Requests in this set may land on different models or tiers
PrefixAudit rule model-churn · v0.1.0
Why it breaks your cache
The cache is per-model. A router, A/B test, fallback or changed service tier that moves a request to a different model produces a cold miss even though the prompt is identical. OpenAI reports "model_changed" and "service_tier_changed" as distinct miss reasons.
How to fix it
Pin the model for calls that are meant to share a prefix. If you must route, use sticky routing keyed on the conversation so a session stays on one model and one shard.
How common is it?
This rule needs request or runtime context that an extracted prompt cannot supply, so it is excluded from the corpus study. It still runs in the auditor when you supply that context.
Estimated impact
Modelled as invalidating 0% of the cached prefix.
Is your prompt affected?
Paste it and find out. Static analysis, in-browser, nothing leaves the page.
Run the audit →