Local context token quality
Measures whether compact local context preserves useful tool state and decision evidence for agent-style workflows.
Qwen 7B 1K ratio report Tool-inclusive reportBenchmarks
Context quality comes first; scale and parity reports support it.
This page links the current benchmark families. The headline benchmark story is local context quality for agent memory; scale, replay parity, and storage reports support that thesis.
Control-state scale
Tracks high-volume control-state behavior so the project can separate serving semantics from raw storage throughput.
Scale report JSONShared-store blob append parity
Compares append implementations and recovery-visible offsets across current release variants, with HTML and JSON artifacts.
Authoritative offset index Incremental journal 1889 Incremental journal 636 Incremental journal 64a Incremental journal c11Benchmark README
The source README remains the canonical place for adding new benchmark runs, raw artifacts, methodology notes, and release comparisons.
Open benchmark README