Prepare Solstice-1 public research release

FILE_MANIFEST.json

@@ -22,16 +22,16 @@
       "sha256": "0e6865f1a81cfd69ca48e6987658854ce0c0d23e91d72ae545393cd5fa032375"
     },
     "README.md": {
-      "bytes": 13160,
-      "sha256": "75e97ec9ef7579fbcf5e29a9615f4468592a9baea0d9f00d82fa63ba6f1fffb3"
+      "bytes": 13125,
+      "sha256": "fb8c647b3adb0d7fdbd8585efac0db74137179c8fcbe293eed09e881d5a2cba3"
     },
     "REPOSITORIES.json": {
-      "bytes": 716,
-      "sha256": "152a87d0098752350d54fdd3cde3861a77eb497c53dc973ee1a8370d12a13ff6"
+      "bytes": 723,
+      "sha256": "291c0f7dc01cd1c0aeb5f49bb8822e0ab32617c72a6886615673ab08e9431891"
     },
     "RUN_GUIDE.md": {
-      "bytes": 3339,
-      "sha256": "fd05e1595bcdc52942e9425b6efb2c7e89553e6b0ad7d5f93d9075969cd75dba"
+      "bytes": 3149,
+      "sha256": "cf237498a98efb229977939cfceaf78a112f25658f32559cad41aab3b8880202"
     },
     "SOURCE_NOTICE.md": {
       "bytes": 1054,
@@ -138,8 +138,8 @@
       "sha256": "7a818c365f1c047b349479829396ba5cd841ca8d2f04fa14fc1489083a1fd4b3"
     },
     "docs/TECHNICAL_REPORT.md": {
-      "bytes": 13515,
-      "sha256": "f98d45b39b9595510fa80d495a3758b211d8c770adc739d381763d3cfb2a2a0b"
+      "bytes": 13512,
+      "sha256": "5fab4741ab86f46c97f8094b75ab3b560197b1a8c9e3d9e4c0783a0cd9de14eb"
     },
     "docs/figures/benchmark-comparison.png": {
       "bytes": 125053,

README.md

@@ -152,7 +152,7 @@ Greedy FP32 outputs reproduced in the release environment. The incorrect example
 
 ## Run locally
 
-Use Python 3.11 on Apple Silicon. While the repository is private, sign in with `hf auth login` after installing `huggingface_hub`.
+Use Python 3.11 on Apple Silicon. The public model package can be downloaded without signing in.
 
 ```bash
 python3.11 -m venv .venv

REPOSITORIES.json

@@ -3,16 +3,16 @@
   "organization": "OpenSML",
   "parameter_label": "175M",
   "license": "Apache-2.0",
-  "release_status": "private review",
+  "release_status": "public research release",
   "repositories": {
     "github": {
       "url": "https://github.com/williamzebrowskI/Solstice-1",
-      "visibility": "private",
+      "visibility": "public",
       "role": "Native inference code, documentation, training provenance, evaluation evidence"
     },
     "huggingface": {
       "url": "https://huggingface.co/OpenSML/Solstice-1",
-      "visibility": "private",
+      "visibility": "public",
       "role": "Complete runnable package, including merged FP32 weights and tokenizer"
     }
   },

RUN_GUIDE.md

@@ -2,7 +2,7 @@
 
 Use Python 3.11 on Apple Silicon macOS with the pinned requirements. The tested public MLX wheel is 0.32.3; it requires a compatible macOS version. See PACKAGING_VERIFICATION.json for the actual tested system. Inference is offline after downloading the package and installing dependencies.
 
-While the model repository is private, run `hf auth login` after installing `huggingface_hub` below. Use a Hugging Face account with read access to OpenSML/Solstice-1. Tokens belong in the terminal sign-in prompt, never in a shared command, notebook, or repository file.
+The model package is public. No Hugging Face sign-in is required to download it.
 
 ```bash
 python3.11 -m venv .venv

SHA256SUMS

@@ -3,9 +3,9 @@ f09ed498cc42b482dae88d9fb5026bf74e1d8b9187f9ea301aeda0db76d6c859  BACKUP_AND_CLE
 af646a190d135ec25a6ee2c76d1594c3f9ba8005cd3ca6df908606a15023273e  CITATION.cff
 cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30  LICENSE
 0e6865f1a81cfd69ca48e6987658854ce0c0d23e91d72ae545393cd5fa032375  PACKAGING_VERIFICATION.json
-75e97ec9ef7579fbcf5e29a9615f4468592a9baea0d9f00d82fa63ba6f1fffb3  README.md
-152a87d0098752350d54fdd3cde3861a77eb497c53dc973ee1a8370d12a13ff6  REPOSITORIES.json
-fd05e1595bcdc52942e9425b6efb2c7e89553e6b0ad7d5f93d9075969cd75dba  RUN_GUIDE.md
+fb8c647b3adb0d7fdbd8585efac0db74137179c8fcbe293eed09e881d5a2cba3  README.md
+291c0f7dc01cd1c0aeb5f49bb8822e0ab32617c72a6886615673ab08e9431891  REPOSITORIES.json
+cf237498a98efb229977939cfceaf78a112f25658f32559cad41aab3b8880202  RUN_GUIDE.md
 87fa33719fe1182d467a05ab417be2a362f50eea59f0ac6496d40e677544e27c  SOURCE_NOTICE.md
 85525fd4cfd74dbabeb8fe767074cf02a8834dc40a3fc2e648180ef2a3ff35d9  THIRD_PARTY_NOTICES.md
 a4bbaa7901cce3f88cdcac1ffd30ecdf60d4999c198054dce6bb7ef1d72607b2  TRAINING_AND_REPRODUCIBILITY.md
@@ -32,7 +32,7 @@ ae708eca3c09d5a27eab3fddbe4106d0658787beebc9393de57f75234480a284  configs/sft3/p
 e0b0c730321360c40f20e53a22890c52cd5294bad2a5777a0a891096557a2f0a  docs/DATA.md
 fc5860bdad9ac24a5931c15fdb0931caeac1751352377a8ef5d57b5240425bb6  docs/EVALUATION.md
 7a818c365f1c047b349479829396ba5cd841ca8d2f04fa14fc1489083a1fd4b3  docs/EXAMPLES.md
-f98d45b39b9595510fa80d495a3758b211d8c770adc739d381763d3cfb2a2a0b  docs/TECHNICAL_REPORT.md
+5fab4741ab86f46c97f8094b75ab3b560197b1a8c9e3d9e4c0783a0cd9de14eb  docs/TECHNICAL_REPORT.md
 37438bfe291e2cc5b7513ec912cab9eaf7ce3d6352543f8e63d44da3a4f7d5b3  docs/figures/benchmark-comparison.png
 64297355167c23c84cc4227bf54eb6be8d66c5837d45e201ed9304594df36d85  docs/figures/benchmark-comparison.svg
 08388301b65a94ce3bd166de02dd6aa887515f9ed617a8972359b742334cb422  docs/figures/chat-diagnostics.png
@@ -473,3 +473,4 @@ ff056da7218dfbcae9d5c30e1a7770739683866bff398eebc05f3cfd686b1957  training/sourc
 ff84dc0dfacceb506e54b6af73c1a76d65fcbf9cf75c7d96fc4489b8ab4f79e6  training/source_blobs/ff84dc0dfacceb506e54b6af73c1a76d65fcbf9cf75c7d96fc4489b8ab4f79e6.py
 ff9cd7e99a7b66c901137a7f61c97c9ecbd3e5effec879cb82a5c68e9ced35f7  training/source_blobs/ff9cd7e99a7b66c901137a7f61c97c9ecbd3e5effec879cb82a5c68e9ced35f7.py
 ffc3d395e98ab0f55c9d38060d65ff8dcdae4b954a25aafb40c14f513d8067ba  training/source_blobs/ffc3d395e98ab0f55c9d38060d65ff8dcdae4b954a25aafb40c14f513d8067ba.py
+b5db6ca19833c36e95806b520265c1a27dd10485405c2cd1912dcc21c7a3949c  FILE_MANIFEST.json

docs/TECHNICAL_REPORT.md

@@ -14,7 +14,7 @@ The complete tensor shape inventory is in [TENSOR_INVENTORY.json](../provenance/
 
 The model is a decoder-only causal transformer with 44 residual blocks and hidden width 576. Attention uses a combined QKV projection, nine query heads, three KV heads, and head dimension 64. Each query/key head is RMS-normalized. Rotary position embeddings use base 10,000. Blocks use pre-normalization and SwiGLU; the actual FFN intermediate width is 1,536. The packed gate/up matrix has shape 3,072 × 576. Input embeddings and output projection are tied. There are no linear biases or dropout.
 
-The model has a deeper, narrower shape than the earlier 24-layer, width-768 JSONLite ancestor or 20-layer OpenSML-150M. These runs also differ in data, optimization, and training exposure, so their results do not isolate an architectural advantage.
+The model has a deeper, narrower shape than the earlier 24-layer, width-768 JSONLite model or 20-layer OpenSML-150M. These runs also differ in data, optimization, and training exposure, so their results do not isolate an architectural advantage.
 
 Training uses MLX fast attention and packed Metal FFN/cross-entropy paths where specified by each frozen recipe. Release inference and the reported FP32 benchmarks use vanilla attention and reference FFN/cross-entropy paths. No quantization is applied to the release weights.