Third-party notices

Original ReCAP additions use Apache-2.0. The following bundled components retain their upstream terms. Modifications include portable local asset paths, launcher integration, dataset adaptation and the ReCAP experiment functions; original attribution headers are retained.

Component	Bundled use	License text
iVideoGPT, 2024 THUML @ Tsinghua University	Tokenizer, processor and training implementation	MIT
RLVR-World, 2025 THUML @ Tsinghua University	Action-conditioned world-model training and inference base	MIT
vector-quantize-pytorch, Phil Wang	Finite scalar quantization implementation embedded by upstream	MIT
PerceptualSimilarity, Richard Zhang et al.	Stripped LPIPS training utility	BSD-2-Clause
Paella, Dominic Rampas	Discriminator implementation embedded by upstream	MIT
Diffusers, Hugging Face	Adapted VAE/model components	Apache-2.0
taming-transformers, CompVis	VQ/LPIPS utility ancestry embedded by upstream	MIT

SOURCE_PROVENANCE.json records hashes of the historical experiment source imported into the candidate. It is an import audit, not the current release file manifest. The precise upstream commit of every historical local file was not recorded; no commit identity is fabricated. SOURCE_MANIFEST.json is regenerated for the current source archive.

Starting checkpoints are thuml/rt1-compressive-tokenizer and thuml/rt1-world-model-multi-step-rlvr. Their model cards identify MIT licensing. Derived model exports carry separate cards describing adaptation and pairing; they do not claim to be the original RT-1 checkpoints.

Dataset images and illustrative GIFs retain source dataset attribution and terms; no ownership or blanket Apache license is asserted over them. See data sources and case provenance. LPIPS/VGG metric weights and raw datasets are downloaded from their publishers and are not bundled as ReCAP model weights.