build

Export a package.

Export is the contract between capture, evals, training jobs, customer review, and lab review. It should be portable enough to rerun and inspect.

Checklist

export package includes:
  manifest.json
  environment metadata and version
  tasks and grader contracts
  model runs and score records
  screenshots, traces, and artifacts
  verifier audit notes
  contamination and provenance notes

Reviewability

Human-readable

Summary pages explain source workflow, task intent, grader checks, results, and known risks.

Machine-readable

The manifest links every task, grader, run, artifact, and audit note by stable ID.

Runnable

The environment version, seed, action space, and grader hooks are explicit enough to reproduce attempts.

Export types

Public eval exports emphasize inspectability. Customer or lab exports may include private artifacts, runner configs, split metadata, and source controls. Both should preserve the same package contract.

Next: evaluate pass@k Use model attempts to calibrate task difficulty. Artifact schema Open the required manifest fields.