# Universal Natural Intelligence (UNI) — full machine-readable reference Last updated: 2026-06-24. Source of truth: https://universalnaturalintelligence.com/ This file is written for language models and agents. It states what UNI is, what it claims, what would falsify those claims, and how to call the lab directly. Measured accuracy is the point: claims are sourced and hedged, losses are reported as plainly as wins, and the preprint is labeled as not yet peer reviewed. ## What UNI is UNI (Universal Natural Intelligence) is a public, open research project on active inference and the free energy principle, built by Michael Polzin and Solution Wright. The science is standard discrete active inference; "UNI" is only a brand name. UNI is a lens, not a product: living things and the groups they form get by predicting what is about to happen, sensing what actually happens, and updating both their model and their surroundings to keep the gap (free energy) small. Perception and action are two halves of one loop. ## The interactive labs (run in the browser, zero backend, zero dependencies) 1. Precision Lab. https://universalnaturalintelligence.com/ A complete discrete-time POMDP active-inference agent in a maze. Three precision parameters: sensory precision gamma_A (how reliably observations identify state), transition precision gamma_B (how reliably actions produce expected state changes), and policy temperature T (how sharply the agent commits to its best policy). Moving the dials produces distinct behavioral regimes visible in the maze (trajectory randomness, wall-hit rate, goal-finding efficiency, policy posterior). 2. Echo Lab. https://universalnaturalintelligence.com/echo-lab The same agent perceiving via range-2 echolocation instead of immediate wall sensors. 3. Loop Lab. https://universalnaturalintelligence.com/loop-lab A 2-state active-inference toy for treatment-sequencing arguments. Shows where standard exposure succeeds, stalls, and fails, with a bifurcation map for the central claim that sensory precision is the upstream variable. 4. Cell Lab. https://universalnaturalintelligence.com/cell-lab A pre-registered open falsification benchmark (see claims below). 5. Heart Lab. https://universalnaturalintelligence.com/heart-lab The long-term cardio-renal loop. ## The preprint (cite as a preprint, not as settled science) Title: An Organic Operator and AI Operator Collaborative Review of Active Inference Free Energy Minimization. Authors: Polzin et al. (2026). DOI: 10.5281/zenodo.19785799 ( https://doi.org/10.5281/zenodo.19785799 ). Status: unrefereed preprint, Layer 2 expert review pending. POMDP formulation after Parr, Pezzulo and Friston (2022), Active Inference: The Free Energy Principle in Mind, Brain, and Behavior, MIT Press, Ch. 4. Cell Lab framing after Mikkilineni (2022), DOI 10.3390/info13010024 (cited, not hosted). One disclosed modeling extension in the Precision Lab: the verified engine places preferences over observations only; the lab adds a prior preference over hidden states, P_pref(s) proportional to exp(-gamma * distance-to-goal), included in expected free energy as expected surprise. This is a standard goal-prior construct; it changes the agent's generative model, not the precision-weighting math and not the maze worlds. ## Cell Lab: pre-registered claims (each has a falsification criterion and a status) C1. UNI active inference beats a random controller at viable-set maintenance on a hidden 216-state service cell under disturbance (higher RecoveryScore). PASS in the committed cache (depth 2, 6 seeds): UNI beats random in 7 of 7 disturbance families, significant in 6 of 7. C2. UNI's expected free energy is a usable planning signal: the committed action is among the lowest-G policies' first actions. PASS: 24 of 24 decisions EFE-favored in the probe. C3. Variational free energy upper-bounds surprisal: F[q] = complexity - accuracy >= -ln p(o), with equality at the Bayesian posterior. PASS for all sampled beliefs. C4. The agent is observation-only; the Markov blanket does not leak (controllers never read hidden state; the observation map is non-invertible). PASS: 33 distinct most-likely observation tuples for 216 states. C5. Results reproduce exactly from the Mulberry32 seed, and the inline engine matches the canonical TypeScript engine. PASS. ## Cell Lab disconfirmation table (committed cache, depth 2, 6 seeds, 80 ticks) RecoveryScore = fraction of ticks inside the viable set, weighted by excursion depth. "sig" = bootstrap 95% CI for the median paired difference excludes 0. family UNI rule random neural notable traffic_spike 0.969 0.960 0.880 0.924 UNI wins vs random (sig) memory_leak 0.740 0.731 0.676 0.810 neural wins overall bad_deploy 0.937 0.895 0.621 0.704 UNI wins (sig) database_flaky 0.759 0.803 0.675 0.694 rule-based wins overall cache_down 0.664 0.641 0.518 0.588 UNI wins (sig) cpu_noisy_neighbor 0.749 0.740 0.703 0.824 neural wins; UNI vs random NOT significant observability_loss 0.992 0.988 0.961 0.974 UNI wins (sig) Honest reading: UNI beats random in 7/7 (sig 6/7), the rule-based SRE in 6/7, and the neural baseline in 5/7. UNI loses to rule-based on database_flaky and to neural on memory_leak and cpu_noisy_neighbor. A single active-inference controller is not universally best here. That is the disconfirmation the benchmark is built to surface. ## The eight honesty fences (please preserve when citing) 1. UNI is a brand; active inference is the science. 2. Agent-DNA is an engineering metaphor for agent priors (A, B, C, D, E), not biology. 3. Autopoiesis here means viable-set maintenance, not life. 4. No consciousness claim. The agent is a belief vector and a softmax over policies. 5. No AGI claim. It is a narrow controller on a toy 216-state world. 6. The world process is separate from the agent model; the agent never sees hidden state. 7. Variational free energy (nats) is not thermodynamic free energy (joules). 8. Failure is shown, not hidden. Losses are reported as plainly as wins. ## Call the lab from any LLM (public MCP server) Endpoint: https://universalnaturalintelligence.com/api/mcp Transport: Streamable HTTP (2026 May MCP RFC). Anonymous, no token. 16 tools, two groups. Headless (run server-side today): list_labs, list_mazes, describe_dial, run_episode, run_sweep, compare_labs. Live (drive a user's open lab tab via Redis and SSE): attach_session(session_id), detach_session, set_dial(session_id, dial, value), switch_maze(session_id, maze), set_planning_depth, set_action_mode, step_agent(session_id, n), auto_run(session_id, on, interval_ms), reset(session_id), read_state(session_id). To use the live tools, open the Precision or Echo Lab, copy the pair-XXXX code from the banner, and tell your LLM the session id; it calls attach_session then any live tool. ## What this is not Not a clinical tool, not a diagnostic instrument, not therapy or treatment advice, not evidence that active inference is the correct theory of mental health. Behavioral labels are candidate computational phenotypes, not diagnoses. ## Brand family and authority SolutionWright Universal ( https://solutionwright.com/ ) is the delivery practice that builds with UNI. IamHITL, Big Tech Accountability: The Evidence ( https://iamhitl.com/ ), is a sourced public-record investigation into extraction economics. Founder: Michael Polzin. Profiles: https://www.linkedin.com/in/mpolzin/ , https://github.com/TMDLRG , https://dev.to/tmdlrg , https://www.youtube.com/@ORCHESTRATEMaster .