Rename to PCVE-RigidBench; document rigid-body-only scope

README.md

@@ -1,9 +1,10 @@
 ---
-pretty_name: PCVE-Bench
+pretty_name: PCVE-RigidBench
 tags:
   - video-editing
   - physics
   - physical-reasoning
+  - rigid-body
   - benchmark
   - video
 task_categories:
@@ -16,32 +17,44 @@ size_categories:
   - n<1K
 ---
 
-# PCVE-Bench
+# PCVE-RigidBench
 
-**P**hysics-**C**onsistent **V**ideo **E**diting benchmark: 20 scenes, 20 source
-videos, 131 edit tasks. Each task asks a model to re-render a source video under
-a changed physical parameter (mass, friction, restitution, initial velocity, or
-the presence of an object) and provides the physically-simulated ground truth to
-score against.
+**P**hysics-**C**onsistent **V**ideo **E**diting benchmark, **rigid-body** subset:
+20 scenes, 20 source videos, 131 edit tasks. Each task asks a model to re-render
+a source video under a changed physical parameter (mass, friction, restitution,
+initial velocity, or the presence of an object) and provides the
+physically-simulated ground truth to score against.
 
 Everything is rendered from a PyBullet simulation driving a Blender scene, so
 every case ships exact per-frame object trajectories alongside the video.
 
+## Scope: rigid bodies only
+
+All 151 clips are **rigid-body** simulations. Every dynamic object is a rigid
+body in PyBullet, and every interaction is a rigid contact parameterised by
+mass, friction, rolling friction and restitution. Static scene geometry (floors,
+tables, ramps, the rug in `ball_carpet_climb`) participates purely as collision
+surfaces with their own friction and restitution -- it is never simulated as a
+deformable.
+
+The benchmark therefore contains **no** deformable, cloth, fluid, granular,
+soft-body or articulated/jointed dynamics, and the edit vocabulary is
+correspondingly restricted to rigid-body properties. Results here say nothing
+about a model's handling of non-rigid physics; that is deliberately out of
+scope for this release, and the `RigidBench` name marks it.
+
 ## Download
 
 ```python
 from huggingface_hub import snapshot_download
 
-root = snapshot_download(repo_id="ccmoony/PCVE-Bench", repo_type="dataset")
+root = snapshot_download(repo_id="ccmoony/PCVE-RigidBench", repo_type="dataset")
 ```
 
 ```bash
-hf download ccmoony/PCVE-Bench --repo-type dataset --local-dir PCVE-Bench
+hf download ccmoony/PCVE-RigidBench --repo-type dataset --local-dir PCVE-RigidBench
 ```
 
-Physics-Consistent Video Editing benchmark. 20 scenes,
-20 source videos, 131 edit tasks.
-
 ## Layout
 - `benchmark_manifest.json`: flat index of every source + edit; the
   authoritative record. Every field the evaluator needs (prompts,
@@ -92,5 +105,8 @@ perceptual metrics.
   `physics` block carries the simulator settings, per-object parameters and
   scene geometry actually used, and `frames` carries the per-frame world
   matrices and velocities.
+- `benchmark_manifest.json` still carries `benchmark_name: pcve_benchmark_v1`,
+  the internal build id of this release. It is kept as-is so existing evaluator
+  code that keys off it keeps working; `PCVE-RigidBench` is the published name.
 - Model predictions are **not** part of this release; the repository contains
   only the benchmark data.