Generation: TRAIRC-style mini-binder search with physics-informed regularization, foldability (first-pass) scoring, and ESM-2–based sequence plausibility (normalized within the candidate pool). Final sequence selection is score-driven (no LLM prompts for ranking).

Boltz-2 validation: Each retained core predicted as a complex against four RBX-1-related structural contexts; ensemble metrics = mean over four runs (ipTM, model confidence, interface iPDE). Per-metric values min–max normalized within the Boltz-scored pool; lower iPDE mapped to higher score.

Structural core rank (22 slots): 0.6·ipTM_norm + 0.2·confidence_norm + 0.2·iPDE_norm (ensemble means).

Pipeline core rank (8 slots): 0.30·ipTM_norm + 0.10·confidence_norm + 0.10·iPDE_norm + 0.20·physics_norm + 0.15·foldability_norm + 0.15·ESM2_norm; filled from remaining binders with mean ensemble ipTM ≥ 0.38 (neutral 0.5 if ESM2_norm missing). If fewer than eight qualify, pad from next structural ranks.

Portfolio: 30 cores total + 35 single-point variants of rbx1_v2_1390 (RNG seed 42) + 35 single-point variants of rbx1_v2_2336 (RNG seed 43); cysteine and proline disallowed as replacement residues.

ColabFold cross-check (orthogonal): AlphaFold2-multimer v3, MMseqs2 MSA mode, num_models=1, num_recycle=3, on the fifteen elite cores matching submission ranks 1–15 (heteromeric FASTA binder:target). ColabFold iptm/ptm reported for that subset only; not numerically equivalent to Boltz ipTM.