Use JSONLite Instruct model name in the model card

README.md

@@ -21,9 +21,9 @@ datasets:
 - databricks/databricks-dolly-15k
 ---
 
-# OpenSML-JSONLite-175M
+# OpenSML-JSONLite-175M-Instruct
 
-**Author:** William Zebrowski · **Selected checkpoint:** recreated SFT-093/update 160 · **Status:** private repository preparation / experimental extraction candidate
+**Author:** William Zebrowski · **Model:** OpenSML-JSONLite-175M-Instruct · **Status:** private repository preparation / experimental extraction candidate
 
 [Technical report](TECHNICAL_REPORT.md) · [Results](JSONLITE_RESULTS.json) · [Checkpoint provenance](checkpoint_provenance.json) · [Native inference verification](INFERENCE_VERIFICATION.json) · [Source repository](https://github.com/williamzebrowskI/OpenSML-JSONLite-175M)
 
@@ -86,7 +86,7 @@ The frozen byte-level BPE tokenizer was fitted for the earlier OpenSML project a
 ## Training and model lineage
 
 ```text
-Stage A pretrained/132005 → SFT-065/469 → SFT-068/379 → recreated SFT-093/160
+Stage A pretrained/132005 → SFT-065/469 → SFT-068/379 → SFT-093/160
 ```
 
 Only those selected stages are inherited. The many intervening research trials and failed RL/JSON experiments are not additional ancestors of the final weights.
@@ -140,7 +140,7 @@ All three SFT stages update the full model on Mac1 with assistant-output-only me
 
 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.
 
-The original SFT-093 trial had 167/196 development exact and was deleted under the failed-gate cleanup policy. The owner requested its recreation and preservation. The retained recreation has a different weight hash and **168/196** exact; bit-identical GPU training reproduction is not established. This card and its result files describe the recreated artifact, not the deleted original.
+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
 
@@ -189,7 +189,7 @@ The model produced separate labeled `track`, `artist` and `album` answers; the a
 
 ### Complete-request correctness
 
-| Panel | Parent SFT-068/379, same adapter | JSONLite recreated SFT-093/160 |
+| 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%** |
@@ -218,7 +218,7 @@ The model produced separate labeled `track`, `artist` and `album` answers; the a
 
 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 recreated 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.
+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>
@@ -243,7 +243,7 @@ General dialogue, coding, multilingual quality, calibrated confidence, prompt-in
 
 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 source repository's snapshot verifier checks shipped code, tokenizer, docs and evidence; its native loader separately verifies the supplied weights.
 
-The model card, metrics and technical report describe one fixed recreated 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.
+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.
 
 ## Licensing and source notices
 

SHA256SUMS

@@ -2,7 +2,7 @@
 82b9908c9fd21fd8c9d10ad809e7c48446485ee98a56bfd9d64bf9fb353761c2  INFERENCE_VERIFICATION.json
 d29cf5952a82db444a6df56446806b1f4e2ccbee1b4aa6b93cb7eeeabcab8536  JSONLITE_RESULTS.json
 c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  LICENSE
-2248cd1826c6054133611e7e24856dc2eda52217b7ca1335e997d4d549b48513  README.md
+0f6af13be03ee3e6896e76737f395e323e9690ad92ce41a1361cfb287544f838  README.md
 64519f25aded3e1fa74b0b9eafdf35df4490e7bb550027338a9bf1a25f9203bb  SOURCE_NOTICE.md
 4f8c90fad555e380c0e99b61582fe6f3bcb9e6f43a1f17d253769384471f8374  TECHNICAL_REPORT.md
 423f0d1e0e2690505c24e6d31308db728fba60657c5166e177bb269d6a7198d0  THIRD_PARTY_NOTICES.md