The mental model

Traditional evals answer one question: did my agent get better? They vary the model, the prompt, or the tools, and watch a score.

Kno asks a different question: which data made it better?

Traditional eval

Prompt ─────┐
Model ──────┼──► Agent ───► Score
Tools ──────┘


Kno

                    ┌─ Document A ──► +12%
                    ├─ Example B ───►  +8%
Agent + Evals ──────┼─ Policy C ────►   0%
                    └─ Asset D ─────►  -3%
                           │
                           ▼
              impact × cost × destination

Data is the variable

A document, an example, a policy, a conversation, each is a candidate asset. Kno treats each one as a variable in a controlled experiment: measure the agent without it, inject it, measure again, compare against fresh controls, and report the delta with its uncertainty.

The agent, the evals, and the goal stay fixed. Only the data changes. That is what makes the result attributable.

The stages

Stage	Question	Status
Init	How is this run set up?	Shipped
Mine	Where do the evals come from?	Shipped
Baseline	How good is the agent now?	Shipped
Value	Which assets improve it?	Shipped
Select	Which combination should I keep?	Shipped
Validate	Does the combination still work on untouched evals?	Planned
Export	Where should each asset go?	Shipped
Report	What does the whole run say?	Shipped

The holdout is the discipline

A slice of your cases is sealed at baseline and nothing reads it until validate. Without that separation, selection could silently overfit your dev set, and every number after it would be inflated.

The full argument, including what is deliberately out of scope: DESIGN.md and the mental model page.