Pixels to Proof

Journals now publish the data behind a paper — tables, and sometimes high-speed videos — but almost nobody checks a paper's printed numbers against them. Pixels to Proof does. You give it exactly what the journal published, nothing more. Muse Spark 1.3 reads everything and proposes a plan: what each file and table column means, every number the paper claims, what is printed on each video. Code checks every proposal; rejected ones go back to Muse Spark with the reason. Statistics are recomputed from the paper's own table, by arithmetic only. When a video is usable, SAM 3.1 tracks the animal, the agent finds the matching row of the data blind, from the jump's shape alone, and re-measures the take-off speed from physics, with self-checks on gravity and free flight. On 13 papers audited hands-off, 63 of 319 printed numbers reproduce to the printed digit, and every other one gets a named reason. It recognises angles that only add up as circular means, and values printed truncated instead of rounded. One beetle-larva jump is re-measured and REPRODUCED; two peacock-spider jumps are refused because the flights are not ballistic; for the male, the extra acceleration points against the motion, as air drag would. Twelve kinds of deliberately wrong model answers produced no wrong verdict.