Candidate pipeline results

Phase 7 — the restartable generate → annotate → filter → select → report pipeline run on the per-prompt RLXF specialists of all three FcGPT bases (21 profiles × 200 sequences each). Selection keeps the top 25/profile that are valid, novel, on-profile (verifier full-match), ranked by developability. 1380 final candidates.

Click any profile for the deep view: per-receptor binding vs target, quality-metric comparison across bases, edit-distance + expression distributions, a mutational heatmap, and the top selected sequences with their mutations. The in-training removed row re-confirms the Phase-5 memorization ranking — retraining_v2 is the most novel base.

Per-base funnel

metricpaper_reproductionretraining_v1retraining_v2
generated420042002100
valid415940471972
kept380738661244
selected525525330
selected expression0.8910.8660.802
in-training removed345176120

Profiles (click for the deep per-profile view)

-----01------10---01------10----0100000001000001010001100100011101001000011100000111000101111111101111111100000011000001110011111111000011110001111101111111100111111110