Precision is the quiet lever in active inference. Turn it up on a sensory channel and the agent starts trusting what it sees; turn it down and the same signal fades into background noise. Attention and confidence are not separate faculties in this frame, they are what precision does when you watch it work.
Precision as inverse variance
In a Gaussian generative model, precision is the reciprocal of variance, π = 1 / σ2. A high-precision distribution is narrow: the model is committing to a tight range of outcomes. A low-precision distribution is broad: the model is hedging. Active inference lifts this scalar into a modulator that sits on every likelihood term in the generative model, on the transition dynamics, and on the prior over policies Class E (Parr, Pezzulo, and Friston, 2022).
The consequence is direct. Free-energy minimization weights each prediction error by the precision of the channel that carried it. A precise sensory channel pulls the posterior hard; an imprecise one leaves it almost where the prior placed it. Weighting error by precision is not a metaphor for attention, it is one common formalization of attention within this framework Class E.
Three precisions, three regimes
Our Precision Lab exposes three dials, because those are the three places precision enters the discrete-time POMDP we teach with Class C:
- Sensory precision: how much the agent trusts its observations. Raise it and the agent commits to what its sensors report. Lower it and the agent relies more on its transition model, that is, on what it expected to see.
- Transition precision: how much the agent trusts its own model of how the world evolves. High values make the agent confident in its rollouts; low values make planning tentative and short-horizon.
- Policy temperature (the inverse of the precision on the prior over policies): high policy precision looks decisive, low policy precision looks exploratory or hesitant, because the softmax over expected free energy flattens or sharpens accordingly.
These three dials produce visibly distinct behavioral regimes in the same maze, on the same generative model, with no change to goals or geometry Class C. That is the empirical bite of the precision claim: identical structure, different behavior, driven by one scalar per channel.
Where the modeling choice actually bites
Precision is often introduced as a knob, and the demonstrations make it look tidy. It is not tidy in three specific places:
- Fixed versus inferred precision. If precision is hard-coded, the agent has a personality and cannot adapt to changing signal quality. If precision is itself inferred (a hyperprior with its own update rule), the model becomes more expressive and more brittle: precision estimates can collapse or explode when the data thins.
- Where precision attaches. Sensory precision on a wall-sensor channel and sensory precision on an echolocation channel are the same math with different consequences. The Echo Lab exists to make that visible: the range-2 sensor gives the agent evidence about state that the immediate wall sensor does not, and precision on that channel controls how far the agent looks before committing Class C.
- Confidence versus certainty. High posterior precision means the model is confident, that is, its variance is small. It does not mean the model is right. Confidence in the model is a statement about the posterior, calibration against ground truth is a separate question and requires a separate check.
Attention as precision, held honestly
Calling precision-weighting "attention" is useful pedagogy and it is also a claim with published mapping to neurobiology in the active-inference literature Class E. We use the framing here because it makes the math steerable in a browser. We do not extend the claim past that: our labs are POMDP toys, not models of any specific neural circuit. When we say the sensory-precision dial is a proxy for attention, we mean exactly this: raising it makes the sensor drive inference more strongly, in the same functional sense the free-energy literature uses.
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.
Honesty fence. Precision-as-attention is a framing from the active-inference tradition, cited to Parr, Pezzulo, and Friston (2022). Our contribution is making the three-dial parameterization steerable in a browser and consistent across five labs. We make no claim about consciousness, and no claim that this is artificial intelligence: UNI targets General Natural Intelligence, natural not artificial, as a testable hypothesis, not a finished result.