Use case

Which data is actively hurting my agent?

Some data makes the agent worse, contradictory policies, outdated examples, mismatched tone. Kno measures regression separately from improvement, so harm shows up even when gains exist elsewhere.

The problem

A data change that improves one class of requests can break another, and most evaluation setups never look for that. Harmful data ships quietly and shows up later as support tickets.

The workflow

1. Value

   Every asset is measured against controls, and regression is measured separately, an improvement in one slice does not mask harm in another.

2. Select

   Assets that hurt get rejected with a `regression` reason and never reach the portfolio.

3. Report

   The report records verdicts, the portfolio, the gaps, and the caveat that nothing is validated on holdout yet.

What it looks like

old_refund_policy.md   -9%   reject → harmful

The recipe

Include your outgoing candidates, the policy being replaced, the examples being retired, in the pool alongside the new ones. Kno values them all the same way, and the controls catch what plain A/B scoring misses.

Stages involved

• baseline
• value
• select
• report