Concepts/Concept
ADPS Agent Systems Blue Book · Engineering Concept
Typed Intermediate Representation Chain: Narrow Language into Verifiable Structure
Narrow natural language through typed, verifiable structures before write-back.
Start with one use-case diagram
A user asks for a turnstile use-case diagram under an existing package. A one-shot response that creates elements, relations, and the view may point to names that do not exist or omit elements from the diagram. A failure forces the entire structure to be regenerated.
Definition
A typed intermediate representation chain narrows natural-language intent through a series of verifiable data structures. The AI4MBSE personal research project can be reconstructed as:
ModelingJob -> ElementPlan -> RelationPlan -> ViewPlan -> WriteSet
Each stage reads an accepted upstream result and emits typed candidates with stable references and provenance. A validator checks well-formedness. The writer consumes a gated WriteSet and does not reinterpret the original language.
Boundary
JSON alone is not a typed IR. Schemas or validators must enforce field types, allowed values, references, and stage invariants. Stages may run as separate agents, repeated model calls, or ordinary functions.
Provenance
- Initial research source: the AI4MBSE personal research project by Liangding Yuan; contract names are an ADPS reconstruction of the published mechanism.
- Lineage: compiler IR, typed ASTs, semantic analysis, and database referential integrity.
- ADPS contribution: models draft candidate IRs, programs validate stage contracts, and the writer consumes only accepted structures.
- Definition status: ADPS restatement.
Initial source: AI4MBSE personal research project; authored by Liangding Yuan.