Skip to content
AgentV

Artifact Bundles

Sharing Research Results

Section titled “Sharing Research Results”

For academic and industrial collaboration, AgentV provides a standardized way to package evaluation results into a single, portable, and verifiable bundle. This is powered by the ArtifactPlugin.

The Publication Bundle

Section titled “The Publication Bundle”

A Publication Artifact Bundle is a signed .zip archive containing everything needed to audit a research study.

Generating the Bundle

Section titled “Generating the Bundle”

After certifying your runs, generate the bundle:

Terminal window
agentv bundle --run-id <run_id>

What’s Inside?

Section titled “What’s Inside?”

A research bundle is structured to ensure that every metric can be traced back to raw behavioral evidence.

  • manifest.json: The core metadata linking scenarios, seeds, and run IDs.
  • run.jsonl: The complete behavioral DNA trace of the agent’s actions.
  • leaderboard.html: The professional visual report for offline viewing.
  • audit_manifest.json: The Forensic Evidence Ledger containing SHA3-256 hashes for all files in the bundle.
  • signature.sig: The Ed25519 signature manifest.

Verifying External Bundles

Section titled “Verifying External Bundles”

Researchers receiving a bundle for peer review can verify its integrity without installing the full AgentV framework locally (provided they have the agentv CLI).

Terminal window
# Verify a shared zip bundle
agentv verify --path ./study_results_v1.zip

Verification Checks:

Section titled “Verification Checks:”
  1. File Integrity: Does every file in the .zip match the hashes in the audit_manifest.json?
  2. Signature Validity: Was the bundle signed by the author’s public key?
  3. Trace Consistency: Does the telemetry in run.jsonl actually resolve to the scores reported in the leaderboard.html?