Most agentic QA systems defer semantic work until a query arrives, which becomes costly when the corpus contains revisions, drafts, revocations, deletions, and sources of varying trust. Each read must reconstruct the current governed state and then discard the work.
We introduce ingest‑time fact compilation, an architecture that performs this work once when corpus data is ingested or changed. Raw passages are rephrased into self‑contained facts; rules for revisions, deletions, effective dates, and source trust are resolved a single time; the resulting state is stored as typed records carrying source and revision provenance.
At query time, a cheap model reads the compiled record directly, avoiding reconstruction from noisy candidates. In a controlled synthetic experiment across five seeds, query‑time reconstruction succeeded in only one of 30 trials, whereas the compiled substrate succeeded in all 30, achieving a 12.89× reduction in mean read cost per question.
For simpler revision queries both architectures were exact, but the compiled path used 21.6× fewer tokens. A separate test showed that fact rephrasing halved verbose Federal Reserve dialogue while preserving high source entailment, yet left concise Wikipedia prose essentially unchanged.
These findings support a narrow but practical claim: resolving a corpus state once makes subsequent QA cheaper and more reliable for inexpensive models. We release the implementation under an MIT license along with all experimental artifacts.
Review