World‑action models increasingly employ test‑time scaling, adding inference computation for candidate selection and adaptive allocation to improve control. We audit the three essential premises—whether candidates carry exploitable signal, whether nominal compute savings become real speedup, and whether evaluation remains reliable—on Cosmos Policy and RoboCasa using a premise‑first, pre‑registered methodology.\n\nOn clean data, a learned value selector underperforms random by $1.66$ percentage points; a hindsight oracle gains only $1.40$ points, with episode‑level intervals spanning zero; and about $82\%$ of states produce identical outcomes across candidates, leaving little headroom for selection.\n\nToken pruning reduces per‑forward FLOPs by up to $47\%$, yet the measured end‑to‑end speedup peaks at $1.15\times$ on the deployed path (or $1.25\times$ when decoding is excluded), short of the $1.5\times$ target. The adaptive‑allocation experiment remains inconclusive under a power gate.\n\nWe also identify and fix a silent rendering corruption that escalates sharply along compute‑heavy planning paths while leaving the official direct‑policy result unchanged. We release a detector, the fix, and accompanying scripts.\n\nThese findings demonstrate that test‑time scaling should be audited as a chain of premises—decision signal, realized efficiency, evaluation integrity—rather than judged solely by success rate. Conclusions are scoped to the tested configuration.\n\nReview