pass^k
The probability an agent solves the same task k times in a row — a consistency metric.
pass^k = P(all k independent runs succeed). If per-run success is p, pass^k ≈ p^k, so it decays geometrically with k. It complements pass@k (at least one of k runs succeeds), which rewards lucky attempts. τ-bench introduced pass^k because production users need the agent to work every time, not once: GPT-4o's τ-retail score fell ~60% from pass^1 to pass^8. For a 99% consistency target over 8 runs, per-run reliability must exceed 99.87%.