Experiment Guide¶
Guide Maps¶
graph TD
baseline["Baseline params and metrics"]
change["Declared parameter change"]
run["Experiment run"]
compare["Comparable metrics and outputs"]
decision["Promote or reject"]
baseline --> change --> run --> compare --> decision
flowchart LR
choose["Choose one declared change"] --> run["Run an experiment"]
run --> inspect["Inspect params, metrics, and predictions together"]
inspect --> judge["Decide whether the result is reviewable"]
judge --> promote["Only then ask about promotion"]
This guide exists because experiment support is one of the easiest parts of DVC to use sloppily.
What an experiment is for here¶
In this capstone, an experiment is a controlled deviation from the baseline parameter surface. It is not a license to mutate the baseline story until the result looks good.
Review questions¶
- Which parameter changed, and why does that change stay comparable to the baseline?
- Which metrics moved, and what do those changes mean operationally?
- Which prediction records changed enough to justify closer review?
- What evidence would be required before promotion to
publish/v1/?
Minimum route¶
- Inspect baseline
params.yamlandmetrics/metrics.json. - Run
dvc exp runwith one declared change. - Use
dvc exp showto compare the candidate against the baseline. - Return to the publish contract only if the candidate is worth promotion.
Read CONTROL_SURFACE_GUIDE.md first when the real pressure is not how to run the experiment, but whether the changed params still support honest comparison.