Externalizing Epistemic Governance

Where does authority sit?

A fluent clinical answer can conceal decisions about which evidence counts, how conflicting sources are ranked and what a local rule permits. The paper asks who controls those decisions and how a clinician can inspect or contest them.

Its central distinction is between epistemic authority, what the evidence supports, and deontic authority, what a jurisdiction or institution permits and reimburses. A reimbursement restriction must not be presented as though it established a clinical fact about the treatment itself.

A constrained architecture for oncology

The proposed architecture assigns context classification and evidence assessment to dedicated layers: a Context Utility Layer (CUL) and a Truth-Checker Layer (TCL). Source admissibility, authority rules and claim states are made inspectable, while the language model operates within a bounded extraction and explanation task.

A versioned clinical decision node specifies the fields and constraints relevant to a particular context, such as disease stage, treatment line, jurisdiction and trial participation. Extracted values must be linked to supporting passages. Missing or ambiguous required information becomes an explicit state that can stop downstream output, rather than an invitation to complete the record by inference.

Evidence-supported options excluded by a permission or reimbursement rule remain visible and labelled. Source disagreement, uncertainty and clinician edits are recorded. A concise clinical view and a more detailed audit view derive from the same underlying governance record.

Auditability and contestability

The contribution is both architectural and normative. The design makes explicit who determines source authority, which constraints apply and where an override occurred. This supports attribution and review of the decisions embedded in the system.

The paper also examines the limits of moving governance outside the model. A second unconstrained generative layer would reproduce the original problem. Even an inspectable, stable rule can entrench an unjust allocation decision, so transparency about authority is a prerequisite for assessing that authority, not a guarantee of its legitimacy.

Scope and validation still required

This is a conceptual architecture instantiated in oncology. The paper does not report cohort-level validation, a clinical outcome evaluation or a fully validated decision-support product. It proposes retrospective assessment of missingness, guideline divergence, documented overrides and jurisdiction-aware source prioritization, followed by prospective workflow evaluation.

Auditability alone does not provide an effective route to challenge a decision: contestation also needs a channel and a person with standing to act. The architecture cannot itself represent every patient interest, resolve a conflict between a governance rule and an individual patient, or establish that a recorded human override amounted to genuine control. Those are substantive limits, not merely pending performance tests.