Artifact schema.
The artifact schema is the portable contract between local workflow capture, public eval pages, remote eval runners, SFT/RL jobs, and customer/lab review.
Top-level manifest
{
"schema_version": "usedesktop.workflow_package.v1",
"package_id": "pkg-kca-2026-06-19",
"environment": {},
"tasks": [],
"graders": [],
"runs": [],
"audits": {},
"artifacts": {},
"provenance": {}
}Required fields
Version of the package contract used by exporters, eval runners, and docs.
Environment identity, reset behavior, observation space, action space, and verifier hooks.
Task prompts, seeds, constraints, expected states, and links to grader contracts.
Scoring logic, weights, success checks, violation checks, and known limitations.
Model attempts, traces, scores, rewards, verdicts, and links to replay artifacts.
Verifier FP/FN audits, difficulty calibration, contamination notes, and review samples.
Provenance
Provenance should explain where the workflow pattern came from, how sensitive data was removed, which version was used for train/eval splits, and whether a public benchmark or customer-specific dataset overlaps with the package.
A package is not considered reviewable until every task has a grader, every grader has an audit story, and every published model result points to a run record.