Optimiser Repairs for Living-Cell Self-Driving Laboratories: Destructive Measurement Subsumes Selection-Level Interventions

Closed-loop optimisation is effective for molecular discovery but transfers poorly to living cells. The standard diagnosis is that living samples violate assumptions the optimiser relies on, implying that repairing those assumptions should recover performance. We evaluate three established repairs, one per violated assumption, on a testbed whose cells, heterogeneity and rare-population abundance derive from real single-cell measurements. The time-varying kernel of Bogunovic et al. [4] does not separate from a stationary baseline at any drift rate; a sliding window helps only under severe drift; confirmation of the incumbent helps near an 11% contamination rate and is costly below it; and kriging-believer fantasies are inferior to plain exclusion of in-flight experiments at every non-zero latency. The largest effect we measure comes from none of these: a rule forbidding re-proposal of the previous batch is worth +0.23 in fraction of optimum reached (z = +10.5). Under the full pathology profile that rule has no effect at all, nor do two of the three repairs, for structural rather than statistical reasons. Separating interventions into modellevel and selection-level classes, we prove that under destructive measurement the platform’s own no-repeat constraint subsumes the entire selection-level class, rendering trajectories identical round for round, and an implementation check confirms the prediction in both directions. Destructive measurement, which the platform literature treats purely as a cost, supplies batch diversification at no charge, and migration to non-destructive readouts removes it.