Skip to content

Repository files navigation

seldinger-tasks

Public, versioned OR-Audit task, dataset, and baseline-agent registry.

Each registry row is an exact org/name@version identity backed by:

  • an immutable Git commit;
  • a SHA-256 package tree digest;
  • a relative package path;
  • a port-compatible dataset.toml, task.toml, or agent.toml contract.

The registry preserves vector results. It does not define procedure enums, collapse safety into an overall score, contain private clinical media, or make clinical-validation claims.

Current reproducible rows: site/leaderboard.json and the dependency-free site/index.html.

Use

or-audit datasets list
or-audit agents list
or-audit datasets pull seldingermed/lumen-nav@0 --out ./registry
or-audit run -d seldingermed/lumen-nav@0 \
  -a seldingermed/lumen-linear@0 \
  --out ./runs/lumen-linear

# Exact staged Lumen v2 runs (requires surgeval >= 0.3.0a6 and Lumen)
or-audit run -c jobs/seldingermed/lumen-nav/2/integration-smoke \
  --out ./runs/lumen-v2-smoke
or-audit run -c jobs/seldingermed/lumen-nav/2/pilot \
  --out ./runs/lumen-v2-pilot
or-audit run -c jobs/seldingermed/lumen-nav-qualification/1/qualification \
  --out ./runs/lumen-qualification-v1

The pilot schedules 154 independent simulator episodes across nine anatomies. Its branch scenes intentionally have one available seed each; configured maximum timesteps are ceilings, not extra episodes or evidence.

The qualification job schedules exactly 10,000 episodes across 49 matched geometry-by-interface-condition cells. Its 1,432 geometry-seed groups are the independent cases; nominal and perturbed conditions are nested repeats, not 10,000 independent anatomies. Run it only after smoke and pilot gates pass.

Regenerate public baselines

With OR-Audit installed from its P4 release:

./scripts/regenerate.sh

This runs all three public baseline packages and refreshes the committed static, vector-preserving leaderboard under site/. The first AngioStress run restores about 7 GB of content-addressed public benchmark artifacts into the OR-Audit cache. Full run bundles stay out of Git; every published row remains reproducible from its pinned packages and artifact head.

License

Apache-2.0. Package metadata and derived public benchmark artifacts may retain narrower upstream claim and data boundaries recorded in their contracts.

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages