T
Traqlyte
Get the API spec
← All posts Fundamentals Draft

Why you can never edit a running experiment

8 min read · Coming soon

It feels harmless: the test is running, one variant has a typo, you fix it. But the users assigned before and after your edit saw different things under the same label, and your results are now a blend of two experiments neither of which you can analyze. This is why TraqLyte locks a campaign version the instant the first assignment lands.

This article is a stub.

The outline below is what the finished post will cover. Want it sooner? Tell us.

01What mid-flight edits do to your data

  • Variant edits: two treatments, one label
  • Weight changes: Simpson's paradox in traffic ramps
  • Adding variants: unequal exposure windows and seasonality bias

02Peeking, the softer version of the same sin

  • Why checking significance daily inflates false positives
  • Sequential testing as the legitimate alternative

03Versioning as the fix

  • Draft, locked, published: the lifecycle
  • Snapshots: every assignment permanently tied to the config it saw
  • Changing your mind means publishing a new version, not rewriting history

04What to look for in your platform

  • Does it stop you, warn you, or silently allow it?
  • Audit trails as the test of seriousness

Research brief

“Find documented cases and statistical explanations of mid-experiment configuration changes invalidating results: peeking, changing traffic allocation mid-test (Simpson's paradox in ramp-ups), adding variants mid-flight, and how platforms enforce config immutability or version snapshots.”

The deep-research question this article will answer, sources cited in the finished piece.

Run experiments your whole stack can call.

Get the API spec