The engine, explained
The Large Behavioral Model.
The engine underneath the Prodigii Platform — built on Active Inference, a published, peer-reviewed Bayesian framework.
Level one — in one paragraph
The whole thing, in one paragraph.
A Large Language Model predicts the next word. A Large Behavioral Model predicts the next state of a system, and chooses the action most likely to reach the outcome you asked for. That is a different job, and it needs a different machine. The Large Behavioral Model under the Prodigii Platform is one: it is optimized to predict the future states of systems rather than to generate language, it reasons on Active Inference — a published, peer-reviewed Bayesian framework — and it uses no tokens to do it. Language models remain the right tool for language. Behavioral models are the right tool for prediction, planning, and executive reasoning. Prodigii uses each for the function it performs best.
Where it comes from, plainly. The core reasoning technology is licensed rather than built at Prodigii. Prodigii reviewed it line by line over three months, then built the governance layer, the enterprise integration, and the applications around it — including the cybersecurity system, delivered in six weeks. The science underneath is Active Inference, published and peer-reviewed, and the architecture is what distinguishes the engine rather than the name: token-free, small enough to run at the edge, and auditable update by update. “Large Behavioral Model” is a term used elsewhere in the field, including by Toyota Research Institute for robot foundation models. Prodigii claims no coinage in it.
Level two — why the difference matters
Two architectures, ten dimensions.
The two model classes are not competitors for one job. They are built for different jobs, and almost every property that matters to an enterprise follows from that. Read the right-hand column as the reason the platform is architected the way it is.
Large Language Models and Large Behavioral Models compared across ten dimensions. The comparison is between two classes of architecture, not between products.
Scroll the table sideways to reach the Large Behavioral Model column.
| Dimension | Large Language Model | Large Behavioral Model |
|---|---|---|
| What it predicts | The next token in a sequence of language. | The next state of a system. |
| What it is optimized for | Generating language: insights, recommendations, and content. | Evaluating alternatives and identifying the actions most likely to achieve a desired outcome. |
| Cost basis | Tokens consumed, per request and per response. | No tokens, and so no direct token costs. |
| Reproducibility | Generative by construction. The same input does not have to produce the same output. | Reasoning is probabilistic and reproducible: the same inputs and model state produce the same result, and every belief update is recorded. |
| Auditability | Outside the scope of the model. Audit and governance must be supplied by a surrounding system. | Policy enforcement, authority management, compliance controls, explainability, and auditability are integrated into the decision and execution process. |
| When governance applies | After output is produced, by whatever reviews it. | Before execution. The governance gate is deterministic: it evaluates every action against authority structures, business rules, safety constraints, and compliance requirements first, and an action that violates one is rejected every time, with the rule that stopped it on the record. |
| Latency | Scales with model size and with the number of tokens generated. | Planning runs in the millisecond-to-second range for the decision problems the platform handles today. Planning cost grows with how far ahead the system looks, which is the constraint the architecture is designed around. |
| Retraining | Improvement generally requires large-scale training or fine-tuning cycles. | Learns from new observations without requiring initial or continuous large-scale retraining cycles. |
| Computational footprint | Substantial, and concentrated in the data center. | Small enough that in Prodigii’s own lab the inference engine runs on a single mini PC alongside other workloads. Deployable across cloud, private cloud, hybrid, edge, and on-premises environments. |
| Role in the platform | Language understanding and language generation. | Prediction, planning, and executive reasoning. |
Every entry in the Large Behavioral Model column is drawn from Prodigii’s architecture documentation and its own lab deployment. The footprint and latency entries are Prodigii’s own characterizations, not independently benchmarked results, and no comparative multiple is offered anywhere on this site.
Level three — the science it is built on
Active Inference: a system that corrects itself against reality.
The Large Behavioral Model reasons on Active Inference, a framework for designing intelligent systems drawn from biological intelligence and rooted in the free energy principle. Stated without the mathematics, it is four steps that repeat. A system maintains a model of its world. It predicts what happens next. It acts to close the gap between what it predicted and what it observes. It updates the model from what actually happened, and predicts again.
The consequence is that the system is corrected by reality on every pass. It does not require a large corpus assembled in advance, and it does not require a retraining cycle to absorb a change in conditions. It requires observations, and it gets those from the enterprise it is deployed in.
At the center
A model of the world
A continuously updated representation of the entities, systems, processes, and operational environment the enterprise runs on. Everything in the loop is measured against it, and every pass changes it.
01
Predict
From the current model, the system computes what is most likely to happen next, and what each available course of action would lead to.
Throughput on line 3 is projected to fall below target within the hour.
02
Act
It selects the action most likely to close the gap between the predicted state and the intended outcome, and executes it once governance has cleared it.
Reschedule the upstream batch and re-sequence two work orders.
03
Observe
It reads back what actually happened from the systems, sensors, and people the action touched.
Actual throughput, actual queue depth, actual changeover time.
04
Update
The difference between what was predicted and what was observed is used to correct the model. The next prediction starts from a more accurate picture.
Changeover on that line runs eleven minutes longer than the model held.
The same loop, on a factory floor
The world model holds the plant: lines, work orders, materials, crews, changeover times, and the current state of each. The system predicts that throughput on line three falls below target within the hour. It evaluates the alternatives available to it, selects re-sequencing two work orders as the action most likely to hold the target, and puts that action to the governance gate, which checks it against authority, safety, and compliance rules before anything moves. The action runs. The system then observes what happened: actual throughput, actual queue depth, actual changeover time. Changeover on that line took eleven minutes longer than the model held. The model is corrected. The next prediction is made from the corrected model, an hour later and slightly more right than the last one.
Nothing in that sequence is language generation, and nothing in it is a workflow somebody wrote down in advance. It is prediction, governed action, observation, and correction, running continuously.
Level four — what this buys the enterprise
Four consequences a buyer can act on.
Reproducible, explainable reasoning
Every parameter is explicit and every belief update is recorded, so the same inputs and model state produce the same result and a decision can be reconstructed afterwards from state, evidence, and policy. That is what makes an autonomous decision reviewable.
Governance before execution
The platform’s governance gate is deterministic: it evaluates every action against authority structures, business rules, safety constraints, and compliance requirements before it runs, an action that violates one is rejected every time, and the rule that stopped it is logged. Governance sits in the execution path, not in a review afterwards.
Learning without retraining cycles
The model is corrected by observation on every pass of the loop. New conditions are absorbed without initial or continuous large-scale retraining.
Deployment at the edge
The computational footprint is small and uses no tokens, so the same system runs across cloud, private cloud, hybrid, edge, and on-premises environments. In Prodigii’s own lab the inference engine runs on a single mini PC alongside other workloads, ingesting a rolling 200 GB network feed.
Level five — where the science comes from
Published, peer-reviewed, and available to read.
Active Inference is not a proprietary idea. It is a research program with two decades of published literature behind it, and the primary papers are listed below so a reader who wants them does not have to email anyone. Dr. Karl Friston, whose work established the free energy principle, advises Prodigii on the scientific foundations of Active Inference. The papers cited here are his published research; Prodigii’s implementation is its own, and nothing in this literature is an endorsement of it.
What Prodigii built is the governance layer, the enterprise integration, and the applications: the platform the licensed reasoning engine sits inside. The science underneath is public, and checking it is encouraged.
The free-energy principle: a unified brain theory?
The canonical statement of the free energy principle.
Active Inference: A Process Theory
The step from principle to mechanism: what a system actually computes.
Active Inference: The Free Energy Principle in Mind, Brain, and Behavior
The book-length treatment, published open access.
A step-by-step tutorial on active inference and its application to empirical data
The most approachable entry point, worked through end to end.
Active inference on discrete state-spaces: A synthesis
The formal synthesis, for a reader who wants the mathematics to implement.
Designing Ecosystems of Intelligence from First Principles
Active Inference as a design program for artificial agents rather than a brain theory.
Level six — what it is not
Four things this is not.
Said plainly, because each of them is a reasonable first guess and all four are wrong.
Not a chatbot
There is no conversation at the center of this. The system takes in the state of an enterprise and returns action, on a loop, whether or not anyone is typing.
Not an LLM wrapper
A Large Behavioral Model is a different class of model with a different objective. It predicts system states rather than tokens, and it uses no tokens to do it. Prodigii runs language models where language is the job, and that is the whole of their role.
Not workflow automation
Automation executes a path someone wrote down in advance. The platform directs the response and activity of other systems in real time, against a model of what is happening now. That is a control layer, not a script.
Not a replacement program
The Prodigii Platform operates above existing systems, applications, workflows, AI models, and operational technology. It is deployed incrementally, alongside what is already running.
Talk to the people who built it
Prodigii walks through the architecture with any organization or investor that asks.
Contact Prodigii