Guardrail cost

Not populated

Real per-node latency and token cost need a live LLM -- evals/canonical.py stands in a deterministic oracle for exactly that reason, so its cost is always $0.0000 (see docs/evidence/evaluation.md). The table below is the design, not a measurement; it fills in once A3 has a live key.

"What does all this safety cost me?" is the second question every buyer asks, and it usually gets an opinion for an answer. This page gives it a number: added p50 and p95 latency, and added tokens, per layer.

The layers are not equally expensive, and the ordering is deliberate:

Layer	Model calls	Expected cost
screen_rules	0	Microseconds, free
classify_intent	1 (small model)	Low, and it runs in parallel with the screen
resolve_policy	0	Free
critical_value	0	Free — a table comparison
citation_coverage	0	Free — a reference check
scope_judge	1 (larger model)	The only expensive control

Because the two deterministic post-flight checks run first and can short-circuit the judge, a run that escalates on a critical value costs less than one that passes cleanly. Safety here is not a tax applied uniformly to every request.

This page is also the direct precursor to A4's guardrail cost accounting.