Source-linked AI summary
Generated Context versus Governed State: Functional Conditions for Accountable Longitudinal Clinical Reasoning
Augusto Bernardo Pissarra, Victor Lorena de Farias Souza
TL;DR
Longitudinal clinical AI lacks an explicit, persistent, governed representation of patient state, despite fluent reconstruction from records. This paper frames reasoning as state estimation, specifies the information architecture and maturity framework needed for accountability, and concludes with an analytic audit instrument whose empirical benefits remain open.
Problem
LLM-centric clinical systems do not inherently maintain an explicit, accountable, continuously governed representation of the patient across longitudinal reasoning.
Method
The paper distinguishes generated context from governed state and formalizes patient belief using clinical-time and awareness-time indices within a broader accountability architecture.
Results
The paper specifies four accountability information requirements and a maturity framework separating what clinical AI systems make governable from what they can compute.
Takeaways & Limitations
The framework turns accountable clinical AI from a slogan into an audit instrument while leaving whether the proposed artifacts improve clinical reasoning as an empirical question.
Takeaways & Limitations
The paper's decomposition is analytic rather than a necessity theorem, and whether maintaining its four artifacts improves clinical reasoning remains an empirical question.
Abstract
from arXiv · showhide
Large language models (LLMs) have become the dominant interface of clinical artificial intelligence, yet the interface they expose (text in, text out, one context window at a time) maintains no explicit, persistent, governed representation of what is currently true about a patient. This paper argues that longitudinal clinical reasoning is a state-estimation problem under partial observability, and that the axis on which clinical AI succeeds or fails is not the fluency of the model reading the record but the governance of the patient state it reasons over. We distinguish generated context from governed state; separate five objects that clinical AI habitually conflates (true state, observations, evidence, belief, and simulated state); define a tiered governance standard against which any clinical AI system can be audited; and show that an operational definition of accountability decomposes into four information requirements: an immutable evidence ledger with awareness-time versioning, a belief state distinct from accumulated evidence, an observation-process model, and claim-level causal typing. We are explicit that this decomposition is analytic rather than a necessity theorem, and that its value is conceptual hygiene: it converts "accountable clinical AI" from a slogan into an audit instrument. A six-level maturity framework separates what a system makes governable from what it can compute, locating current LLM-centric practice at high capability but low maturity. The paper is fully self-contained: the four research questions the framework poses are stated in the introduction, and the conclusion records what the paper establishes toward each; future work develops the buildable core of the architecture and the research program toward full Clinical World Models. No empirical result is claimed here.
1 Introduction
The introduction argues that clinical AI must move from reconstructing patient information on demand to maintaining an explicit, accountable, continuously governed patient state. It frames longitudinal reasoning as state estimation under partial observability and presents the paper’s research questions, conceptual vocabulary, accountability requirements, and maturity framework.
- Problem: Clinical information systems have progressively improved document creation, exchange, and display without directly maintaining the patient’s current state.The paper traces this pattern from paper and scanned charts through CDA and FHIR bundles.
- Problem: LLMs can reconstruct patient information fluently on demand, but standard LLM-centric architectures do not inherently maintain an explicit, accountable, continuously governed patient representation.The paper identifies confusing reconstruction-on-demand with governed state as its central problem.
- Motivation: The paper’s operative thesis is that governance of patient state, rather than fluency of the model reading it, determines accountable longitudinal clinical reasoning.The contrast is generated context versus governed state, not neural versus symbolic representation.
- Novelty: Longitudinal clinical reasoning is framed as state estimation under partial observability, centered on what a system must maintain so reasoning can be reconstructed, corrected, audited, and separated from simulation.This reframing organizes the paper and distinguishes its question from research focused on learning patient dynamics.
- Contributions: The paper states four research questions and contributes a tiered audit standard, a vocabulary separating five conflated objects, an accountability decomposition, and a six-level maturity framework.It presents itself as a self-contained conceptual and architectural contribution with no companion reading required for evaluation.
2 Background: where clinical AI stands
Clinical AI already demonstrates substantial medical knowledge and broad practical applications, but its dominant retrieval-and-agent architectures still reason over transient context rather than a shared, governed patient state.
- Existing capabilities: General-purpose LLMs show substantial medical knowledge, including near-expert performance on licensing-exam-style questions, with applications in documentation, triage, and decision support.The paper does not deny these capabilities or applications.
- Architectural limits: RAG grounds answers in a specific chart, but retrieval searches documents rather than estimating patient state.RAG is described as foundational to many clinical AI products.
- Architectural limits: Agent chains add tools and memory, yet each agent still reasons over transient context instead of a shared, governed state object.Chaining language models alone does not create governed shared state.
3 Generated context versus governed state
The section distinguishes transient generated context from a persistent, externally addressable, versioned, clinically governed patient-state object. It frames governed state as a tiered engineering standard whose requirements depend on clinical stakes, not on whether the underlying representation is symbolic or neural.
- Generated context versus governed state: At inference time, patient-specific knowledge enters through a transient context window and leaves as generated text, regardless of any latent internal representation.The interface does not expose persistent state that another system can query or independently audit.
- Generated context versus governed state: Accountable longitudinal reasoning requires an externally addressable, persistent, versioned, clinically governed patient-state object with independently verifiable invariants.The distinction is located in the governed object rather than the neural network, since the same machinery may participate in both regimes.
- Tiered governance standard: The requirements are a normative, architecture-neutral engineering standard organized into tiers matched to stakes, with each tier inheriting the requirements below it.The definition is not presented as a necessary-and-sufficient condition.
- Tiered governance standard: Core governed state requires identity, persistence, explicit semantics, provenance, update governance, and historical query as the minimum for authoritative patient representations.Safety-critical state adds deterministic validation, correction propagation, and explicit uncertainty; predictive state further adds calibration, model-version binding, and monitoring under distribution shift.
- Architecture and system properties: Language models can maintain governed state at the perception boundary, while symbolic systems without provenance and protections against silent overwrites can fail the core tier.Persistence and auditability are properties of the whole system, not consequences of using symbolic representations.
4 Five objects that must not be conflated
Clinical AI must distinguish five objects—true state, observations, evidence, belief, and simulated state—because each represents a different stage in how information is produced, recorded, interpreted, or hypothesized. The framework preserves uncertainty, provenance, multitemporal awareness, and quarantine between governed records and hypothetical branches.
- The five objects: The true state is the patient’s latent physiological and clinical condition, while observations are selective measurements produced by the care process rather than neutral samples.Neither system nor clinician directly possesses the true state; observations reflect what care happened to measure and record.
- The five objects: Evidence consists of immutable, provenance-wrapped observation units with event, documentation, ingestion, and awareness timestamps, while corrections create linked replacements rather than in-place edits.The evidence available by awareness time τ forms the conditioning set for belief.
- The five objects: Belief is the system’s probability distribution over the true state at clinical time t given evidence known by awareness time τ, requiring both clocks to support replay.Comparing b_τ1(x_t) with b_τ2(x_t) asks what the system believed at different awareness times about the same clinical time.
- The five objects: Simulated state is belief evolved under hypothetical actions for prospective “what if?” analysis, and its values must remain quarantined from governed evidence and belief.Branches can be created, compared, and discarded without contaminating the record of what actually occurred.
- The five objects: Conflating evidence with belief makes retrieval fragments appear to answer clinical questions, whereas a belief state preserves competing evidence chains and represents unresolved uncertainty explicitly.In a conflicting medication-list example, uncertainty is the answer; hiding it does not resolve the conflict.
5 One question, asked of both architectures
The contrast-safety question exposes why longitudinal clinical reasoning is not mere retrieval: it requires governed representations of current state, care-process state, prior reactions, and missing observations. A six-year case shows that generated context can produce fluent confidence, whereas governed state represents typed ambiguity and uncertainty-reducing actions.
- Component-level question: Contrast safety requires current renal function, metformin care-process state, prior reactions, and awareness that relevant records may be missing.An order is care-process state, “no reaction documented” differs from “no reaction,” and outside imaging can make record absence differ from world absence.
- Component-level question: The governed architecture can represent typed gaps and actions that would close them, rather than reducing the answer to “yes” or “no.”A governed observation model can represent a missing measurement as a typed information gap and justify obtaining it as uncertainty reduction.
- Case-based illustration: The six-year case includes changing creatinine, CKD, AKI, delayed documentation, inaccessible outside labs, and a fourteen-month measurement gap before the query.Creatinine was 0.9 mg/dL in 2019, peaked at 2.4 during the 2021 AKI, recovered to 1.1 in January 2022, and then lacked measurement for fourteen months.
- Case-based illustration: Documentation lag can expose future information, outside-lab absence is an availability failure rather than absence in the world, and measurement gaps should remain typed ambiguities.The 2022 discharge summary was written thirty days after the events, while 2023 outside measurements existed but never reached the system.
- Case-based illustration: Generated context may retrieve the stale 2022 creatinine of 1.1 and answer confidently, because it lacks the governed objects that account for staleness, unavailable data, and gaps.The illustrative case attributes every architectural difference to objects that exist in the governed architecture but not the generated-context architecture.
6 An accountability decomposition
Accountable longitudinal clinical reasoning requires reconstructable, typed information about what a system knew, observed, inferred, predicted, or simulated, and whether its conclusions are associational or causal. These requirements analytically decompose into four persistent functional artifacts, showing why transient context alone cannot provide accountability.
- Definition: Accountability requires third-party reconstruction of available evidence, separation of observation from inference or simulation, and identification of associational versus causal claims.The definition also requires distinguishing observed absence from unmeasured absence.
- Operational decomposition: The requirements decompose into four artifacts: an immutable awareness-time evidence ledger, an epistemically typed belief state, an observation model, and claim-level causal typing.These artifacts may be maintained explicitly or in an informationally equivalent form.
- Derivation: Awareness-time versioning is necessary because event content alone cannot reveal when the system received it, while separating evidence from belief preserves the observed–inferred distinction.A discharge summary can describe day 2 yet arrive on day 3 or day 30.
- Logical status: The decomposition is analytic rather than a strong theorem: it clarifies and coherently decomposes the definition without claiming an empirical improvement in clinical reasoning.Alternative factorizations remain compatible when the same functional properties survive.
- Consequences: A transient context assembled at reasoning time maintains none of the four artifacts, so prompting, retrieval, fine-tuning, or larger context windows cannot alone make longitudinal reasoning accountable.Accountability requires persistent governed information outside transient context.
7 A maturity framework for clinical representations
The framework separates representation maturity—what clinical systems make explicit and governable—from computational capability—what they can infer internally. It orders clinical representations from raw text to a future Clinical World Model, diagnosing LLM-centric systems as high-capability but low-maturity and calling for maturity to meet capability.
- Motivation: Clinical prose carries rich structure, but that structure remains implicit, inconsistently expressed, entangled with authorship, and therefore not queryable, validatable, updatable, or governable.The operational concern is whether textual structure meets the standard of Governed Clinical State.
- Framework distinction: Representation maturity concerns what a system makes explicit and governable, whereas computational capability concerns what it can infer internally.These properties can diverge: models may estimate latent physiological states without building entity graphs, or host mechanistic world-model machinery without NLP.
- Six-level framework: The six levels progress from raw text through canonical entities, typed relations, governed temporal evidence, and governed belief to a Clinical World Model.The sequence is L0 Text → L1 Entities → L2 Relations → L3 Evidence → L4 Belief → L5 World Model.
- Six-level framework: Level 3 makes assertions time-qualified and provenance-tagged, while Level 4 represents a governed belief state over the patient’s condition.Level 5 extends belief with governed dynamics, an observation model, and causal qualification, and is identified as future work.
- Implication: LLM-centric systems combine high computational capability with low representation maturity, motivating the program of raising maturity to meet capability.The framework is an architectural maturity framework, not a claimed law of intelligence.
8 Positioning, novelty, and what this paper does not claim
The paper locates its novelty not in individual components such as provenance, ontology grounding, or patient-state modeling, but in an architectural invariant governing their relationship. Its contribution depends on replay being deterministic only under a fully pinned ledger, operator, model, ontology, configuration, and stochastic state.
- What this paper does not claim: Provenance, ontology grounding, neurosymbolic decision support, and patient-as-state modeling are established in adjacent clinical-AI work.The paper therefore rejects novelty claims based on these components in isolation.
- Positioning: The framework distinguishes model capability from enforced output governance, including causal claim typing as an artifact requirement rather than merely an evaluative hierarchy.Qazi et al.’s hierarchy covers temporal prediction, action-conditioned prediction, counterfactual rollouts, and planning.
- Novelty: The paper’s novelty rests on awareness time, deterministic belief replay, quarantined simulation, and a governed update operator without a reverse simulation-to-evidence or belief path.It identifies the architectural invariant, rather than any individual component, as the contribution.
- Qualification: Replay is deterministic only relative to the ledger prefix and pinned operator, model, ontology, rule, configuration, and sampled stochastic-state versions.Without model-version binding, the same ledger can produce a different posterior after a model update.
9 Conclusion
The conclusion reframes clinical AI as governed patient-state maintenance: clinical reasoning is state estimation under partial observability, requiring an architectural accountability remedy rather than generated context alone. It answers the four research questions provisionally, identifies the operational belief state and observation model as critical future-work dependencies, and states that no empirical or clinical result is claimed.
- Conclusion: The paper relocates clinical AI from what language models generate to what clinical systems must maintain: governed state for reasoning under partial observability.The proposed remedy is architectural and decomposes accountability into four information requirements, although the supplied passage truncates the list.
- Research questions: RQ1 and RQ4 are predicted, not demonstrated; RQ2 is not yet answerable without an operational belief state, while RQ3 is now well-posed for verification.The architecture predicts improved temporal consistency, provenance, and transportability, but the relevant quantitative comparisons and measurements remain future work.
- Future work: Future work will address RQ1–RQ4 with preregistered quantitative thresholds, beginning with RQ1 and RQ3 before RQ2 and RQ4.The later questions require the operational belief state and observation model first; specification work will publish their engineering treatment and assumptions, though the supplied passage truncates the details.
- About MyndwareMed: MyndwareMed implements the governed evidence layer in production at maturity Level 3, is engineering the governed belief layer at Level 4, and is researching the Clinical World Model at Level 5.The paper presents its architecture as the accountability standard for that program.
- Scope and competing interests: This position-and-specification paper uses no patient data, reports no experiments, and claims no empirical or clinical result.The authors are MyndwareMed employees, and implementation statements concern engineering artifacts rather than demonstrated clinical outcomes; the supplied passage truncates the final sentence.