From governing cases to governing distributions, the decoupling of influence and responsibility, the integrity of execution conditions.
The public debate frames hospital-without-walls care as a liberation. The building dissolves, the patient stays home, the system gains capacity. The reading is not wrong, it is incomplete. Removing care from its walls does not abolish the clinical authority that was exercised within them. It redistributes the technical, informational, and operational conditions under which that authority continues to be exercised.
The thesis of this work is precise. The risk is not the disappearance of a decisional center but the emergence of distributed architectures in which influence over a trajectory, the visibility of what tilts it, and assignable responsibility cease, gradually, to be aligned. This misalignment is not a scandal. It is a structural property, and it can be governed, on condition that it be named before it sediments.
The wall did not hold care as a container holds its contents. It made care legible. Within its walls, three things were in the same place without anyone having to think them separately: who knew, who saw, who answered. An adverse event had a ward, a head of department, a medical commission, an institution that answered for it before the patient and before the regulator. Responsibility was not lighter than it is today. It was localizable.
The morbidity and mortality conference, the institutional medical commission, the on-call rotation, the discharge report: as many devices as presupposed a place where a patient’s trajectory was observed by an identified collective, and where an error could be replayed, discussed, imputed. Hospital medicine took more than a century to build these forms of shared responsibility. All of them rested on a tacit condition: that the trajectory and its observation take place in the same location.
Distributed care removes this condition. It has not rebuilt the forms that depended on it. That finding, and not any architectural nostalgia, opens the problem.
Hospital-without-walls care is not a coming threat. It is a reality installed at scale, whose public indicators are measured in hundreds of thousands of patients per year. In France, home hospitalization treated 184,400 patients in 2024, accounting for 7.7 million patient-days, delivered by more than 280 institutions, at a cost to the public health insurance of approximately 1.94 billion euros. Medical remote monitoring entered ordinary law through two decrees of March 3 and March 31, 2026. The coordination of complex pathways has, since the law of July 14, 2019, a dedicated public operator, the coordination support unit.
The European and American picture follows, at heterogeneous governance rhythms. England counted approximately 11,635 virtual ward beds in March 2025, against a target of 24,000 and an initial allocation of 450 million pounds over two years. The Nordic countries are scaling Hospital@Home at full deployment in 2026. Germany remains centered on psychiatric home treatment, with no national equivalent in acute care. In the United States, the federal Acute Hospital Care at Home waiver was extended by five years through the Consolidated Appropriations Act of 2026, with a deadline of September 30, 2030, covering 366 programs across 139 health systems and 37 states.
The British evidence is clear and limpid. 11,635 home care beds are being deployed, and the minimum dataset that will allow them to be evaluated is planned for 2026. That is to say, afterward. The available systematic review situates the medico-economic effects of home hospital care on an interval ranging from savings above 8,000 euros per patient to additional costs above 2,000 euros. This heterogeneity does not say that the model is bad. It says that it is not economical by nature, and that it is so only under organizational conditions. And organization is precisely what is distributing itself.
The wall delimited a physical perimeter, made of concrete and corridors, visible to the naked eye. In its place installs an informational and operational perimeter, made of what is captured, seen, triggered, escalated, and prioritized. The question is not technical. It bears on visibility and on action.
The inclusion boundary shifts. The wall excluded by space, and everyone knew it. The informational perimeter excludes by absence of signal, and no one decided it. The system sees only what it has rendered observable. The patient without a connected device, or whose pathology does not figure among the inscribed remote-monitoring activities, is not refused. They are invisible to the layer that orients. Care abandonment, which we have analyzed elsewhere as a silent underuse of the trajectory, finds here a new mechanism of emergence: one does not exit care by a decision, one exits by a default of capture.
A second form of exclusion is added, more insidious. The perimeter can exclude by silent alteration of the signal itself. Seeing falsely sometimes becomes more dangerous than not seeing. Incomplete telemetry, a desynchronized threshold, an altered datum, a loss of availability do not produce only an IT incident. They modify the very conditions under which a trajectory becomes visible, and therefore orientable. The system does not see badly by accident. It sees according to a perimeter that no one has instituted as an act, and whose integrity no one guarantees as an act, although it produces all the effects of one.
Distributed care does not hold by magic. It is coordinated. Control tower, command center, orchestration platform, support unit: the names vary, the function describes itself without needing personification. And describing it correctly requires dismissing at once the polemical formula according to which the platform would govern in the physician’s stead. That is not what happens. Maintaining it would weaken the analysis.
Two too-seductive qualifications must be ruled out. Speaking of a “clinical function” of orchestration pulls the subject toward the clinical, and therefore toward the frameworks that already govern the medical act and the software: medical devices regulation, software qualification, AI Act, product liability. Speaking of a “function of structuring clinical trajectories” makes the same error in a more cautious form. The operative qualification lies elsewhere. Orchestration is an operational orientation function with clinical effect. Operational, because it performs no medical act. With clinical effect, because its settings determine care trajectories. This point, a clinical effect without a clinical act, is what escapes existing frameworks, which think responsibility as the attribute of a decision.
The mechanism is without mystery. The device captures a flow of remote monitoring and medical objects. It qualifies this flow through thresholds, which transform a measurement into an alert. It prioritizes alerts according to a window of actionability. It routes the demand toward home, specialist, hospital, or wait. Each of these four operations falls under informational logistics. None has the form of a clinical decision. Orchestration does not decide. It tilts.
A classical governance governs cases. It admits or refuses this patient, treats or abstains, orients here or there. Each act bears on an individual and leaves a decision to display. Orchestration does not govern cases. It governs distributions of probability of access. It does not decide that a given patient will stay home, it adjusts a parameter that renders a fraction of the population more probably maintained at home, and another more probably summoned. The governed object is no longer the trajectory, it is its probability.
Operational definition. A probabilistic governance acts not on identifiable individual decisions, but on the parameters that modify the distribution of probabilities of access, orientation, or care within a population.
This definition distinguishes itself from three neighboring notions with which it is confused. Capacitive triage sorts existing cases under bed constraint, through an explicit decision rule. Populational optimization pursues an aggregate objective. Statistical governance steers through indicators and targets, but via identifiable decisions. Probabilistic governance recognizes itself by the fact that its effect cannot be attached to any individual decision, and that its signature is the absence of a contestable decision. In Valenciennes, a capacitive control system predicts hospital activity at 48 hours with a reliability of about 95 percent. It refuses no patient. It tilts flows, and a tilted flow, repeated at the scale of a territory, ends up determining who will be seen, when, and in what order.
Within the walls, three things held together: influence over the trajectory, visibility of what tilted it, responsibility for what resulted. Distributed care does not abolish them. It desynchronizes them, and it is this desynchronization, not any dispossession, that constitutes the institutional risk.
The drift has a direction. Influence migrates toward layers that do not decide: rule engines, routing logics, queue prioritization, visibility filters. These layers sign nothing, yet they determine what surfaces, what becomes prioritized, what stays invisible. Monday morning, someone raises a threshold. Friday, a physician validates a maintenance at home without knowing that the bar has moved. Responsibility remains anchored to the signatory, influence has passed to the parameter setter. Whoever sets the threshold is not whoever answers for the trajectory. The asymmetry is human before being technical. It opposes those who bear the burden of the error to those who compose its conditions.
To the asymmetry, dilution adds itself. In a distributed chain, an unfavorable trajectory has no single author. The system may remain everywhere compliant with its specifications and produce, at scale, an attrition of access that no decision ordered. Medicine knew how to treat diagnostic error. It does not yet know how to treat the orientation error produced by a system, and its most fearsome case, the unwarranted absence of orientation, leaves no trace in activity indicators.
The strongest objection must be faced head-on. Responsibility frameworks already exist: prescriber, HAD coordinating physician, GDPR data controller, operator of the remote-monitoring device. That is exact, for formal decision-makers. But these frameworks assign responsibility to whoever decides, when influence has migrated to whoever sets parameters. The gap is not the absence of responsibility. It is its decoupling from the actual locus of influence.
The consequence for the decision-maker is not to rebuild walls. The lost territoriality cannot be restored. Nor is it to aim at perfect alignment, which would be paralyzing. The target is more exact, and more demanding: to re-align what must be re-aligned, and to render governable the misalignment that subsists.
Five enforceable principles fix this work. First, every critical orchestration function must be explicitly qualified, in the sense of an enforceable executory qualification, and not left in the state of an implicit technical parameter. Second, orientation thresholds must be auditable, for a threshold is an institutional act disguised as a setting. Third, causal visibility must be reconstructible, which presupposes at minimum the version of the model or rule, its execution configuration, the clinical context, the trace of its modifications, and the enforceable integrity of the conditions under which these parameters were executed. Fourth, contestability must exceed simple logging, for a log is not a recourse. Fifth, which commands the other four, responsibilities must follow the actual capacities of operational influence, and not organization charts or signatures alone.
The reconstructibility targeted is not total observability, impossible and paralyzing. Real distributed systems will remain partially opaque, probabilistic, emergent, incomplete. The criterion is not to see everything. It is to be able, when a trajectory has gone wrong, to trace back to the parameters that tilted it and to designate who answers for them.
The framework is today structural; it must become measurable, on pain of remaining a well-built intuition. Four observables can be proposed, as program more than as result. Detecting a probabilistic governance, by identifying the parameters whose variation displaces, at constant population, the distribution of orientations. Measuring the migration of influence, by estimating the share of the variance of access trajectories explained by orchestration settings, compared to that explained by formal clinical decisions. Objectifying the decoupling, by verifying, for each critical parameter, whether the one who sets it figures in the responsibility loop. Detecting excessive misalignment, by tracking the rate of unfavorable events that cannot be traced back to a parameter and a responsible party, and the rate of unwarranted silences not reviewed. None of these measures exists today in stabilized form. Their absence is not a detail of implementation; it is, in itself, an indicator of the present low governability.
The British lesson is limpid. One does not measure after having deployed, one measures before extending. The minimum dataset awaited for 2026, while the beds have been deployed for years, is the example of what must not be reproduced.
PREDICARE, within the French perimeter, constitutes a relevant ground on which to test these principles, and not a proof that they work. The precision is doctrinal, not rhetorical. What makes it an interesting ground is not that it predicts better, but that it explicitly poses, in its design, the separation of critical functions and patient contestability, that is to say the very conditions of re-alignment. The ground does not demonstrate the doctrine. It tests its feasibility.
Insofar as orchestration parameters govern distributions at scale, their alteration ceases to be a local technical incident. An orientation threshold is not a simple setting; it is a point of systemic modulation, and as such, an object of power. The layers that tilt trajectories become institutional attack surfaces. A silent modification of a threshold, a routing corruption, a desynchronization between rule versions, a partial unavailability of flows immediately displace probabilities of access to care, without any individual decision appearing irregular.
The cybernetic problem is therefore no longer only that of data confidentiality or of system availability. It becomes that of the integrity of conditions of visibility and orientation. Cybersecurity no longer protects an infrastructure alone; it protects the conditions of execution of an architecture of influence. The wall produced an implicit trust by physical co-location. Distributed care must produce an explicit trust, through execution proofs, enforceable traceability, and reconstructible integrity.
Part of the dissociation observed is functionally necessary. It is what enables scale, continuity of follow-up, smoothing of pathways, coordination of a territory. Requiring that a clinician re-decide explicitly each threshold, each routing, each prioritization, for every patient, would amount to dissolving orchestration, and with it the gains that justify distributed care. Perfect alignment is a regulatory myth.
The misalignment described here is a tendency, not a law. Some architectures maintain alignment, by design or by regulation. The assertion that influence has, on average, migrated toward layers that do not decide is not measured at population scale, any more than the efficacy of the British deployment is. It remains a structural hypothesis, with an explicit falsifier: if, for an unfavorable trajectory, one can reconstruct which parameters tilted it and attach an enforceable responsibility to whoever set them, then orchestration remains governed, and the thesis falls.
Finally, the fine distribution of legal responsibility between suppliers and the coordinating operator, in home hospitalization as in remote monitoring, falls under specialized legal counsel that this text does not pretend to formulate. The question is signaled, the law is not settled.
The wall rendered authority localizable without one having to think about it. Distributed care does not render it unfindable at a stroke. It renders it gradually more difficult to reconstruct, as influence lodges in layers no one inspects. There is no single tipping point, no spectacular point of no return, but a slow drift: thresholds modified without trace preserved, routings whose responsible party is not identified, absences of alert that nothing revises.
Governing distributions without a contestable decision, letting influence decouple from responsibility, and neglecting the integrity of execution conditions are not three distinct problems. They are three faces of the same renunciation. Leaving the walls does not compel us to dissolve authority. It compels us to re-index it on actual influence, and to render governable the gap that subsists. Failing which, authority does not disappear in one piece. It becomes, trajectory after trajectory, progressively unfindable.
Distributed care does not redistribute trajectories alone. It redistributes the technical conditions that render those trajectories visible, orientable, and contestable. From then on, cybersecurity ceases to be a peripheral computing layer. It becomes a constitutive property of the very possibility of a distributed clinical authority.
The full PDF is available below. For the French version, see Sortir des murs sans dissoudre l’autorité.