Overseer: Governed Graph Engineering for Source-Traceable Bitemporal Knowledge Graphs

In multi-agent AI systems, the output of one agent becomes the input of the next: a conclusion extracted today is reused in tomorrow’s analysis, and later decisions silently depend on both. When an upstream source is subsequently corrected or retracted, every downstream result that reused it is called into question — yet a system that retains only prompts and outputs cannot say which results those are, what evidence each rested on, or what had actually been recorded when each was produced. Conventional logging does not answer these questions: it records what happened, not what depended on what, and not what was known when. We present Overseer, a governed graph-engineering substrate that makes these dependencies explicit: assertions, their evidence, and their direct parents form a content-addressed dependency graph, and every change is recorded in an append-only bitemporal event history, so historical states can be replayed exactly and corrections can be traced to their dependents without rewriting prior state. Semantic and derivational objects are represented separately from historical assertion occasions. We evaluate the substrate across six corpora using deterministic lineage, historical reconstruction, mutation, governance, and adversarial-control experiments. The evaluated artifact reconstructed 30,192 of 30,192 tested historical states, admitted zero of 2,636 prohibited reuse cases in the frozen retraction test, and supported cross-language verification over a 75,113-assertion graph. Negative controls mark an equally important boundary: structural provenance does not establish semantic truth, and a resolvable citation does not establish that the cited source supports a claim. Progressively strengthened controls then separate what the experiments establish from what they do not. Graph topology explains traversal; content addressing and parent commitments explain the tested integrity properties; a generic bitemporal append-only log matches Overseer on historical state reconstruction and correction with retained history. A preregistered successor experiment likewise found that Overseer and a strengthened generic governance control reproduced all 25 historical parent-eligibility queries. The evaluated fail-closed reuse predicate is therefore portable rather than Overseer-specific. Overseer’s demonstrated contribution is the executable integration of these mechanisms into a source-traceable, temporally reconstructible, and auditable graph substrate.