From the observations of Tycho Brahe to GW250114: three results in physics re-measured with the answer withheld

We ask whether AI agents can re-measure historical results of physics from the primary records under a written protocol that keeps the expected answer out of the agents' instructions and fixes in advance the comparison with modern values. In practice each analysis also fixed its plan before the reported result was computed, froze the result before that comparison (by hash in episode 1 and, for the decision-setting result, in episode 5; in episode 3 only by commit order, beside a control carrying the same physics, and its headline was refitted after the checker's comparison), passed each measurement to a separate re-analysis agent and kept a record of every run; Section 2 says where the record falls short. The paper contains three measurements, one comparison and one feasibility test. Given Tycho Brahe's Mars logs and no model family, an agent chose two Keplerian orbits about one common point, which predicted the withheld years 1596–1601 to a median of 2.46′. Newcomb's Mercury transit contacts of 1677–1881, fitted with a Newtonian model, require an extra perihelion advance of 43.71 ± 1.04″ per century, 0.7σ above general relativity. For GW250114 every one of the 40 analysis cells, over two independently run inspiral-only posteriors, gave P(ΔA > 0) of at least 0.846; once a single-mode bias in the earliest ringdown fits had been found and those cells withdrawn, the retained cells of the headline lane gave P = 0.86–0.98, with a minimum of 0.86 at 20 M_f, below the pre-registered 0.90, so the rule withheld the claim. A brief that leaked the shape of the answer, an offset attributed to the wrong cause, the single-mode bias and a failed run were each caught and reported. The code, plans, reports and run logs are in the paper's repository, github.com/FavioVazquez/tycho-to-gw250114, whose remeasured/ folder is a byte-identical copy of the time-stamped record: the road-remeasured branch of github.com/FavioVazquez/roadtoreality at commit b6b0b05. The git-ignored ringdown posterior samples and a snapshot of that commit are in the dataset record doi:10.5281/zenodo.23410453.