Evidence Eligibility Shapes Reward Comparisons in Scientific Experiment Allocation

Scientific agents that allocate experiments are often compared by the reward they optimize, but such comparisons also depend on which measurements may count as evidence. Across 8,192 Gaussian campaigns, a confirmation-only evidence rule gives certificate reward a 23.01 percentage-point (pp) advantage in true findings over maximum-observation reward. Admitting every valid observation, a post hoc control, shrinks the gap to 2.46 pp after matched reselection; it replicates on fresh seeds (2.24 pp), and a tuned evidence-progress heuristic stays stronger. A pre-declared study then embeds the controllers in a scripted evidence-contract executor, a non-LLM program agent that works through typed tools, a claim validator and a ledger, deciding when to open, search, confirm, claim or abandon campaigns under a shared 1,536-credit budget. With eligibility controlled, the certificate-reward agent yields an extra 2.47 findings per program (realization-resampled 95% CI 1.07–3.85), through speed and reinvestment, not a better reward. A pre-declared matched-spend follow-up still detects 1.67 and 4.94 findings per program at 13 and 19 campaigns, but only through observation controllers whose fixed search phase exceeds the allowance; the controller whose phase fits gives 0.37, below the margin. Eligibility adds 2.03 findings per program for the search-heavy observation agent, and the tuned heuristic still leads by 0.83. Separately, repeatedly inspected fixed-sample tests make 37.54% false claims versus 2.12% for a Gaussian e-process, which fails under a shared offset. Reward comparisons are interpretable only when evidence eligibility and spend are both fixed; we establish neither a superior learning method nor biological efficacy.