Figure 16.1: One agent, two honest summaries of the same ten attempts, drawn here for an illustrative per-attempt success rate of 90 percent
Figure 16.1 One agent, two honest summaries of the same ten attempts, drawn here for an illustrative per-attempt success rate of 90 percent. The chance that at least one of k attempts succeeds (pass@k) climbs toward certainty; the chance that all k succeed (passk) decays toward zero. The widening gap between them is the reliability envelope: the wider it is, the more capable-but-erratic the agent—and a demo only ever shows you the top curve.
