The value of early feedback in AI-guided experiment selection

An AI system can choose promising experiments without making good use of their results. We asked whether receiving results earlier actually improves its subsequent choices. In each paired campaign, the same tool-enabled AI started with the same 32 measured candidates. One branch saw their results immediately; the other saw them one decision later. Both then received all results from their own choices. Each branch selected 128 new candidates under the same evaluation budget. We ran 32 paired campaigns on each of two published laboratory datasets. Earlier results improved average selection quality for 2,679 Bcl-xL peptide-display records by 0.101 on the assay's 0–11 scale (adjusted p=0.0000143). They reduced average quality for 149,284 GB1 protein variants by 0.101 in selection-fitness units (adjusted p=0.00106). These scales measure different outcomes. Neither dataset established better final choices. A later descriptive analysis located most of each overall effect in the first batch; subsequent differences partly offset it. Earlier computational studies distinguished strong search performance from benefits attributable to feedback. A separate docking calibration failed its preset accuracy gate. All selection studies used existing measurements or computational scores; no new molecules were tested. The practical result is that feedback timing can help or hurt selection under the tested conditions. Evaluating every choice in a campaign reveals consequences that assessing only its final batch can miss. Preprint. Not peer reviewed. Supplementary information accompanies the manuscript. The deposit licenses these documents only; confidential source data, raw model traces and reviewer code/data packages are not included.