Skip to content

reference

Run Details and Metadata

Inspect the identity, recorded settings, source, telemetry, and status that explain an evaluation Run.

Run Details and Metadata

Definition

A Run's details explain which candidate was measured, against which Benchmark Version, under which execution conditions. Read the recorded Run rather than today's project defaults.

Fields, states, or lifecycle rules

  • Managed Runs identify a saved Harness Version and captured runtime.
  • Imported Runs identify the external submission and its exact Cases.
  • Planned Trials, machine allocation and evaluation settings describe the recorded execution.
  • Status distinguishes output progress from rubric evaluation progress.
  • Labels and descriptive metadata help interpretation but do not replace issued identities.
  • Tokens, latency and cost are nullable measurements. Missing capture remains unknown.

Run history and score use

Use Run history to inspect independent executions of the same candidate. Used for score identifies the reference selected before outcomes were known. Run evaluation reopens matching work; Run again deliberately adds an independent Run.

A later label or metadata change does not make two different execution populations identical. Retain the exact Run and Trial links when recording a comparison or asking an expert about a response.

Settings versus observations

Machine allocation and planned Trial count are accepted execution commitments. Measured latency, token use and actual responses are observations produced afterward. Keep both when explaining an operational difference.

Historical descriptive fields remain attached to their Runs. They do not imply that there is a current project-level Run Metadata configuration tab or that editing today's settings changes accepted history.

Using details in follow-up work

Link an expert to the relevant response and criterion rather than only an aggregate. When copying an External Agents task, confirm whether its optional reference Run represents the code the agent will change. When starting Lemon Goal, inspect the selected starting candidate and available measurements.

Never put credentials in labels, metadata or Case content. Grant secrets through the owning settings and Harness controls.

Found something unclear?

Report outdated, unsupported, or confusing docs so we can fix the source page.

Report a docs issue