The NOVA framework models iterative generate-verify-accumulate-retrain loops as adaptive sampling over knowledge space and proves that finite-domain coverage occurs only under specific non-contamination conditions while violations trigger distinct failure modes including a contamination trap driven by shrinking valid mass and fixed false-positive rates [arXiv:2605.15219]. Under tail-equivalence to Zipf laws with exponent alpha greater than 1, cumulative generation cost satisfies R_cum(D) equals Theta of c_gen times D to the alpha, quantifying asymptotic diminishing returns.

This bound extends prior analyses of automated scientific search by showing Good-Turing estimators capture only local batch diversity rather than historically undiscovered valid mass, a distinction absent from earlier empirical reports on theorem-proving and molecule-generation pipelines. Human amplification via targeted guidance and verification is shown to be most efficient precisely at autonomous exploration barriers.

The result formalizes fundamental ceilings on scalable scientific automation, linking the derived cost exponent directly to limits observed in related self-improvement systems where verification noise eventually dominates genuine signal.