Researchers proposed observable-specific validity records for scientific AI. These link a model to its reference, error measure, tolerance, domain, and interventions. Certificates for a perturbed oscillator recovered distinct convergence orders. Hubbard dimer relations resolved differing coupling domains for energy and response. Dimensional homogeneity cut energy scale transfer error from 52.3% to 0.0768%. It raised the median fraction satisfying gap and response tolerances from zero to 100%. Paired transformations verified a protected scaling relation. A coupling crossover identified extra information needed for response transfer. A causal model recovered positive intervention response under strong negative observational association. The team derived an exact control query requirement scaling as x^-2k for mechanisms differing at order x^k. They verified its success curve numerically. A 256-task interface enables independent checks.

• 52.3% to 0.0768% energy scale transfer error reduction
• 16 repeated neural fits
• 256-task evaluation interface
• scaling as x^-2k for control query requirement