Generated Context versus Governed State: Functional Conditions for Accountable Longitudinal Clinical Reasoning

For decades, the digital records of our health have been nothing more than digital filing cabinets. When a doctor writes a note about a patient's blood pressure, or a nurse records a medication change, that information is stored as a document, much like a letter in a drawer. To understand a patient's current condition, a human must read through these letters, piece together the story, and hold the full picture in their mind. This has always been the work of the clinician: to reconstruct the truth of a person's health from a scattered collection of papers. In recent years, powerful computer systems known as large language models have arrived, capable of reading these documents with incredible speed and fluency. They can summarize a patient's history in seconds, answering questions that once took a human hours to research. But while these systems are excellent at reading the text, they do not inherently keep a separate, organized record of what is actually true about the patient at any given moment. They treat the patient's history as a stream of words to be processed, rather than a living state to be tracked.

This gap between reading fluently and knowing accurately is the central problem addressed by a new paper from researchers at MyndwareMed. The authors argue that treating a patient's medical history as a simple collection of text is a fundamental flaw when trying to build artificial intelligence that can be trusted with long-term care. They propose that clinical reasoning is not a reading comprehension test, but a problem of tracking a changing reality. Just as a robot navigating a room must constantly update its map of where walls and doors are, a medical system must maintain a continuous, governed record of a patient's state. The researchers distinguish between two very different ways of handling information. The first is "generated context," where the computer gathers relevant documents for a single question and then forgets them once the answer is given. The second is "governed state," where the system maintains a persistent, versioned, and auditable record of what it believes to be true, updated only by strict rules that separate what was observed from what was inferred.

The paper does not claim that these new computer models are bad at reading. In fact, they are often better than humans at finding patterns in text. The problem is that reading the record is not the same as knowing the patient. The researchers show that current systems often confuse a written note with a medical fact, or a missing document with a missing condition. For example, if a patient had a kidney problem ten years ago but no tests have been done in the last year, a standard system might confidently state that the patient's kidneys are fine because the most recent note says nothing is wrong. A system with a governed state, however, would recognize that the last known measurement is old and that the current status is unknown. It would explicitly flag the gap in information rather than filling it with a confident guess. This distinction is vital because in medicine, the difference between "we know this is true" and "we have not seen evidence of this" can determine whether a treatment is safe or dangerous.

To solve this, the authors break down the messy process of medical reasoning into five distinct objects that must never be mixed up. First, there is the true state of the patient, which is the actual physical condition of their body; no system can ever possess this directly, only estimate it. Second, there are the observations, which are the specific measurements or notes taken by the care team. Third, there is the evidence, which is the raw data ingested by the system, complete with a record of when it was written, when it was received, and who wrote it. Fourth is the belief, which is the system's current best guess about the patient's condition based on all the evidence it has seen so far. Finally, there is the simulation, which is a hypothetical scenario used to test "what if" questions, such as what would happen if a patient took a different drug. The paper argues that most current artificial intelligence systems collapse these five separate things into one, treating the accumulated notes as if they were the current truth. This leads to errors where old information is treated as new, or where a guess is mistaken for a fact.

The researchers propose a new standard for how these systems should be built, organized into a tiered framework that measures how well a system governs the patient's state. At the lowest level, a system might just store text. At higher levels, it begins to organize that text into structured facts with clear sources and timestamps. The most advanced level requires the system to maintain a separate "belief" object that is distinct from the raw evidence. This belief object must be able to show uncertainty, such as admitting that a measurement is missing or that two different doctors have given conflicting reports. Crucially, this system must be able to replay its own history. If a doctor asks, "What did you believe about this patient's heart condition three months ago?", the system must be able to reconstruct the exact state of its knowledge at that time, based on the evidence available then, not the evidence available today. This ability to look back and see how the system's understanding changed over time is what the authors call accountability.

The paper also introduces a specific way to handle missing information. In a standard system, if a record does not mention an allergy, the system might assume the patient has no allergies. In a governed system, the absence of information is treated as a specific type of data point. The system learns to distinguish between "no allergies documented" (meaning the question was never asked), "patient reports no allergies" (a reported negative), and "allergy testing negative" (an observed negative). Each of these carries a different level of certainty and requires a different clinical response. By typing these absences correctly, the system avoids the dangerous habit of assuming that silence means safety. The authors demonstrate this with a detailed example of a patient with kidney issues over six years. A standard system, looking only at the most recent note, might recommend a risky medical procedure. A governed system, tracking the timeline of measurements and the gaps between them, would correctly identify that the patient's current kidney function is unknown and recommend a new test before proceeding.

This work is presented as a conceptual blueprint rather than a finished product. The authors are clear that they have not yet proven that this new architecture leads to better patient outcomes in a real-world trial. Instead, they have defined the necessary conditions for such a system to exist and provided a way to audit any future system to see if it meets those conditions. They argue that the current focus on making language models smarter and faster is missing the point. The real challenge is not in the intelligence of the reader, but in the governance of the state it reads. They suggest that for artificial intelligence to be truly accountable in a hospital setting, it must move beyond simply generating text and begin to maintain a persistent, versioned, and auditable record of what it knows and what it does not know. The paper concludes by outlining a research program to build these systems, starting with the basic infrastructure for tracking evidence and moving toward a full model of how patients change over time. The goal is to turn the vague promise of "accountable AI" into a concrete set of engineering requirements that can be checked, tested, and trusted.