Unify CSVAE checkpoint naming and Phoenix policy metadata

Model tensor values preserved; previous release remains available at its recorded revision.

NOTICE.md

@@ -8,11 +8,11 @@ DINOv3 and derivative DINO Materials are distributed under the [DINOv3 License](
 
 Release preparation selected the complete EMA student at step 110000 and removed training-only state. It preserved the selected inference tensors. This is a derived checkpoint and is not an official Meta checkpoint.
 
-## STVAE150k
+## CSVAE
 
-The STVAE checkpoint contains the FP32 1024-to-48 encoder at step 150000. Its provenance binds it to the accompanying DINO110k teacher; release preparation retained its parameters and updated the teacher container reference. The encoder implementation credits the FastWAM and StarVLA projects, whose code notices remain in Phoenix-MWA.
+The CSVAE checkpoint contains the FP32 1024-to-48 encoder at step 150000. The saved teacher SHA-256 `ec83a00d2e65fab98a7f5c15a56700386dbf672dcde275ed7d99f16940650c0c` identifies the original full training checkpoint. The accompanying `dino110k.pt` is its EMA inference export, with a different container SHA-256 `81a0082487879332968b5ec458a8d493bf30da601b855cfbaf0e09e250619053`; these two file hashes must not be compared as if they described the same container. The CSVAE format migration retained all 118 FP32 tensors exactly and preserves the original teacher provenance. The encoder implementation credits the FastWAM and StarVLA projects, whose code notices remain in Phoenix-MWA.
 
-The available provenance does not establish a separate license grant for Phoenix-owned STVAE weight contributions. This notice therefore does not infer an MIT or Apache license for those weights from implementation code, nor does it expand any applicable upstream rights.
+The available provenance does not establish a separate license grant for Phoenix-owned CSVAE weight contributions. This notice therefore does not infer an MIT or Apache license for those weights from implementation code, nor does it expand any applicable upstream rights.
 
 ## Scope
 

README.md

@@ -0,0 +1,14 @@
+# Phoenix Encoders
+
+Shared DINO and CSVAE encoders for Phoenix pretraining and Phoenix-MWA.
+
+- `dino110k.pt`: existing DINO inference weights, unchanged.
+- `csvae.pt`: CSVAE weights; `model_type=causal_semantic_vae`, `format_version=2`.
+
+CSVAE retains the original model configuration and all 118 tensor values. Use the matching CSVAE code (`dino_csvae`, `csvae_encoder`, `csvae`) and `CSVAE_CHECKPOINT` / `--csvae` interfaces. The saved training step is metadata, not part of the filename.
+
+Previous release revision: [`07d7189d56c45ffb5ace1472b2023242d85584ec`](https://huggingface.co/OpenPhoenix/Phoenix-Encoders/tree/07d7189d56c45ffb5ace1472b2023242d85584ec). Use that revision when reproducing the previous code interface.
+
+Current `csvae.pt` SHA-256: `fbb4d3f32309b77a5f068634816752d4497fb01798c4be2292bb25690f5baafd`.
+
+See [NOTICE.md](NOTICE.md) and the included licenses for attribution and license scope. This metadata migration does not change any license grant.

stvae150k.pt → csvae.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7e41b75082444011bb6f6451c102e4a01bf2066faf20a7b0615fe05e0472d643
-size 289373779
+oid sha256:fbb4d3f32309b77a5f068634816752d4497fb01798c4be2292bb25690f5baafd
+size 289356611