Performance

Budgets

Path	Budget	Gate
Decision engine only (no classifiers), 50 tiers, 5 policies, data class active	p95 < 1 ms	crates/decision/tests/latency_gate.rs
Engine + local ONNX classifiers (phase 4)	p95 < 30 ms added latency	crates/classify/tests/onnx.rs::perf_gate via ROUTED_PERF=1 make onnx

Running the gates

ROUTED_PERF=1 make test                       # engine gate (skipped without ROUTED_PERF)
ROUTED_PERF=1 ROUTED_PERF_SLACK=1.5 make test  # noisy VMs: multiply the budget
ROUTED_PERF=1 make onnx                       # classifier gate (downloads libonnxruntime into .cache/)
scripts/cargo-in-podman.sh cargo bench -p routed-decision   # criterion trend numbers
scripts/cargo-in-podman.sh cargo bench -p routed-classify   # heuristic classifier trend numbers

The classifier gate runs against the committed fixture model (crates/classify/tests/fixtures/, regenerable with trainer/scripts/make_classifier_fixture.py), so it bounds the tokenize -> run -> extract pipeline, not a production model. Re-calibrate when a real trained classifier lands (phase 6).

Reference runner class

Gates are calibrated for GitHub-hosted ubuntu-latest (x64, 4 vCPU). Numbers from the local podman machine (arm64 VM, 6 vCPU) are informational only.

Recorded results

Date	Environment	p50	p95	p99
2026-08-22	podman machine arm64, 6 vCPU, debug build	see CI log
2026-08-22	classifier gate (fixture model), same VM: 200 runs in 0.43 s, p95 well under budget