Evidence comparison · 4 August 2026

One architecture. Two kinds of digital twin. Different evidence boundaries.

Two public preprints by Emanuel Shirbint and Alexander Rybalov apply governed abductive reasoning to different targets: a multiple-sclerosis patient trajectory and a person model. Their shared architecture is conceptual; their permitted interpretations are not interchangeable.

Preprints Not peer reviewed Conceptual research No clinical deployment claim Not diagnostic

The common problem is plausible narrative without sufficient support. In both papers, an LLM can produce a fluent account while missing required premises, relying on a weak mechanism, or discounting counter-evidence. The proposed response is to make uncertainty, provenance, contradiction, role, and governance explicit parts of the system architecture.

Claim boundary. These are Version 1 public preprints. The hosting record identifies each as not peer-reviewed. The medical paper describes a blueprint and staged validation programme, not a clinically deployed system or demonstrated patient-outcome improvement. The human paper proposes a model and research programme, not an empirical personality theory, diagnostic instrument, therapeutic intervention, or licence for covert person inference.

The shared architectural move

Both papers separate three layers that a monolithic generative interface can collapse:

Semantics: what does the observation mean here?

The semantic layer turns narrative fragments into typed events, relationships, uncertainty, and contradictions rather than leaving them as persuasive prose.

Ontology: what is the observation in a constrained model?

The ontology layer tests whether a proposed explanation has the required premises, mechanism, temporal order, and relationship to counter-evidence.

Semiosphere: in whose role and institutional code is it read?

The same fact can carry different meanings for a clinician, patient, researcher, regulator, system operator, or modelled person. The architecture treats that translation as a safety boundary rather than a cosmetic change of tone.

Where the two proposals diverge

Dimension	Medical digital twin preprint	Human digital twin preprint
Primary target	Evidence synthesis around a multiple-sclerosis patient trajectory.	Self-modelling and AI systems that construct models of a person.
Central risk	A fluent clinical narrative can outrun mechanistic, temporal, or counter-evidential support.	A fluent account of personality can harden into premature identity closure.
Proposed controls	Required-premise states, disease ontology, evidence tiers, temporal checks, provenance, discordance detection, and role-specific data boundaries.	Pressure-event provenance, an attractor registry, role-code separation, staleness, contestability, and refusal of covert pressure engineering.
Permitted output	Governed research hypotheses, clinician-review signals, or patient-controlled summaries, depending on subsystem.	Revisable, conditional person-model inferences that remain open to contest and counter-evidence.
Explicit non-claim	No diagnosis, prescribing, autonomous therapy change, clinical deployment, device approval, or validated patient-outcome benefit.	No diagnostic or therapeutic status, no metaphysical claim about the self, and no permission to infer or engineer personality without consent.
Current evidence status	Conceptual architecture and validation protocol awaiting prospective evaluation.	Conceptual framework and research programme awaiting operationalisation and empirical testing.

Why status labels matter

A digital-twin concept can sound operational because it names subsystems, runtimes, data flows, and evaluation stages. Architectural specificity is not equivalent to completed deployment or validated effectiveness. The two papers themselves draw these boundaries; a faithful public account must preserve them.

The same discipline applies across Emanuel Shirbint’s public record. Peer-reviewed clinical studies are reported as peer-reviewed empirical work. Conference work remains separate from journal publication. Public preprints remain non-peer-reviewed. Conceptual models remain separate from production systems, employment, and advisory work.

Primary records

From Plausible Narrative to Sound Abduction: A Governed Abductive Architecture for Medical Digital Twins in Multiple Sclerosis Care — Preprints.org Version 1, posted 12 May 2026; labelled “Preprint,” “Concept Paper,” and “not peer-reviewed.” DOI: 10.20944/preprints202605.0728.v1.
Human Digital Twins and Self-Knowledge: A Governed Attractor Architecture for Sound Self-Abduction — Preprints.org Version 1, posted 2 June 2026; labelled “Preprint” and “not peer-reviewed.” DOI: 10.20944/preprints202606.0132.v1.

Evidence checked 4 August 2026 (Asia/Jerusalem). Citation counts are intentionally omitted because current registry coverage is not reconciled into the canonical claim ledger. For the complete publication-status map, see status, version lineage, and evidence.