senkani bench

Reproduce the published 80.37× compression figure on your hardware. The suite drives 10 task fixtures through 7 feature configs, default 256 iterations.

Syntax

Behavior

Reproduces the 80.37× fixture-suite figure on your hardware. Writes JSON results; prints a summary.

bench hosts two subcommands. savings is the default — bare senkani bench runs the token-savings suite exactly as before. pii is the headless PII-classifier bench: it times the classifier decode path (synthetic one-hot logits through the real BIOES/Viterbi decoder) with an injected stub backend — no model load, no network — and emits three rows in the standard bench report format: pii_cold_start (backend load + first call), pii_warm_call (median of N reuse calls), and pii_output_shape (span-shape assertion sweep with a quality gate). Cold/warm rows carry estimated confidence — stub latency shape, not real-model perf.

Example

$ senkani bench --iterations 256
$ senkani bench pii --warm-iterations 32 --json pii-bench.json

Flags

Flags — bench pii