An SLI is good events over valid events, and both are choices

The numerator and the denominator are both decisions, and getting the denominator wrong is what makes an objective meaningless.

good / valid

valid    requests that COUNT. this is where the argument is.
         health checks?          no
         bot traffic?            probably not
         requests from a client
           that timed out first?  yes — the user saw it
         4xx from a bad request?  no, that is the client's fault
         429 from our rate limit? yes, we chose to refuse

the denominator decides whether the number means anything.

Excluding health checks and monitoring traffic is uncontroversial and is usually forgotten, which inflates every availability number by a few points. The 4xx question is the one that produces a real argument: a 400 caused by a malformed client request is not the service failing, and a 400 caused by validation that was tightened last week is. Writing the exclusions down alongside the objective is what stops the number being renegotiated during an incident.