UNSYS MODEL
The unsys model is a generative model of a concept the user is trying to learn or verify: hidden states, state factors, observations, priced edges, and policies. Our system generates a learning simulation of that model and uses it to calibrate sessions, verify knowledge, and steer learning.
Karl Friston’s Free Energy Principle says living systems keep a generative model of the world, then act to reduce uncertainty about its hidden states. A mismatch between the observation a hidden state would generate and what the learner produces is a surprise. We use those Concepts for Calibrate, Learn, and Verify. We do not compute variational free energy, and we do not claim the full formal theory.
Before a session starts we freeze a generative model of the concept. Related hidden states share a state factor. Each hidden state carries an observation: what we would expect while focusing on it. An action is travel along a priced edge from the current hidden state to another, with a cost: its expected-free-energy price. A policy is a sequence of those hidden states, or the cheapest outgoing edge. Taps, speech, and canvas work are projected onto a policy. Learn shows those costs so the policy can steer. Verify hides them and scores distance from the cheapest policy.
Integrals is a frozen generative model of the concept the user is trying to learn or verify. Three state factors structure it: Meaning, Technique, and Setup. A hidden state can open several finer hidden states at the same depth. Each hidden state carries an observation: what we would expect while focusing on it. Travel from a hidden state is a priced edge: the expected-free-energy cost of going to the next hidden state. Costs change with the hidden state you are on. Switch the explorer between Learn and Verify. In Learn the costs stay visible and the cheapest edge is marked as the policy, so the policy can steer. In Verify the costs stay hidden; each step is scored as distance from that cheapest policy. Click or drag a hidden state, then take an edge.
G* · Integrals
Drag a hidden state. It returns.
Edge costs stay visible. The cheapest outgoing hidden state is the policy.
On IntegralsWe are focusing on this if the observation is a quantity accumulated from a rate. The user is trying to learn or verify what an integral is.Cheapest policy · Accumulation 0.3
You are on Integrals, the frozen generative model. Click or drag a hidden state, then take a priced edge.
G*
The generative model represents a concept's hidden states, our observations of those states, and the priced edges that travel between them. Our system generates a learning simulation of this model that is then used to calibrate the learner's sessions, verify someone's knowledge, and serve as a learning compass to help someone learn more optimally.
s
A hidden state represents one hypothesis about the concept that the user is trying to learn or verify. The system infers which hidden state the user is focusing on while calibrating, learning, or verifying.
f
A state factor represents one dimension of the concept, grouping the hidden states that answer the same kind of question. The system uses state factors to structure the generative model so calibration can find where understanding is thin, and so learning and verification stay oriented to that dimension.
o
An observation represents what we would expect while the user is focusing on a hidden state. The system keeps that observation with the hidden state and uses it, in calibration, learning, and verification, to project what the learner produces onto the generative model.
u
An action represents travel from the current hidden state to another along a priced edge, and how that travel changes what we can learn about the concept. Each edge carries a cost: its expected-free-energy price. The system shows those costs while learning so the cheapest policy can steer, and records the same costs as distance while verifying.
π
A policy represents a sequence of hidden states through the generative model, or the cheapest outgoing edge from the current hidden state. The system uses that cheapest policy as the learning compass. Calibration writes the prior the policy will use; learning shows the cheapest edge; verification scores distance from it.
The same frozen generative model. Calibrate builds a prior over hidden states. Learn follows a policy: edge costs stay visible, and the cheapest outgoing hidden state is the compass. Verify hides those costs and measures distance from the cheapest policy.
Calibrate
The user taps hidden states they would not trust themselves on, for the concept they are trying to learn or verify. Taps that stay on one state factor are a hole. Speech about how they would look next is projected onto nearby hidden states of the same state factor. The result is a prior over hidden states: the starting point Learn and Verify later price against.
Learn
The user talks, writes, and draws. Live inference projects that work onto hidden-state travel on priced edges. The cheapest outgoing edge is the preferred next policy, and those costs stay visible so the policy can steer learning.
Verify
The same projection, on a clock, for someone whose knowledge we are verifying. Edge costs stay hidden. Each projected step is compared with the cheapest priced edge on that hidden state. Distance is the average of how much more those steps cost. Distance 0 means every step followed the cheapest policy. The score is stored with the identity they claimed.
Projection
Taps, canvas work, or one spoken plan are projected onto a nearby hidden state on the frozen generative model. That travel is an action: a priced edge. The check is whether that projection matches the cheapest policy. Expected free energy is the influence on those costs. Learn shows the costs. Verify hides them and records distance from the cheapest policy. The system records which policy the user followed while learning or verifying.
The Free Energy Principle is an influence on the unsys model. We do not claim that Uncertain Systems is Friston’s full formal theory, and we do not claim completed FEP empirical results.
