference. Custom inputs should remain unscored unless an external reference is supplied. Preserve cancellation between GPU calls. Report JSON assembly accurately rather than labeling an external serializer as a token-level JSON grammar.
+See `../provenance/LINEAGE.json`, `../results/JSONLITE_RESULTS.json`, `../reports/model-design/SFT-100_native_json_20261014/`, `../pretraining/native_json_100/`, `../examples/cases.json` and `../inference_native/INFERENCE_VERIFICATION.json` in the source archive. `scripts/verify_snapshot.py` checks the source manifest; the HF root runner checks the source archive and weight identities. Root tokenizer artifacts mirror the archive's frozen tokenizer.
 
-## 18. Dataset notices and release status
+## 14. Attribution and licensing
 
-The following are source declarations and provenance records, not a determination that all downstream license obligations have been resolved. Dataset metadata, underlying content rights, dependency licenses and the owner's model-weight license are separate matters.
+Source code carries the repository's Apache-2.0 license; upstream data/dependency notices retain their terms. SNIPS original entity annotations are CC0-1.0. SQuAD/Dolly and the pretraining mixture have separate source notices recorded in `THIRD_PARTY_NOTICES.md`. Source licensing does not automatically assign a trained-weight license; the owner has not assigned one in this update. No blanket commercial-use or redistribution determination is claimed.
 
-| Source | Recorded notice / primary evidence |
-| --- | --- |
-| SmolLM corpus / FineWeb-Edu / Cosmopedia | ODC-BY in pinned card; preserve underlying source notices. [Card](https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus/blob/3ba9d605774198c5868892d7a8deda78031a781f/README.md) |
-| DCLM-Edu | CC-BY-4.0 source declaration. [Pinned card](https://huggingface.co/datasets/HuggingFaceTB/dclm-edu/blob/dbad8ad71224482740cd9c9d353591adbf62fe04/README.md) |
-| FineWiki | Local specification records CC-BY-SA-4.0/GFDL; source-card metadata should be retained verbatim. [Pinned card](https://huggingface.co/datasets/HuggingFaceFW/finewiki/blob/8bd13e72e6a002407649b3e898535f42ceb1aeb9/README.md) |
-| Python mirror / Python-Edu | Mirror metadata ODC-BY; underlying code comes from repositories and The Stack v2/SWH provenance. Pinned mirror card retrieval failed; no blanket code rights claim. [Upstream subset](https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus) |
-| SQuAD 2.0 | CC-BY-SA-4.0 on original source. [Source](https://rajpurkar.github.io/SQuAD-explorer/) |
-| Dolly 15k | CC-BY-SA-3.0 dataset card; selected closed QA only. [Card](https://huggingface.co/datasets/databricks/databricks-dolly-15k) |
-| SNIPS | CC0-1.0 repository license; retain author citation. [Pinned license](https://github.com/sonos/nlu-benchmark/blob/b86ac7f1577868c42158d0dec77db50956046696/LICENSE) |
-| Controlled authored QA | Programmatically authored fictional examples with checked oracles; no independent human certification claim |
-| Weights / current release code | Final release license and public notices not assigned by this report |
 
-The initial base report described an Apache-2.0 code license, but a root `LICENSE` file was not present in the examined working checkout. That historical statement does not assign a final code or weight license here. The public release needs explicit license files and notices. This technical report does not establish blanket commercial-use or redistribution clearance.
-
-Completed work includes preservation of the checkpoint, verification of its lineage and hashes, recording of development, complete observed-test and playground metrics, documentation of the field/schema adapter, and verification of the local API and UI. The original and recreated model identities are recorded separately.
-
-The frozen source package is OpenSML-JSONLite-175M. Its source code uses the included Apache-2.0 license; dependency/source notices, a native inference runner, pinned public runtime dependencies, a 40-case clean-environment parity record and a repository integrity manifest are included. The Hugging Face companion package additionally includes the selected weight artifact and tokenizer; the model-weight license remains unassigned. Broader external evaluation and full-suite parity still need separate verification. Evaluation beyond the reused SNIPS panel and 12 authored cases is still needed to establish transfer to intended production inputs. An experimental release can report these results and limitations; preserving this checkpoint does not override its failed retention flags or prove reliable general extraction.
-
-## 19. Artifact and evidence index
-
-| Evidence | Repository-relative path |
-| --- | --- |
-| Final model / metadata / optimizer | pretraining/runs/sft_093_json_approaches_20261014/step_0000160_82ffab052730/ |
-| Preservation receipt | pretraining/runs/sft_093_json_approaches_20261014/user-selected.json |
-| Recreated DEV 160 outputs | pretraining/runs/sft_093_json_approaches_20261014/evaluations/development_0000160.json |
-| Original deleted diagnostic metrics | reports/model-design/JSON-090_099_approaches_20261014/SFT-093_update160_all_metrics.json |
-| Recreation identity comparison | reports/model-design/JSON-090_099_approaches_20261014/recreated-093-playground.json |
-| Full parent serving test | reports/model-design/JSON-090_099_approaches_20261014/release-fulltest-extract-068.json |
-| Full recreated serving test | reports/model-design/JSON-090_099_approaches_20261014/release-fulltest-sft-093.json |
-| 40-case live results | reports/model-design/JSON-090_099_approaches_20261014/playground-093-40-results.json |
-| 40 fixtures | pretraining/playground/fieldwise-examples.json |
-| SFT-093 data and manifest | pretraining/data/sft_093_json_approaches_20261014/ |
-| SFT-093 training | pretraining/json_approaches_090/{data,evaluate,train,verify}.py |
-| Recreation runner | pretraining/scripts/recreate_sft_093_playground.py |
-| Serving adapter | pretraining/scripts/json_fieldwise_playground.py |
-| API/server | pretraining/scripts/{json_playground,start_json_playground,completion_playground}.py |
-| SFT-065 source | pretraining/qa_065/ |
-| SFT-068 source | pretraining/extract_068/ |
-| Base implementation | pretraining/sml_v1/ |
-| Frozen tokenizer | reference/opensml-150m/tokenizer/bytebpe32k_v1/ |
-| Research history | reports/model-design/JSON-001_text-to-json_training-plan_20261009.md |
-
-The public release should include the tested serving dependencies and source version explicitly. Workspace paths, ignored files and historical reports will not necessarily be present in a Git or Hugging Face checkout. The paths in this index locate audit evidence; they do not assert that those artifacts are already publicly hosted.
-
-## Appendix A. Complete supported canonical field glossary
-
-Exactly 39 canonical definitions are loaded from the training renderer. Aliases require an exact, unique matching description. These labels follow SNIPS domain conventions and do not establish general named-entity recognition coverage.
-
-| Canonical field | Trained description |
-| --- | --- |
-| album | album name |
-| artist | performer or musical artist |
-| best_rating | maximum rating scale |
-| city | city or town |
-| condition_description | weather condition |
-| condition_temperature | temperature description |
-| country | country |
-| cuisine | cuisine style |
-| current_location | expression referring to current location |
-| entity_name | named item to add to a playlist |
-| facility | requested restaurant facility |
-| genre | music genre |
-| geographic_poi | geographic point of interest |
-| location_name | screening location name |
-| movie_name | movie name |
-| movie_type | kind of movie |
-| music_item | word specifying the requested music category, such as artist, track, song, or album |
-| object_location_type | type of screening venue |
-| object_name | name of the work being rated or searched for |
-| object_part_of_series_type | type of book series |
-| object_select | selection expression referring to the work |
-| object_type | kind of work or object |
-| party_size_description | description of who is attending |
-| party_size_number | explicit party size |
-| playlist | playlist name |
-| playlist_owner | person owning the playlist |
-| poi | named nearby point of interest |
-| rating_unit | rating unit |
-| rating_value | given rating |
-| restaurant_name | restaurant name |
-| restaurant_type | kind of restaurant |
-| served_dish | food or dish |
-| service | music service |
-| sort | requested ranking, ordering, or preference |
-| spatial_relation | spatial or proximity expression |
-| state | state or region |
-| timeRange | date, time, or time-range expression |
-| track | track name |
-| year | year expression |
-
-## Appendix B. All 40 playground cases and actual recreated outputs
-
-The first 28 cases are observed SNIPS development views; cases 29–40 were newly authored. Array order and string case are preserved. Reference answers are not sent to model generation. A correct badge checks complete parsed JSON equality, not just schema validity.
-
-| Case | Source text | Expected JSON | Actual JSON | Exact |
-| --- | --- | --- | --- | --- |
-| FIELD-01 | In my playlist Rock Gaming add Mariem Hassan con Leyoad | `{"playlist_owner":["my"],"playlist":["Rock Gaming"],"entity_name":["Mariem Hassan con Leyoad"]}` | `{"playlist_owner":["my"],"playlist":["Rock Gaming"],"entity_name":[]}` | No |
-| FIELD-02 | In my playlist Rock Gaming add Mariem Hassan con Leyoad | `{"artist":[],"music_item":[]}` | `{"artist":["Mariem Hassan"],"music_item":[]}` | No |
-| FIELD-03 | Add track to my Digster Future Hits | `{"playlist_owner":["my"]}` | `{"playlist_owner":["my"]}` | Yes |
-| FIELD-04 | Add track to my Digster Future Hits | `{"artist":[],"entity_name":[],"playlist":["Digster Future Hits"]}` | `{"artist":[],"entity_name":[],"playlist":["Digster Future Hits"]}` | Yes |
-| FIELD-05 | I would like to book a brasserie for nine in AR | `{"restaurant_type":["brasserie"],"party_size_number":["nine"],"state":["AR"]}` | `{"restaurant_type":["brasserie"],"party_size_number":["nine"],"state":["AR"]}` | Yes |
-| FIELD-06 | I would like to book a brasserie for nine in AR | `{"party_size_number":["nine"],"timeRange":[],"sort":[]}` | `{"party_size_number":["nine"],"timeRange":["nine"],"sort":[]}` | No |
-| FIELD-07 | Book a best restaurant in Wrigley. | `{"sort":["best"],"city":["Wrigley"]}` | `{"sort":["best"],"city":["Wrigley"]}` | Yes |
-| FIELD-08 | Book a best restaurant in Wrigley. | `{"timeRange":[],"poi":[]}` | `{"timeRange":[],"poi":[]}` | Yes |
-| FIELD-09 | Is it going to be colder in Louisiana | `{"condition_temperature":["colder"],"state":["Louisiana"]}` | `{"condition_temperature":["colder"],"state":["Louisiana"]}` | Yes |
-| FIELD-10 | Is it going to be colder in Louisiana | `{"current_location":[],"geographic_poi":[],"condition_temperature":["colder"]}` | `{"current_location":[],"geographic_poi":[],"condition_temperature":["colder"]}` | Yes |
-| FIELD-11 | was the weather colder last August at Stunt Ranch Santa Monica Mountains Reserve | `{"timeRange":["last August"],"geographic_poi":["Stunt Ranch Santa Monica Mountains Reserve"],"condition_temperature":["colder"]}` | `{"timeRange":["August"],"geographic_poi":["Stunt Ranch Santa Monica Mountains Reserve"],"condition_temperature":["colder"]}` | No |
-| FIELD-12 | was the weather colder last August at Stunt Ranch Santa Monica Mountains Reserve | `{"current_location":[],"geographic_poi":["Stunt Ranch Santa Monica Mountains Reserve"],"state":[]}` | `{"current_location":[],"geographic_poi":["Stunt Ranch Santa Monica Mountains Reserve"],"state":[]}` | Yes |
-| FIELD-13 | Play the last theme by Behzad Mirkhani. | `{"artist":["Behzad Mirkhani"],"music_item":["theme"],"sort":["last"]}` | `{"artist":["Behzad Mirkhani"],"music_item":["theme"],"sort":["last"]}` | Yes |
-| FIELD-14 | Play the last theme by Behzad Mirkhani. | `{"service":[],"music_item":["theme"],"playlist":[]}` | `{"service":[],"music_item":["theme"],"playlist":[]}` | Yes |
-| FIELD-15 | Play a track by Yui on Vimeo | `{"artist":["Yui"],"music_item":["track"],"service":["Vimeo"]}` | `{"artist":["Yui"],"music_item":["track"],"service":["Vimeo"]}` | Yes |
-| FIELD-16 | Play a track by Yui on Vimeo | `{"playlist":[],"sort":[]}` | `{"playlist":[],"sort":[]}` | Yes |
-| FIELD-17 | I would give Things Fall Apart 3 stars | `{"rating_value":["3"],"rating_unit":["stars"],"object_name":["Things Fall Apart"]}` | `{"rating_value":["3"],"rating_unit":["stars"],"object_name":["Things Fall Apart"]}` | Yes |
-| FIELD-18 | I would give Things Fall Apart 3 stars | `{"object_name":["Things Fall Apart"],"object_select":[],"object_type":[]}` | `{"object_name":["Things Fall Apart"],"object_select":[],"object_type":[]}` | Yes |
-| FIELD-19 | I'd rate this book 1 of 6 stars | `{"rating_value":["1"],"object_name":[]}` | `{"rating_value":["1"],"object_name":[]}` | Yes |
-| FIELD-20 | I'd rate this book 1 of 6 stars | `{"object_part_of_series_type":[],"object_name":[]}` | `{"object_part_of_series_type":[],"object_name":[]}` | Yes |
-| FIELD-21 | Show me Happiness Was Free. | `{"object_name":["Happiness Was Free"]}` | `{"object_name":["Happiness Was Free"]}` | Yes |
-| FIELD-22 | Show me Happiness Was Free. | `{"object_type":[],"object_name":["Happiness Was Free"]}` | `{"object_type":[],"object_name":["Happiness Was Free"]}` | Yes |
-| FIELD-23 | Find me the Martin Morning saga | `{"object_name":["Martin Morning"],"object_type":["saga"]}` | `{"object_name":["Martin Morning"],"object_type":["saga"]}` | Yes |
-| FIELD-24 | Find me the Martin Morning saga | `{"object_name":["Martin Morning"]}` | `{"object_name":["Martin Morning"]}` | Yes |
-| FIELD-25 | find animated movies in the neighborhood | `{"movie_type":["animated movies"],"spatial_relation":["in the neighborhood"]}` | `{"movie_type":["animated movies"],"spatial_relation":["in the neighborhood"]}` | Yes |
-| FIELD-26 | find animated movies in the neighborhood | `{"movie_type":["animated movies"],"timeRange":[],"movie_name":[]}` | `{"movie_type":["animated movies"],"timeRange":[],"movie_name":[]}` | Yes |
-| FIELD-27 | where can I watch Mabel’s Dramatic Career at twenty two o'clock | `{"movie_name":["Mabel’s Dramatic Career"],"timeRange":["twenty two o'clock"]}` | `{"movie_name":["Mabel's Dramatic Career"],"timeRange":["twenty two o'clock"]}` | No |
-| FIELD-28 | where can I watch Mabel’s Dramatic Career at twenty two o'clock | `{"timeRange":["twenty two o'clock"],"spatial_relation":[],"object_location_type":[]}` | `{"timeRange":["twenty two o'clock"],"spatial_relation":[],"object_location_type":[]}` | Yes |
-| FIELD-29 | Play the track Blue Lantern by Mira Sol. | `{"track":["Blue Lantern"],"artist":["Mira Sol"],"album":[]}` | `{"track":["Blue Lantern"],"artist":["Mira Sol"],"album":[]}` | Yes |
-| FIELD-30 | Book River Moon in Denver for four people tomorrow. | `{"restaurant_name":["River Moon"],"city":["Denver"],"party_size_number":["four"],"timeRange":["tomorrow"]}` | `{"restaurant_name":["River Moon"],"city":["Denver"],"party_size_number":["four"],"timeRange":["tomorrow"]}` | Yes |
-| FIELD-31 | What is the weather in Oslo on Friday? | `{"city":["Oslo"],"timeRange":["Friday"],"country":[]}` | `{"city":["Oslo"],"timeRange":["Friday"],"country":["Norway"]}` | No |
-| FIELD-32 | Give the book Glass Harbor four out of five stars. | `{"object_name":["Glass Harbor"],"rating_value":["four"],"best_rating":["five"],"rating_unit":["stars"]}` | `{"object_name":["Glass Harbor"],"rating_value":["four"],"best_rating":[],"rating_unit":["stars"]}` | No |
-| FIELD-33 | Find showtimes for Silver Orchard at Maple Cinema. | `{"movie_name":["Silver Orchard"],"location_name":["Maple Cinema"],"timeRange":[]}` | `{"movie_name":["Silver Orchard"],"location_name":["Maple Cinema"],"timeRange":[]}` | Yes |
-| FIELD-34 | Add Night Steps to my playlist Evening Walks. | `{"entity_name":["Night Steps"],"playlist":["Evening Walks"]}` | `{"entity_name":["Night Steps"],"playlist":["Evening Walks"]}` | Yes |
-| FIELD-35 | Play songs by Mira Sol and Theo Lane. | `{"artist":["Mira Sol","Theo Lane"],"album":[]}` | `{"artist":["Mira Sol"],"album":[]}` | No |
-| FIELD-36 | Give me the weather in Bergen and Oslo. | `{"city":["Bergen","Oslo"],"country":[]}` | `{"city":["Bergen"],"country":["Norway"]}` | No |
-| FIELD-37 | Find a Thai restaurant in Boston. | `{"cuisine":["Thai"],"city":["Boston"],"restaurant_name":[]}` | `{"cuisine":[],"city":["Boston"],"restaurant_name":[]}` | No |
-| FIELD-38 | Play the album Quiet Signals by Mira Sol. | `{"album":["Quiet Signals"],"artist":["Mira Sol"],"track":[]}` | `{"album":["Quiet Signals"],"artist":["Mira Sol"],"track":["Quiet Signals"]}` | No |
-| FIELD-39 | I am looking for the book Paper Bridge. | `{"object_name":["Paper Bridge"],"rating_value":[]}` | `{"object_name":["Paper Bridge"],"rating_value":[]}` | Yes |
-| FIELD-40 | Please play some music. | `{"country":[],"timeRange":[],"artist":[]}` | `{"country":[],"timeRange":[],"artist":[]}` | Yes |
-
-## Appendix C. Recreated all-source metric counts
-
-Complete metric definitions are in Section 13. These are the current artifact’s metrics, not those of the deleted original. The linked JSON evidence retains every category’s raw counts and generated outputs. `count.*` labels are integer counts or, for token F1, the accumulated per-request score; `rate.*` labels are fractions on a 0–1 scale. NLL is a loss, not a correctness rate.
-
-### JSON development panel
-
-| Metric | Value |
-| --- | --- |
-| count.absent_correct | 174 |
-| count.absent_total | 191 |
-| count.exact | 168 |
-| count.false_abstention | 4 |
-| count.field_correct | 437 |
-| count.fields | 468 |
-| count.fn | 14 |
-| count.format_ok | 196 |
-| count.fp | 26 |
-| count.json_parse | 196 |
-| count.missed_abstention | 17 |
-| count.natural_stop | 196 |
-| count.nonrepetitive_nonempty | 196 |
-| count.predicted_values | 290 |
-| count.present_correct | 263 |
-| count.present_total | 277 |
-| count.requested_order | 196 |
-| count.schema_valid | 196 |
-| count.token_limited | 0 |
-| count.tp | 264 |
-| count.unsupported_values | 1 |
-| rate.absent_accuracy | 0.9109947643979057 |
-| Requests | 196 |
-| rate.exact | 0.8571428571428571 |
-| rate.f1 | 0.9295774647887325 |
-| rate.field_accuracy | 0.9337606837606838 |
-| rate.json_parse_rate | 1.0 |
-| Output NLL | 0.02902049323023008 |
-| rate.precision | 0.9103448275862069 |
-| rate.present_accuracy | 0.9494584837545126 |
-| rate.recall | 0.9496402877697842 |
-| rate.schema_valid_rate | 1.0 |
-
-### LABELED development panel
-
-| Metric | Value |
-| --- | --- |
-| count.absent_correct | 174 |
-| count.absent_total | 191 |
-| count.exact | 164 |
-| count.false_abstention | 11 |
-| count.field_correct | 431 |
-| count.fields | 468 |
-| count.fn | 20 |
-| count.format_ok | 196 |
-| count.fp | 25 |
-| count.missed_abstention | 17 |
-| count.natural_stop | 196 |
-| count.nonrepetitive_nonempty | 196 |
-| count.predicted_values | 283 |
-| count.present_correct | 257 |
-| count.present_total | 277 |
-| count.requested_order | 196 |
-| count.token_limited | 0 |
-| count.tp | 258 |
-| count.unsupported_values | 1 |
-| rate.absent_accuracy | 0.9109947643979057 |
-| Requests | 196 |
-| rate.exact | 0.8367346938775511 |
-| rate.f1 | 0.9197860962566845 |
-| rate.field_accuracy | 0.9209401709401709 |
-| Output NLL | 0.029267541360946333 |
-| rate.precision | 0.911660777385159 |
-| rate.present_accuracy | 0.927797833935018 |
-| rate.recall | 0.9280575539568345 |
-
-### FIT development panel
-
-| Metric | Value |
-| --- | --- |
-| count.absent_correct | 20 |
-| count.absent_total | 21 |
-| count.exact | 31 |
-| count.false_abstention | 0 |
-| count.field_correct | 79 |
-| count.fields | 80 |
-| count.fn | 0 |
-| count.format_ok | 32 |
-| count.fp | 1 |
-| count.json_parse | 32 |
-| count.missed_abstention | 1 |
-| count.natural_stop | 32 |
-| count.nonrepetitive_nonempty | 32 |
-| count.predicted_values | 60 |
-| count.present_correct | 59 |
-| count.present_total | 59 |
-| count.requested_order | 32 |
-| count.schema_valid | 32 |
-| count.token_limited | 0 |
-| count.tp | 59 |
-| count.unsupported_values | 0 |
-| rate.absent_accuracy | 0.9523809523809523 |
-| Requests | 32 |
-| rate.exact | 0.96875 |
-| rate.f1 | 0.9915966386554621 |
-| rate.field_accuracy | 0.9875 |
-| rate.json_parse_rate | 1.0 |
-| Output NLL | 0.008540522438109594 |
-| rate.precision | 0.9833333333333333 |
-| rate.present_accuracy | 1.0 |
-| rate.recall | 1.0 |
-| rate.schema_valid_rate | 1.0 |
-
-### QA development panel
-
-| Metric | Value |
-| --- | --- |
-| count.abstained | 4 |
-| count.all_requested_values_present | 8 |
-| count.exact | 34 |
-| count.false_abstention | 3 |
-| count.missed_abstention | 23 |
-| count.natural_stop | 136 |
-| count.nonrepetitive_nonempty | 136 |
-| count.relation_correct | 34 |
-| count.token_f1 | 65.24622858028253 |
-| count.token_limited | 0 |
-| count.wrong_value_mentioned | 4 |
-| Requests | 136 |
-| rate.exact | 0.25 |
-| Output NLL | 0.5983618814429706 |
-| rate.token_f1 | 0.47975168073737157 |
-
-### Full serving test pooled counts
-
-| Metric/count | Parent | Recreated |
-| --- | --- | --- |
-| exact | 953 | 1132 |
-| format_ok | 1384 | 1385 |
-| json_parse | 1384 | 1385 |
-| schema_valid | 1384 | 1385 |
-| field_correct | 2796 | 3047 |
-| fields | 3347 | 3347 |
-| tp | 1759 | 1740 |
-| fp | 524 | 223 |
-| fn | 120 | 139 |
-| present_correct | 1759 | 1740 |
-| present_total | 1879 | 1879 |
-| absent_correct | 1037 | 1307 |
-| absent_total | 1468 | 1468 |
-| false_abstention | 18 | 69 |
-| missed_abstention | 431 | 161 |
-| unsupported_values | 43 | 5 |
-| predicted_values | 2283 | 1963 |
-| natural_stop | 1388 | 1388 |
-| token_limited | 0 | 0 |
-| nonrepetitive_nonempty | 1388 | 1388 |
-| requested_order | 1384 | 1385 |
-
-## Appendix D. Citation/provenance practice
-
-When reporting this artifact, cite OpenSML and the project author, this report, the exact recreated weight hash, the adapter version and the decoding protocol. Include SQuAD, Dolly, SNIPS and pretraining dataset attribution as appropriate. Label 168/196 as observed development, the full test as observed SNIPS annotation agreement and 29/40 as a local diagnostic. Scores from OpenSML-150M, SFT-036, earlier JSON models or other representation trials do not belong to these weights. Report a test as complete only when all expected requests are present, and distinguish structural validation from correct values.
-
-The repository edition preserves the original report results and replaces private workspace paths in published JSON records. Original metadata/result hashes remain provenance; sanitized files have new hashes in `provenance/SNAPSHOT_MANIFEST.json`. This report was prepared from retained local artifacts and checked against source implementations on October 14, 2026. Primary source links were checked where accessible; retrieval limitations and unresolved release items are declared above.
-
-## Appendix E. Original workspace evidence hashes
-
-These hashes identify the original workspace source/data manifests and serving/evaluation artifacts checked for the report. Published JSON records replace private workspace paths and sometimes omit QA output payloads; their new hashes are recorded in `../provenance/SNAPSHOT_MANIFEST.json`. The hashes below remain original-evidence identities, not integrity claims for sanitized files. They complement the training-time code/input hashes embedded in each checkpoint contract; they do not replace those original contracts. The serving pipeline was added after model training. The full test JSON files record weight and metadata identities verified against the preserved artifacts after generation.
-
-| Artifact | Original workspace SHA-256 |
-| --- | --- |
-| [configs/corpus_20b.json](../provenance/LINEAGE.json) | `d31deb465fd3ba8dcd4590ecf9f01375246784df91217442ae5fcbf0f7ab252b` |
-| [pretraining/data/sft_065_targeted_qa_20261013/manifest.json](../provenance/LINEAGE.json) | `a43ddd52c79ce4fae2ecb507b402fd21f6aaa666a578b7e57da1f8b7048a82a8` |
-| [pretraining/data/sft_068_snips_extraction_20261013/manifest.json](../provenance/LINEAGE.json) | `6268800c46a3405f0fb8849bf71d5f23b8d395e9713bbbb0b8998b8236b77cfd` |
-| [pretraining/data/sft_093_json_approaches_20261014/manifest.json](../provenance/LINEAGE.json) | `3edfaf21bfcce93e2ab7c83f3195222eeea6fdbda15543a63e1c9aa02fe9e7e9` |
-| [pretraining/scripts/json_fieldwise_playground.py](../provenance/LINEAGE.json) | `50ef9208864d0f075a6013130d879577cae42fced1f5fec539f155e01b6f0106` |
-| [pretraining/scripts/start_json_playground.py](../provenance/LINEAGE.json) | `894b4502ee4a1c013e9a0ade3aed64b1f32662efff167838cdc3ce7da0949ca1` |
-| [pretraining/json_approaches_090/evaluate.py](../provenance/LINEAGE.json) | `02b5c87148ad1aa41e44b0994d2f3fe0027a81eb11ba2ddf6524bc4cf816f5d8` |
-| [pretraining/extract_068/data.py](../provenance/LINEAGE.json) | `497b0b12c8896c2045f470c1c2de0330472663f1fd5b67746734a3f7907ee971` |
-| [pretraining/sft/data.py](../provenance/LINEAGE.json) | `70348a9df1473da88330695d5409f17b929b9bc1dc79f77525272f6d87317aa1` |
-| [pretraining/sml_v1/model.py](../provenance/LINEAGE.json) | `1eee6f8afe78767ff6b9c604ccf2328c98b24bc94dbad6291a0ad699dd7345fe` |
-| [reports/model-design/JSON-090_099_approaches_20261014/release_fulltest_runner.py](../provenance/LINEAGE.json) | `b837e34dabcb7389e9c1de328e7c95277364f4e759c74c5a2ea502cde77e0d61` |
-| [reports/model-design/JSON-090_099_approaches_20261014/release-fulltest-extract-068.json](../provenance/LINEAGE.json) | `0bc2bcf68a7ce3e6dce9a79556de812b42e51e62174d0053d678d876434c550f` |
-| [reports/model-design/JSON-090_099_approaches_20261014/release-fulltest-sft-093.json](../provenance/LINEAGE.json) | `800e762ffc88f39e57c15dcaa3fab1ae9d1e94fce64ba7a876099ed4cca7dad2` |
-
-The saved full-test runner requires the pinned workspace, source panel, dependencies and already running local playground service. It is not a standalone model loader. Running it again generates new predictions and overwrites its two result files; retain the published result hashes separately when reproducing it. No training runs are launched by that runner.
-
-## Appendix F. Dedicated repository packaging verification
-
-The included `native_bundle.zip` contains the standalone inference source. The portable root `inference.py` reads a relative config/tokenizer and an explicitly supplied matching weight file. It has no dependence on workspace model-registration receipts or private path metadata. Its greedy fieldwise prompt/parser/model path reproduces all 40 saved playground values and stop reasons with stock MLX 0.32.3, NumPy 2.4.2, Tokenizers 0.22.2 and jsonschema 4.26.0 in a clean temporary environment. This verification has the same 29/40 reference accuracy as the original service. It checks inference packaging; it does not change weights, retrain the model, establish full-suite cross-runtime parity or make the test a fresh blind benchmark.
-
-The source snapshot includes no weight or optimizer binaries, raw training pools, packed arrays or private cluster/network receipts. Public checkpoint JSON has redacted workspace paths and separate published hashes. Original metadata/weight identities remain in `../provenance/LINEAGE.json`. Source-file inventory and the integrity manifest describe the actual package.
+Hugging Face edition: archive-relative source/evidence references resolve inside `native_bundle.zip`; root `RUN_GUIDE.md` explains download and native inference.

THIRD_PARTY_NOTICES.md

@@ -10,7 +10,7 @@ The root Apache-2.0 license covers author-controlled source contributions. Insta
 | [Python-Edu mirror](https://huggingface.co/datasets/Avelina/python-edu-cleaned) | Educational Python pretraining | Local metadata records ODC-BY; underlying code repositories and The Stack v2/SWH provenance retain their conditions. The pinned mirror card was not accessible during the report's external audit. |
 | [SQuAD 2.0](https://rajpurkar.github.io/SQuAD-explorer/) | Grounded QA SFT | Pranav Rajpurkar, Robin Jia and Percy Liang; CC-BY-SA-4.0. Original answer offsets were checked and a curated subset retained. Raw training corpus is not distributed. |
 | [Dolly 15k](https://huggingface.co/datasets/databricks/databricks-dolly-15k) | Selected closed QA SFT | Databricks dataset contributors; CC-BY-SA-3.0. Raw training corpus is not distributed. |
-| [SNIPS NLU benchmark](https://github.com/sonos/nlu-benchmark/tree/b86ac7f1577868c42158d0dec77db50956046696) | Exact field extraction training and adapted held-out evaluation | SNIPS/Sonos contributors; CC0-1.0. Original chunk annotations transformed into requested-field views; [included license](licenses/SNIPS-CC0.txt). |
+| [SNIPS NLU benchmark](https://github.com/sonos/nlu-benchmark/tree/b86ac7f1577868c42158d0dec77db50956046696) | Exact field extraction, native JSON SFT and adapted held-out evaluation | SNIPS/Sonos contributors; CC0-1.0. Original chunk annotations transformed into requested-field views; [included license](licenses/SNIPS-CC0.txt). |
 | OpenSML authored fictional facts | Controlled QA examples | Programmatically authored and checked; not independent human certification. |
 
 MLX, NumPy, Tokenizers and jsonschema are installed dependencies, not vendored distributions in this archive. Their upstream distributions supply their license files. This repository includes author-controlled MLX model/kernel and orchestration source from the OpenSML workspace; the original first release's public source repository is [opensml-150m](https://github.com/williamzebrowskI/opensml-150m).

checkpoint_metadata.json

@@ -0,0 +1,174 @@
+{
+  "extraction_sft": {
+    "contract": {
+      "code": {
+        "pretraining/extract_068/data.py": "497b0b12c8896c2045f470c1c2de0330472663f1fd5b67746734a3f7907ee971",
+        "pretraining/extract_068/evaluate.py": "a7ff05b0d2383421e83d0247e613017715485b3a5fffdcc7240b90387a756c06",
+        "pretraining/extract_068/runtime.py": "7b97df4835d03210c77da36f53340aa83d5915b9cfa7aae968aef6961422c2f1",
+        "pretraining/extract_068/train.py": "f44af1b00a7e36c3dd4c42c30dc63b73bc0d5917495f1985cc6bc3c5b902b6c5",
+        "pretraining/extraction_001/__init__.py": "f1e824c48131bc441020c864929875bfa44daebba75b15181cc401f7b73ae8a2",
+        "pretraining/extraction_001/data.py": "59843f3beca3aecbe8e905af8c969fa9fd6a42554da8d2b1054d027fb771a659",
+        "pretraining/extraction_001/evaluate.py": "b22da13e22b5ebffeb217200e5ff431773d85e6ae394045e96c97ed1bc7e7c54",
+        "pretraining/extraction_001/train.py": "071ef9293bb825753b47b3704e1df1e69f33b792c45c5748f226e40c6c054e26",
+        "pretraining/json_evidence_binding/__init__.py": "357162457d07712f819559c1f1a6769323941355531cb6a04f657ace44d5c52c",
+        "pretraining/json_evidence_binding/data.py": "fd828fcaa202085ea38a4a927e31f7bfc09e19a8965cd170f095408744e3a8a0",
+        "pretraining/json_evidence_binding/evaluate.py": "25fc997a4035ab44f6e5ac47988abc38886efb2c23d8cc270f014e8c0a04cc71",
+        "pretraining/json_evidence_binding/objective.py": "5bcfe1d57e445ca79a82daf81a8d664bb1c8178aaf3ff4189665e205a95537aa",
+        "pretraining/json_evidence_binding/order.py": "37759caee925d4eb2a6e8273075524090516017c199fbc9ebe75edee9b1a7cd6",
+        "pretraining/json_evidence_binding/runtime.py": "4d6695edd7eccfd26ca31a282e76ae663fe149b86f6fc42e8edc4e7625ca15ec",
+        "pretraining/json_evidence_binding/storage.py": "a93d3af323904983f016870865e90ee89a02df4c66d15e502ef9a2ebfff99424",
+        "pretraining/json_evidence_binding/train.py": "2e20b5a09501145a68500994eed4a8e5134148d805dafd93726ea09afca23ebf",
+        "pretraining/json_sft/__init__.py": "aa9dcf15611647559bbc4ab1d8dfa3746cf731f30e4b5fd35bf53a1ed00bea95",
+        "pretraining/json_sft/adapters.py": "09dd78767c9f7e0dbaa0184e64d0011afadd7a326986766026dfcc64106ab8d9",
+        "pretraining/json_sft/benchmark.py": "75ae0df6b0d6676805d6c1ef3b7b73893f473df001dcc2a542f31fb074dd3601",
+        "pretraining/json_sft/business.py": "5c31088adce5f85ce8980d159fa0f9c39a0e1d9a7725560838a2852ac5cfbff4",
+        "pretraining/json_sft/constrained.py": "dc21aa142fe525f3559cd5638e46ef287555efba8239a06072a092fc10700cf5",
+        "pretraining/json_sft/contrasts.py": "e504c5f879011963b82728d73fe334585a640d419c4e3d47d919a4cc89f396fd",
+        "pretraining/json_sft/corpus.py": "37d7b74e0478853903953934b5d2006fd059e1ca2a613cc69f6788e9f34d0e6d",
+        "pretraining/json_sft/decomposed.py": "46f863e628bcc5fedfd6f34f7c759d4b0998091d8252b48729027ce9ef2a01b9",
+        "pretraining/json_sft/diagnostics.py": "92f750807aebb9cb38154c4599d468850ce25ef86496de8e660a77e538e08f03",
+        "pretraining/json_sft/evaluate.py": "2c3076f69ef2d7721f25a79618563d102484bedbf2912f684b644b9135d25d17",
+        "pretraining/json_sft/expansion.py": "231274b7e6d486d39bf962ff310ad4557ccfd84f8cac329cf57ed5890aed7cc7",
+        "pretraining/json_sft/grounded.py": "f668379c5f904ee36c267f6897f810cdb04f80e19a8fd5fdfb791b31f902e8c9",
+        "pretraining/json_sft/narrative.py": "9d507dbc7c2eff7a4b77064393dc65e65c6e3a1866b4375df3d8b66c755d5be1",
+        "pretraining/json_sft/ordered.py": "0ac1b6dd3139425c06094cf5e185ffab5389a377ecaf58b6320978908df4b7ef",
+        "pretraining/json_sft/ordered_serialization.py": "14c23f875b45d17d8d07cb106afc29a6b72cb9b7bb5a393ee0ee55bf84b5aa1a",
+        "pretraining/json_sft/plain.py": "64dd1060d5a674ef04bba3ea3965b7e89ab4faa58bbc3c437816ed2023023843",
+        "pretraining/json_sft/precise.py": "7696328b55290548f05562a9562ea898301a992b4161000f5f2dd7d1bbaf621a",
+        "pretraining/json_sft/reader.py": "ff288fddabd0525f8fc6a8f2a6aeaedf309cfb954448d8c6aa57e122e8628d6e",
+        "pretraining/json_sft/relational.py": "d2e5d3f850b5b543876fafac8b3d57c3ea7c6983437ba6744026ed86701e929e",
+        "pretraining/json_sft/schedule.py": "50ec8a06b0f881714451501b6325a0a64c7cbff159b57f1214f0a3a6ac83014c",
+        "pretraining/json_sft/serialization.py": "d04338a44b569c3f38ed42b0e1135738c1c119387134945335d51bb354c2213f",
+        "pretraining/json_sft/serving.py": "d533a59791bd2516dd1ec4cd2493c554161c0fe1e3603d55fa12631491cddbf8",
+        "pretraining/json_sft/slots.py": "8db7eb31a53899b9e4306af97fad6219ff6c8054a0df67ceffe0dbe927e5df24",
+        "pretraining/json_sft/train.py": "abe0613af70c559dca777cc1fc3c905be0f9583fb6aa30442881da8d0a95f316",
+        "pretraining/json_sft/transfer.py": "ff5bc65fae8da93164f830cff5a20fcff42132bcbf02a94de672af9653cf340d",
+        "pretraining/json_sft/unquoted.py": "969263db40f1a5561acbafdfb21dbb74ce85dd394c2bc9309cd208f61c977289",
+        "pretraining/json_sft/value_weights.py": "043cde3b06edcbc0bd332eb8ace4257e1864af77c239204c120261b95696598b",
+        "pretraining/native_json_100/__init__.py": "8d4b1be2f7d75122c982cd881c19a41b1d39594c5e30fbbe2bc87fcccf9c0242",
+        "pretraining/native_json_100/data.py": "594a83b846c97836be009e61d9fcf9a1b4ceb0d0f9bb491bbb5b56f02a433edf",
+        "pretraining/native_json_100/evaluate.py": "3c03d63aeb85ee85d857f219ee3aa4c190f34532ab6643cb487ca8e41d9779c3",
+        "pretraining/native_json_100/runtime.py": "9b04f8b33fceb717fa6e9290a75bcc7c8365ad6162d3f59de19e08b913a034c4",
+        "pretraining/native_json_100/train.py": "63823c5bc7a4b44cb57d7cd4bead7e97ecf309b09816e4d81327a9b7d795ea72",
+        "pretraining/native_json_100/verify.py": "aa1452e3994eb534f30185c69ba1fcac5dc47081a51e547c31db132e8c369118",
+        "pretraining/qa_064/data.py": "a5d86a1c73285f420c728b5e8331da468eb8846086a8bbe69ce61df6e64fef7f",
+        "pretraining/qa_064/evaluate.py": "b0b8bd42b4842065e3ef8ba3c1fee42c95464f0f8ec49d030e4ab4f1998f94ae",
+        "pretraining/qa_064/runtime.py": "7b97df4835d03210c77da36f53340aa83d5915b9cfa7aae968aef6961422c2f1",
+        "pretraining/qa_064/train.py": "9400aa40b0f5b332334e11fc8b62f5d06464fe1188603bb65585302461d3bf19",
+        "pretraining/qa_065/controlled.py": "55799050001b55a642705064bf79ce45625eca81ebfe77d7f70cba8435726f51",
+        "pretraining/qa_065/data.py": "1e4878ce7bd9e48d8d5c79eafac31093eaf7cdcdd4a6aeb5f9b6519967e47e57",
+        "pretraining/qa_065/evaluate.py": "f8e4a209a7cdb42cded3cb5ffdb2aabfa0d12cc13c10c2f40f33907c532f74e5",
+        "pretraining/qa_065/runtime.py": "7b97df4835d03210c77da36f53340aa83d5915b9cfa7aae968aef6961422c2f1",
+        "pretraining/qa_065/train.py": "52b11949018bc87c6c80dc3bee7571b2ca95967ebb47fe450a5f5552c46d4a16",
+        "pretraining/qa_066/controlled.py": "fc101c488049fd836e53422a72d13d8d02f253153726e9e4d5eda0eb19ab9baa",
+        "pretraining/qa_066/data.py": "4b89e4107bd7076270151989871805af4162d76a299e72d40881fdc26750015a",
+        "pretraining/qa_066/evaluate.py": "d998fb255daabc15b6070bffaf9adeeebdbe631e54be05b83b473f79b3c8642c",
+        "pretraining/qa_066/runtime.py": "7b97df4835d03210c77da36f53340aa83d5915b9cfa7aae968aef6961422c2f1",
+        "pretraining/qa_066/train.py": "3da0be266aa1f047addba8d0c696245ed71e4fde236ecfc8b383c69e8b32e783",
+        "pretraining/scripts/run_sft_100_native_json.sh": "40be750725ffaae7acb81fa03c592cea1b862dba409532ad45d9e7e64fc34fce",
+        "pretraining/sft/data.py": "70348a9df1473da88330695d5409f17b929b9bc1dc79f77525272f6d87317aa1",
+        "pretraining/sft/evaluate.py": "80624551512112850145ccb57d6435d7a1bae86d0b95f8f52d56b8cb6f102a76",
+        "pretraining/sft/quality.py": "2abf5926a22a33b4f29c699855b24f7e459b6e631aa6ef780876ddc205f17758",
+        "pretraining/sml_v1/__init__.py": "7e9e02294e03e6b8a12fd202ab288f2f9b0073bb11b8cb2a5fa82211b0d1dfa4",
+        "pretraining/sml_v1/batch_transition.py": "e6778ba8a8215f0b3c20386e768ea3e15b5939e7919935046689733eb96eba28",
+        "pretraining/sml_v1/checkpoint_bundle.py": "73cbf0c17cf74a44bb7c74a8f9e5bb63671057a1f664a6a4af0e0777b2c8f1a4",
+        "pretraining/sml_v1/checkpoint_metadata.py": "0aaa7019ab53fd8b5e6b0e397f970026b4e8d38e12bf5d5f0fcb7f686de9b82c",
+        "pretraining/sml_v1/common.py": "ec4dda551cbe45cbd30d64c7c82379a8849ce02a3ad1a597a293844b5dc86351",
+        "pretraining/sml_v1/continuation.py": "3c5248646c27bcdf6bd4010742fe6e6d0aa9f059b543e183b60e00e3dbe034a3",
+        "pretraining/sml_v1/corpus.py": "be77769180342f6a8b1d5619e5b3b64f20ca40d3424219168cccf4789bb02038",
+        "pretraining/sml_v1/data.py": "411284cda81b07f10bc624ebd8f7fe9debd2911e950971f96b204dfd1cee6a6d",
+        "pretraining/sml_v1/data_review.py": "a71eabcc526344b6502cae999020da63c950b2d80c13077629bb86c8757a18ed",
+        "pretraining/sml_v1/diagnostic_eval.py": "317c05d4d17239e62d1316797892288f5e4eae467b3e66a4ca4f15d189960a9b",
+        "pretraining/sml_v1/fast_loss.py": "f99edeedb6cb328ef728d05643f65ea20371437f7e5806009ae3ad0885bafbf1",
+        "pretraining/sml_v1/fast_swiglu.py": "59d1adb31400e3c288914b4c0cc8a21681a73ca3dfa7c722ea4e13d21dc38528",
+        "pretraining/sml_v1/inference.py": "4255b93bbc18d120318d83fb26b2fda1fe16d19ca1423250c77188940cc4f008",
+        "pretraining/sml_v1/jaccl_control.py": "a3dfbd61de43fda14daf4bbd921ccb382ea8af2c8342457fe16d1ad0908e88fe",
+        "pretraining/sml_v1/launch.py": "41e6a98d9b41dc844063fc28ce81a08d83b919306b9daf0716b76a0bc21365fc",
+        "pretraining/sml_v1/lr_scheduler.py": "e141132efbd5f1beab4b187d4eab15efd46fcf9710d6a2ba7c882f62362a77cf",
+        "pretraining/sml_v1/lr_trial.py": "2f643abaa6ac3f86e47b37632df4d25fa34d0b25e77b59ffecc0290b2385e7dc",
+        "pretraining/sml_v1/model.py": "1eee6f8afe78767ff6b9c604ccf2328c98b24bc94dbad6291a0ad699dd7345fe",
+        "pretraining/sml_v1/playground_publish.py": "bffd8f8c5a74a0bf9c3ca070e6511cc1ab8ceb8343c0092a3d711b4fb9ece21c",
+        "pretraining/sml_v1/precision.py": "b9f535fca724153703ed57a300eceb5b34a8f8381fc65bf5c4c5a0f7e066303c",
+        "pretraining/sml_v1/prepare_data.py": "3d78ed45dbab52a671d889fe937a210a832e55426bbdc4f905ddbd202ccf2336",
+        "pretraining/sml_v1/pretrain.py": "456a498d50e991cd83363588392b91e58a8bdeec7ceaf9950671925b6a86a46b",
+        "pretraining/sml_v1/recipe.py": "d354731910ac16868b21717a089fce5199a898dec39c6561b3d10c89bfa8a2d1",
+        "pretraining/sml_v1/research_stage_b.py": "5234bc6b824e4b07f63dda48e5353bbdb5f5abac4c98c299b642d1815ce85d94",
+        "pretraining/sml_v1/sample_prompts.py": "1173c8c94dc11c6b8e446c26804a543b68e366986f01d249384082629dca369b",
+        "pretraining/sml_v1/stable_schedule.py": "631451a45b020120610b7116e8e63276ba5043011359fbd13e49e63ec8e2b02b",
+        "pretraining/sml_v1/stage_b.py": "f80f065f6892eea951e94a7ff6925d5a1bff9419a7384e0756cf2faeb92e4775",
+        "pretraining/sml_v1/stage_b_data.py": "c065be118b0c6b8948a15d4b81e7fd7371f4e699ea3c49d34fa0abde258e0505",
+        "pretraining/sml_v1/stage_b_eval.py": "943fd969994e896faff14481b94eb121d23b6fef00244874d5d5361df6bd8cce",
+        "pretraining/sml_v1/stream.py": "70e36e12d33547bfce505bd0302605e3d66c9ee65fc0facb26ccf5341bde8b4b",
+        "pretraining/sml_v1/tokenization.py": "254e1f5d60cc4498a9f1209955752db28b91839f48d3121d674f3f5ac42e52e5",
+        "pretraining/sml_v1/topology_transition.py": "adc3a0a6f19e7cfb3073f4c1e0432f664e5fb709141ceb4148aca7cda08ddaab",
+        "pretraining/sml_v1/train_tokenizer.py": "90c58a30a50cb625e36222825768a134aba5c81df87de8655ab95de2bd47402c",
+        "pretraining/tests/test_native_json_100.py": "00b4d10d22d7497dd14444cf6a669f9dc195f6d1b52dc7dcb679e76f7057483b",
+        "reference/opensml-150m/tokenizer/bytebpe32k_v1/corpus.json": "e4c47b7b6aba08f948de7c2c7cb94c58dbd09dde2ded6ac185c60856ee686d27",
+        "reference/opensml-150m/tokenizer/bytebpe32k_v1/manifest.json": "1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb",
+        "reference/opensml-150m/tokenizer/bytebpe32k_v1/report.json": "e871f690eabebf82da173f0e406f38b79a10cd7d2647bb3592c3fc35ae9a2f16",
+        "reference/opensml-150m/tokenizer/bytebpe32k_v1/tokenizer.json": "e095f0413a11fa91fbb58f37d00e4c6bcda59beb89b28f6f680db1436a7e8891",
+        "reference/opensml-150m/tokenizer/bytebpe32k_v1/tokenizer_config.json": "619f500553367eb76f5be00a98625130ec857b9c367561334984096c029ba2a9"
+      },
+      "manifest": "f0d23b86eeba71d7f616ab630da8914aed4f5586f7dc149426b487581c833f00",
+      "parent": "[workspace]/pretraining/runs/sft_093_json_approaches_20261014/step_0000160_82ffab052730",
+      "parent_metadata": "32a380899c455449db066ef317de9dd8e80f6977bfb42a6fe03b67c5f23963f8",
+      "parent_sha256": "d552c1380c71c7c0569e9d692ec9c46c5c94bb87a7b024e56df0f0e36df4d798",
+      "plan": {
+        "batch_size": 32,
+        "betas": [
+          0.9,
+          0.95
+        ],
+        "epochs": 1,
+        "eps": 1e-08,
+        "eval_steps": [
+          0,
+          100,
+          200,
+          300,
+          469
+        ],
+        "final_lr": 1e-06,
+        "grad_clip": 1.0,
+        "max_tokens": 256,
+        "microbatch": 2,
+        "padding_policy": "micro_multiple128",
+        "peak_lr": 5e-06,
+        "seed": 20261014,
+        "warmup_updates": 20,
+        "weight_decay": 0.01
+      },
+      "versions": {
+        "jsonschema": "4.26.0",
+        "mlx": "0.32.3.dev20260920+7241f12",
+        "numpy": "2.4.2",
+        "tokenizers": "0.22.2"
+      }
+    },
+    "order_hash": "76518047ebfc78beb886d78322635ee275acdb6ad479a5d722bb871be3c89961",
+    "records": 15000,
+    "serialization": "native-sft-chat-v1",
+    "total_steps": 469
+  },
+  "recipe": {
+    "model": {
+      "attention_impl": "fast",
+      "bias": false,
+      "ce_impl": "metal",
+      "d_model": 768,
+      "ffn_impl": "packed-metal",
+      "loss_dtype": "float32",
+      "max_seq_len": 2048,
+      "mlp_multiple_of": 256,
+      "mlp_ratio": 4.0,
+      "n_heads": 12,
+      "n_kv_heads": 4,
+      "n_layers": 24,
+      "qk_norm": true,
+      "rope_base": 10000.0,
+      "vocab_size": 32000
+    }
+  },
+  "step": 469,
+  "tokens": 4634126
+}

checkpoint_provenance.json

@@ -36,10 +36,23 @@
       "weight_bytes": 351242342,
       "original_metadata_sha256": "32a380899c455449db066ef317de9dd8e80f6977bfb42a6fe03b67c5f23963f8",
       "published_metadata_sha256": "32a380899c455449db066ef317de9dd8e80f6977bfb42a6fe03b67c5f23963f8"
+    },
+    {
+      "stage": "sft-100",
+      "step": 469,
+      "tokens": 4634126,
+      "records": 15000,
+      "weight_sha256": "656f45015aab548465d93bed0b57e687497163741885da7364872851fa17e9a4",
+      "weight_bytes": 351242342,
+      "original_metadata_sha256": "3864a0eaa43015d286e4da9327b0d3cea4fa4f5b80f336baabe2a549379e2e36",
+      "published_metadata_sha256": "5ecf9a8850aef6cc75add6bacda880648e72de1b38089d585e237bbce828c2c3"
     }
   ],
   "weights_in_git": false,
   "public_weight_location": null,
   "model_weight_license": "not yet assigned",
-  "scope": "Only these selected checkpoints are inherited. Sanitized metadata has separate hashes. Local paths become [workspace]."
+  "scope": "Stage A/132005 -> SFT-065/469 -> SFT-068/379 -> SFT-093/160 -> SFT-100/469. Only those stages are inherited. Final native JSON targets are directly supervised.",
+  "selected_stage": "sft-100",
+  "selected_update": 469,
+  "weight_location": "https://huggingface.co/wzebrowski/OpenSML-JSONLite-175M"
 }

config.json

@@ -18,11 +18,17 @@
     "rope_base": 10000.0,
     "vocab_size": 32000
   },
-  "weight_sha256": "d552c1380c71c7c0569e9d692ec9c46c5c94bb87a7b024e56df0f0e36df4d798",
+  "weight_sha256": "656f45015aab548465d93bed0b57e687497163741885da7364872851fa17e9a4",
   "serialization": "native-sft-chat-v1",
-  "output_mode": "fieldwise-assembly",
+  "output_mode": "native-json",
   "context_limit": 2048,
-  "max_output_tokens_per_field": 160,
   "weight_format": "BF16 safetensors",
-  "supported_runtime": "native MLX; not a Transformers or mlx-lm auto-loading export"
+  "supported_runtime": "native MLX; not a Transformers or mlx-lm auto-loading export",
+  "max_output_tokens": 256,
+  "model_calls_per_request": 1,
+  "json_assembly": false,
+  "output_repair": false,
+  "constrained_decoding": false,
+  "selected_stage": "sft-100",
+  "selected_update": 469
 }

inference.py

@@ -2,7 +2,7 @@
 """Run JSONLite offline with the bundled native MLX implementation.
 
 Defaults to model.safetensors beside this script. The verified source bundle
-includes the exact model architecture, tokenizer, field adapter and examples.
+includes the exact model architecture, tokenizer, native JSON prompt/decoder and examples.
 """
 from pathlib import Path
 import hashlib
@@ -12,7 +12,7 @@ import tempfile
 import zipfile
 
 ROOT = Path(__file__).resolve().parent
-BUNDLE_SHA256 = "5df3ada361b6fe7226170a73cd8d9cf08a70dd3cdee3ea7c78cb751e7922c94c"
+BUNDLE_SHA256 = "88e6a957ec915d23a4be199b8107cd4812577f8f86d446b0e890fef4a3d22da8"
 
 def main():
     archive = ROOT / "native_bundle.zip"

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d552c1380c71c7c0569e9d692ec9c46c5c94bb87a7b024e56df0f0e36df4d798
+oid sha256:656f45015aab548465d93bed0b57e687497163741885da7364872851fa17e9a4
 size 351242342

native_bundle.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5df3ada361b6fe7226170a73cd8d9cf08a70dd3cdee3ea7c78cb751e7922c94c
-size 1980238
+oid sha256:88e6a957ec915d23a4be199b8107cd4812577f8f86d446b0e890fef4a3d22da8
+size 5606613