Active Inference (Parr, Pezzulo, Friston, MIT Press, 2022) is the textbook this project reads from. Here is a chapter-by-chapter map of where the UNI build leans on it hardest, where it departs, and why. No endorsement is claimed by any author of the book: the citations are ours, the readings are ours, the responsibility for how we use them is ours.
The frame we adopt
Chapters 1 through 3 lay out the free-energy view of self-organising systems: a Markov blanket separating internal and external states, a generative model on the internal side, and variational free energy as an upper bound on surprise that both perception and action minimise Class E. UNI adopts this frame as-is. Every lab on the site is a discrete-time POMDP under this bound, and every dial on those labs is a term inside it Class C.
Chapter 4 introduces the discrete-time POMDP formulation: categorical distributions over hidden states, an A matrix for likelihoods, a B matrix for transitions, a C vector for prior preferences, a D vector for initial state, and expected free energy as the policy-scoring rule with an epistemic term (KL divergence between posterior and prior) plus a pragmatic term (log-preference over observations) Class E. This is the chapter our five labs implement most literally. If you open the Precision Lab and read the source, the variable names are the textbook's Class C.
Where we lean hardest
Chapters 4 and 7 (discrete-state POMDPs and message passing on categorical models) carry most of the weight. The Cell Lab controller is a depth-2 planner over categorical states with an explicit A, B, C, D and an expected-free-energy tie-break, and the pre-registered benchmark reports RecoveryScore against random, rule-based, and neural baselines with the failure rows shown, not hidden Class B.
Chapter 2 on precision (the inverse variances that weight sensory, transition, and policy signals) is the source of every dial the labs expose. The Loop Lab bifurcation map is a direct empirical read on that chapter: change sensory precision, watch the behavioural regime change, everything else held fixed Class B.
Where we depart, and why
Chapters 5 and 6 develop continuous-state generalised-coordinate formulations with predictive coding as the standard neural implementation. UNI does not use those in the public labs. Not a disagreement: a scope choice. Discrete POMDPs are steerable in a browser with no backend and no dependencies, and the pedagogical payoff per line of code is higher. The continuous-state material is where we point readers who want the neural-implementation story Class C.
Chapter 10 on the mechanics of policy selection (softmax over negative expected free energy with a temperature) is used, but the policy space in our labs is deliberately small (depth 2, a handful of primitive actions). Deep tree search over long horizons is a place where the textbook is more ambitious than our current build, and we say so on the science page rather than dressing the labs up.
The textbook, correctly, is agnostic about applications. UNI is not: the Cell Lab framing pulls in Mikkilineni (2022, DOI 10.3390/info13010024) for viable-set maintenance in service cells, and the Heart Lab uses the same math on the cardio-renal loop Class E. Those application choices are ours, not the book's, and we cite accordingly.
What we do not claim
We do not claim the textbook authors have reviewed, endorsed, or verified this build. We do not claim the preprint (DOI 10.5281/zenodo.19785799) is peer reviewed: it is a Zenodo preprint with expert review pending Class C. We do not claim the labs prove active inference is the correct theory of anything. They demonstrate that a specific discrete-POMDP construction, faithful to Chapter 4, produces measurable behavioural regimes we can steer and falsify.
Themesis publishes a resource map of active-inference pathways in 2026 (Where to Start with Active Inference, A Resource Map for 2026), which lists SolutionWright among five entry points. Our one-line frame: it is a factual reference to publicly visible work, not an endorsement of any claim on this site.
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.