Good afternoon, San Francisco! Today, we're releasing Test-Time Structure-Space Search, a new OpenDDE inference algorithm for more effective co-folding when designing and selecting antibody drug candidates 🧬. 16% → 78%, that's top-1 success on antibody–antigen targets where OpenDDE usually fails, using Test-Time Structure-Space Search with just five correct contacts. Frozen model. No retraining. No gradients through the network. Test-time compute isn't only for LLMs. Rather than only picking the best of many samples, we search for better structures while the sampler is still denoising. Same OpenDDE, same five contacts, same budget of 25 candidates per target: • re-rank with the contacts: 33% • steer the sampling with them: 78% Re-ranking recovers only about a quarter of the gain. The rest comes from changing the sampling trajectory itself. How it works: • At scheduled denoising steps (about three dozen per sample), we check OpenDDE's predicted clean complex against the contacts. • If they fail, a solver moves the antibody as a rigid body: Refine (nudge locally) → Search (broad pose search, only if still failing) → Refine (polish the best pose). • A clash-aware energy refuses moves that create severe overlaps, and fused GPU kernels make thousands of pose evaluations per call practical. • The corrected structure goes back to the sampler. If the contacts already hold, nothing moves. Results: 805 held-out SAbDab targets where the base model usually fails (median DockQ < 0.23 over five samples). Top-1 success by number of known contacts: 0 (unguided): 16% and 5: 78%. With 5 contacts, mean top-1 DockQ goes 0.14 → 0.43 and the CAPRI-medium rate 11% → 39%. Pocket mode (5 epitope residues, nothing on the antibody side): 17% → 33%, versus 26% for re-ranking. A smaller gain, same pattern. Try it: Put contact pairs or epitope residues in the "constraint" field of your input JSON and run with --use_tfg_guidance true. Checkpoints and output format are unchanged, and inputs without a constraint behave as before. With the optional trunk cache, inputs that differ only in their constraint share one trunk. Examples for Fab, Fv and VHH are in examples/tfg. Fine print: Constraints are simulated from reference structures, so they are correct by construction. This measures what accurate interface knowledge is worth, not robustness to experimental noise, and wrong contacts can push success below unguided. SAbDab only. Success = DockQ ≥ 0.23, 25 candidates per target, 95% bootstrap CIs. Validated on one GPU with 200 diffusion steps. Code: github.com/aurekaresearch/Op… Tech Report: github.com/aurekaresearch/Op…