Decision path
P50 0.06 ms · P95 0.14–0.19 ms · P99 0.32–0.37 ms across 1k–100k calls, ~14,600 ops/s sustained. Parse, normalize, secret scan, risk, policy, substitution — excludes the upstream tool round trip.
CirvixBenchmarks
Numbers with
a machine attached.
Decision latency, throughput, memory, startup, and audit overhead — measured, recorded, and reproducible from a clean checkout.
Run them yourself
$ npm install $ node benchmarks/decision.mjs # decision-path latency $ node benchmarks/system.mjs --json # startup, memory, throughput, audit overhead
Numbers only count when attached to a machine. Every published figure carries the platform, CPU, memory, Node version, and rule count it was measured with — quote the block, not just the digit.
Latest measured results (Windows 10 x64, Core Ultra 5 225U, 15 GB, Node v24.14.1)
System steady-state
16,118 decisions/sec · RSS ~58 MB · durable audit append adds 0.48 ms per decision (disk-bound, reported separately).
Method
- Warmup excluded from measurement; percentiles computed over all measured calls.
- The audit path is timed as its own run — never averaged into the no-audit latency claim.
- Rows not yet measured on a platform print not measured. Estimates are never printed as numbers.
- Hardware and version context are recorded in
benchmarks/results.jsonvia--record.
Install Free Escape Benchmark Attack Lab Metrics Research Labs Book a walkthrough