The Dormant Machine

Two very different traditions have been circling the same idea without quite meeting. One is Critical Realism, a philosophy of science that starts from ontology - it asks what exists? The other is LeJEPA, a recent line of work on self-supervised world models that starts from representation - it asks what must be represented to explain what we observe?

They are walking towards each other from opposite ends of the same tunnel. Critical Realism says reality is layered: beneath the events we record sit mechanisms - real, structured causal powers that generate those events. LeJEPA says a model trained the right way can recover latent variables - the hidden factors that generate the observations. Set the two vocabularies side by side and the rhymes are hard to miss.

The interesting part is not where the two line up. It is where they almost do - and the most valuable mismatch of all is a single Critical Realist idea that today’s world models mostly ignore: a mechanism can be entirely real while lying completely dormant.

Where This Began - The Substrate Beneath the Software

I did not arrive at world models through machine learning. I arrived through a more stubborn question: why does the same intervention succeed in one hospital and fail in another that looks identical on paper? That question led me, almost by accident, to Critical Realism - and to a handful of thinkers who quietly reorganised how I see technology altogether.

Roy Bhaskar gave me the foundation. His insistence that reality is stratified - that beneath the events we observe lie real mechanisms with causal powers, whether or not they happen to be firing - is the single idea this entire blog rests on. Ray Pawson and Nick Tilley then took that ontology and made it operational for the messy social world: their Context-Mechanism-Outcome formula - what works, for whom, in what circumstances - is, if you squint, the exact shape of the dormant-capacity model in this post. A mechanism produces an outcome only when the context fires it. Justin Jagosh sharpened the method, showing how retroduction lets you reason backwards from an outcome to the hidden mechanisms that must have produced it. And Margaret Archer reminded me that structure and agency are not a tug-of-war but a process unfolding over time - her morphogenesis - the continual making and remaking of social structure over time - is a moving picture, not a snapshot.

What these authors share is a conviction I have come to hold tightly: technology never runs in a vacuum. It runs in society - an open system of mechanisms, tendencies, and powers, most of them dormant at any given moment. Build a clinical AI that models only the clinical signal and you have modelled the foam on the wave. The substrate is social. That is precisely why a healthcare world model has to reach for the real, and not merely the actual.

The Idea Machine Learning Keeps Missing

Critical Realism insists that a causal power exists even when it is not firing. A hospital can hold strong leadership that shows no measurable effect until a major transformation programme begins. The mechanism is present; the evidence is absent. A purely data-driven system, trained only on what manifested, would never see it. A Critical Realist would never miss it.

In machine-learning language this is conditional activation, or state-dependent dynamics. A self-driving car’s world model carries a “road condition” variable even on a dry day - its influence appears only when the rain starts. The variable was always there. The trigger was not. Carry that into a health system, and the stakes change. Dormant capacities in healthcare are everywhere:

• A hospital looks stable while carrying a hidden bed-allocation fragility that surfaces only during an ambulance surge.
• A claims platform appears healthy while holding a latent denial propensity that activates when coding and payer rules collide in unusual ways.
• A pharmacy service runs normally while harbouring a stockout sensitivity that manifests only under a supply-chain shock.
• A triage pathway encodes an escalation bottleneck that appears only when wait time, acuity mix, and staffing constraints coincide.

So the research question that sits at this intersection is the one neither community has fully answered: can a world model learn latent variables that represent not only the current state of a healthcare system, but also its dormant causal capacities - the mechanisms that have not yet manifested but could be triggered under future conditions?

Why Symphonix-Health Is the Right Test Bed

Symphonix-Health^† is already built more like a distributed model of a health system than a single application. GHARRA is a federated, zero-trust registry for healthcare AI agents, with a non-PHI control plane, an ABAC-style policy engine, event streaming, and trust routing. Nexus A2A is an inter-agent protocol with 25 clinical agents, a 13-point route-admission process, and more than 7,000 scenario-driven tests. The observability work maps, across 34 repositories, where metrics, traces, audit trails, and health probes already live - and where they do not.

Read those artefacts together and a pattern appears: the tools are the mechanisms. ambulance-ems is pre-hospital flow; appointment-system and scheduling-gateway are scheduling friction; lis and pacs-ris are diagnostic bottlenecks; eps and pharmacy-system are prescribing and medication safety; insurance-eclaims is adjudication; triage-api is escalation. Symphonix-Health now has a verified platform-spine world model. What remains is the full clinical operational world model: longitudinal patient state, care-pathway state machines, ontology-bound clinical semantics, care-team authority, temporal reasoning, outcome feedback, and live telemetry-backed validation.

The Core Move: Separate Active State From Dormant Capacity

Here is the design idea in one sentence. Active state is what is presently expressed - arrivals, beds occupied, queue lengths, pending claims, stock levels. Dormant capacity is a property that exists before it manifests - surge fragility, denial propensity, burnout susceptibility, interoperability brittleness.

In Critical Realist terms, the model should represent both the actual (what is happening now) and the real (what can happen when the right conditions obtain). In machine-learning terms, that means carrying a slow-moving latent (a hidden internal variable the model learns, but never reads off directly) for each mechanism, plus a trigger gate - a switch that decides when that hidden variable is allowed to affect what happens next:

where, reading the symbols left to right:

• x_t - active state: what the system is doing right now (arrivals, beds occupied, queue lengths, stock).
• m_t - dormant capacity: the latent mechanisms not yet expressed (surge fragility, denial propensity, burnout susceptibility).
• z_t - the full latent state, simply the active state and dormant capacity stacked together.
• c_t - context: the surrounding conditions (policy, season, staffing pressure) that can switch a mechanism on.
• u_t - action or intervention: a decision applied to the system (add staff, change a rule).
• g_t - the trigger gate: a value between 0 and 1 that decides whether a dormant mechanism is currently switched on.
• σ - the sigmoid function that squashes the gate into that 0–1 range; W and b are its learned weight and bias.
• ⊙ - element-wise multiply: dormant capacity m_t only enters the dynamics when its gate g_t opens.
• f - the transition function that maps the current state to the next; ε_t is random noise, and x_t+1 is the resulting next active state.

The point is not merely to predict the next observation. It is to make the dormant variable m_t identifiable enough that probing the latent tells you what kind of system you are currently inside - even when the harmful mechanism has not yet fired.

A dormant mechanism will not appear just because the latent cloud is Gaussian, and that single sentence in panel four is the whole lesson for clinical AI: the Gaussian regulariser in LeJEPA (Lean Joint Embedding Predictive Architecture) tidies the model’s internal representation of whatever it is currently encoding, so the active state of a system becomes identifiable, easy to probe linearly and well suited to planning - yet it cannot conjure causal mechanisms that the training data never actually exercised.

For healthcare, the implication is a warning against a seductive false signal. A clinical model whose latent space looks clean, balanced and statistically well-behaved may still be entirely blind to a rare disease pathway, an uncommon adverse drug response, a comorbidity interaction or a subgroup effect - simply because those mechanisms stayed dormant in the data. Organisational neatness is not causal completeness, and it is certainly not safety.

To make such a dormant mechanism learnable you must design the data and the training to make it legible, which in practice means deliberately capturing its triggering conditions, its longitudinal patient trajectories and relevant context variation across diverse populations and care settings - rather than trusting a regulariser to surface it for free. For artificial intelligence more broadly, the takeaway is that representation-quality metrics are necessary but never sufficient evidence of capability: a model must be evaluated by actively exercising the capacities you care about, because whatever you never trigger, you never truly learn - and what you have not learned, you cannot trust.

The architecture follows from this: one encoder for recent event context, one predictor for masked future embeddings, a shared latent bottleneck regularised towards that Gaussian, and two heads - one for active-state probing, one for dormant-capacity and trigger prediction. The decisive evaluation is not forecasting accuracy. It is this: after a trigger fires, does the model behave as though it had already encoded the mechanism before manifestation? If yes, it has learned structure, not surface correlation.

Four Honest Limitations

A proposal is only as trustworthy as its stated weaknesses. There are four, and they are serious.

...And Four Answers

The encouraging part is that each limitation has a concrete, literature-backed response - and several are enforced by machinery Symphonix-Health already owns.

The Test That Matters

If this works, the payoff is not a better forecast. The hierarchical, trigger-aware variant should recover active state about as well as strong sequence baselines - and then beat them at pre-trigger mechanism identification and at planning under regime change. In plain terms: it should be better at knowing what kind of system it is in, not just where the next point on the curve will land.

That is the empirical bridge between a JEPA-style world model and a Critical Realist view of mechanisms. Current world models focus on what is active. Critical Realism reminds us that reality contains causal powers that have not yet manifested - and that a system which can only see the firing mechanisms is blind to the ones quietly waiting for their trigger.

† About Symphonix-Health. Symphonix-Health is a healthcare-AI infrastructure company building agent-first clinical systems: GHARRA, a federated, zero-trust registry for healthcare AI agents; the Nexus A2A inter-agent communication protocol (25 clinical agents, a 13-point route-admission process, and 7,000+ scenario-driven tests); BulletTrain; and a portfolio-wide observability layer. This blog treats that estate as the empirical domain for a healthcare world model. ↩
symphonix-health.com · github.com/Symphonix-Health

References: Klindt, D., LeCun, Y., & Balestriero, R. (2026). When does LeJEPA learn a world model? arXiv. · Assran, M., et al. (2023). Self-supervised learning from images with a Joint-Embedding Predictive Architecture. CVPR. · LeCun, Y. (2022). A path towards autonomous machine intelligence. OpenReview. · Bhaskar, R. (1975). A Realist Theory of Science. · Archer, M. S. (1995). Realist Social Theory: The Morphogenetic Approach. Cambridge University Press. · Pawson, R., & Tilley, N. (1997). Realistic Evaluation. London: Sage. · Jagosh, J. (2020). Retroductive theorizing in Pawson and Tilley’s applied scientific realism. Journal of Critical Realism, 19(2), 121–130. · Johnson, A., et al. (2024). MIMIC-IV v3.1. PhysioNet. · The MITRE Corporation (2025). Synthea synthetic patient population simulator. · Symphonix-Health (2025–2026). Global agent registry, Nexus A2A protocol, and observability gap analysis & implementation plan. GitHub.