No blind traffic increases.
The rollout can require minimum candidate samples and healthy runtime metrics before the next exposure step becomes available.
Keep the stable baseline active, route a controlled share to a candidate and increase exposure only after real production evidence exists.
At 5%, eligible production requests remain mostly on the stable baseline while a small share is routed to the candidate. Increasing traffic changes that future routing share; it does not generate documents by itself.
Baseline and candidate stay explicit
Traffic increases only with evidence
Rollback remains a first-class action
The rollout can require minimum candidate samples and healthy runtime metrics before the next exposure step becomes available.
A staged version only receives production samples when eligible document requests are routed to it.
Increasing traffic does not generate documents; production usage creates the evidence used for rollout decisions.
Keep traffic at 5% until the minimum healthy sample threshold is reached.