LegoFlow

BlocksEvaluator

Dashboard

Evaluator's dashboard is the benchmark result and analysis browser. Run /evaluator:dashboard to inspect jobs, scores, failure modes, post-eval analysis, and step-by-step trajectories.

A demo dashboard is available here. Stepping into a single trajectory is inert there, because that view streams chunked data only the live deployment serves.

Deployment mode

Both modes read the same files, so the numbers never depend on which one you pick.

  • Local (default): run /evaluator:dashboard. You get a summary of the latest or selected job, and the web board on request. Trajectories, verifier logs and model output stay on the machine that produced the job.
  • Cloud: run /evaluator:dashboard publish to Cloudflare. Set CLOUDFLARE_API_TOKEN, CLOUDFLARE_ACCOUNT_ID and the target Pages project first. For public trajectory browsing at scale, switch on the chunked or R2-backed trajectory mode before publishing.

Where it reads from

Everything the board shows comes out of one directory tree:

PathWhat it holds
artifacts/jobs/Harbor eval jobs, one per immediate subdirectory.
artifacts/jobs/<job>/<task>/agent/The agent side of one task: trajectory, tool calls, and the model exchange behind them.
artifacts/jobs/<job>/<task>/verifier/The verifier side of that task: reward and verifier output, which is what pass and fail are computed from.
artifacts/jobs/<job>/analysis/Post-eval analysis from scripts/analyze_job.sh, which fills the failure attribution and task breakdown panels.

Evaluator is registry-driven and stages no tasks locally, so a job appears on the board as soon as it has run here. A job that was never analyzed simply has no analysis/ directory, and the board leaves those panels out rather than drawing them empty. Run scripts/analyze_job.sh <job_dir> at any time, including on old jobs, to fill them in. See Job Analysis.

Import External Data

The board only reads artifacts/jobs/. Symlink an outside job there, and the link name becomes the job name.

What it monitors

MonitorWhat it tells you
Benchmark jobsWhich dataset, model endpoint, agent scaffold, and config profile each job used.
Resolve ratePass/fail reward summaries and per-job success rates.
Failure patternsError types, verifier failures, unresolved categories, empty patches, and timeouts.
Task breakdownsPerformance by language, area, topic, bug class, difficulty, and benchmark slice.
Agent trajectoriesMessages, tool calls, observations, file edits, and final answers for each task.
Job comparisonSide-by-side comparison across models, checkpoints, scaffolds, or benchmark runs.
Analysis artifactsReports produced after a job finishes, including instance and trajectory analysis.

When to Use It

Use /evaluator:dashboard when you want to answer questions like:

  • What is the model's benchmark resolve rate?
  • Which task categories are failing?
  • Did failures come from verifier errors, agent behavior, endpoint issues, or timeouts?
  • How does this checkpoint compare with another model or scaffold?
  • Which trajectories should be inspected manually?

For Evaluator mechanics, see Design. For artifact contracts, see Output Format.

On this page