eviction concept

Definition	Choosing what to drop when a cache is full. Chapter 13. Also evict.
Example	A cache of 4 entries receiving a fifth drops the least recently used one.
Book	Data Mining as Observation, draft 0.2, commit `f3914f0`; entry id `eviction`, kind concept.
Status	measures [demonstrated]. Corrections: none recorded.
Defining equation	none
Assumptions and scope	Choosing what to drop when a cache is full. Evicting tokens replaces attention over all tokens with attention over the kept subset, so the output lies between the kept values, equals them when they agree, and the whole cache is the case of no eviction. Which tokens can go is consumer-relative, since a head reads a key only through its score, and the serving-stack measurement of ninety-seven percent eviction at a five percent keep against an oracle is the ledger’s number, after one attempt whose control did not exist.
Prior art	none recorded
Evidence	`geometric-observation/claims/LEDGER.md:81`, `geometric-observation/claims/LEDGER.md`, `geometric-observation/prereg/GO-P-2026-077-kv-consumer-relative.md`, `geometric-observation/results/GO13-kvaw2-governed.json`, `lean/DataMiningAsObservation/Eviction.lean`
Reviewed	not yet reviewed; generated 2026-09-10 from records at the commits on the provenance page.

Choosing what to drop when the cache is full.

Equation

none

Conditions

Choosing what to drop when a cache is full. Evicting tokens replaces attention over all tokens with attention over the kept subset, so the output lies between the kept values, equals them when they agree, and the whole cache is the case of no eviction.
Which tokens can go is consumer-relative, since a head reads a key only through its score, and the serving-stack measurement of ninety-seven percent eviction at a five percent keep against an oracle is the ledger’s number, after one attempt whose control did not exist.

Conditions are curated in entries.toml rather than read from a record.

Ledger

measures. GO-2/GO-12/GO-13 operational (KV serving, 077) [demonstrated]. Consumer-relative access width measured on a production serving stack (Qwen2.5-7B KV-cache eviction, matched budget): task quality tracks measured predictive uncertainty u about the consumer’s future reads, not nominal scorer width — the … geometric-observation/claims/LEDGER.md:81.

First stated

Chapter 13 section 13.6 of Data Mining as Observation, with the program’s serving-stack case in geometric-observation/prereg/GO-P-2026-077-kv-consumer-relative-eviction.md and the ledger row that carries it.

Measurements

Where the book states it	Numbers, as the book’s sources table records them	Source
chapter 13 section 13.6	attempt three, windows 1024, 256, 32, uncertainty 0.982 to 0.892, 5 of 6, 0.4375 with SE 0.070 at 5 percent keep vs 0.30, 97 percent eviction, oracle-miss 0.370 vs 0.25, V4 0.078 vs 0.0625, contrast 0.359 with SE 0.068, n 64, seed 20260812, 89 duty cycles	`geometric-observation/claims/LEDGER.md` row GO-2/GO-12/GO-13 operational; `geometric-observation/prereg/GO-P-2026-077-kv-consumer-relative.md`; `geometric-observation/results/GO13-kvaw2-governed.json`

Failures and corrections

none

Invariance envelope

none declared

Machine checked

lean/DataMiningAsObservation/Eviction.lean, theorems kept_weights_sum, keptOutput_le_max, min_le_keptOutput, keptOutput_const, keptOutput_univ, at observation-data-mining f3914f0; what the check covers is stated in the book’s appendix C.

Used in

Data Mining as Observation chapters 13.

KV cache; attention; coherence time; budget.

Status

Generated 2026-09-10 by encyclopedia/generate.py; book at observation-data-mining f3914f0; the commit of every record is listed in the encyclopedia’s provenance.

← eventexpectation →