One runner for the open (vs PDB deposition) and in-house (vs XDS) arms,
replacing the scattered scripts outside the repository. Each run writes a
new, read-only directory named <date>_<rugnux git>_<label> holding a copy
of the binary with its sha256 and build flags, manifest.json (inputs,
references, options, host, GPU sharing) and results.json in one schema
for all arms, plus per-set work directories and a Markdown and a
self-contained HTML report (per-arm and per-population pass rates,
resolution gain, ISa/R_meas distributions, timing, failures, delta vs a
baseline run).
- Dataset lists and references are data files: open.json (151 public
sets, deposited group/cell/resolution, pinned inputs) and inhouse.json
(standard test crystals and no-crystal controls, XDS references). The
private arm's manifest stays outside the repository; a private run gets
its own run directory and report.
- Scoring is setting-invariant: lattice by Niggli-reduced primitive cell
(beta vs 180-beta, C2 vs I2), symmetry via sgequiv on the open arm and
by point group against XDS, then merge usability. Re-scoring the rc169
and rc170 battery outputs reproduces their verdicts.
- Input discovery (`discover`) ranks sweeps by frame count, never by
file size or sort order; a run never searches, it runs what the
manifest names.
- `compare A B` flags per-set changes beyond noise; `--rerun-changed`
reruns them with A's saved binary to separate real changes from
run-to-run noise.
- `--tier smoke` selects a 22-set subset spanning the populations; a
full-arm run takes a lock; `--gpulock` queues on the shared GPU and
marks the timing as not a reference.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
leonarski_fandClaude Opus 5
5e7b53dd7b