Skip to content

openclaw-baseline-evaluation.md

Latest commit

 

History

History
68 lines (46 loc) · 1.91 KB

openclaw-baseline-evaluation.md

File metadata and controls

68 lines (46 loc) · 1.91 KB

OpenClaw Baseline Evaluation

Use this workflow to generate a repeatable OpenClaw-first baseline snapshot for the v3 learning loop.

If the baseline remains too cold to say anything meaningful about candidate creation or repeated-task intervention, use the higher-signal companion workflow:

Preconditions

  • openclaw CLI is installed and the gateway can load experienceengine
  • ee doctor openclaw reports a healthy host wiring state
  • the current local ExperienceEngine database already contains OpenClaw task data

Command

ee evaluate openclaw-baseline

Optional flags:

ee evaluate openclaw-baseline --lookback-hours 168
ee evaluate openclaw-baseline --output-dir ./artifacts/evaluations/openclaw/manual-run

Outputs

By default, ExperienceEngine writes local-only artifacts to:

artifacts/evaluations/openclaw/<timestamp>/

Each snapshot contains:

  • summary.json
  • summary.md

What The Snapshot Covers

  • input record totals and outcome distribution
  • injection coverage
  • candidate lifecycle distribution
  • distillation job status distribution
  • node state and feedback distribution
  • latest observed record / candidate / node pointers

How To Use It In The Current WSL Baseline

  1. Run one or more real OpenClaw tasks in the current workspace.
  2. Run:
ee doctor openclaw
ee evaluate openclaw-baseline
  1. Record the generated snapshot path.
  2. Compare later snapshots after distiller/profile/gating changes.

Interpretation Notes

  • This snapshot is a baseline, not a trend report.
  • Injection coverage being high is not automatically good.
  • A growing discarded candidate count usually means either the gate is too wide or the distiller profile needs work.
  • OpenClaw is the current baseline host. Claude Code and Codex remain regression or reuse hosts for this stage.