Routing

Routing is the step where a variant you saved in Testing starts meeting real users. Move the sliders, save, and it applies from that moment. No deploy, no config file edits.

Success metricsdefine the bar Testingtry candidates Routingroute traffic Comparejudge outcomes

The ratio graph and the sliders

Area	What you see
Ratio graph on top	A picture of how traffic is currently divided across the variants
Variant rows	One slider per variant. Change one row and the others adjust automatically so the total always stays 100. Nothing applies until you press Save
”Results in Compare”, top right	Takes you to the results for the split you are currently running

Adding variants and rolling back

Action	How
Create a variant	The recommended path is validating in Testing and saving as a variant. A new variant always starts at zero
Turn a variant off	There is no on/off switch. Zero is off
Roll back	Return the new variant to zero and the previous one to 100, and it is done. It takes seconds
How many variants	A policy holds at most five variants. Dropping a ratio to zero keeps the slot, so use “Remove from routing” to free one

Requests routed to a provider whose key is missing simply fail

If you add a model from a provider you have not used before, the code that calls AI has to send that provider’s API key along with the request. Without it, only the requests routed to that variant fail, and those failures never show up on Requests. Check the key before you raise the ratio.

Assignment is sticky per person

Which variant a user gets is computed from their device, so the same person keeps getting the same variant call after call. Their experience stays steady, and conversion in Compare splits cleanly at the person level. Once you change the ratios, assignment follows the new ones from that point, so some users move to a different variant.

How to raise ratios

Stage	Ratio	What to confirm
Right after saving a new variant	10%	The new variant’s error rate and p50 latency over a day or two
If nothing looks off	30 to 50%	Whether the new variant is level with or ahead of the incumbent
Once the difference is clear	100%	Drop the variant you are not keeping to zero. After the switch, watch cost and outcome trends on Overview’s 30-day range

You confirm each stage in Compare, and how much sample is enough to judge follows that screen’s rule too. Change one thing at a time on a feature: swap the model and the prompt together and you cannot tell which change made the difference.