> ## Documentation Index
> Fetch the complete documentation index at: https://docs.selftune.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Evidence and health

> Understand how SelfTune turns observations and evaluations into an honest health state.

Evidence is information tied to a specific skill revision and behavior. Health
is SelfTune's current summary of that evidence.

Evidence may come from observed agent sessions, routing examples, execution
evaluations, replay, baselines, or post-change monitoring. These sources answer
different questions and should not be treated as interchangeable.

## Health is contextual

A health conclusion depends on the skill revision, harness, time window,
sample coverage, and evaluation conditions. SelfTune should report uncertainty
when those inputs are missing or not comparable.

An unknown or no-data state means “there is not enough evidence.” It does not
mean the skill failed.

## Evidence before change

An improvement should cite the evidence it addresses and reuse comparable
conditions when evaluating the candidate. After a change, new observations can
confirm the result or reveal a regression.

See [Understand skill health](/guides/understand-skill-health) for the practical
workflow.

For every health or improvement claim, name the skill revision, evidence type,
coverage, time range, and relevant evaluation conditions when available.
Separate observed facts from inference. Refuse to convert missing or
non-comparable evidence into a winner.