Model seam

Where a language model is permitted to influence a decision, and the shape of the check that sits after it.

#Pinning

Every agent declares the model it runs on, and the console shows an unpinned agent as a gap. An agent whose model can move is an agent whose evals describe a system that no longer exists - the evidence and the thing it licenses have to be the same thing.

#The verifier is reject-only

There is no approve path

A verifier that can approve is a second model whose confidence becomes the reason to act. This one can only refuse: it examines a proposal against deterministic checks and either rejects it with a named reason or stays silent. Silence is not endorsement, and it never substitutes for the human release.

#Confidence is computed, never self-reported

A model asked to score its own confidence produces a number correlated with fluency rather than correctness. The gate here is computed from observable properties - whether required fields resolved, whether the rules that applied were servable, whether identity reconciled, whether anything abstained - and a proposal below the threshold is held regardless of how the model described itself.

#Untrusted content never enters the system prompt

Retrieved documents, payer notes and tool output are passed as clearly-fenced data. See Untrusted content.

#Budgets

Per-agent and per-practice, in two windows. A runaway agent is a cost problem and, where it is calling a payer, a disclosure problem - so the ceiling is enforced at the gateway rather than in the agent loop that would have to notice its own runaway.