From real Sessions to controlled promotion
This guide follows one honest evolution loop. It separates diagnosis from promotion and keeps three questions visible in every chapter: what you do, what Harbor records, and what the result does not prove.
Contents
Improvement requires a fixed task, valid evidence, one controlled change and a policy outside the Optimizer.
Understand Dataset, Generator, Evaluator, Optimizer, train/validation/test splits and meta-evaluation in business language.
Use a disclosed Historical Job to find recurring failures before building a promotion Dataset.
Convert recurring failure patterns into reviewable Tasks without leaking raw private history.
Freeze identities, establish a comparable baseline, change one surface and inspect Trial-level evidence.
Interpret PROMOTE or REJECT as a policy recommendation and hand deployment authority to external CI/CD.