Primary-source research explainer · 7 August 2026

Evidence can support an explanation. It cannot grant authority to act.

A new public preprint by Emanuel Shirbint proposes a governed closed loop for medical digital twins in multiple-sclerosis care. Its central rule is deliberately narrow: an explanation, however plausible or well supported, must not move directly into clinical intervention.

Preprint · Version 1 Not peer reviewed Conceptual architecture No clinical deployment claim Not medical advice No autonomous treatment authority

Claim boundary. The primary record describes a conceptual architecture and proposed validation path. It does not report a deployed medical device, a clinical trial, autonomous prescribing, validated patient-outcome improvement, regulatory approval, or permission to bypass qualified clinical judgment and patient consent.

The paper separates two questions that fluent AI systems can easily blur: What is probably happening? and What should be done, by whom, under whose authority? The first is epistemic. The second is clinical, institutional, legal, and personal.

The governed path

The proposed loop keeps each transition visible so that a recommendation cannot acquire borrowed authority merely because it appears inside a sophisticated system.

1. Evidence
Typed, time-stamped observations with provenance.

2. Patient state
A temporal representation that preserves unknowns and contradictions.

3. Hypotheses
Competing explanations, counter-evidence, and an explicit unknown state.

4. Causal review
Separation of association, mechanism, and intervention effect.

5. Governance gates
Eligibility, evidence quality, scope, consent, and escalation checks.

6. Authorized decision
A legitimate human decision state, not an approval ornament.

7. Action contract
Owner, timing, monitoring, stop rules, and accountability.

8. Outcomes
Observed consequences return as new evidence without overwriting history.

What the gates are designed to prevent

Gate	Question it preserves	Unsafe shortcut it blocks
Evidence integrity	Are source, identity, measurement, and transformation records trustworthy?	Treating generated or mislinked data as observed fact.
Temporal validity	Is the evidence recent enough for this decision?	Turning stale records into current clinical certainty.
Competing hypotheses	What contradicts the leading explanation, and what remains unknown?	Premature closure around a fluent narrative.
Causal support	Does the evidence support an intervention effect, not only an association?	Presenting correlation as a treatment rationale.
Eligibility	Do indication, risk, contraindications, alternatives, and monitoring support consideration?	Confusing a plausible option with a suitable action.
Authority and consent	Who may decide, and has the patient authorized the relevant use?	Borrowing legitimacy from a model, institution, or interface.
Safe learning	Can outcomes update the system without erasing the original recommendation and decision?	Rewriting history or learning silently from ungoverned action.

Human oversight must be observable

The preprint proposes an oversight-quality profile rather than treating a clicked approval button as proof of meaningful review. Relevant signals include whether evidence was inspected, whether alternatives were considered, how long the review took, whether reasoned dissent occurred, and whether escalation remained possible. A low override rate alone cannot distinguish an accurate system from ceremonial approval.

Urgency changes the route, not the invariant

The architecture allows a bounded emergency route through pre-authorized delegation and mandatory retrospective ratification. That is not an exemption from authorization. It changes who authorizes and when the authorization is reviewed while keeping an authorized decision state on the path to action.

What the public evidence establishes

The Preprints.org primary record identifies Emanuel Shirbint as the sole author, labels the work Version 1 and not peer-reviewed, records submission on 30 July 2026 and posting on 31 July 2026, and exposes the conceptual abstract. Crossref's publisher-deposited record confirms the title, author, preprint subtype, DOI, and publication date. Europe PMC independently indexes the same DOI and author record.

Those records establish authorship, public availability, version, status, and the stated conceptual contribution. They do not establish clinical effectiveness, implementation, regulatory standing, production use, employment history, or advisory work. Peer-reviewed journal articles, conference proceedings, preprints, conceptual research, production systems, employment, and advisory claims remain separate evidence classes.

Primary records

Preprints.org Version 1 record — submitted 30 July 2026, posted 31 July 2026, explicitly marked not peer-reviewed.
DOI 10.20944/preprints202607.2365.v1 — publisher-deposited preprint metadata.
Europe PMC record PPR1292259 — matching DOI, title, sole-author string, publication date, and abstract.

Evidence checked 7 August 2026 (Asia/Jerusalem). Citation counts are intentionally omitted from this asset because current registry counts and coverage are not reconciled into the canonical claim ledger.