Describe 035 training recipe without predecessor references

035_step200000/README.md

@@ -2,7 +2,7 @@
 
 PT checkpoint and matching G1 / A3-fast ONNX and RK3588 RKNN exports. The `g1` name is a historical interface name; these artifacts control A3. Full policies accept `obs_dict: float32[1,1570]` and return `action: float32[1,29]`. Standalone encoder / decoder ONNX exports are included for advanced integration.
 
-035 continues the 024 checkpoint with the VR3 / no-ankle-orientation recipe. Use the adjacent portable configuration files. The PT retains the original policy, value, optimizer and scheduler states exactly; internal training arguments, run metadata and motion-library state are omitted. It supports inference and further training, but is not a bit-exact resume snapshot of the original training environment.
+035 uses the VR3 / no-ankle-orientation training recipe. Use the adjacent portable configuration files. The PT retains the original policy, value, optimizer and scheduler states exactly; internal training arguments, run metadata and motion-library state are omitted. It supports inference and further training, but is not a bit-exact resume snapshot of the original training environment.
 
 Weight licensing will be handled separately; no new weight license is declared in this upload.
 

035_step200000/SHA256SUMS

@@ -1,4 +1,4 @@
-12c5e11b18ea61104858dbe16512882fb2ef6fc44157ff0cba9d5ac29a47f620  README.md
+c8545b3f15eba4f1685e5002ebeab52e231a8b514800b799b70d90aba1b735b1  README.md
 9f7e8ae75b130e8696020c5f7d4e078898283caf12a46fb195fb3d7c3e08dfa8  config.yaml
 c0dddddcaeec5211d0e996331540e08162e69ff1208b62be0c361d6310006bc5  meta.yaml
 869b3a24140fe05a468128f222c44ac04008f2cc871ab5e86625de1ec1166533  model_config.yaml