Design note. This page explains the rationale behind LangState. It is not a claim that a complete application loop or production integration is bundled.
- What might the user mean?
- How strongly does the system support each candidate?
- Which value is authoritative enough for an application action?
LangState’s dual-state design gives those questions different representations.
Two roles
Interpretive state is evidence-oriented. It keeps candidate values, confidence, and the latest inference for a field.
Canonical state is decision-oriented. It keeps the resolved values that downstream validation or business logic can consume.
The boundary allows an application to inspect uncertainty, request confirmation, apply deterministic constraints, or compare alternative resolvers before committing a value.
Why not only a transcript?
A transcript preserves conversation but does not itself provide stable field paths, typed boundary models, dependency relationships, or a clear answer to “what is the current value of this field?” LangState treats conversation as potential evidence and state as an explicit artifact.
Immediate writes can be appropriate for deterministic inputs. For inferred inputs, retaining candidates makes correction and evaluation more explicit. The application can choose its resolution policy instead of embedding it invisibly in extraction.
Costs and tradeoffs
The split adds model types, conversion logic, and lifecycle decisions. It does not automatically improve accuracy, latency, or cost. Those outcomes depend on the mutator, projector, prompts, models, data, and application rules and must be measured for the actual system.
For shipped behavior, return to Interpretive state, Canonical state, and Implementation status.