Function evaluate 

pub fn evaluate(g: &Guardrail, outcomes: &[bool]) -> GuardrailVerdict

Evaluate a trailing window of resolved outcomes. outcomes is oldest-first; true means the served answer was correct.

Uses the same bound as crate::conformal::calibrate — rate + sqrt(ln(1/delta) / 2n) — so the guardrail defends exactly the quantity the published guarantee describes. A guardrail computing a different bound than the advertised one would be defending a number nobody was promised.