Add HF 0420 balanced validation package README

validation/hanoi_0420_balanced/README.md

@@ -0,0 +1,226 @@
+# GNN Hanoi 0420 Balanced Validation Package
+
+This folder contains a self-contained validation-data package for paired evaluation of a pure fine-tuned video world model versus a graph-conditioned world model on Tower-of-Hanoi disassembly episodes.
+
+This artifact is for validation only. It does not contain model checkpoints, generated predictions, inference outputs, benchmark tables, or final metric claims.
+
+## Files in this Hugging Face folder
+
+Expected files:
+
+```text
+validation/hanoi_0420_balanced/
+  README.md
+  GNN_Hanoi_0420_balanced_validation_package.zip
+  GNN_Hanoi_0420_balanced_validation_package.zip.sha256
+  GNN_Hanoi_0420_balanced_validation_package.zip.test.txt
+  GNN_Hanoi_0420_balanced_validation_package.zip.contents.txt
+```
+
+The ZIP contains:
+
+```text
+GNN_Hanoi_0420_balanced_validation_package/
+  README.md
+  MANIFEST.json
+  episodes.csv
+  metadata/
+    session_metadata.json
+    session_plan.json
+    perception_log.jsonl
+    setup_snapshot.yaml
+  data/
+    episode_00.zip
+    episode_07.zip
+    episode_10.zip
+    episode_12.zip
+    episode_14.zip
+    episode_15.zip
+  config/
+    val_0420_episode_manifest.jsonl
+    test_split_0420_balanced.txt
+```
+
+## Source
+
+The package is built from this same Hugging Face dataset repository:
+
+```text
+ChangChrisLiu/GNN_Disassembly_WorldModel
+```
+
+Source session:
+
+```text
+hanoi/session_hanoi_0420_132840
+```
+
+Selected validation episodes:
+
+```text
+episode_00  classical    6 moves
+episode_10  classical   12 moves
+episode_14  single_ring  3 moves
+episode_07  single_ring  7 moves
+episode_12  rearrange    4 moves
+episode_15  rearrange    6 moves
+```
+
+This gives a compact balanced validation subset:
+
+- 2 classical tasks
+- 2 single-ring tasks
+- 2 rearrange tasks
+
+## Why this package exists
+
+This package is intended to freeze a clear validation subset for later model comparison. It avoids using local 0423 data because that local run has known action/proprio recording issues. HF 0420 is preferred because the selected episode ZIPs include synchronized robot states, side RGB/depth frames, and graph annotations.
+
+Before reporting benchmark numbers, verify that `session_hanoi_0420_132840` was held out from training and model selection for every compared checkpoint. If either checkpoint trained on these episodes, do not call the resulting metrics validation performance.
+
+## Data inside each episode ZIP
+
+Each included episode ZIP was locally validated for ZIP integrity and required schema. Each selected episode contains:
+
+```text
+metadata.json
+robot_states.npy
+timestamps.npy
+side/rgb/frame_*.png
+side/depth/frame_*.npy
+annotations/side_graph.json
+annotations/side_masks/frame_*.npz
+annotations/side_embeddings/frame_*.npz
+annotations/side_depth_info/frame_*.npz
+annotations/side_robot/frame_*.npz
+```
+
+Important action-field note:
+
+These selected episode ZIPs do not bundle a separate `robot_actions.npy`. For any action-conditioned inference or diagnostics, derive per-frame action deltas from robot states:
+
+```python
+robot_action_t = robot_states[t + 1] - robot_states[t]
+```
+
+Use the same source/eval frame interval consistently when deriving action deltas for model inputs.
+
+## Graph information
+
+Yes, graph information is included. Each selected episode ZIP contains:
+
+```text
+annotations/side_graph.json
+```
+
+This is the ground-truth graph source for the Hanoi objects/rings and should be used for graph-conditioned validation inputs and graph-based scoring. Do not use HSV proxy labels as the final graph ground truth.
+
+Graph modes must be labeled explicitly in later experiments:
+
+- `gt_anchor_graph`: only current/anchor GT graph is used
+- `oracle_future_graph`: future GT graph is used; upper-bound only
+- `predicted_graph`: deployment-style recursive graph prediction
+
+## Later inference protocol
+
+This ZIP is data-only. Later inference should use a paired protocol:
+
+1. Build a window-level inference manifest from the selected episodes.
+2. For each window, run both models with identical:
+   - episode/window
+   - anchor/context frames
+   - goal prompt
+   - seed
+   - output FPS
+   - resolution
+   - sampler/chunking settings
+   - generation length
+3. Generate one recursive 10-second rollout per window.
+4. Evaluate cumulative prefixes from that same rollout:
+   - 1s
+   - 3s
+   - 6s
+   - 10s
+
+Do not generate separate independent 1s/3s/6s/10s clips unless explicitly labeling that as a separate non-recursive protocol.
+
+Recommended timing contract:
+
+```text
+source FPS: 30
+eval/model FPS: 5
+frame interval: 6 source frames per eval frame
+10s rollout: 50 eval frames
+1s prefix: 5 eval frames
+3s prefix: 15 eval frames
+6s prefix: 30 eval frames
+10s prefix: 50 eval frames
+```
+
+## Later metric protocol
+
+Valid immediately after predictions exist:
+
+- ROI PSNR
+- ROI SSIM
+- ROI LPIPS
+- ROI motion error
+- ROI color metric
+- ROI PSNR min/std/drift slope
+- whole-frame backup metrics
+
+Recommended ROI:
+
+```text
+x = 25-75%
+y = 20-80%
+```
+
+Graph metrics require generated-frame labeler validation before final claims:
+
+- ring position accuracy
+- exact state match
+- held-state F1
+- edge/is_locked F1
+- state edit distance
+- illegal-stack rate
+- task success
+
+## Local verification result before upload
+
+The package ZIP was created and verified locally.
+
+Package SHA256:
+
+```text
+6a9a42ed346df7a1139344e351ce24daa2b82957dd1feeb0ff41c8ff2e168c3d  GNN_Hanoi_0420_balanced_validation_package.zip
+```
+
+ZIP test result:
+
+```text
+test of GNN_Hanoi_0420_balanced_validation_package.zip OK
+```
+
+Per-source-file checksums are included in the package manifest and in the original local download checksum log.
+
+## How to use
+
+Download and verify:
+
+```bash
+sha256sum -c GNN_Hanoi_0420_balanced_validation_package.zip.sha256
+unzip -t GNN_Hanoi_0420_balanced_validation_package.zip
+unzip GNN_Hanoi_0420_balanced_validation_package.zip
+```
+
+Then inspect:
+
+```bash
+cd GNN_Hanoi_0420_balanced_validation_package
+cat README.md
+cat episodes.csv
+cat config/test_split_0420_balanced.txt
+```
+
+The package is ready for validation dataset preparation and later paired inference/evaluation. It is not itself a final benchmark-result artifact.