Holo3 · Theoretical-optimum KV simulation
fake the prior KV as a cache hit · prefill only the new screenshot + text · replay every evict baseline step
fake-prefix patch ↗
KV optimum ↗
acceleration ↗
live A/B ↗
status
The modification — prefill stage only
Measured prefill rate vs context depth — the optimum reads off this curve, the live sweep verifies it
model:
live sweep (fake cache + real screenshot, measured on server)
optimum operating band (bounded depths 3k–27k)
Per-step wall clock — one representative task, evict vs optimum
Per task — evict baseline vs theoretical optimum