Update rebuttal evaluation artifacts and documentation

Add common-interface SC results, Matrix SC-versus-horizon analysis, Endpoint FID/KID evaluation, frozen manifests, tests, and reproduction commands.

README.md

@@ -1,86 +1,96 @@
 ---
-pretty_name: Matrix-Game 2.0 Self-Consistency Reproduction
+pretty_name: Reasoning-Structured Videos Evaluation Artifacts
 tags:
   - world-model
   - video-generation
   - benchmark
+  - compositional-consistency
   - reproducibility
 ---
 
-# Matrix-Game 2.0 Self-Consistency Reproduction
+# Reasoning-Structured Videos: Evaluation Artifacts
 
-This README explains how to reproduce the Matrix-Game 2.0 self-consistency (SC) row in *Reasoning-Structured Videos: A Stratified Diagnostic Suite for Compositional Consistency in World Models*.
+This page hosts evaluation artifacts for **Reasoning-Structured Videos: A Stratified Diagnostic Suite for Compositional Consistency in World Models**.
 
-All experimental data, generated videos, and evaluation files required for this reproduction are available on the current dataset page under **Files and versions**:
+The benchmark tests whether action-conditioned video world models respect three trajectory relations:
 
-<https://huggingface.co/datasets/VideoWorldmodel/Evaluation>
+- **Inverse:** a path followed by its inverse should return to the initial state.
+- **Loop:** a closed path should return to the initial state.
+- **Equivalence:** two different paths reaching the same state should produce matching endpoints.
 
-No external dataset or ground-truth video is required for this reproduction.
+We report **self-consistency (SC)** as the primary cross-model diagnostic and use GT-anchor and distributional metrics as complementary evidence.
 
-## What is reproduced
+## Files
 
-The statistical unit is one relation graph:
+| File | Description |
+| --- | --- |
+| `matrix_game.zip` | 1,371 released Matrix-Game 2.0 SC rollout videos |
+| `evaluate_matrix_game_sc.py` | Standalone Matrix-Game SC reproduction script |
+| `Reasoning-Structured-Videos-Rebuttal-main.zip` | Frozen R20/R50 manifests, minWM/HY-WorldPlay camera adapters, metric evaluators, tests, and result summaries |
 
-- **Inverse-SC:** generated first frame versus generated final frame.
-- **Loop-SC:** generated first frame versus generated final frame.
-- **Equivalence-SC:** generated final frame of branch A versus generated final frame of branch B.
+Reference videos and trajectory metadata are hosted separately in [VideoWorldmodel/ReasoningStructureTestset](https://huggingface.co/datasets/VideoWorldmodel/ReasoningStructureTestset).
 
-The released data contains:
-- 448 Inverse graphs
-- 445 Loop graphs
-- 239 Equivalence A/B pairs
+## Main results
 
-Before evaluation, the script verifies the expected file counts and checks the integrity of the Equivalence A/B pairings.
+### Common camera-trajectory SC (frozen R20)
 
-## Download the experimental data
+minWM and HY-WorldPlay receive the same frozen graph IDs and camera trajectories through their official pose-control pathways. Each model completed 80/80 rollouts with zero evaluation failures.
 
-The MP4-only reproduction archive is available under **Files and versions** on the current dataset page.
+| Model | Inverse LPIPS / PSNR | Loop LPIPS / PSNR | Equivalence LPIPS / PSNR |
+| --- | ---: | ---: | ---: |
+| HY-WorldPlay 1.5 AR Distill | **0.4038 / 16.12** | **0.4562 / 15.48** | **0.3542 / 17.37** |
+| minWM | 0.7229 / 10.46 | 0.7442 / 11.08 | 0.4227 / 14.05 |
 
-Install the Hugging Face CLI and download the archive:
-```bash
-python -m pip install -U huggingface_hub
+R20 contains 20 graphs per relation: 60 graph units and 80 videos because every Equivalence graph has A/B branches. The code archive includes graph-bootstrap intervals and the corresponding GT-anchor summaries.
 
-hf download VideoWorldmodel/Evaluation \
-  --repo-type dataset \
-  --include "MatrixGame2_SC_videos.zip" \
-  --local-dir .
-```
+### Matrix-Game Inverse SC versus revisit horizon
 
-Extract the archive:
-```bash
-unzip MatrixGame2_SC_videos.zip -d data/matrix_game_sc
-```
+All 448 Inverse graphs are evaluated at matched-state revisits with increasing action separation. “Physical frame” denotes the zero-indexed position in the saved Matrix MP4. Matrix exports 357 frames for the complete 40-action schedule, so action boundary `k` is mapped to physical frame `round(356k/40)`, anchoring both endpoints and preserving outward/return symmetry.
 
-The expected directory layout is:
-```
-data/matrix_game_sc/
-|-- inverse_easy/      # 246 MP4 files
-|-- inverse_hard/      # 202 MP4 files
-|-- loop_easy/         # 247 MP4 files
-|-- loop_hard/         # 198 MP4 files
-`-- equivalence/       # 478 MP4 files = 239 A/B pairs
-```
+| Revisit horizon | LPIPS (95% CI) | PSNR dB (95% CI) |
+| ---: | ---: | ---: |
+| 10 actions | 0.4851 [0.4652, 0.5053] | 14.03 [13.53, 14.53] |
+| 20 actions | 0.6106 [0.5976, 0.6234] | 11.46 [11.14, 11.77] |
+| 30 actions | 0.6685 [0.6584, 0.6784] | 10.71 [10.45, 10.96] |
+| Full rollout (40 actions) | 0.7062 [0.6973, 0.7151] | 10.45 [10.21, 10.69] |
 
-If the dataset was cloned using Git LFS, verify that the downloaded files are real MP4 files rather than small Git LFS pointer files.
+The paired full-minus-10 change is +0.2211 LPIPS [0.1988, 0.2434] and -3.58 dB PSNR [-4.02, -3.15]. Adjacent boundary-rounding schemes preserve the monotonic trend and change intermediate means by at most 0.005 LPIPS / 0.10 dB.
 
-## Environment
+### Endpoint distribution fidelity
 
-Python 3.10 or newer is recommended.
+Endpoint FID/KID compare generated and matched GT logical `raw359` endpoint sets using clean-fid Inception-v3 pool3 features.
 
-Create and activate a virtual environment (Linux/macOS):
-```bash
-python -m venv .venv
-source .venv/bin/activate
-```
+Matrix-Game, complete available output set:
 
-On Windows:
-```powershell
-.venv\Scripts\activate
-```
+| Relation | Graphs / images | Endpoint FID ↓ | KID ×1000 ↓ |
+| --- | ---: | ---: | ---: |
+| Inverse | 448 / 448 | 134.94 | 23.018 |
+| Loop | 445 / 445 | 141.33 | 24.362 |
+| Equivalence | 239 / 478 | 145.55 | 22.943 |
+
+minWM, frozen R50 SC outputs:
+
+| Relation | Graphs / images | Endpoint FID ↓ | KID ×1000 ↓ |
+| --- | ---: | ---: | ---: |
+| Inverse | 50 / 50 | 242.46 | 47.860 |
+| Loop | 50 / 50 | 249.40 | 54.173 |
+| Equivalence | 50 / 100 | 240.23 | 50.359 |
+
+Endpoint FID is an image-set metric and is not numerically comparable to the clip-level FVD reported elsewhere. The Matrix analysis also includes 20/50/100/200/full graph sensitivity and a GT-vs-GT finite-sample floor. These endpoint metrics complement, rather than replace, paired SC LPIPS/PSNR.
+
+## Quick reproduction: Matrix-Game SC
+
+This path requires only the files on the current page; no GT dataset is needed.
 
-Install the required dependencies:
 ```bash
-python -m pip install --upgrade pip
+python -m pip install -U huggingface_hub
+
+hf download VideoWorldmodel/Evaluation \
+  matrix_game.zip evaluate_matrix_game_sc.py \
+  --repo-type dataset \
+  --local-dir .
+
+unzip matrix_game.zip -d data/MatrixGame2_SC_videos
 
 python -m pip install \
   "numpy>=1.26,<3" \
@@ -88,90 +98,60 @@ python -m pip install \
   "torch>=2.2,<3" \
   "torchvision>=0.17,<1" \
   "lpips==0.1.4"
-```
 
-For a specific CUDA build, first install the matching torch and torchvision wheels by following the official PyTorch installation instructions. Then install NumPy, OpenCV, and lpips==0.1.4 without reinstalling PyTorch.
-
-The first LPIPS run downloads the official AlexNet weights.
-
-## Reproduce the paper results
-
-Run the complete evaluation:
-```bash
 python evaluate_matrix_game_sc.py \
-  --data data/matrix_game_sc \
+  --data data/MatrixGame2_SC_videos \
   --output results/matrix_game_sc \
   --lpips \
   --device auto \
   --check-paper
 ```
 
-Available device options include:
-- `--device auto`
-- `--device cpu`
-- `--device cuda`
-- `--device cuda:0`
+If the archive creates one additional top-level directory, point `--data` to the directory that directly contains the five relation folders.
 
-Example: run the evaluation explicitly on CPU:
-```bash
-python evaluate_matrix_game_sc.py \
-  --data data/matrix_game_sc \
-  --output results/matrix_game_sc \
-  --lpips \
-  --device cpu \
-  --check-paper
-```
+Expected values:
 
-For a faster PSNR-only audit, omit `--lpips` and `--check-paper`:
-```bash
-python evaluate_matrix_game_sc.py \
-  --data data/matrix_game_sc \
-  --output results/matrix_game_sc_psnr \
-  --device auto
-```
+| Relation | Graph N | Recomputed LPIPS / PSNR | Paper LPIPS / PSNR |
+| --- | ---: | ---: | ---: |
+| Inverse-SC | 448 | 0.7061 / 10.4476 | 0.71 / 10.45 |
+| Loop-SC | 445 | 0.7173 / 10.6227 | 0.72 / 10.62 |
+| Equivalence-SC | 239 | 0.5918 / 12.5732 | 0.59 / 12.57 |
+
+The evaluator verifies file counts and Equivalence pair integrity, and writes per-graph scores, graph-bootstrap confidence intervals, an audit, and a paper-value check.
 
-## Expected results
+## Reproduce the additional analyses
 
-| Relation       | Graph N | Recomputed LPIPS / PSNR | Paper LPIPS / PSNR |
-|----------------|---------|-------------------------|--------------------|
-| Inverse-SC     | 448     | 0.7061 / 10.4476        | 0.71 / 10.45       |
-| Loop-SC        | 445     | 0.7173 / 10.6227        | 0.72 / 10.62       |
-| Equivalence-SC | 239     | 0.5918 / 12.5732        | 0.59 / 12.57       |
+Download and unpack the code artifact:
 
-The `--check-paper` option succeeds when all six computed values reproduce the precision reported in the paper.
+```bash
+hf download VideoWorldmodel/Evaluation \
+  Reasoning-Structured-Videos-Rebuttal-main.zip \
+  --repo-type dataset \
+  --local-dir .
 
-Two Equivalence pairs are exact pixel matches. Their PSNR values are infinite, so they are recorded in `psnr_exact_match_n` and excluded from the finite PSNR mean. This produces the reported Equivalence PSNR.
+unzip Reasoning-Structured-Videos-Rebuttal-main.zip
+cd Reasoning-Structured-Videos-Rebuttal-main
 
-## Outputs
+python -m pip install -r requirements-eval.txt
+python -m unittest discover -s tests -v
+```
 
-The output directory contains:
-- `per_graph.csv`: One row per Inverse/Loop graph or Equivalence A/B pair
-- `summary.csv`: Means, standard deviations, and graph-bootstrap 95% CIs
-- `paper_check.csv`: Computed values compared with the reported paper values
-- `audit.json`: Metric definitions, counts, versions, and bootstrap settings
-- `report.md`: Compact Markdown results table
+The repository README documents four reproducibility paths:
 
-Confidence intervals use 10,000 graph-level bootstrap resamples with seed 2026.
+1. recompute SC/GT LPIPS and PSNR from minWM or HY-WorldPlay rollout directories;
+2. recompute Matrix-Game Inverse SC versus revisit horizon directly from `matrix_game.zip`;
+3. recompute relation-wise endpoint FID/KID;
+4. run the full Matrix endpoint audit using `matrix_game.zip` and the GT dataset.
 
-### Metric settings
-- LPIPS uses `lpips==0.1.4` with AlexNet.
-- LPIPS inputs are RGB images scaled to [-1, 1].
-- PSNR is computed on decoded RGB uint8 endpoint frames.
-- PSNR uses MAX=255.
-- No resizing is applied.
-- Exact pixel matches are reported separately from the finite PSNR mean.
+The code archive contains frozen result summaries but not the large minWM/HY-WorldPlay rollout videos or model checkpoints. Regenerating those videos requires the official [minWM](https://github.com/shengshu-ai/minWM) or [HY-WorldPlay](https://github.com/Tencent-Hunyuan/HY-WorldPlay) repository, its checkpoint, its official environment, and suitable GPUs.
 
-## Tests
+## Frozen identifiers
 
-Run the test suite with:
-```bash
-python -m unittest discover -s tests -v
-```
+- Public subset seed: `2357`
+- R20 manifest SHA-256: `f8307b78ffb4633b1d6ca6340498a9db615dca4cc45a045d16139abf4375abdf`
+- Nested R20/R50 manifest SHA-256: `c1b3e051ec861f6babd00f67cf7c31d0ceba590d93a6fac1ac6a768b40448db4`
+- `matrix_game.zip` SHA-256: `caec3d8deb20cffbc645f0aeeaa2af425536715eef935b17347cdfab728a88d4`
 
 ## Scope
 
-This code reproduces the Matrix-Game 2.0 SC endpoint metrics only.
-It does not claim to reproduce the paper’s GT-anchored metrics or FVD, which require additional generation-context and clip-sampling details.
-
-All data required for the SC endpoint reproduction described in this README is available on the current dataset page.
-```
+The released experiments evaluate deterministic camera trajectories in static scenes. They support relation-specific diagnostic conclusions under the stated protocol; they are not intended as a universal ranking of world-model architectures.

Reasoning-Structured-Videos-Rebuttal-main.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:531f9241e61808d39f491165872fc710a4a9b1b2e48d7ec60c3dbd4faedc2310
-size 205510
+oid sha256:c3e6ee7b98ee35ee79259da1e3419211d6c049403a33ebb9d691292dc94bf834
+size 461247