Words do work. When we write "generative model" or "free energy" on this site, each phrase is doing a specific job, and getting the job wrong is how confusion enters a research program. This glossary is not exhaustive. It is opinionated, cross-linked, and kept short on purpose.
A note on how to read it. Each entry says what we mean, and where useful, what we do not mean. Inline claims are tagged by evidence class: E for expert citation, C for configuration or integration observable in the code, B for tool-observed artifact state. UNI is a working hypothesis on an attainable path toward General Natural Intelligence: a natural, active-inference approach whose evidence is growing, evidence-classed, and tested in the open. Do not take the claim on faith. Test the build, inspect the gates, and help us find where it fails.
The core vocabulary
- Active inference
- A theoretical framework in which a system minimizes variational free energy through both perception (updating beliefs about hidden states) and action (changing the world so it fits the beliefs) (Class E, Parr, Pezzulo, Friston, 2022). On UNI it is the math the labs make steerable, not a metaphor for anything else.
- Free energy (variational)
- An upper bound on surprise, the negative log-evidence a system assigns to its observations under its generative model (Class E). Measured in nats. This is an information-theoretic quantity, not a thermodynamic one. We say "free energy" because the field does; we do not mean joules.
- Generative model
- The probabilistic model an agent carries about how hidden causes produce observations: a joint distribution over hidden states and outcomes, factored by prior and likelihood (Class E). It is the agent's model of its world, not a neural network by default. See Active Inference Fundamentals, a Working Map.
- POMDP
- Partially Observable Markov Decision Process. A discrete-time formalism in which the agent chooses actions under uncertainty about the underlying state, receiving observations rather than states (Class E). Every lab on this site is a POMDP with dials exposed (Class C, from the Precision, Echo, Loop, Heart, and Cell labs).
- KL divergence
- D_KL(Q || P) is the expected extra nats you pay to describe samples from Q using a code built for P (Class E). Asymmetric, non-negative, and zero exactly when the distributions match. It is not a distance. The direction is the whole story. See KL Divergence, What It Actually Measures.
- Bayesian inference
- Updating a prior belief P(z) by a likelihood P(x|z) to a posterior P(z|x) (Class E). In UNI we usually cannot compute the posterior exactly, so we approximate it variationally. That approximation is why KL direction matters at all.
- Markov blanket
- The set of variables that render internal states conditionally independent of external states, given sensory and active states (Class E). We use it as a working assumption for defining the boundary of a system, not as a metaphysical claim about selves. Where the blanket sits is a modeling choice, discussed frankly.
- Precision
- The inverse variance of a probabilistic signal (Class E). In the Precision Lab, three precision dials (sensory, transition, policy temperature) produce distinct behavioral regimes (Class C, observable in the lab). Precision is how the framework encodes confidence.
- Expected free energy (EFE)
- The free energy the agent expects to pay under each candidate policy, decomposed into an epistemic term (information gain) and a pragmatic term (goal-directed value) (Class E). Policies are selected in proportion to the negative of their expected free energy. EFE is not a reward; it is what an active-inference agent uses instead of one.
- Prior preferences
- P(observations) encoded as what the agent prefers to observe, folded into the pragmatic term of expected free energy (Class E). This is how goals enter an active-inference agent without a value function.
- Autopoiesis (as we use it)
- Viable-set maintenance: the agent keeps its state inside a bounded region under disturbance (Class C, operational definition used in the Cell Lab benchmark). We do not mean life. We do not mean consciousness. It is the narrow, testable sense the Cell Lab measures.
- Falsifier
- A pre-registered condition that, if observed, would count against the hypothesis (Class B, see the Cell Lab pre-registration in the preprint). Every UNI claim ships with at least one falsifier; when a falsifier fires, we log it and update.
- Evidence class (A-G)
- A discipline for tagging what kind of evidence backs a claim: A direct runtime observation, B tool-observed artifact, C code-indicated, D test-defined expectation, E test outcome or expert citation, F human or documentation claim, G inference. We tag inline so readers can weight each statement without guessing.
- Stratified Palimpsest benchmark
- Our layered-temporal-task benchmark that stresses generative models with overlapping timescales (Class B, referenced in the Zenodo preprint). It is designed to be falsified on specific claims, not to flatter the framework.
Two terms we treat with care
"Intelligence" and "understanding" appear in the literature we cite, and we use them, but we do not treat either as an achievement claim about UNI. UNI is a working hypothesis on an attainable path toward General Natural Intelligence, natural not artificial. We do not claim to have general intelligence, and we do not claim to have anything artificial. The lab either behaves as predicted in the pre-registered runs, or it does not. That is the level we work at.
Further reading, and one honest recommendation
For learners who want the information-theory scaffolding several entries above (entropy, log-evidence, partition function) sit on top of, Themesis Top Ten Terms in Statistical Mechanics for AI covers the vocabulary at the layer just below this glossary. We recommend it as preparation for the UNI Workshop for math-hungry learners. This is a factual recommendation, not an endorsement in either direction.