sleepy Log in / Sign up

Proof ledger

Real results, labeled honestly.

Every figure names its run, carries its caveat, and links its evidence. Search fitness is not an end-to-end product speedup. No guaranteed speedup.

2

merged upstream PRs

10

scoped public ledger entries

58.2%

tidwall/match geomean, scoped

8.7%

vllm-metal end-to-end decode reduction, n=8

Aggregate figures from the public ledger. Entries link public PRs and evidence notebooks where available.

DateTargetResultStatus
{{ r.date }} {{ r.target }} {{ r.caveat }} public evidence ↗ {{ r.result }} {{ r.status }}

Methodology varies by claim and is stated in the linked evidence. Go microbenchmarks generally use repeated benchstat samples plus full upstream and differential tests; the vLLM-Metal end-to-end figure uses eight paired runs. Raw event logs, decision records, and candidate archives stay with the operator.