C.33:2 - Problem

Architecture work depends on partial carriers. Diagrams, views, relation graphs, ADRs, model queries, code-agent probes, neural-network architecture reviews, eval reports, method descriptions, and operation observations can carry enough structure for one action while losing structure needed for another action.

The practical problem is not “is the carrier good?” The problem is: what selected structure can be recovered from it for this declared architecture use, and what missing-structure return is needed before relying on it further?

Common failures:

• a diagram, model, generated graph, ADR, or benchmark trace starts acting as architecture by presentation;
• structural information is confused with a score, entropy value, epiplexity estimate, dashboard reading, or eval result;
• hidden structure becomes invisible exactly when a later candidate, decision, or work method depends on it;
• source labels such as layer, router, expert, cache, memory, block, gate, SSM, pruning, distillation, or architecture search are copied as FPF ontology instead of being recovered through current FPF rules that define the relevant structures and relations;
• partial-observation outputs from code agents or AI tools are treated as internal belief proof, safe-change authority, evidence sufficiency, or release confidence.