Episode 03 · Planned
Prefix reuse
Control exact shared-prefix ratios and compare cold and warm cache behavior.
Within each runtime, compare cache enabled and disabled where supported. Control exact shared-prefix ratios and separate cold from warm behavior. Keep the model revision, tenant or isolation context, tokenizer, adapters and other eligibility conditions fixed. Exact-token reuse is not semantic caching.
Evidence: eligible and reused tokens, cache-hit definition, KV capacity, evictions or recomputation, prefill duration, TTFT and SLO-qualified goodput. Reset cache state between the appropriate arms.