Skip to content

Evaluate and improve

Begin with evidence from actual tasks, then build a stable comparison.

Four useful concepts

ConceptQuestion
DatasetWhat should be tested?
GeneratorWho produces the answer or artifact?
Evaluator and criteriaWhat counts as good, and who checks it?
OptimizerWho proposes the next controlled change?

The bundled Harbor Skill guides this setup. Invoke evolve-agent-with-harbor from the conversation and explain your task. The desktop includes the plugin, Skill, and matching Python adapter; do not repeat the standalone package setup command. Candidate execution requires its Docker environment and usable model services; run Doctor before starting.

Without a dataset

Ask to diagnose recent completed sessions in the current working directory. Preview the selected records, evaluator, estimated requests, and evidence-storage notice. Confirm before evaluation. Inspect coverage, evidence, and abstention reasons in the Harbor workbench.

Historical diagnosis does not rerun the Candidate and cannot be used as a promotion baseline. A completed but unscored record is not a zero score. If there is no eligible history, complete a task or supply an explicit Query or Dataset.

With repeatable tasks

Confirm the four concepts, fix the Candidate model, and establish a valid baseline. Change one factor, rerun comparable tasks, and inspect regressions as well as gains. A changed model or evaluator identity can require a fresh baseline.

A single-Query quick diagnostic checks wiring rather than establishing quality under its drafted rubric. Formal promotion requires an accepted policy; evaluator reliability requires independent Ground Truth meta-evaluation. Deployment remains a separate action.

See Harbor Evolution for costs, credentials, and limits.