Evaluation Dashboard
Dashboard summarizes benchmark performance through Progress, Leaderboard and Frontier. Use Results when a chart needs Case-level explanation.
Prerequisites
- Runs for one identified Benchmark Version.
- A clear understanding of the selected metric and Regime.
- Complete results for claims that require a rankable comparison.
Read the dashboard
- Confirm the Benchmark Version and candidate identities.
- Choose Progress to inspect movement over recorded executions, Leaderboard for the selected performance ordering, or Frontier to inspect visible tradeoffs.
- Read the metric's population, aggregation and completeness.
- Follow a surprising value into Results and inspect its Cases and Rubrics.
- Preserve regressions and resource tradeoffs alongside gains.
A chart is a summary, not a new approval state. The newest candidate or the highest displayed aggregate does not automatically become the selected implementation.
Repeated behavior
Repeated Trials support different questions. At-least-one success and all-Trials success are different reliability statements. Read pass@n or pass^n only with the recorded Trial count and cohort. Ordinary scores and means answer different questions again.
Uncertainty belongs to the measured sample. A small numerical lead does not remove missing judgments, a failed mandatory gate or concentrated regressions.
Object and state changes
Dashboard controls read existing results. They do not launch a Run, alter representative score selection, correct a judgment or retain a Lemon Goal candidate.
The reference Run used for a candidate score is selected before outcomes are known. Additional independent Runs remain available in history rather than silently changing that reference population.
Success criteria
- The metric, population and repeated-Trial rule are explicit.
- Incomplete results remain visibly incomplete.
- A meaningful change can be traced to exact responses and criteria.
- Cost, tokens and latency are used only when actually captured.
Common failure modes
- Treating a ranking as an approval.
- Substituting later successful Trials for missing reference results.
- Reading a negative point score as an impossible result.
- Treating absent telemetry as zero consumption.