The headline is about "Cost Per Verified Success" and how exit-0 denominators lie. Let me craft a punchy headline.

Short summary

Cost-per-task dashboards divide spend by 'successful tasks,' but success is usually self-reported (exit_code==0 or HTTP 200), meaning silent failures inflate the denominator and understate true cost. The author proposes 'cost per verified success' where each task must be re-confirmed by an independent witness (DB row, file manifest, HTTP body token). The math guarantees verified_cost >= naive_cost, and the understatement factor (M/M') quantifies how far off your dashboard is.

• •Exit-code-0 success counts include silent failures, making cost-per-success a lower bound on reality
• •Verified success requires independent re-confirmation of each task's effect, not just the agent's self-report
• •Understatement factor M/M' is always >= 1 and measures the gap between naive and verified cost

Generated with AI, which can make mistakes.