# The Boundary Chooses The Theory

Healthcare has no single theory deep enough to govern it. Healthcare has boundaries where the next action forces a choice among theories.

The clinician anchors on physiology. The patient anchors on lived experience and trust. The coder anchors on specificity. The payer anchors on medical necessity. The quality team anchors on measurable process. The compliance team anchors on defensible record. The future clinician anchors on memory that still makes sense later.

Each anchor is real. Each becomes dangerous when it acts alone.

The product problem is to make partial truths correct each other while correction can still change the next action.

That is the point of care.

Friston's free energy principle and active inference belong here as a grammar for boundary selection. Signal crosses in. A model updates. Action goes out. Prediction error returns. The useful question is where the system can still reduce the uncertainty that would otherwise become repair.

That grammar keeps the theory modest. Product does not need the whole metaphysics of life. It needs a way to ask where an anchor should enter the loop.

Scientific pluralism gives the better posture. Working science often uses many models, methods, and explanatory aims because different inquiries need different handles. Healthcare intensifies that plurality because its truths are biological, personal, institutional, legal, financial, and temporal at once.

The chart is healthcare's internal model. A clinical conversation becomes orders, referrals, patient instructions, codes, authorizations, quality measures, claims, audit evidence, legal record, and future clinical memory. When the chart is vague, every downstream actor receives a different failure. A coder queries. A payer denies. A patient misunderstands. A future clinician inherits ambiguity.

Those are correction loops arriving after the action window closed.

Retrospective CDI is the cleanest example. The system needed diagnostic specificity while the patient and clinician were together. It asked later, after context cooled. Point-of-care CDI moves the missing anchor back to the live boundary.

This is the deeper category underneath ambient AI. The scribe was the first sensor. It listened where the old system lost signal. Once the product hears the encounter, the note is only one output. The same signal can clarify diagnosis, support code, ground medical necessity, prepare an order, produce patient instructions, create audit evidence, and warn that a missing phrase will become a later denial or query.

Ambience matters because its public posture is closer to this plural-anchor boundary product. It presents documentation and coding as one platform, and emphasizes Epic/FHIR integration, point-of-care revenue integrity, specialty-specific documentation, CDI, audit trails, compliance, patient summaries, referrals, and inpatient CDI around diagnostic specificity during care. That looks like a product trying to coordinate healthcare's anchors while the encounter can still answer.

Abridge is the serious countercase because it is moving toward the same boundary from a stronger adoption base. Its public materials emphasize clinically useful and billable notes, Linked Evidence, contextual reasoning, problem prediction aligned with billing codes, orders, revenue-cycle surface area, and point-of-conversation prior authorization partnerships. Distribution can beat architectural clarity if it reaches the live boundary first.

The prediction stays conditional. If buyers keep naming the category "ambient scribe," Abridge has the obvious path: trusted notes, clinician adoption, auditability, EHR integration, and enterprise deployment. If buyers learn to name the deeper category, Ambience has the cleaner starting shape: point-of-care coordination across clinical, coding, CDI, payer, patient, and audit surfaces.

The most important thing healthcare does not yet understand is that AI will make its boundaries visible. Over the next thirty years, every denial, query, delay, appeal, duplicate intake, missing order, unclear instruction, and orphaned referral will become legible as a correction loop that fired too late. The old vocabulary will call this workflow transformation, revenue integrity, quality, prior authorization, patient engagement, or risk capture. The product science underneath is anchor selection at a live boundary.

There are real ways this breaks. Clinicians may reject live coding and CDI support if it feels like billing pressure inside care. Regulators may punish products that blur clinical judgment and revenue optimization. Epic, Microsoft, or payer-owned infrastructure may claim the boundary. Abridge may converge fast enough that Ambience's initial architectural advantage disappears.

The durable claim sits below the company call. Healthcare AI will move from generating artifacts after work to coordinating partial truths while work is happening. The winning product will know which anchor matters before the system pays the cost of being wrong.

Point of care eats the scribe because the encounter is where healthcare's truths can still correct each other.

## Source Notes

Scientific-pluralism sources pressure the posture of the node: successful inquiry often works through multiple theories, models, methods, practices, and aims, while unity remains a debated ideal rather than a default operating assumption. The node's product translation is anchor selection, not relativism.

Friston and collaborators' Markov blanket and active inference work frames living systems as bounded systems whose internal and active states reduce uncertainty across a sensory/action boundary. The use here is intentionally narrow: boundary, model, action channel, prediction error, correction loop.

Public criticism of the free energy principle often centers on overbreadth or tautology. That critique is part of the argument: a formalism can be too broad as a total explanation while still yielding a useful product primitive.

Ambience's public materials position the company as a documentation-and-coding platform for clinicians and health systems, with Epic/FHIR integration, point-of-care revenue integrity, coding/CDI, patient summaries, referrals, specialty tuning, and an inpatient CDI product aimed at diagnostic specificity, present-on-admission designations, complication tracking, audit trails, and compliance.

Abridge's public materials position it as an enterprise AI platform for clinical conversations, with clinically useful and billable notes, Linked Evidence, contextual reasoning, problem prediction aligned with billing codes, real-time orders, revenue-cycle surface area, and prior-authorization collaborations with Highmark and Availity at the point of conversation. That is the strongest counterweight to the prediction: Abridge is moving toward the same convergence.