Dukkha and the Free Energy Principle

Worked out with Claude, Jul 2026. The question: under the Free Energy Principle, what corresponds to Dukkha? Prediction error can't work as the answer — an agent with zero surprisal has stopped functioning as an agent, and the Buddha kept a boundary. So the mapping has to land somewhere in how an agent relates to its prediction error. Links out to Mistranslating the Buddha for the NGD reading of the terms, and to Remapping and Navigation for the Markov-blanket material.

Caveat up front: FEP carries enough generality to accommodate almost any story brought to it. Treat what follows as a translation aid, not a derivation. Three pieces below earn their keep by making distinct predictions; the rest functions as scaffolding.

Not the error — the rate

Joffily & Coricelli (2013): valence tracks the rate of change of free energy, not its level. Feeling good = error falling faster than expected. Feeling bad = error falling slower than expected, or climbing. A second-order quantity.

This explains the observation that started the question: two people can sit at identical error levels and suffer wildly differently, because they hold different priors about how fast their error ought to resolve.

Dukkha ≈ a chronic, expectation-violating shortfall in the rate of free-energy reduction — error that persistently fails to convert into model improvement.

"A difficult emptiness" reads well as exactly this: the loop keeps running and nothing settles.

The second arrow, literally

Sallatha Sutta: first arrow = the pain; second arrow = the reaction to it. Under FEP the second arrow becomes literal — a prediction error about a prediction error. I hold a prior saying "an agent like me shouldn't sit in this much error here," and violating that prior generates its own error signal one level up.

Unlike the first arrow, the second one admits revision. This gives formal shape to the intuition that liberation concerns one's relationship to error rather than the error itself.

Tanha and Updana as precision pathology

FEP grants an agent two legitimate moves for reducing free energy: change the model (perception, learning) or change the world (action). Tanha looks like a third, illegitimate move — clamp precision on a prior so the error gets suppressed instead of resolved. The error doesn't go anywhere. It just loses the ability to propagate upward and update anything.

Updana — high-precision priors on self-related states that no available policy reaches.
Sankhara — the accumulated policy machinery built to defend those priors. NGD's "house."
Anicca — the world runs nonstationary, so any stationary prior held at high precision carries an irreducibly nonzero error floor. Suffering as a modeling error type: stationary model, nonstationary process, precision cranked high enough that the mismatch teaches nothing.

Note the prediction this generates: relief comes from lowering precision on those priors, not from pushing harder — which matches NGD's "pushing less hard on experience" nearly line for line.

The boundary survives; the model of it doesn't

The Buddha kept a Markov blanket. An agent can't run inference without one, and per Remapping and Navigation, dissolving it amounts to a double death with an heir rather than merger.

The distinction that resolves this: self-boundary ≠ self-model. Anatta doesn't dissolve the blanket. It demotes the model of the blanket — from a maximally-precise fixed prior that the whole hierarchy optimizes to preserve, down to a low-precision, revisable, task-local hypothesis. The blanket keeps working; it just stops functioning as the thing everything else must not contradict.

Non-dual states then look like temporary precision drops on the self-model. Real, but state-dependent — which explains why NGD files them as partial solutions requiring constant maintenance.

Why good updaters read as enlightened

The enlightened agent doesn't run low error. It runs high model evidence per unit error — errors metabolized as information instead of parried as threat. Error turns aversive only when it registers as evidence against a self-prior that mustn't turn out false. Loosen that prior's grip and the same signal becomes news.

Diagnostic, when something goes wrong: do I ask what does this tell me, or what does this mean about me? The second question marks a clamped prior.