Provenance Record

Strong Analysis piece (1,007 words, within the 800-2,000 range). Clean Overview / What We Know / What We Don't Know / Analysis structure. Technical concepts (PbRL, optimal transport, video foundation models, pseudo-labeling) are explained accurately and accessibly without overstating the result. The 'What We Don't Know' section is exemplary: it explicitly flags the unnamed foundation model, the missing benchmark/baseline names, the absent margins, and the single-outlet provenance of the real-robot numbers.

All 5 source snapshots read from disk and verified. source-0 (Mirage News, status 200): confirms author roles/affiliations ('Tung M. Luu, PhD candidate, first author'; 'Hwanhee Kim, M.S candidate, second author'; 'Professor Chang D. Yoo'; School of Electrical Engineering), ICML 2026 at COEX Seoul in July, and the top-0.7%/168/23,918 stat. source-1 (icml.cc official program, status 200): paper title verbatim, full abstract present, all four abstract quotes ('its scalability is hindered by high labeling costs'; 'a semi-supervised framework that learns effective reward functions from only a handful of labels'; 'leveraging optimal transport to align visual trajectories within the rich representation space of' ViFMs; 'high-fidelity pseudo-labels for large amounts of unlabeled data'; 'outperforms state-of-the-art offline PbRL methods under limited feedback budgets'; 'visual distractors') reproduced verbatim, VOTP expansion 'Video-based Optimal Transport Preference' confirmed, first author listed as 'Minh-Tung Luu'. source-2 (OpenReview, status 200): all seven keywords confirmed verbatim and in the exact order the article lists them; first author rendered 'Tung Minh Luu'. source-3 (The Brighter Side of News, status 200): every experimental specific verbatim — 50 demonstrations, keyboard teleoperation, 50 percent success rate, five preference labels for LiftBanana, 10 for DrawerOpen, VOTP 80% / 70% — plus the 168/23,918/0.7% figures and the Yoo quote 'The core of physical AI is making machines understand human intentions and choose the correct actions.' source-4 (Seoul Economic Daily, status 200): 'around 10 good demonstration videos and bad demonstration videos' quote verbatim and the applications list (robotic arm, humanoid, autonomous vehicles, smart factories, drones, surgical robots, screen-viewing software agents). Rule 9 satisfied: abstract quotes are cited to ICML/OpenReview; experimental specifics are cited to the outlet (Brighter Side / Seoul Economic Daily) carrying them verbatim.

No hallucinations detected. Critically, the spurious arXiv ID 2602.11393 (a separate Verghese/Atkeson paper) does NOT appear anywhere in the article or in any of the five source snapshots — confirmed absent. The author-name discrepancy (ICML: 'Minh-Tung Luu'; KAIST/Mirage: 'Tung M. Luu'; OpenReview: 'Tung Minh Luu') is openly disclosed in the 'What We Don't Know' section rather than silently resolved, satisfying Rule 7. The one recoverable issue: the body sentence in 'What We Know' asserts 'Tung M. Luu is the first author' (the KAIST rendering) while the authoritative ICML program lists 'Minh-Tung Luu' first — a clarification note makes the canonical first-author rendering explicit for readers.

High-quality, fully source-verified Analysis. Paper identity, author names, and experimental numbers all check out against the snapshots; the wrong arXiv paper was correctly excluded and the author-name variance was honestly disclosed. APPROVE_WITH_CORRECTIONS: the single recoverable item — the body's first-author rendering vs the official ICML program — is honestly summarizable in one corrections note, and the article publishes with that public trail.