UWSim LLM+PCG Underwater RGB-D Dataset v2-r2

LLM+PCG-generated underwater RGB-D transition corpus with fail-closed publish gates and per-file SHA-256 pinning (companion release of dataset-v2-r1, DOI 10.5281/zenodo.21819582). **Key Statistics:** - Collection: 20 episodes requested -> 19 COMPLETE / 1 collection-period rendering failure (llmpcg-013, record kept in the collection root); all 19 passed the fail-closed dataset-v2 gate (0 gate FAILED) - Frames: 9139 (481/episode x 19, 24 s @ dt=0.05 s) - Resolution: 256x256 RGB-D (center-cropped + area-resized from 1280x720 renders) - Archive size: 4211.0 MB (archive SHA-256 0c79b9c730bcec02aa67ea518b916f9d045daaf9eca5afe084d8673de356ce65) - Sealed source dataset_sha256: 9fc6c6c9132060eb4e749ff8618f3070c05ea8969a9892c1d0c1a5ba56d641ee - Independent verifier: tools/verify_dataset_v2.py (27,436 files re-hashed PASS) **Generation pipeline:** real LLM (DeepSeek deepseek-v4-flash) per-episode intents -> validated PCG compilation (r39 scene binding) -> Python Fossen dynamics -> packaged UE5 real-RHI rendering -> dataset-v2 fail-closed gates. **Honest boundaries (mandatory reading):** - sim-only; physics_authoritative=false; frames are UE-rendered procedural scenes, NOT real underwater imagery and NOT ground truth. - Frames are not re-renderable byte-for-byte (GPU/driver dependent); the shipped bytes + hashes are the reference. - This dataset evidences PIPELINE CAPABILITY only. The controlled test of "LLM improves world-model training value" (H3/H3-R2, 3 seeds x 30, real RHI) returned NOT_SUPPORTED as a power-sufficient informative null and is sealed with PI countersignature. No LLM-value claim may be derived from this dataset. - URI-neutralized distribution copy; migration logged in uri_rewrite_log_r2.json. **License:** CC-BY-4.0