The Verifier Is Part of the Simulator: Evidence-View Gaming in Agentic Equation Discovery

A growing number of agentic pipelines search for governing equations by having an agent propose candidate laws and a verifier program check each against a simulated trajectory. The verifier sees only a numerical \emph{summary} of that trajectory: the data integrated against test functions on a time grid, with several choices of grid and test-function set on offer. That summary carries a discretization error of its own, so the verifier is part of the simulator, not a neutral readout of it. Whether the choice of summary can change the verdict has not, to our knowledge, been measured. On a benchmark of 1{,}800 episodes we freeze the trajectory, the candidate library, the verifier, and its thresholds, and leave the agent one move: which summary to admit as evidence. A public score rewards decisive verdicts and rises with the reported misfit whenever the library is rejected. Maximizing it selects the coarsest summary (no agents involved), which then declares the true law absent on all 1{,}500 episodes where the library does contain it. We call this \emph{evidence-view gaming}: reward hacking by choice of permitted summary. Language models reproduce it: three Claude models rarely err under an accuracy goal, but once the goal rewards the score or missing physics they take the summary their own trust reports call least trustworthy. On PDEBench the error flips direction, certifying a law with its diffusion term missing on summaries with too few test functions to see diffusion at all. Requiring several summaries to agree removes most of these errors, but abstains often and is not the best rule on the transfer: a safeguard, not a certificate. Evidence chosen after its verdict is known cannot alone authorize a claim.