Finalize source archive, report links and integrity manifests

PACKAGING_VERIFICATION.json

@@ -5,7 +5,7 @@
   "weight_file": "model.safetensors",
   "weight_bytes": 351242342,
   "weight_sha256": "781597baeb9b51d535fc27cd6f941dbc202e01d62c8d963587a3f3d565128560",
-  "native_bundle_sha256": "dd156bba973c85afa13ee348b24ce2f220243718c1d3c0f3298eeaf1f3cda859",
+  "native_bundle_sha256": "c54fd408f6b7ec0c6e80377ee8b6eeffe3bfbf76389e56fdad4c78f54828e2d1",
   "runtime": "stock MLX 0.32.3; Apple Silicon; Python 3.11",
   "requests": 60,
   "raw_runtime_matches": 60,

SHA256SUMS

@@ -3,19 +3,19 @@
 9a4fb5f559d82ed0120a57139eb1e1c616bfdb18c39cca7b1e9bcab24ef968bc  INFERENCE_VERIFICATION.json
 74a074f90bd5e7a26d5f6cd5757ddc1fbfcc72e259e0b23a210fa48d91b35fdb  JSONLITE_RESULTS.json
 c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  LICENSE
-0a701e1ea095ced1be14abfd3223b8fd11504dfb5c40091ae3ea6d2f6fb7357a  PACKAGING_VERIFICATION.json
+f53194e53a2af5ec857337e8213397ea4f81a7e02d2dbcf0384a00cf1cf7f76f  PACKAGING_VERIFICATION.json
 2728c5efa01d9d59fb1def12a543980c300edcf557fabfa921b08dc0712e68cb  README.md
 5ee8f11f0e87becf1a29c44e2b0f3dacc2c779712b635558be57dd7ef44455d3  RUN_GUIDE.md
 c6191baf937bce61f1b747acfbcb975a17eef6e8a598682a751a1031d1cfee23  SOURCE_NOTICE.md
-f99732ce374682389f742c81ef7bda0b2da38bb12066df05aa2f57956618971e  TECHNICAL_REPORT.md
+6088771ab13fa1080c0ca0f31649deb9f417597d7c2055b0f388e86f41f2c415  TECHNICAL_REPORT.md
 e4e44da050abbd930748ec91731ddf1be16be83a4aa59bd8c4fc63478b384a03  THIRD_PARTY_NOTICES.md
 663087d32872824f7f2fec425ad4054f308913c43da7c3f83521c7011967fdbe  checkpoint_metadata.json
 a365a7697b463bacc6e44e26d11a1e62d43d96b5d52c70c2e9df8d4513f86537  checkpoint_provenance.json
 ecec6f191822d16a3da0b571fa129f5016e1243ca7d567de476a9d0e8ad66646  complete-request-exact.svg
 981e02e7777980705770cacf19f16d76ea74e85e7edb018f875ddb4766afe19d  config.json
-8cf99e9c97e59dfd23932693bb4e81fd2581c970051c4c92e0cc70d1b1f42c15  inference.py
+8c0f0f5a91079699d090986d5d241385309b10579f34f271c02f05dca6ddd4bf  inference.py
 781597baeb9b51d535fc27cd6f941dbc202e01d62c8d963587a3f3d565128560  model.safetensors
-dd156bba973c85afa13ee348b24ce2f220243718c1d3c0f3298eeaf1f3cda859  native_bundle.zip
+c54fd408f6b7ec0c6e80377ee8b6eeffe3bfbf76389e56fdad4c78f54828e2d1  native_bundle.zip
 641fe860d6fd24547fe4d15bd852190f00ddec31bd74eb7ca9eeae74031c2f0c  pretraining-training-loss.png
 c7b1a3c0efb09ef3b5416a24072e48b25e2cc1f122fb5ed638c0f95b72594af9  requirements.txt
 e095f0413a11fa91fbb58f37d00e4c6bcda59beb89b28f6f680db1436a7e8891  tokenizer.json

TECHNICAL_REPORT.md

@@ -58,7 +58,7 @@ The same architecture is retained across pretraining and all inherited SFT stage
 | FFN implementation | Packed Metal SwiGLU |
 | Training CE implementation | Metal cross entropy, FP32 reductions |
 
-Implementation evidence: [`model.py`](../provenance/LINEAGE.json), [`precision.py`](../provenance/LINEAGE.json), and the final bundle's `recipe.model` metadata. Weight count and file size are consistent with BF16 parameters plus the safetensors header. GPU allocation readings are not equivalent to file size, total process RSS or system RAM requirements.
+Implementation evidence: [`model.py`](checkpoint_provenance.json), [`precision.py`](checkpoint_provenance.json), and the final bundle's `recipe.model` metadata. Weight count and file size are consistent with BF16 parameters plus the safetensors header. GPU allocation readings are not equivalent to file size, total process RSS or system RAM requirements.
 
 The base initialization is recorded as `normal-0.02-residual-scaled-v1`. Earlier OpenSML-150M **weights** were not used to initialize the base; its frozen tokenizer was reused.
 
@@ -102,7 +102,7 @@ Source weights below are intended processed-token sampling weights, not measured
 
 Public source cards: [SmolLM corpus](https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus/blob/3ba9d605774198c5868892d7a8deda78031a781f/README.md), [DCLM-Edu](https://huggingface.co/datasets/HuggingFaceTB/dclm-edu/blob/dbad8ad71224482740cd9c9d353591adbf62fe04/README.md), [FineWiki](https://huggingface.co/datasets/HuggingFaceFW/finewiki/blob/8bd13e72e6a002407649b3e898535f42ceb1aeb9/README.md). The Python-Edu upstream subset identifies files; this run used the configured text mirror. The mirror's pinned README could not be retrieved during report preparation, so its provenance statement here comes from the local corpus specification, not a newly completed external audit.
 
-The exact corpus specification is [`configs/corpus_20b.json`](../provenance/LINEAGE.json). Training preserves source text and code indentation. NFC/whitespace normalization is used for identity hashing and partitioning, not to rewrite model input.
+The exact corpus specification is [`configs/corpus_20b.json`](checkpoint_provenance.json). Training preserves source text and code indentation. NFC/whitespace normalization is used for identity hashing and partitioning, not to rewrite model input.
 
 The content-split policy hashes identity into 10,000 buckets: 97% training, 1% validation, 1% final and 1% tokenizer development. Identity checks do not establish that a supervised benchmark article never appeared elsewhere in pretraining web data.
 
@@ -168,9 +168,9 @@ The four inherited foundation stages remain unchanged: grounded evidence QA (15,
 
 The native JSON foundation has 12,115 distinct SNIPS TRAIN requests, with 12,115 canonical string-array schemas, 1,443 renamed-key views and 1,442 nullable-string views. Source text, field definitions and the complete schema are input; complete JSON and stop tokens are the assistant target. Original CC0 SNIPS chunk annotations provide source spans. Metadata fidelity and schema validation do not establish perfect semantic annotation. Frozen tokenizer and architecture are unchanged.
 
-### Additional extraction SFT
+### Prior cardinality/replay extraction SFT
 
-This completed continuation starts from the native JSON foundation. It adds matched zero/one/two/three-value supervision for artists and cities, including zero-primary-value cases where an album or country is explicitly present. It retains original JSON TRAIN replay and adds frozen-parent distribution distillation on replay assistant tokens. Every model batch has 24 replay and 8 new views. Public checkpoint name: **main**; the neural tensor is `model.safetensors`.
+This completed continuation starts from the native JSON foundation. It adds matched zero/one/two/three-value supervision for artists and cities, including zero-primary-value cases where an album or country is explicitly present. It retains original JSON TRAIN replay and adds frozen-parent distribution distillation on replay assistant tokens. Every model batch has 24 replay and 8 new views. This is an inherited ancestor of the current main checkpoint.
 
 | Setting | Value |
 |---|---|
@@ -188,7 +188,7 @@ This completed continuation starts from the native JSON foundation. It adds matc
 | Weight decay / clipping |0.01 /1.0|
 | Seed |20261119|
 
-The main checkpoint inherits only the selected prefix; updates after it are not part of these weights. The endpoint's actual learning rate is recorded by the cosine schedule, not the planned final LR. Presentation counts include multiple schemas for the same text and do not mean distinct user queries.
+The current weights inherit only this ancestor’s selected prefix; updates after that ancestor are not part of these weights. The endpoint's actual learning rate is recorded by the cosine schedule, not the planned final LR. Presentation counts include multiple schemas for the same text and do not mean distinct user queries.
 
 For every row, CE is averaged over supervised assistant tokens within each sequence, then averaged across sequences. On original TRAIN replay add 1.0×KL(frozen foundation || student), using the full 32,000-token distribution, temperature 1 and teacher-forced ground-truth histories. Only assistant JSON and turn/document-stop positions participate. The reference network is frozen/stop-gradient; new authored rows use CE alone. This regularizer is training-only: no teacher is needed for inference, and it does not guarantee retention or repair parent errors.
 

inference.py

@@ -12,7 +12,7 @@ import tempfile
 import zipfile
 
 ROOT = Path(__file__).resolve().parent
-BUNDLE_SHA256 = "dd156bba973c85afa13ee348b24ce2f220243718c1d3c0f3298eeaf1f3cda859"
+BUNDLE_SHA256 = "c54fd408f6b7ec0c6e80377ee8b6eeffe3bfbf76389e56fdad4c78f54828e2d1"
 
 def main():
     archive = ROOT / "native_bundle.zip"

native_bundle.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dd156bba973c85afa13ee348b24ce2f220243718c1d3c0f3298eeaf1f3cda859
-size 12544225
+oid sha256:c54fd408f6b7ec0c6e80377ee8b6eeffe3bfbf76389e56fdad4c78f54828e2d1
+size 12831619