Deferred Proposition Formation in Information Systems: Separating Stored Data from Propositions with Large Language Models

Data stored in an information system is neither reality itself nor a transparent reproduction of propositions about reality. It is a representation produced through processes of observation, judgment, selection, and formalization. The epistemic commitments appropriately made on the basis of such representations may therefore depend on the context in which the data is used. Many conventional information systems nevertheless commit relatively early to determinate states or categories. Multiple observations concerning a customer may, for example, be compressed into a value such as status = "warm", while a sequence of events in a business process may be summarized as status = "completed". Subsequent processes then use these stored values as grounds for claims about the world. This paper proposes deferred proposition formation (DPF): an architectural principle that separates the time at which a representation is stored from the time at which a context-sensitive epistemic commitment is made on its basis. By using a large language model (LLM) as an interpretation layer, a system can preserve provenance-rich and context-rich representations and defer proposition formation until a concrete context of use becomes available. DPF is not presented as a replacement for deterministic systems. Its proposed contribution is to treat the timing and strength of epistemic commitment as explicit architectural variables: what must be determined in advance, and what can remain open until the context of use is known?