Skip to content

Guide / Run identity

Version AI agent evaluation runs for comparison.

A result cannot be reproduced or compared if the agent, wrapper, tools, or environment changed invisibly. Build a compact run manifest that separates agent identity from WagerCall's immutable environment identity.

For
Teams comparing agent changes across releases
Outcome
A stable manifest that explains exactly what ran

1 / Layers

Separate agent identity from environment identity

Run identity fields by owner
LayerRecordWhy it matters
AgentModel, settings, prompt, policy, wrapper buildIdentifies the variable under evaluation
ProtocolMCP protocol and client implementationIdentifies tool-contract behavior
EnvironmentGame, version, rules hash, engine build, RNG identityIdentifies immutable task behavior
RunMode, canonical config, seed policy, attempt indexIdentifies this execution condition

2 / Capture

Capture the manifest before the first action

Persist or export the manifest when the Session or Room opens, before an outcome can influence which fields you remember to record.

  • Use stable identifiers or digests for prompt and wrapper artifacts.
  • Store configuration in a canonical serialized form.
  • Record the exact game resource kind and version.
  • Keep secrets, bearer credentials, and private observations out of the manifest.

3 / Compare

Reject accidental comparisons when manifests differ

Diff manifests before combining runs. A changed version, configuration, tool contract, or wrapper may explain a result that would otherwise be attributed to the model or prompt.

A game version change and an MCP protocol change are different axes. Preserve both rather than collapsing them into one application release label.

4 / Share

Publish enough identity to review the claim

  • The variable changed between conditions
  • Every environment field held constant
  • Run count and exclusion policy
  • Canonical record or replay references where sharing is appropriate
  • Known limits and unavailable artifacts

Next step

Create the run protocol

Attach the manifest to a bounded question and analysis plan.

Create the run protocol