JP Journal · deterministic local artifact
KV cache eviction policy lab
Replay one fixed synthetic trace through leaf-LRU and priority-aware LRU. Active blocks are pinned, parent ancestry stays resident, identities include tenant/model/adapter scope, and every total derives from the same immutable events.
Ready. The bounded trace contains 10 events and capacity is 6 blocks.
Decision ledger
| Policy | Event | Request | Hit blocks | Victim | Reason | Modeled saved / miss | Invariant |
|---|
Full candidate ranking receipts
Every eviction decision lists every safe inactive-leaf candidate and its expiration, priority, reuse cost, recency, and stable rank tuple.
Truth limit: This educational simulator uses synthetic prefixes and modeled prefill costs. It is not a serving-engine benchmark, GPU allocator, future-reuse predictor, or universal policy recommendation. Prefix-block reclamation here is lossless and distinct from lossy attention-token eviction.