FREE · prove

Behavior Baseline

Record approved behavior snapshots that future changes can be compared against without causing real-world side effects.

01 · THE SITUATION

“You are about to ship an AI workflow, but there is no safe record of how its important behavior looked at delivery.”

The delivered AI workflow has no safe, repeatable record of what its important behavior looked like at go-live.

REPOSITORY → ACTION → ARTIFACT → OUTCOME

02 · WHAT IT READS

acme-support-agent/fictional fixture

repo/evals/probes/refund-routing.yamlApproved hosts, cases, and side-effect restrictions.
repo/prompts/refund-triage.mdThe instructions exercised by the representative cases.
repo/config/routing.tomlThe threshold and model settings captured with the result.
repo/tools/refund_api.pyA non-production path that cannot issue a real refund.

03 · WHAT THE FDE DOES

cliVerified public-main primitive

                      adopt probe manifest validate evals/probes/refund-routing.yaml

Available validation primitive; baseline recording remains planned.

workbenchIllustrative planned interaction

Planned workflow — illustrative.

04 · WHAT GETS PRODUCED

outputs/baseline.jsonIllustrative fixture

Versioned probe baseline

A safe reference result for refund routing, tied to the prompt, configuration, and probe manifest used.

05 · WORK AVOIDED

You leave a repeatable reference for detecting behavior change instead of relying on memory.

06 · CREDIBILITY BOUNDARY

A baseline is not production monitoring or proof of correctness; it covers only approved probes and cannot trigger real side effects.

07 · NEXT STEP

Use the fictional repository demo to inspect the whole free-layer sequence before touching a real project.