An architecture and research-program paper. Autonomous systems increasingly generate plausible research-level mathematics; the unresolved problem is not generation but admission — under what protocol a generated claim may become a load-bearing dependency of future work. This paper reorganizes the Bourbaki Engine around a dual thesis: reliable autonomous mathematics requires two coupled but institutionally separated systems — a Discovery Plane that expands the space of mathematical possibilities (and may be exploratory, redundant, speculative, and internally contradictory) and a Trust Plane that governs which outputs may become reusable knowledge, under a small explicit trusted computing base. It preserves claim promotion as the unit of trusted progress, edge-level soundness obligations, an epistemic separation of powers, certificate discipline (search != evidence != certificate != checker), an event-sourced knowledge substrate with dependency-driven invalidation, and the False Theorem Promotion Rate as the primary trust metric. Two of its seven specified modules are implemented and separately published: MIRADOR (representation) and ProofContext (retrieval). The sixteen-round Bang affine plank pilot is recast as a retrospective requirements-elicitation slice (n=1), and evaluation is gated by a falsifiable R0–R9 readiness ladder. The paper makes no promise of solving a Millennium Prize Problem; a Millennium problem is a gated frontier stress test whose success cannot be self-declared. Includes the full auditable process trail: claim ledger, primary-source related-work matrix, and four fresh-context adversarial reviews with responses.
Lucius et al. (2026) studied this question.