Random-split rankings poorly predicted microfluidic surrogate rankings under shift
Source check: Source checked. Not independent scientific confirmation.
Finding: Across 107 held-out groups from six published tables, a random-split ranking of six learned prediction models showed little agreement with their rankings in a simulated adaptive evaluation. Mean Kendall correlation was 0.08, with a reported interval from −0.04 to 0.19.
Technology: XGBoost, random and quantile forests, Gaussian processes and neural surrogate models
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- Six fitted surrogates made predictions for previously recorded measurements. Comparing their errors under different evaluation procedures quantified how poorly random-split standings carried across the tested settings.
- Test and result
- The analysis compared rankings from random-split error with rankings from held-out group performance under the same offline simulated evaluation, using previously published measurements.
- Limitation / still unknown
- The comparisons reused finite public tables and modeled lab activity offline. The study did not perform a new laboratory experiment, and the interval does not rule out modest agreement.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.36395v1 · https://arxiv.org/html/2610.36395v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: Primary §6.4.3 and Table 20, frozen protocols and v1 front matter.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.