v0.6.0

[0.6.0] - 2026-08-01

ESMFold2 now folds at the paper's protocol by default: 10 recycling loops and 100 requested sampling
steps, of which 68 execute after the sigma-schedule clip. The old default under-recycled by 3.3x and
burned roughly twice the diffusion compute for no measured quality gain. Folds that draw several
samples per batch pick up --max_parallel_samples on Protenix-v2, OpenDDE and ESMFold2, where the
flag had been silently ignored, and single-card folds sharing a host can now be given a CPU budget
with the new --host_threads — four co-resident folds previously aggregated only 2.56x of a possible
4.00x because each sized its thread pools to every core. A fold error is no longer truncated before
its tail, which matters more than it sounds: a clipped out-of-memory message had hidden the real
evidence and produced a wrong "Wormhole cannot run this" verdict.

Fixed

• c145a1a6 worker: stop truncating away the tail of a fold error.
• b62301f5 boltz2: width-based diffusion chunking. Upstream's N % mps + 1 chunk-count formula
  produced a single 1000-sample chunk at N=1000 / mps=5 — about 0.9 GB of L1 forward buffers — and
  ran out of memory at both mps 5 and mps 3. It now matches protenix.edm_sample's width<=mps
  semantics. 9f31125fc refuses to run rather than derive a wrong sibling seed past two chunks.
• 364bdd46 predict: --max_parallel_samples was a silent no-op for protenix-v2, opendde and
  esmfold2.
• f0fe2a72 esmfold2: thread the fold seed into the diffusion sampler's private RNG.
• fce156d4 protenix/opendde: broadcast the sample-invariant diffusion biases instead of
  replicating them.
• 45036687 tests: a script dropped into tests/ ran its scenarios at import and ended in
  sys.exit(), so pytest aborted collection with INTERNALERROR and the entire 209-test host suite
  silently never ran. Caught by this release's own gate.

Changed

• f8a00aed esmfold2 / esmfold2-fast default to the paper protocol (10 recycling loops, 100
  requested sampling steps). Explicit flags are honoured verbatim; other models are unchanged.
• 228dc8e5 rfd3: raise the design batch clamp to the memory bound it stands for.

Performance

RFdiffusion3 (43 commits, p19-p32): resident pair tables, a one-kernel head merge, build-once sparse
pair-bias construction, and an opt-in RFD3_TUNE_MATMUL calibration path. Two trace levers were
measured and recorded as negative rather than shipped (5a32a203), and 5945a944 fixes a traced
decoder handing out the buffer its own trace replays into. Measured effect at release: rfd3 runs at
0.1185 designs/s, +5.3% against its committed baseline.

Documentation

• 54682d89 RELEASING.md's accuracy floor table had drifted from release_gate.py:MODELS — it
  listed ESMFold2 at 4.0 A / TM 0.65 while the code has enforced 8.0 A / TM 0.40 since 88c14f3b.
  Re-synced and annotated with the source of truth.
• 6c63537b, 2fcf0bad document --host_threads and --max_parallel_samples, the two
  user-facing CLI flags that had shipped without ever reaching the options table.
• Profiling instrumentation (6 commits): measured Blackhole p150a roofline, ttnn.graph capture
  characterised as an instrument, real-model op counts (ESMFold2-Fast 1290 ops per diffusion step,
  whole-fold 43,291 ops), and a worked example showing ESMC-300M is DRAM-bandwidth-bound rather than
  matmul-bound.

Release gate (Blackhole P150a on tt-quietbox, card 2)

Host suite, run one process per test file: 200 passed, 22 skipped, 1 failed. The single failure,
test_protenix_confidence.py::test_confidence_device_resident_parity, is pre-existing: it produces
a bit-identical PCC of 0.9807124853748275 against its >0.99 bar at both this commit and v0.5.0, so it
predates this release. Packaging guard: 15/15 expected data files and 31/31 declared dependencies
present in the wheel and sdist. UX gate: PASS on all 11 surfaces.

Accuracy gate — every shipped fold architecture folded end-to-end with production sampling and
checked against a per-model ground-truth floor, not self-consistency:

model	RMSD (A)	TM	floor	result
boltz2	1.555	0.942	<=3.0 / >=0.75	PASS
esmfold2	1.834	0.906	<=8.0 / >=0.40	PASS
esmfold2-fast	1.811	0.909	<=4.5 / >=0.60	PASS
protenix-v2	1.458	0.945	<=6.0 / >=0.50	PASS
opendde	1.350	0.953	<=6.0 / >=0.50	PASS

BoltzGen scRMSD 1.300 A at a 100% pass rate (floor <=2.0 A, >=50%); OpenDDE-abag DockQ 0.856 with
fnat 0.922 (floor >=0.50); capacity leg peaked at 5.90 GiB against a 7.0 GiB budget on a
1095-token target, writing all 50 samples.

Parity gate (scripts/full_parity_gate.py, 23 legs): 21 PASS, 2 GAP, 0 DRIFT. Both GAPs
reproduce a committed GAP-evidenced verdict rather than drifting — boltz2-prot-nomsa at
ratio 2.07 and boltz2-affinity-fkbp12-nomsa at 0.0033. Highlights: ESMC-300m/600m minimum
per-residue PCC 0.99918 / 0.99938, SaProt-35m/650m embedding PCC 0.99914 / 0.99964, ESMFold2 within
its noise floor on all four targets, Boltz-2 trp-cage envelope ratio 0.59 and HSA 0.71, Protenix-v2
7ROA 1.05 / ubiquitin 0.00 / HSA 0.48, both OpenDDE structure legs bit-identical at 0.0000, and the
RFD3 featurizer 43/43 keys bit-exact. Five of six Boltz-2 affinity legs pass and the sixth
(fkbp12-msa, 0.0028) improves on its recorded gap.

Performance gate: PASS on all 13 models within a +/-15% threshold — boltz2 +0.2%, esmfold2 +0.7%,
esmfold2-fast -0.9%, protenix-v2 -4.5%, opendde +0.9%, opendde-abag -0.3%, esmc-300m +0.3%,
esmc-600m +0.0%, esmc-6b +1.0%, saprot-650m -1.1%, boltzgen +0.0%, boltz2-affinity -4.9%,
rfd3 +5.3%. opendde-abag is a first seeded baseline at 1.208666 structures/s, not a compared
number: v0.5.0 reported seeding one but the JSON edit was never committed, leaving the model with no
baseline on any card for two releases.

Both the perf sweep and the host suite must be run one model (or one test file) per process.
tt-bio supports one device context per process, and iterating models inside a single long-lived
process produces false failures — it reported rfd3 at -42.5% and failed five host-suite tests that
each pass when run alone.