18.9 An evidence architecture for visual assurance

The preceding benchmark describes the intended destination. The evidence reported here isolates narrower obligations: whether a frozen diffusion system can maintain, diagnose, and extend a small typed visual declaration as less of that declaration is supplied. This is not a chronological claim that each later experiment is globally stronger. It is an architecture of evidence in which distinct controls answer distinct questions.

Seven obligations organize the architecture:

1. learn a name for a registered visual concept;

2. represent that concept as a typed object with parts and relations;

3. localize a structural counter-witness and repair only its support;

4. estimate the declaration rather than receiving its exact mask;

5. transfer the procedure to photographs evaluated by sealed annotations;

6. represent an unreliable relation by an interval and abstain;

7. acquire another view only where present evidence is insufficient.

The dependencies among these obligations matter. Exact masks isolate the repair operator before declaration error is introduced. Estimated masks test whether the declaration can be recovered. Natural photographs test transport outside the procedural renderer. A failed regime is retained as a typed obstruction; another regime changes the declaration or repair language rather than retroactively weakening its threshold.

The following table separates textual-inversion calibration from natural-image transport.