Variability-Aware Trust in Agentic Clinical AI: Modeling and Measuring Trajectory-Level Uncertainty

Variability-Aware Trust in Agentic Clinical AI: Modeling and Measuring Trajectory-Level Uncertainty | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Variability-Aware Trust in Agentic Clinical AI: Modeling and Measuring Trajectory-Level Uncertainty Yunguo Yu This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9326470/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Large language models (LLMs) are increasingly deployed within multi-step agentic workflows in healthcare, where sequential decisions shape downstream clinical actions. However, existing evaluations primarily assess single-turn accuracy and fail to capture how stochastic variability propagates across multi-step trajectories. In this work, we introduce Variability-Aware Agentic AI Trust (VA-AAT), a framework for modeling and measuring trajectory-level uncertainty in agentic clinical systems. We formalize variability propagation through trajectory divergence and derive conservative bounds that characterize how per-step stochasticity may accumulate under simplifying assumptions. Building on this formulation, we define repeatability-based metrics that quantify consistency at both step and trajectory levels, and we propose a risk-informed trust framework linking repeatability to graded oversight strategies. We empirically evaluate VA-AAT on simulated and real-world clinical workflows, demonstrating that small per-step variations can lead to substantial divergence in multi-step trajectories. Variability was strongly influenced by model selection and decoding temperature, and we identified instruction-following failures (“unknown actions”) as an additional source of instability. Importantly, we show that repeatability provides a useful but incomplete indicator of system reliability: highly repeatable systems may still produce consistently incorrect outputs, while less repeatable systems may reflect underlying clinical ambiguity. Accordingly, repeatability should be interpreted as a diagnostic signal of variability-induced risk rather than a guarantee of safety. These findings highlight the need for trajectory-level evaluation and variability-aware governance in clinical AI. The proposed framework provides a foundation for assessing stochastic reliability in agentic systems and supports the design of risk-sensitive deployment strategies in healthcare settings. Artificial Intelligence and Machine Learning Bioinformatics Medical Informatics Information Theory Information Retrieval and Management large language models clinical AI agentic AI repeatability variability calibration trust escalation Software as a Medical Device FDA regulation Jensen–Shannon divergence Full Text Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9326470","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":617796760,"identity":"ce168638-0931-43de-a3d5-56ab8ba31a52","order_by":0,"name":"Yunguo Yu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAw0lEQVRIiWNgGAWjYDADfhCRUECKFskGkBYDUrQYHACTxKi8kWMmXfDrsJzx+dWJHx4YMMjzix0gQsvMvsPGZjfebpYAOsxw5uwEIrTw9hxO3Hbj7AaQlgSD28Rq2Tzj7OYfxGvh+XE4cQN/7zbibJE886zYmrch3VjiBu82iwQDCcJ+4TuevPE2zx9rOf7+s5tv/qiwkeeXJqBF4QCHAQNjG5AlAVYpgV85CMg3sD9gYPgDZPEfIKx6FIyCUTAKRiYAAAlWRzOowwZ5AAAAAElFTkSuQmCC","orcid":"https://orcid.org/0009-0006-2003-3813","institution":"Zyter|TruCare","correspondingAuthor":true,"prefix":"","firstName":"Yunguo","middleName":"","lastName":"Yu","suffix":""}],"badges":[],"createdAt":"2026-04-05 13:23:03","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9326470/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9326470/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":106403640,"identity":"8950e8d0-e1a7-4607-8e30-5c39de1417fe","added_by":"auto","created_at":"2026-04-08 09:14:40","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1668536,"visible":true,"origin":"","legend":"","description":"","filename":"YuAgenticVaribility2026.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9326470/v1_covered_d29e3210-a5c7-4499-8370-2bbec876ee02.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eVariability-Aware Trust in Agentic Clinical AI: Modeling and Measuring Trajectory-Level Uncertainty\u003c/p\u003e","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Zyter|TruCare","isAcceptedByJournal":false,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"large language models, clinical AI, agentic AI, repeatability; variability, calibration, trust escalation, Software as a Medical Device, FDA regulation, Jensen–Shannon divergence","lastPublishedDoi":"10.21203/rs.3.rs-9326470/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9326470/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLarge language models (LLMs) are increasingly deployed within multi-step agentic workflows in healthcare, where sequential decisions shape downstream clinical actions. However, existing evaluations primarily assess single-turn accuracy and fail to capture how stochastic variability propagates across multi-step trajectories.\u003c/p\u003e\n\u003cp\u003eIn this work, we introduce Variability-Aware Agentic AI Trust (VA-AAT), a framework for modeling and measuring trajectory-level uncertainty in agentic clinical systems. We formalize variability propagation through trajectory divergence and derive conservative bounds that characterize how per-step stochasticity may accumulate under simplifying assumptions. Building on this formulation, we define repeatability-based metrics that quantify consistency at both step and trajectory levels, and we propose a risk-informed trust framework linking repeatability to graded oversight strategies.\u003c/p\u003e\n\u003cp\u003eWe empirically evaluate VA-AAT on simulated and real-world clinical workflows, demonstrating that small per-step variations can lead to substantial divergence in multi-step trajectories. Variability was strongly influenced by model selection and decoding temperature, and we identified instruction-following failures (“unknown actions”) as an additional source of instability.\u003c/p\u003e\n\u003cp\u003eImportantly, we show that repeatability provides a useful but incomplete indicator of system reliability: highly repeatable systems may still produce consistently incorrect outputs, while less repeatable systems may reflect underlying clinical ambiguity. Accordingly, repeatability should be interpreted as a diagnostic signal of variability-induced risk rather than a guarantee of safety.\u003c/p\u003e\n\u003cp\u003eThese findings highlight the need for trajectory-level evaluation and variability-aware governance in clinical AI. The proposed framework provides a foundation for assessing stochastic reliability in agentic systems and supports the design of risk-sensitive deployment strategies in healthcare settings.\u003c/p\u003e","manuscriptTitle":"Variability-Aware Trust in Agentic Clinical AI: Modeling and Measuring Trajectory-Level Uncertainty","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-07 05:26:44","doi":"10.21203/rs.3.rs-9326470/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"64cdfede-f8e6-4a1f-9cb0-e73bf85c24c2","owner":[],"postedDate":"April 7th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":65747025,"name":"Artificial Intelligence and Machine Learning"},{"id":65747026,"name":"Bioinformatics"},{"id":65747027,"name":"Medical Informatics"},{"id":65747028,"name":"Information Theory"},{"id":65747029,"name":"Information Retrieval and Management"}],"tags":[],"updatedAt":"2026-04-07T05:26:44+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-07 05:26:44","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9326470","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9326470","identity":"rs-9326470","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}