Different responsibilities, different tools
Models can help organize evidence, surface alternative explanations and propose tests. Their output remains uncertain. The system that enforces authorization, position limits and order eligibility should not depend on how convincingly a model describes its own proposal.
This separation is central to the SumTwo design: specialized research agents produce structured proposals; deterministic checks decide whether an action is permitted within an approved mandate. Human governance sets that mandate and approves material model changes.
A decision must leave a record
An actionable proposal needs an instrument, direction, horizon, sources, timestamp, version and invalidation conditions. Missing information should remain visible. Reviewer disagreement is part of the decision record, not a detail to summarize away.
Recorded events support operational review. Independent retention, compliance supervision and formal governance procedures are separate capabilities that must be established before a system can claim institutional assurance.
Learning without silently changing the mandate
New evidence may justify a new candidate model or strategy. That candidate should be evaluated as a new version against a defined baseline. Learning is not permission to rewrite live limits, bypass approval or reinterpret an earlier result.