Continuity Is All You Need

The question

When a clinical language model is updated, what must remain true for the original decision to rely on it to remain justified? The paper calls the distance between controlling a modification process and demonstrating the preservation of validated properties a preservation gap.

The analysis concerns systems with distributed, globally plastic representations: a modification intended to improve one capability can also affect others. It asks what follows for governance when preservation cannot be established from the architecture itself.

The proposed preservation requirement

The Monotonic Preservation Constraint (MPC) requires a property that justified clinical authorization or institutional reliance at an earlier time to persist in the next version, unless it is revalidated under equivalent evidentiary standards.

This is a requirement about the properties supporting reliance, rather than a demand that every output remain identical. A system may gain capabilities while still needing to preserve its validated safety limits, performance bounds, uncertainty handling or conditions of use.

The paper also develops a Risk × Verifiability model. It considers both whether a particular output can be checked and whether preservation across versions can be established. High clinical risk, limited verifiability and frequent updates increase the recurring burden of reassessment.

How the argument is developed

The work combines conceptual analysis of model architecture with an examination of clinical AI governance and change-control frameworks. Examples from deployed systems illustrate the difference between an initial performance assessment and continued justification for reliance.

Its central proposal is to make the continued validity of authorization-relevant properties an explicit object of evaluation. Documentation of a change, monitoring and successful performance on a new benchmark do not automatically establish that every property supporting the earlier authorization was preserved.

Scope and limits

The paper presents a normative and architectural argument. It creates no new dataset, reports no clinical experiment and does not empirically estimate the cost of reassessment. Its strongest conclusion concerns high-risk delegated clinical autonomy under the combination of global plasticity, open-ended scope, frequent modification and absent preservation guarantees.

The analysis leaves room for bounded, inspectable uses such as documentation, information retrieval and advisory support. Clinical usefulness and the justification for delegating autonomous clinical authority are treated as separate questions. The MPC and the risk–verifiability model require operational definitions and empirical work before they could serve as validated assessment instruments.