Multi-agent debate (DyTopo)

LexCore doesn't ask one model for an answer. It runs a tiered bench of specialists — a root orchestrator, domain managers, and epistemic/conceptual/writing agents — each specialised by its instructions rather than by being a different model.

• Specialists include a contract-risk analyst, clause comparator, precedent researcher, corporate-governance and data-protection agents, materiality and enforceability analysts, a redliner, and a summary writer.
• Each specialist's knowledge lives in a versioned, reviewable file — not a prompt buried in code — carrying Nigerian grounding: CAMA 2020 for governance, the NDPA 2023 (not the earlier NDPR) for data protection, the Arbitration and Mediation Act 2023 for dispute clauses.
• That material is orientation, never a citation. A specialist is told where to look; it is forbidden from stating a section number it cannot see, and must route an ungrounded point to a human gate instead.
• Agents run in parallel and blind to one another each round, so their findings are independent commitments — nobody bandwagons on the loudest output.
• A Q-learning loop orders which specialists run, and an optional recruitment mode lets domain managers pick a domain-scoped subset per round (deterministic, no extra model call).
• A Precedent Board surfaces prior outcomes across the debate.

Findings are produced, challenged in debate, put through the CitationGate, and only then synthesised. The debate is the reason a finding is more than a single model's first guess.