Use AI normally. Put reliability under separate control.

Start with the work, not a control checklist.

Prompt normally. Add a file or project context. Choose a model, or let Coherence configure the workflow. Search and specialist tools can join when the task needs them.

Models and tools produce candidate material.

AI services, retrieval, documents, computation, and specialist tools do what they do best. Their output remains candidate material: no provider, citation, or tool result awards itself a reliability verdict.

A separate deterministic boundary decides what happens next.

CRE can pass sound work, constrain a repair, keep the answer within a defined bound, request specific outside input, or contain unresolved output.

It evaluates what the response asserts, what support those assertions require, which uncertainty remains, whether the premise or scope changed, and whether a source or authority actually carries the conclusion attributed to it.

A replacement has to earn its way through the boundary again.

When repair is possible, Coherence generates a replacement within CRE-defined constraints. Better wording, a new source, or a second model does not inherit approval. The new candidate is checked again.

When the repair needs a current fact, source, specialist tool, or clarification, Guided Repair asks only for that missing material and returns it as constrained candidate input. The original candidate, intervention, replacement, and any unresolved limit can remain connected for review.