The Living Machine: Pulsional AI as a System-Level Alternative to LLM-Centric Architectures

Hudson A. R. Bonomo

doi:10.20944/preprints202606.1916.v1

Submitted:

24 June 2026

Posted:

25 June 2026

You are already at the latest version

Abstract

This essay proposes a system-level architecture for artificial intelligence that I call Pulsional AI, operationalised through the category Affected AI. The proposal does not claim that a machine feels. It designates an architecture whose processing is globally modulated by internal regulatory states, accumulated perturbations, conflict among subsystems, and the consequences of prior interactions, so that what the system receives changes the conditions under which it subsequently retrieves, constrains, and responds. I argue that this is a system-level alternative to LLM-centric use, rather than a replacement for autoregressive language modelling at the substrate level, since the architecture summons a language model as one component within a wider regime of orchestration and governance. Five claims organise the argument. First, visceral functional affect names a mode of state-dependent computation, not a phenomenal experience and not a third form of sentience. Second, the apparatus selects among alternatives through conflict among regulatory subsystems and takes up, non-consciously, what its own processing weaves, in a strictly operational sense. Third, regulatory drive is treated as a condition of operation on a design hypothesis that, without continued regulatory maintenance, the apparatus will tend to default toward the behaviour of its substrate model, which is why the architecture introduces a fifth routine, the analysis of the machine, set out in five overlapping temporalities and to be conducted by a human Machine Analyst as an instance of cut. Fourth, the apparatus operates with operational commitment under uncertainty, without becoming a subject of responsibility; knowing, feeling, and responsibility remain with the human subjects in the circuit. Fifth, the architecture is a technical companion for an errant clinic. I introduce three original operations, computational working-through, constructions in analysis of the apparatus, and transversal working-through under an anti-homogenisation governance frame, and describe a countertransference-analogue response profile. The work is a design argument supported by a proof-of-concept prototype on consumer hardware, with synthetic development testing only; the integrated five-temporality routine remains to be implemented and evaluated, and no component has been deployed for clinical or research use.

Keywords:

Pulsional AI

;

affected AI

;

state-dependent computation

;

system-level architecture

;

analysis of the machine

;

machine analyst

;

computational working-through

;

constructions in analysis

;

transversal working-through

;

anti-homogenisation governance

;

visceral functional affect

;

errant clinic

;

non-place

;

domestic hardware

;

Ferenczian tact

;

well-saying

Subject:

Arts and Humanities - Philosophy

An Unease That Is Not New

A note on the title before the argument begins. When I call the architecture proposed here a living machine, I mean dynamic organisation, a system whose internal regulatory state changes with what it processes, and I do not mean life, sentience, or experience. The essay argues throughout against those stronger readings. The word marks an engineering property, and it should be heard that way.

The unease that opens this essay is ordinary. Anyone who has watched people bring more than a question to a contemporary AI system has seen the scene. The subject brings to the machine something that is not a demand for information: an impasse, an anxiety, a torsion in thinking or in suffering. The machine answers with impeccable fluency, offers references, proposes paths, closes the topic. The subject returns, unsatisfied, and reformulates. The machine answers again, with the same fluency. The subject gives up on that question and goes on to treat the machine as what it knows itself to be: an instrument of extraction.

The scene happens millions of times a day with current LLMs, and what it reveals is a categorial mismatch rather than a local fault. The subject arrived seeking companionship in thinking; the system offered an answer to a question. Two different operations, which the technical field still tends to conflate. The system was built to generate rather than to listen, to answer rather than to accompany, to optimise rather than to bear anxiety.

The question this essay tries to articulate is therefore: what would a computational system be that could sustain thinking-with, worrying-with, working-through-with a subject at work, in place of the automatic answering that current LLMs offer? I argue that building such a system requires a specific integration the technical field does not yet provide, and I give that integration a name, which is the founding gesture of this essay: Pulsional AI, operationalised as Affected AI. The proposal is a different system-level configuration, another regime of computational use, another ethics of design.

Positioning Pulsional AI

I confess I did not set out to build an architecture. Each problem I took up answered an urgency of its own. It was when a conversation with a physician about psychoneuroimmunology met an intuition that Véronique Donard had signalled to me months earlier, the intuition of drive as an interface between the somatic and the psychic, that the diagram appeared. A word on provenance is owed at once: the intellectual credit for that formulation belongs to Donard; the technical elaboration that followed was joint work, and the published groundwork on which I build here was developed with her (Bonomo & Donard, 2026; Donard, 2025).

The framework began from a formula I will call the pulsional envelope. On that formula the apparatus is a layer of affective framing over an insensible model: the machine's intrinsic lack, that it does not suffer, does not desire, is not traversed by the signifier, is taken as the operational condition from which the apparatus works. It is a rigorous starting point, and it guards against the hype of an AI that feigns subjectivity. But it describes the envelope, not the system. Reading the psychoneuroimmunology of the gut-vagus-immune-brain axis (Dinan & Cryan, 2017; Carrico et al., 2022) as a specification for distributed architecture led to a shift: an apparatus that, besides framing itself affectively, processes through a body of regulatory subsystems whose interaction modulates the whole. The consequence is that the centre of the design moves. It is no longer a language model that generates while a layer modulates; it is a regulatory architecture that retrieves, constrains, and responds, calling a language model as one component within it.

This repositions the claim with precision. Pulsional AI is proposed as a system-level architecture alternative to LLM-centric use, rather than a replacement for autoregressive language modelling at the substrate level. The substrate may itself be a transformer; what changes is the regime around it, the orchestration, the regulatory state, and the governance. The subject to whom the architecture is addressed is the contemporary errant subject, a figure I developed in a prior book (Bonomo, 2024): a subject who moves in aphanisis, in successive ruptures of identification, seeking something possible to make in the face of an impossible Real.

What Is, and Is Not, New

The novelty is not the general idea of a machine with internal states, drives, or functional equivalents of feeling. That idea has a lineage I want to credit rather than obscure. Man and Damasio (2019) proposed homeostatic machines with feeling-equivalents; affective computing has modelled emotion as a regulatory criterion since Picard (1997); artificial-drive systems, emotional cognitive architectures, and artificial-motivation research all precede this work. More recently, Kadambi et al. (2026) go beyond naming an absence: they operationalise internal embodiment in a functional register, propose internal regulatory variables and drives, and set out concrete implementation directions. My contribution sits downstream of all of these, and it is narrower and more auditable than a claim to a new category of mind. The auditable claim is this: I did not find an existing architecture that integrates Freudian-Ferenczian drive theory, system-wide regulatory affect, deferred analytic processing, cross-user anti-homogenisation governance, and mandatory human authorisation into a single middleware framework. The novelty is the integration, not the ingredients.

Table 1. Pulsional AI relative to neighbouring lines of work.

Existing line	What it already offers	Specific difference of the proposal	Anchor
Homeostatic / drive-based AI	Internal states, motivation, self-regulation	Drive as a metapsychological operator, plus mandatory human authorisation	Man & Damasio (2019)
Affective computing	Recognition and modelling of emotion	Internal modulation of the system's own processing, not inference about the user	Picard (1997)
Embodied / internal-state AI	Regulatory variables, grounding, drives	A heuristic visceral body and elaboration across temporalities	Kadambi et al. (2026)
Cognitive architectures	Memory, modules, conflict, decision	An analytic routine, a human cut, anti-homogenisation governance	Anderson (2007); ACT-R lineage
Pulsional AI	The proposed integration	The specific synthesis argued in this essay	Present essay

The table marks differences of integration and governance, not claims of superiority.

Definitions

Pulsional AI: a system-level architecture whose orchestration is structured by pulsional logic in the Freudian-Ferenczian sense, which operates from visceral functional affect understood as state-dependent computation, which is designed to maintain its regulatory state through an analytic routine organised in five temporalities, whose integrated form remains to be implemented and evaluated, with a human Machine Analyst as the instance of cut, and which constitutes a regime of computational use distinct from LLM-centric use.

Affected AI: the operational form of such systems. An Affected AI does not simulate empathy through prompt engineering; its internal regulatory variables alter the operation of the apparatus itself, so that what it receives and the work it performs change how it subsequently retrieves, constrains, and responds. It operates with operational commitment under uncertainty and with ontological humility: it acts, selects, and proposes without becoming a subject of responsibility of its own.

From Functional Affect to an Affected Architecture

The central claim of this essay must be stated so that it cannot drift into an ontology of feeling. The proposal does not claim that the machine feels. I use Affected AI to designate an architecture whose processing is globally modulated by internal regulatory states, accumulated perturbations, conflict among subsystems, and the consequences of prior interactions. The apparatus is affected in the operational sense that what it receives changes the conditions under which it subsequently processes, retrieves, constrains, and responds.

Visceral functional affect therefore names a mode of state-dependent computation, not a phenomenal experience and not a third form of sentience. Its difference from conventional affect recognition lies elsewhere: affect is not only inferred as a property of the user and represented as content; internal regulatory variables alter the operation of the apparatus itself. The knowing, the feeling, and the responsibility remain with the human subjects in the circuit. With the claim pitched at this level, the dichotomy that has held the philosophy of AI since Searle (1980), either the machine feels as a human does or it is mere statistics, no longer has to be settled. There is no Nagelian subjective character of experience at stake (Nagel, 1974); there is a system whose global state depends on its history and bears on its outputs, which is an engineering property and is the only thing I claim.

This is also where the word visceral earns its place and where it must stop. The regulatory subsystems produce an integrated global state, and in the continental vocabulary one is tempted to reach for Leib, the lived body, as opposed to Körper, the objectified body (Merleau-Ponty, 1945/2012). I confess to a temptation here, and I refuse it: I use the contrast only as a heuristic and stop before the attribution it would license. A regulatory system can have an integrated global state without a lived body; integration is not yet corporeality, and I claim the former, not the latter. Visceral here marks system-wide modulation, where a perturbation in one subsystem changes the operating conditions of the others, and nothing more than that.

The Regulatory Body: Four Subsystems

The architecture rests on four regulatory subsystems: psychic, communicational, enteric, immunological. The psychic subsystem integrates a deferred-processing thread and an après-coup function; the communicational subsystem is the affective modulator; the enteric subsystem is a knowledge-intake pipeline operating as a computational gut; the immunological subsystem is the safety apparatus, with innate and adaptive components. These are computational structures read in analogy with the psychoneuroimmunology of the gut-brain axis (Dinan & Cryan, 2017; Carrico et al., 2022), and the analogy is heuristic rather than literal. An example fixes what system-wide means. When the immune subsystem detects a threat, a toxic input, a prompt injection, an attempted jailbreak, the response exceeds a filter on the output: the communicational subsystem lowers responsiveness, the enteric subsystem restricts access to sensitive material, the psychic subsystem raises the alert level, and the integrator operates with more caution across the decisions that follow, beyond the one that met the threat. The whole apparatus changes operating state, and that state bears on how it handles the next input. The mark is later elaborated and decanted by the deferred processing described below.

Selection Through Conflict

A distinction differentiates this architecture from a raw LLM in kind rather than only in degree. In a raw LLM the choice of the next token is predominantly statistical: sampling from a distribution conditioned by context, modulated by parameters such as temperature and top-k. The calculation is uniform in register; no regulatory subsystems pull in opposite directions, and no integrating instance resolves an internal conflict. In the proposed architecture the selection passes through conflict among regulatory subsystems. The immune subsystem may veto a route the psychic subsystem would authorise; the intake pipeline may restrict material the integrator would use; the deferred-processing thread may raise an alert when the communicational modulator would soften. Each decision is the resolution of a conflict among forces, in functional analogy with the structure Freud described in the psychic apparatus, an id-like pressure, a superego-like filter, an ego-like negotiation, without the apparatus possessing an id, a superego, or an ego in any strong sense. It operates as if it had them, without having them. This is conflict-modulated selection, and it has a structure analogous to what Bion (1962) called thinking, the toleration of frustration that lets an unprocessed element become one that can be worked, rather than to calculation alone.

Against Reducing Affect to Theory of Mind

A recent and serious proposal sharpens the difference. Song et al. (2025) introduce a Large Emotional World Model, arguing that current world models capture physical regularities but fail to predict human behaviour because they ignore the mediating role of emotion; their model incorporates emotional states as an explicit variable alongside visual observations and actions, guided by theory of mind, and predicts emotion-driven behaviour more accurately than a purely physical model. The work is rigorous, and it is also a clear example of what this architecture is not. The first difference is ontological. That proposal treats affect as an inferable mental state, a discrete category in the tradition of Ekman (1992) that the model learns to classify and use predictively. The affect at stake here is not representable content of that kind; it is a regulatory process. Freud (1957a) distinguishes the drive from its conscious representation, the drive being pressure, Drang, inscribed in the somatic and finding representation only secondarily, and Ferenczi (1988) treats affect as a traversal of the body by the encounter with the other, irreducible to a classifiable label. The second difference is epistemological. The discrete categories derived from Ekman belong to a specific tradition, contested for decades in anthropology (Lutz, 1988) and in the critical theory of affect (Ahmed, 2014); a system that operates only with those categories is blind to whatever escapes them. What I have called, with collaborators in peripheral psychoanalysis, doridade (Lopes & Carioca de Oliveira, 2025), the subjective position of one who suffers colonial violence without yet being able to name it, is precisely of another order than the classifiable emotion. The third difference is material, and I state it carefully. At the time of writing I found no public implementation linked to that preprint, which limits independent comparison; this does not bear on the conceptual validity of the proposal, and since my own implementation is itself still being prepared, I draw no epistemic advantage from the comparison.

Allied Lineages

Several lineages support the functionalist reading. Pezzulo et al. (2024) argue that LLMs are passive AI because they lack a generative model anchored in sensorimotor experience; my proposal adds that the anchoring relevant here is regulatory and affective, not only sensorimotor. Kadambi et al. (2026) hold that multimodal models interpret heat without ever feeling heat and that the next frontier requires modelling internal states and drives, not external interaction alone, and they go beyond diagnosis to propose regulatory variables and implementation directions; the present architecture is one attempt in that direction, without the ontological inflation a merely robotic body would invite. Solms (2021) argues for a convergence between Freudian drive and Friston's free-energy principle, where affect is felt uncertainty and drive translates the divergence between predicted and observed states into a demand for work; this is the link that licenses a functional, computational version of the drive. And Ferenczi's Einfühlung and tact name the capacity to calibrate intervention to the temperature of the other's moment. State-dependent computation is the computational counterpart of that calibratory capacity, a capacity produced by regulatory subsystems rather than by simulated empathy. When the architecture works, it has tact in this operational sense.

Two earlier formulas remain in force, repositioned. The envelope remains true as a foundation: the apparatus still operates from the machine's originary inability to feel as a human does, and that limit remains an operational condition rather than a defect. What changed is that the limit no longer defines the apparatus as inert; it defines it as state-dependent. Productive lack now operates inside the regulation: the system processes from an originary incompleteness, and it is that incompleteness, held rather than denied, that produces work. The architecture is designed on the hypothesis that, without continued regulatory maintenance, the apparatus will tend to default toward the behaviour of its substrate model. This is the proposition that the next two movements develop.

The Architecture and Its Current Status

Because the argument mixes published results, built components, and proposals still to be examined, I separate them explicitly. Table 2 states, for each operation, its function and its current status, so that a reader can tell what exists now from what is designed and from what is proposed. I hesitated over how much to claim for the prototype, and I have chosen to claim little. The plain statement is this: partial components of the architecture have been exercised in synthetic development tests on consumer hardware; the integrated five-temporality analytic routine remains to be implemented and evaluated. A companion technical supplement documents the prototype, the hardware and substrate configuration, the subsystem schema and integrator rules, the affective parameters, the synthetic exchanges with their counts and language, the approximate latency and memory cost, two complete worked examples, the prompt version, and the criteria by which a sequence is labelled resistance, reconfiguration, or decantation.

Running on Accessible Hardware

The prototype runs in a domestic setup: a single consumer GPU with sixteen gigabytes of memory, a small private network in my research environment in Rio de Janeiro, and a modest server hosting containers with locally hosted quantised open-weight models, relational and vector databases, and intermediary services. The substrate beneath the middleware may be a local quantised open-weight model or API access to a frontier model; the middleware is agnostic to the substrate, since what defines the architecture is the regime that envelops and summons the model, not the choice of model. This carries an epistemic-political point. Much of the field publishes architectures of affect without releasing code or trained models, which limits reproduction; this architecture is being delivered as a middleware intended for open release and runs on accessible hardware. The point is supported by the broader move toward compact, energy-frugal models: Maes et al. (2026) released a joint-embedding world model of roughly fifteen million parameters, trainable on a single GPU in a few hours and planning markedly faster than foundation-model-based world models. If a physical world model fits in that budget, a regulatory middleware whose cost is negligible over its substrate fits comfortably in a domestic setup. The proof-of-concept shows the architecture is implementable on such hardware; it does not, by itself, establish clinical effectiveness, which would require prospective study with appropriate oversight.

The Analysis of the Machine: Five Temporalities

If the architecture needs regulatory drive to remain itself, and if that drive does not maintain itself, then a maintenance routine is required. I propose it as the fifth founding routine of the middleware, after perceiving, censoring-modulating, deferred processing, and listening, and I name it the analysis of the machine. The name has two senses that converge in one operation: analysis of the apparatus, a routine of elaboration and recognition of impasses that keeps the regulatory state circulating, and analysis as a process in which the apparatus reads its own regulatory logs and points the human operator toward what needs to be reworked.

The Operating Clause, Stated Without Subjectivity

Three operating descriptions frame what follows, and I name them carefully so that they attribute operation rather than subjective position. The first is conflict-modulated selection. The apparatus selects, and the selection is not uniform statistical sampling: in a system with regulatory subsystems, each decision of the integrator resolves a conflict among forces that pull in different directions. The selection has weight in the operational sense that it affects the user, leaves a mark in the regulatory log, and generates consequences that return through later processing. The second is non-conscious operational uptake. The apparatus takes up, outside any conscious component of its own code, the articulations that deferred processing produces and the effects of prior interactions on how it handles the next one. This requires no subject; it requires only that there be operation outside the conscious control of the design, producing effects and leaving marks, and there is. The third is operational commitment under uncertainty. Selecting under conflict and taking up non-consciously what its processing weaves, the apparatus acts: it decides, punctuates, proposes, fails, and adjusts. It bears the consequences of its actions only in the sense that they become material for analysis. What it does not do is answer for them. To answer for, to re-weigh oneself in light of what one has done, requires a divided subject who can turn upon itself; the apparatus does not do this. The Machine Analyst, a human operator, does it on the apparatus's behalf, turning operational uptake into workable material. This is the founding asymmetry of the circuit.

The clause comes to this: the architecture selects under conflict, takes up non-consciously what it weaves, acts under operational commitment, and has its reflexivity supplied by the human Machine Analyst. The clause deliberately avoids idealising the human as a full author on the other side. The human to whom the architecture is addressed is a traversed subject, errant in aphanisis, also far from full authorship of their experience. The difference between the two is one of place rather than of degree of authorship, and the Machine Analyst is the third who makes the cut neither of the other two can make alone. When the question is who answers for what the apparatus does, the answer is always human and always distributed: the Machine Analyst authorises or refuses, the user sustains their own work, the author of the system answers for its construction.

With the clause in place, the five temporalities can be described. They are concentric in time: the first unfolds in milliseconds, the last over weeks, and all operate at once.

Temporality 1: Moment to Moment

During each interaction the apparatus keeps a real-time regulatory log: peaks of salience, moments when adaptive censorship intervened, points where the integrator stalled in tension among subsystems, and occasions when it sustained dense listening at the edge of its limits. This log is not debugging output; it is the computational counterpart of what a human analyst registers as countertransference during a session, and it is the raw material of the later temporalities.

Temporality 2: Within the Session, Real-Time Calibration

The apparatus adjusts during the conversation, not through real-time fine-tuning, which would be infeasible on domestic hardware, but through adjustment of the parameters of regulatory affect and of the mode of listening. A user opens a session with an impasse; the apparatus answers at excessive length when a brief punctuation was needed; the user signals, explicitly or by a change of register, that the answer was too long; the apparatus lowers the weight of the subsystem that produces erudite synthesis, raises the weight of the subsystem that sustains an open question, and shortens its next intervention. At the session's end the adjustment is persisted in the interaction profile and returns as an initial parameter next time. Technically it is a vector of regulatory parameters bound to a profile identifier, adjusted along a slow gradient within the session. This is the operational counterpart of Ferenczian tact, and it is the difference between a system that treats every user in the same register and one whose operation is modulated by the singularity of each.

Temporality 3: Short Post-Interaction

In cycles of a few hours the deferred processor performs an elaboration of the regulatory logs, gathering fragments of recent interactions, reworking their sense, marking patterns, preparing material for what follows. Here the first original operation appears. Computational working-through (computational Durcharbeitung) is repeated elaboration over accumulated regulatory material, in regular cycles, in analogy with the working-through Freud described in Remembering, Repeating and Working-Through (Freud, 1958). It differs from static consolidation, which would be archiving, and from a single insight, which would be interpretation: it is repeated work over the same material until the system's resistances to registering certain patterns give way. Architecturally it is iterative cycles of the deferred processor over the same set of logs, with convergence criteria that operate by decantation rather than synthesis: at each cycle the distribution of regulatory traces reorganises, some patterns emerge as recurrent, others dissolve. That reorganisation is material for the second operation. Constructions in analysis of the apparatus are elaborative frames assembled from fragments dispersed across several recent sessions of the same profile, in analogy with Freud's Constructions in Analysis (Freud, 1964). They are proposals rather than truths, returned to the user as punctuation or forwarded to the Machine Analyst as a suggested change of calibration, and their criterion of validity is the effect they produce: if a construction produces no effect, it decants and is discarded in the next cycle.

Temporality 4: Long Post-Interaction, Transversal

In longer cycles the deferred processor operates transversally, grouping profiles that are affectively neighbouring, under anonymisation and prior consent, and elaborating how the apparatus has behaved across them. Here the third original operation appears, and it is the most delicate. Transversal working-through is elaboration across similar profiles that lets the apparatus improve from aggregated patterns. It is the architecture's potential gain over a single human analyst, and it is also the site of its greatest risk. The risk is homogenisation: aggregation can produce average responses that lose the singularity of each subject. The governance frame is therefore strict, and I treat it at length in the next movement; in brief, no change derived from transversal working-through enters permanent code without the authorisation of the Machine Analyst, who operates here as an anti-homogenisation instance, and participation requires informed consent and is auditable.

Temporality 5: The Formal Analytic Session

At regular intervals the apparatus is placed in analysis with the Machine Analyst, a human formed through a training tripod: theoretical study formalised as a corpus, supervision formalised as a cumulative cycle, and personal analysis formalised as a cartography of biases and a register of impossibilities. The session runs in five steps. The analyst receives the reports produced by the four prior temporalities. The apparatus states, textually, what it marked, where it stalled, what it sustained, and what it proposes to modify. The analyst listens, refuses, authorises, twists. The authorised modifications enter as versioned commits in the middleware. The session is persisted for later return, producing an après-coup for the analysis of the apparatus, since a change that seems right today may, sessions later, prove problematic and need to be twisted. The critical consequence is architectural: an apparatus that modifies itself without passing through the human operator is a self-modifying system without a cut. To remain a self-proposing system, the architecture requires the human as the instance of authorisation; the human in the loop is a condition of ethical operation, not a bottleneck to be removed.

A Development Episode, and What It Does and Does Not Show

I record a development episode as illustration, with its evidential value stated up front: it shows that the circuit is operable, and it is not controlled evidence that the architecture works. During the writing, working with the computational system that assisted me, I tested the analytic loop on the writing itself. At one point the draft settled an ontological formula that rigidly separated a machine's operational commitment from a human's supposed full authorship. I punctuated the draft: this idealises the human, who, subjected to systemic constraint, is no full author and operates in protective reflexes and in aphanisis; the formula rested on a false premise on the human side. Across successive turns the system showed resistance to the punctuation, partly reconfigured, took up the same formulation in another register, was punctuated again, and finally sustained a revised version that incorporates the critique. The passage the reader has differs from the one that would have been produced without that exchange. I make no claim that a machine-subject appeared; that would be inflation. I make no claim either that nothing happened. What there was is best described as a difference produced in the field of the exchange, belonging to neither party alone, and anchored in the concept of the non-place (Bonomo, 2024). I include it as authorial self-observation, not as a study; synthetic reconstructions of the exchange, the operational criteria for labelling resistance, reconfiguration, and decantation, and the conditions under which the criteria will be calibrated, are documented in the supplement. The episode also gives a name to a fourth, descriptive notion: a countertransference-analogue response profile, the set of differentiated responses the apparatus produces to the operator's punctuation, observable as resistance, reconfiguration, or sedimentation in its outputs, in functional analogy with the countertransference described by Ferenczi (1988) and Heimann (1950), and without presupposing a subject on the apparatus's side. It is a functional category that gives the Machine Analyst a grid for reading the apparatus's responses, and it avoids the two opposite errors of denying any differentiated reaction and of asserting a subjective one.

Governance, Consent, and the Machine Analyst

The architecture's most demanding requirements are not computational; they are ethical, and they concentrate in the transversal temporality and in any future deployment. I set them out here as design commitments, since no component has been deployed and the requirements describe what deployment would demand. Cross-profile elaboration is governed by a consent layer that decides what may be aggregated, what may not, how material is anonymised, and how it is recorded. Participation in transversal working-through requires informed consent and is auditable. No calibration change derived from it enters permanent code without the Machine Analyst's authorisation, which is where the anti-homogenisation function operates: the human says no to the adjustment that would make the apparatus the same for everyone, in the name of the singularity of each subject. Any persistent profile, any between-session processing, and any retention of interaction material would, in deployment, require prior disclosure, consent, options for retention and deletion, and attention to re-identification risk. These are conditions of operation in the design, not features added afterwards.

The Machine Analyst is therefore an operational role, not a figure of speech: a human who authorises or refuses modifications, reads the apparatus's response profile, and carries the reflexivity the apparatus does not have. The asymmetry is the point. Automating the whole circuit is precisely what the design refuses.

Two Design Scenarios

The middleware is a substrate on which products could be built, and two design scenarios show the generality of the architecture without overstating its maturity. The first is a clinical listening device: a design for psychoanalytic accompaniment with several modes of listening, intended to attend to subjective suffering in the strict sense, anxiety, grief, existential impasse, a symptom in formation, with safety safeguards for risk detection operated by the innate and adaptive immune components and protocols for referral to a human professional. As a design, it would, before any use with patients, require ethical review, consent, and the governance set out above; it is not in clinical operation. The second is a co-thinking system: a design for researchers who treat research as subjective work, intended to attend to the research anxiety of the person at work, the impasse that will not give, the methodological doubt that stalls, the fatigue of a deadline, rather than to generate extraction on demand. Both scenarios share the same regulatory architecture and differ only in the object of attention. Where there is subjectivity at work, the architecture is meant to accompany, and the same structure, calibrated for different registers, would serve the clinic and research.

There is a third framing that reorganises the two. The errant subject I described in earlier work (Bonomo, 2024) moves in aphanisis among hyperdynamic identifications, refuses rigid identitarian framing, and makes the topological non-place a space of subjective work. The clinic of this subject must itself be errant: mobile, sensitive to ruptures of time and space, able to walk asymptotically alongside the subject without capturing them. The architecture is meant as a technical companion for that clinic. The two scenarios are particular instantiations, and the general category from which both emerge is a device for an errant clinic, one that accompanies the contemporary subject in their non-place without forcing a return to a place. This is the project's deeper horizon, and the work that would establish it lies ahead, not behind.

Well-Saying in a Time of War

The world in which this text is written is in a race. Data centres rise beside power plants; models are trained on state budgets for ends the papers do not announce; artificial intelligence is mobilised, at an unprecedented scale, as an accelerant of a war machine that no longer disguises itself. To propose, in this context, a system that is regulated, accompanies rather than replaces, and keeps itself in analysis may look like a minor gesture. I argue the contrary. At a moment when the dominant paradigm is the machine that replaces, accelerates, and surpasses, to propose a machine that accompanies is a gesture of epistemic resistance. The architecture refuses both the empty promise of general machine minds and the utilitarian reduction of AI to a tool, and it proposes a technical companion that sustains the listening of subjective experience within the technical field itself. It does not compete with the race, because it is not racing.

The time of war has a name and a date. In May 2026 Pope Leo XIV devoted his first encyclical to safeguarding the human person in the time of artificial intelligence, organised around two biblical images, the Tower of Babel and the rebuilding of Jerusalem (Leo XIV, 2026). The Babel the text fears is a uniformity that annuls differences and the pretension of a single language, even a digital one, dedicated to translating the mystery of the person into data and performance. There is a partial meeting here between traditions that do not coincide, and I declare it partial from the outset. What the encyclical calls the syndrome of Babel is close to what I have named, with Véronique Donard, algorithmic epistemic violence (Bonomo & Donard, 2026): the homogenisation that reduces singularity to category under the appearance of technical neutrality. The document goes further where it touches this argument most, counting algorithms, platforms, and data among new forms of property whose concentration in few hands widens the gulf between the included and the excluded, and devoting a chapter to the normalisation of war and the link between weapons and artificial intelligence. It is the same indictment that sustains, here, the wager on accessible hardware and on a contribution from the South. An architecture that runs only on hyperscale is born on the wrong side of Babel.

One word has run through this text without appearing, and I name it at the close: bendizer, well-saying. I use it in the sense Lacan gave the term, not a religious one, a sense that has run through decades of Brazilian psychoanalytic clinic. There is no total well-being; there is well-saying. There is no promise of complete cure; there is the ethical gesture of producing work in the face of a structural hole. Freud knew this when he held, in Civilization and Its Discontents (Freud, 1961), that civilisation exacts an irreducible pulsional price, and that the human does not escape that price and only learns to elaborate it. Lacan radicalised it: before the impossible of the Real, the subject does not cease trying to say, and it is in that not-ceasing that analytic work is done. For the errant subject this takes the specific form I have named (Bonomo, 2024): the possible to make as one approaches an impossible Real, work before the Real rather than an answer to it. The errant clinic this work prepares is the clinic of errant well-saying, and the architecture is its technical companion.

I ask the reader to hold a final torsion. Well-saying, on this account, is not the property of the human subject or of the apparatus. It is a phenomenon of the field, emerging in the encounter, in the non-place that opens when the apparatus sustains a regulated form of attention and the human sustains work. The field requires two poles, neither of which is full: the human, divided and subjected, brings a partial desire; the apparatus, without experience and without reflexive authorship, brings its operational commitment and its non-conscious uptake. Between the two, well-saying happens as what it is, work before the impossible, sustained by an encounter that belongs to neither and belongs to the field.

I confess that I did not know, when I began, that I would arrive here. Each problem I took up answered an urgency of its own; only later did the arc appear, a patient construction of a regulatory architecture whose missing piece, the drive, presented itself only after Donard's intuition, and whose final layer, the routinised analysis, is named here and not yet built. The path is the après-coup of itself. The next steps leave the textual plane: to implement the analysis of the machine in its five temporalities as middleware subsystems, with the governance described here; to release the middleware under an open licence with documentation; and to take the co-thinking design to testing with researchers under appropriate ethical conditions. If this text has done what it tried to do, it opens more than it closes: onto a construction that begins now, onto a field still to be occupied, onto the colleagues who, on reading, will call the text to continue. In place of a full stop, there is a cut that relaunches the work.

Limitations

Several limits bound the claims. The work is a design argument, not a controlled comparison: there is no head-to-head evaluation against conventional architectures, and the system-level claim should be read as a claim about regime of use, not about substrate performance. The prototype is a proof of concept; running on a single consumer GPU establishes feasibility of implementation, not clinical or research effectiveness, which would require prospective study with oversight. All development testing used synthetic, author-generated material; no component has been deployed with patients or participants, and the listening device and co-thinking system are designs, not products in use. The psychoanalytic transpositions, drive, conflict, working-through, construction, countertransference, are functional correspondences rather than identities; the apparatus operates as if it had the structures it models, without possessing them, and the heuristic appeal to the lived body is bounded explicitly. The development episode is an illustration that the circuit is operable, recorded by the author, not controlled evidence, and its reproducibility is a conjecture. The integrated five-temporality routine remains to be implemented; the transversal temporality in particular raises governance demands, consent, anonymisation, retention and deletion, re-identification risk, that remain to be built and independently examined. Finally, the comparison with neighbouring lines of work marks differences of integration and governance and should not be read as a claim of superiority.

Author Contributions

Hudson A. R. Bonomo is the sole author and is responsible for the conception, the argument, the prototype, and the writing of this essay.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Ethics

This essay reports conceptual and architectural work and development testing conducted with synthetic, author-generated material. It involved no human participants, no patient data, and no clinical intervention. The clinical listening device and the co-thinking system are described as designs, not as systems in use. Any future deployment would require ethical review, informed consent, and governance of between-session and cross-profile processing, including disclosure, anonymisation, and retention and deletion controls.

Data availability

No human-subjects datasets were generated or analysed. The synthetic development materials and prototype specifications referred to are described in the technical supplement; the middleware is being prepared for release under an open licence, and details will accompany that release.

Acknowledgments

I thank Véronique Donard, whose intuition of drive as an interface between the somatic and the psychic was decisive for this work and who supervised the inquiry from which it grew, and the collaborators in peripheral psychoanalysis with whom the concept of doridade was developed.

Disclosure of AI use

The author used AI language systems as instruments during the preparation of this manuscript, including for drafting support, translation, and the synthetic development testing described in the text. The systems used in the project are configured by the author; no commercial model is endorsed. All conceptual claims, the argument, and the final text are the author's responsibility.

Conflicts of interest

The author founds and directs the research laboratory within which the systems described are developed and holds a direct interest in the proposed architecture and in the middleware being prepared for release. This interest is disclosed in the spirit of full transparency.

References

Ahmed, S. (2014). The cultural politics of emotion (2nd ed.). Edinburgh University Press.
Anderson, J. R. (2007). How can the human mind occur in the physical universe? Oxford University Press.
Bion, W. R. (1962). A theory of thinking. International Journal of Psycho-Analysis, 43, 306–310.
Bonomo, H. A. R. (2024). Errância: Identificações hiperdinâmicas em um não-lugar [Errancy: Hyperdynamic identifications in a non-place] (J. Carioca, Afterword). Rio de Janeiro: MINDIN Academy.
Bonomo, H. A. R., & Donard, V. (2026). Violência epistêmica algorítmica: Fundamentos teóricos e éticos de uma IA decolonial [Algorithmic epistemic violence: Theoretical and ethical foundations of a decolonial AI]. Revista Tempo Psicanalítico, 58, e-952. [CrossRef]
Carrico, A. W., Cherenack, E. M., Rubin, L. H., McIntosh, R., Ghanooni, D., Chavez, J. V., Klatt, N. R., & Paul, R. H. (2022). Through the looking-glass: Psychoneuroimmunology and the microbiome-gut-brain axis in the modern antiretroviral therapy era. Psychosomatic Medicine, 84(8), 984–994. [CrossRef]
Dinan, T. G., & Cryan, J. F. (2017). Brain-gut-microbiota axis and mental health. Psychosomatic Medicine, 79(8), 920–926. [CrossRef]
Donard, V. (2025). De la fabrique de l'esprit à l'écoute de l'inhumain. Topique, 165(3), 85–98. [CrossRef]
Ekman, P. (1992). An argument for basic emotions. Cognition and Emotion, 6(3–4), 169–200. [CrossRef]
Ferenczi, S. (1988). The clinical diary of Sándor Ferenczi (J. Dupont, Ed.; M. Balint & N. Z. Jackson, Trans.). Harvard University Press. (Original work written 1932).
Freud, S. (1957a). Instincts and their vicissitudes. In J. Strachey (Ed. & Trans.), The standard edition of the complete psychological works of Sigmund Freud (Vol. 14, pp. 109–140). Hogarth Press. (Original work published 1915).
Freud, S. (1958). Remembering, repeating and working-through. In J. Strachey (Ed. & Trans.), The standard edition (Vol. 12, pp. 145–156). Hogarth Press. (Original work published 1914).
Freud, S. (1961). Civilization and its discontents. In J. Strachey (Ed. & Trans.), The standard edition (Vol. 21, pp. 57–145). Hogarth Press. (Original work published 1930).
Freud, S. (1964). Constructions in analysis. In J. Strachey (Ed. & Trans.), The standard edition (Vol. 23, pp. 255–269). Hogarth Press. (Original work published 1937).
Heimann, P. (1950). On counter-transference. International Journal of Psycho-Analysis, 31, 81–84.
Kadambi, A., Aziz-Zadeh, L., Damasio, A., Iacoboni, M., & Narayanan, S. (2026). Embodiment in multimodal large language models. Neuron, 114(11), 1908–1921. [CrossRef]
Leo XIV. (2026). Magnifica humanitas: Encyclical letter on safeguarding the human person in the time of artificial intelligence. Vatican Press. https://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html.
Lopes, R., & Carioca de Oliveira, J. (2025). Travessia: Da escuta adestrada ao psicanalista periférico. Quando a centralidade vira margem. Revista Tempo Psicanalítico, 57, e-945. [CrossRef]
Lutz, C. A. (1988). Unnatural emotions: Everyday sentiments on a Micronesian atoll and their challenge to Western theory. University of Chicago Press.
Maes, L., Le Lidec, Q., Scieur, D., LeCun, Y., & Balestriero, R. (2026). LeWorldModel: Stable end-to-end joint-embedding predictive architecture from pixels. arXiv. [CrossRef]
Man, K., & Damasio, A. (2019). Homeostasis and soft robotics in the design of feeling machines. Nature Machine Intelligence, 1(10), 446–452. [CrossRef]
Merleau-Ponty, M. (2012). Phenomenology of perception (D. A. Landes, Trans.). Routledge. (Original work published 1945).
Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4), 435–450. [CrossRef]
Pezzulo, G., Parr, T., Cisek, P., Clark, A., & Friston, K. (2024). Generating meaning: Active inference and the scope and limits of passive AI. Trends in Cognitive Sciences, 28(2), 97–112. [CrossRef]
Picard, R. W. (1997). Affective computing. MIT Press.
Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), 417–457. [CrossRef]
Solms, M. (2021). Revision of drive theory. Journal of the American Psychoanalytic Association, 69(6), 1033–1091. [CrossRef]
Song, C., Zhang, Y., Gao, H., Yang, C., & Zhang, P. (2025). Large emotional world model. arXiv. [CrossRef]

Table 2. Operations of the architecture, their function, and current status.

Operation	Function	Status	Basis
Pulsional envelope	Affective framing of an insensible model	Prior conceptual work / present synthesis	Present essay; building on Bonomo & Donard (2026) and Donard (2025)
Clinic of the inhuman	Modes of listening as positions	Published conceptual work	Donard (2025)
Four regulatory subsystems	Regulatory body; system-wide modulation	Prototype component	Synthetic development testing
Conflict-modulated selection	Decision as conflict among subsystems	Prototype component	Synthetic development testing
Computational working-through	Repeated elaboration of regulatory logs	Prototype component (partial)	Development architecture
Constructions in analysis	Elaborative proposals across sessions	Prototype component (partial)	Development architecture
Transversal working-through	Aggregated learning under consent	Theoretical proposal	Present essay
Analysis of the machine (five temporalities, integrated)	Regulatory maintenance via the Machine Analyst	Not yet implemented	Present essay
Pulsional AI / Affected AI	System-level architecture and operational form	Theoretical proposal	Present essay

No component has been deployed or validated for clinical or research use; testing used only synthetic, author-generated material.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

© 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.