L4 — Interoceptive / Autonomic + Affect-as-Precision
How to read this page
Three ways to read this page. Precise is the document itself, exactly as it is written in the repository. Plain and Clear were written for this website to help you meet that document — they are about it. They are not it, and they are not evidence.
The Cookbook is the method carried out step by step: 34 pages of recipes for building a developmental active-inference SIMULATION — a bounded peek at a toy world, never a person. The front matter says that word is never softened under any pressure, so it is not softened here. The recipes run from the molecular and cellular rungs up through metabolism, motor control, perception, language and metacognition, and on to rungs that are still open questions. Around them sit a set of kitchen rules, a shared pantry of engines and primitives, and a second family of recipes about nature itself — rocks, water, air, stars, DNA, ants, whales, bats, humans.
It is for the reader asking what building this would actually take. Each recipe names its ingredients, the order of work, the tests to run at that stage, and the point at which a step stops being something already carried out and becomes something proposed.
Begin with the front matter and then the kitchen rules. Those two pages fix the honest position and the fence labels that every later recipe leans on, and without them the status markers on a recipe are easy to skim past. After that the recipes can be read in any order.
The nature recipes sit slightly apart and should be read that way. They cite outside science — geology, chemistry, biology, astrophysics — and a nature citation is never a UNI gate: those chapters contain zero UNI claims and raise no rung.
What it is not: a claim that the whole ladder has been cooked. The book recommends the complete recipe and, on the same page, labels every rung by its real state — that tension is deliberate and is the thing the book is built around. Where a recipe and the claim ledger disagree, the ledger wins.
Your browser cannot switch reading levels, so the document itself is shown.
Precise — the source document
This is the document. Rendered from the repository at the commit above, with nothing rewritten for the web. A gate re-renders it on every deploy and fails the build if a single byte differs.
What you are building. A global affect modulator
Zthat turns interoceptive / autonomic signals into a precision dial: raising arousal / threat sharpens perception (precision-weighting) and flips the pragmatic↔epistemic balance of planning — affect MODELED, never felt. This is one rung of a developmental active-inference SIMULATION, a bounded peek, never a person.
Ingredients
Drawn from the shared pantry, by name:
- The Z affect modulator (uni-gpt / uni-mind) — the global vector
[energy, arousal, valence, fatigue, pain, threat, safety, inflammation]that sets precision / preferences / habits / learning-rate / planning-horizon. Affect modeled, never felt. - The JAX POMDP + EFE + Dirichlet engine (
core.py, uni-mind) — the discreteperceive → EFE-plan → act → learnloop and its precision dial (gamma_asensory,gamma_btransition, policy softmax temperature). float32 host — anchors hold to ~6e-8, NOT the f64 tier (M10). - A grounded reader to test the pragmatic↔epistemic flip under live observations.
- The contains-baseline + load-bearing-discriminator discipline (M7) — supplies the required Z-ablation that must collapse the effect.
This is the WORLD ⊥ BODY ⊥ MIND framing (M12): interoception = hardware self-signals entering the mind through the typed sensory blanket; affect is a latent modulator over that mind's own precision, never a phenomenal state of the substrate.
Method
Wire Z as a precision modulator, not a reward. Couple the eight-channel
Zvector to the engine's control surface: it sets sensory precision (gamma_a), transition precision (gamma_b), policy temperature, preference weighting (the C tensor), habit strength (E), learning-rate, and planning-horizon. Z is a dial over how sharply the existing generative model is read, not a new objective term — keep the no-backprop AST-guard live (M13); learning stayscounts + lr * sufficient_stat.Show the precision-weighting effect. Raise arousal / threat and confirm perception sharpens: higher
gamma_aconcentrates the observation likelihood, so the posterior tightens around the sensed state. This is the textbook claimF[q] ≥ −ln p(o|m)read through a precision lens — affect tunes how confidently the agent reads its world (M12, textbook level only; cite Parr/Pezzulo/ Friston, Active Inference, MIT Press 2022).Show the pragmatic↔epistemic flip. Under the grounded reader, demonstrate that shifting Z re-weights expected free energy between its pragmatic (goal-seeking) and epistemic (information-seeking) components, so the selected policy flips. Raising threat/arousal moves the balance; the planner's chosen action class changes as a function of affect, on the same engine.
Register the Z-ablation discriminator (M7). Pre-register that removing Z collapses the effect: with Z ablated, precision-weighting does not change and the pragmatic↔epistemic flip does not appear under the grounded reader. The gain is credited ONLY if it appears with Z live and disappears under ablation, and survives a control.
Gate
With Z live, precision-weighting and the pragmatic↔epistemic flip both appear; the Z-ablation collapses them; and the effect survives a control.
Exact ledger figures (L4.1, Class C, status proven). Affect-as-precision: emotion modulates
precision so perception sharpens and the pragmatic↔epistemic balance flips; the global Z modulator
[energy, arousal, valence, fatigue, pain, threat, safety, inflammation] sets precision / preferences
/ habits / learning-rate / horizon. This is a dev-gate / held-out (Class C) result, not a
machine-exact anchor. The verdict is the gate, never a point estimate (M2). Do NOT card this above
a Class-C functional result.
Falsifier
An ablation of the Z modulator shows no change in precision-weighting / no pragmatic↔epistemic flip under the grounded reader, OR the effect collapses under control. (Verbatim from ledger row L4.1.) If removing Z leaves perception equally sharp and planning unchanged, the affect-as-precision claim is dead.
Recorded NEGATIVE / sub-bound (first-class, inline)
M10 honest boundary (recorded sub-bound). Neuroticism does NOT change behavior under bimodal surprise without a graded task. Under the grounded reader's bimodal-surprise probe, varying the neuroticism-style affect parameter produced no behavioral change — exposing it would need a graded task, not a bimodal one. This is a real boundary on the affect machinery, carried inline beside the PASS: the Z dial demonstrably modulates precision and the pragmatic↔epistemic balance, but it does not license a claim that every affect-personality parameter drives behavior on every task. The recorded result delimits the positive rather than inflating it.
HONEST FENCE — proven (functional, Class C)
A held, gate-level functional PASS exists, with its Z-ablation falsifier still live. NOT claimed:
- Affect is MODELED, never felt. Phenomenal feeling / sentience is explicitly DISCLAIMED — no falsifier is offered for it because it is disclaimed, not tested. Z is a latent modulator over the mind's own precision, never a quale, never a felt emotion of the substrate.
- This is NOT consciousness, sentience, awareness, a mind, a feeling, AGI, human-level, or "active inference demonstrated." (Functional self-awareness may be described only at L8; phenomenal sentience stays disclaimed program-wide.) "Active inference" is the framing lens here, at textbook level only — no AIF loop is claimed in the Rust crate.
- The gate is Class C (dev-gate / held-out), NOT a machine-exact (Class A) anchor and NOT the f64
<1e-10exact tier — the JAX host is float32 (M10). - No SIGNED consult design is folded into this rung (the 2026-06-27 consults attach at L2, L5, L9, L11, L12); nothing raises L4's status. The rung stands on row L4.1 alone.
Honest program position: ~2 of 11+ developmental rungs earned. The whole program remains a developmental active-inference SIMULATION — a toy world, a bounded peek, never a person. Where this recipe and the ledger disagree, the ledger wins.
sha256 423455401097fcda — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
This recipe adds something like mood to the engine, and is careful about what that means. A small set of body signals is turned into a dial that controls how sharply the system reads its world and how it balances chasing a goal against gathering information. The page says it in italics and repeats it at the end: affect is modelled, never felt — a simulation, a toy world, never a person.
The dial is a modulator, not a reward. It does not add a new goal to the system. It changes how confidently the existing model is read and how the planner weighs its options, and nothing about the learning rule changes.
The test is built around removal. The claim counts only if the effect appears when the dial is live and disappears when it is taken away. That is registered in advance as what would sink it: if taking the dial out leaves perception equally sharp and planning unchanged, the claim is dead.
A boundary is recorded beside the result. One affect parameter produced no behavioural change on a particular kind of probe, and the page carries that inline to delimit the positive rather than inflate it. The closing lines are mostly a list of what is not claimed: no awareness, no mind, no feeling.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 423455401097fcda
Clear — written for this website, not the source document
This chapter builds a global affect modulator and treats it strictly as a dial over precision. A small vector of interoceptive and autonomic channels — things like energy, arousal, fatigue, pain, threat and safety — is coupled to the engine's control surface. There it sets how sharply observations are weighted, how sharply transitions are weighted, and how sharp the policy choice is. It also sets how strongly preferences pull, how strong habits are, how fast learning happens, and how far ahead the planner looks. The framing is stated at the top and never loosened: this is one rung of a developmental simulation, and affect is modelled, never felt.
The method has four steps, and the first is the conceptual one. The dial is wired as a modulator rather than as a reward. It changes how sharply the existing model is read, not what the system is trying to achieve. The learning rule stays the same count-based update, with the no-gradient guard live. The second step shows the sharpening effect — raise arousal or threat and the observation likelihood concentrates, so the posterior tightens around what is sensed. The third shows a flip: shifting the dial re-weights the planner's balance between goal-seeking and information-seeking, so the chosen kind of action changes as a function of affect on the same engine. The fourth step registers the discriminator in advance — remove the dial, and both effects must vanish.
The gate is that pairing, and the chapter is careful about its class. This is a development-gate result rather than a machine-exact anchor, and the page says explicitly not to label it higher. What would sink it is quoted from the row in the ledger itself, the append-only record of claims. If removing the modulator leaves precision-weighting unchanged and no flip appears under the grounded reader, or the effect collapses under a control, the claim is dead.
The recorded boundary is small but the chapter treats it as first-class. Varying one affect-personality parameter produced no behavioural change under a probe that offered only two sharply separated levels of surprise; exposing such an effect would need a graded task instead. The chapter draws the right conclusion from it: the dial demonstrably modulates precision and the planner's balance, but that does not license a claim that every affect parameter drives behaviour on every task. The negative delimits the positive.
The closing lines spend most of their space on what is not claimed. Feeling itself is disclaimed rather than tested, and the page notes that no test that could show it wrong is offered for it precisely because it is disclaimed. The modulator is a latent quantity inside a model, never an experience of the machine. Nothing here is awareness, a mind, a feeling, human-level ability, or a demonstration of the underlying theory. No signed design attaches to this rung, so it stands on its own single ledger row and nothing lifts it.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 423455401097fcda