Frameworks built, named, and pressure-tested across real work. Not borrowed vocabulary. Original positions.
Declared beats inferred. The human stays in the retrieval loop because the loop is where learning happens.
Standard RAG lets the model find relevant chunks. Vectorless RAG inverts this: the human selects, primes, and structures the input. The LLM narrates, not navigates.
A portfolio pattern that keeps the operator workbench visible next to the finished room.
Atelier is the workbench: atomic ideas, sources, the scaffolding. Salon is the finished room visitors walk into. One capture, two renderings. Both layers are first-class.
The same structure keeps appearing across unrelated domains. Name it, test it, share it.
The declared-vs-inferred finding appeared first in I-O Psychology, then construction PM, then consent verification, then compassion journaling. Four domains, one skeleton. That is the practice.
Catalog the atomic affordances. Compose workflows from named actions. The human lays the edges. The model traverses them.
Top-down composition (state the outcome, let the model pick the path) leaves the composition space open. Bottoms-up composition catalogs named affordances and composes deliberately. The model can only traverse edges the human laid down.