Interaction between compiled substrates and agentic read strategies
Characterize how compiled semantic substrates should interact with agentic read strategies, including when an agent should request sibling claims, lineage information, or the raw source.
Background
Ingest-time semantic compilation stores validated claims and their source spans in a maintained semantic substrate, allowing a reader model to consume compact, provenance-backed payloads rather than repeatedly interpreting raw documents. The paper notes, however, that the compiled-payload advantage did not automatically persist when retrieval was mediated by a tool-using agent.
The unresolved issue concerns the agent’s read policy: an agent may need to determine whether a retrieved claim is sufficient or whether it should request related claims, provenance lineage, or the underlying raw source. Characterizing this interaction is necessary for extending the benefits of compiled substrates from direct retrieval pipelines to agentic systems.
References
The interaction between compiled substrates and agentic read strategies --- when the agent should ask for siblings, lineage, or raw source --- is open, and we flag it rather than paper over it.