v0.6.5

[0.6.5] - 2026-08-20

Fixed

• Predictions no longer depend on a target's position in a multi-target job. Two
  byte-identical inputs folded by one worker returned different affinity values
  (0.648724 / 0.722511 / 0.687149 for three copies of one CDK2 + ligand YAML); they now
  return the same value, bit-identical to folding either one on its own. Two causes.
  Ligand conformer generation left ETKDG's seed unset, so RDKit drew from a stream that
  advances on every embedding and the same SMILES got a different reference conformer
  each time it was parsed — that moved the structure prediction too, not just the
  affinity scalar, and it applies to BoltzGen as well. And the affinity checkpoint loads
  lazily inside the first affinity target of a run, which advanced the RNG the diffusion
  samples from, so target 1 sampled differently from every target after it.
  Ligand numbers move once with this fix: a pinned conformer is not the one the old
  stream happened to hand the first target, and the spread across conformers of one
  ligand is about 0.074 log10(IC50). Set TT_BIO_ETKDG_SEED to draw a different
  conformer. scripts/boltz2_affinity_batch_position_repro.py folds N targets in one
  process and fails if identical ones disagree, and tt-bio gates it as
  scripts/release_gate.py --model batch-position. CCD-bound ligands are unaffected.
  If you ran a multi-target SMILES-ligand job on 0.6.4 or earlier, re-run it: every
  target after the first one in that job got the wrong number.

• An affinity run's results.json reported only the structure leg's time, so the
  number was short by however long the affinity trunk ran. It now reports both.

Changed

• RFD3 designs in 91.4 s on a p150a, down from 105.1 s, which is 3.33x an H200's
  27.5 s and inside the 4x bar the performance page holds every model to. The row is
  no longer hidden there. (dff668d9)

Known issues

• Boltz-2 affinity on the trypsin parity leg differs about 2% from the CPU reference
  (2.552 against 2.606 log10(IC50)); the predicted pose is unaffected. The reference
  bound tightened this release because the conformer-seeding fix above removed noise
  that had been inflating it, so this is a pre-existing gap now visible, not a new one.
  See docs/implementation-parity.md.