The authors ground their argument in the JLAC11 code list, Japan’s 17-digit laboratory code standard maintained by the Japan Society of Laboratory Medicine. Drawing on version 1.2 of the publicly available list, they counted rows, distinct codes, and method codes per combination of analyte and specimen, and compared the method name with the product name in each row. The point of this exercise is not bookkeeping for its own sake: it is to show that the same analyte measured on the same specimen can be reported under codes that embed different measurement methods, and that an AI model trained on one method’s distribution may silently misinterpret another’s.

This is where the commentary’s central claim bites. Laboratory-aware AI checklists, such as the EFLM checklist for the assessment of AI and machine learning studies in laboratory medicine published in Clinical Chemistry and Laboratory Medicine, have made real progress in forcing researchers to specify data provenance, preanalytical conditions, and analytical platforms before a study is accepted. But Yuji and Yuji argue that these checklists are, by design, pre-deployment instruments. They ask whether the right things were documented before the model was built. They do not ask whether the meaning of the data flowing into the model will remain stable after the model is running in a live hospital information system.

The authors connect this gap to Japan’s ongoing push for electronic health records and electronic prescriptions. The Ministry of Health, Labour and Welfare set targets in August 2025 for electronic prescriptions and electronic health records, and the Japan Association for Medical Informatics has published the JP-CLINS FHIR implementation guide for the Electronic Health Record Information Sharing Service. An Act amendment consolidated in force from 1 October 2026 added provisions to the Act on Promotion of Comprehensive Securing of Medical and Long-Term Care in the Community. Together, these regulatory and technical moves are pushing laboratory data into shared FHIR-based infrastructure at unprecedented scale, which means the semantic quality of laboratory codes is no longer a local concern but a national one.

Yuji and Yuji’s argument draws on a body of recent work in the same journal and adjacent outlets. Perdomo Lampignano and colleagues wrote in 2026 about operationalising post-deployment surveillance of AI medical devices in clinical practice, arguing that validation alone is insufficient. Ver Berne and Jacobs asked when useful clinical AI exceeds meaningful oversight. Dong and colleagues examined how clinical readiness claims derived from medical AI benchmark results should be governed. Sorin and Klang insisted that clinical AI generators and reviewers must be tested together. Shujaat argued for lifecycle assurance for clinical artificial intelligence beyond regulatory approval. Each of these pieces, the commentary suggests, points toward the same conclusion: the hard part of AI governance in medicine is not what happens before deployment but what happens after.

But the commentary adds a twist that these frameworks do not fully address. Post-deployment surveillance typically monitors model performance—accuracy, drift, calibration—against a fixed reference. Yuji and Yuji argue that in laboratory medicine, the reference itself is unstable. Hemolyzed specimens, a well-documented challenge for emergency departments and clinical laboratories since Lippi and colleagues’ 2011 review, can shift analytical results in ways that depend on the measurement method. Plebani, Laposata, and Lundberg’s brain-to-brain loop concept for laboratory testing, introduced forty years before their 2011 reflection, already recognised that laboratory results are not static objects but part of a communicative process between clinician and laboratory. An AI model embedded in that loop inherits all of its instabilities.

The commentary also engages with the question of metrological and algorithmic traceability. Li and Plebani, writing in the Journal of Clinical Laboratory Analysis in 2026, examined how machine learning in laboratory medicine can be made traceable both metrologically and algorithmically. Padoan and colleagues asked whether metadata and peridata in clinical laboratory data flows could bridge the gap to new AI-based applications. Park, Kim, and Badrick and colleagues addressed property and unit discordance in laboratory interoperability, proposing NPU as a governance layer. Naliyatthaliyazchayil and Stenerson harmonized LOINC codes and units in real-world oncology data. Okada and Yamamoto mapped LOINC and Japan’s JLAC standard using generative AI for international collaborative research. Each of these efforts, the commentary implies, is trying to stabilise meaning—but stabilising meaning is a continuous process, not a checklist item.

Yuji and Yuji’s own prior work frames the stakes. In a 2026 piece in Clinical Chemistry and Laboratory Medicine, they argued that artificial general intelligence and the clinical laboratory represent a paradigm shift toward Lab 2.0, echoing Crawford and colleagues’ 2017 Project Santa Fe report on clinical lab 2.0. If laboratories are moving toward a model where AI is not just a tool but a participant in the diagnostic process, then the semantic quality of the codes that AI reads becomes a first-order clinical safety issue, not a technical detail. The commentary’s title—Semantic Governance Stops at the Model—is a deliberate provocation: the checklists that govern what AI should know about the laboratory stop working once AI is deployed, and the frameworks that govern deployed AI do not know enough about the laboratory.

The authors are explicit that they are not calling for the abolition of checklists. The EFLM checklist and the comprehensive checklist for the self-assessment of medical AI studies introduced by Cabitza and Campagner in 2021 have improved the quality of medical informatics research. Rather, the commentary argues that checklists are necessary but insufficient, and that the remaining work—ensuring that the meaning of laboratory data remains stable and interpretable after deployment—must be carried by post-deployment oversight frameworks that are currently not equipped for it. The authors received no external funding for the submitted work, and the Comment did not involve human participants, human data, or animals.

The commentary’s practical implication is stark. As Japan and other health systems push laboratory data into shared FHIR infrastructure, and as AI models are deployed against that data at scale, the gap between pre-deployment checklists and post-deployment surveillance becomes a live clinical safety issue. Yuji and Yuji’s contribution is to name that gap precisely: semantic governance stops at the model, and everything after the model is left to oversight frameworks that were never designed to carry the weight of laboratory meaning. Whether post-deployment surveillance can be redesigned to carry that weight is the question the commentary leaves open.

Subject of Research: Semantic interoperability and post-deployment oversight of AI in laboratory medicine

Article Title: Semantic Governance Stops at the Model: What Laboratory-Aware AI Checklists Leave to Post-Deployment Oversight

Article References: Semantic Governance Stops at the Model: What Laboratory-Aware AI Checklists Leave to Post-Deployment Oversight. (n.d.). https://doi.org/10.1007/s10916-026-02478-y

Image Credits: AI Generated

DOI: 10.1007/s10916-026-02478-y

Keywords: artificial intelligence, laboratory medicine, semantic interoperability, post-deployment surveillance, clinical decision support, health informatics, JLAC11, FHIR, Semantic, Governance, Stops, Model