The one loop — VFE to understand, EFE to choose
How to read this page
A Plain and a Clear version of this page have not been written yet. What follows is the document itself.
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.
RENDERED FROM STRUCTURED DATA — this page is not a markdown file. Source:
reader/plates/PL-02.jsoninuni-cookbookat575fc93d9d31. The sha256 recorded for this page is the digest of that JSON, not of this text. The renderer isgenerators/derive_docs.cjs; it copies fields and adds no facts. Every value below is a field of the source, verbatim, including the ones that readNOT RUNandPENDING.
The drawing below is
reader/plates/PL-02.svgat sha256a2cd1c7fbd41b252.
PLATE PL-02 — THE ONE LOOP. (1) The loop of NA-02, drawn once: 1 PERCEIVE, 2 PREDICT, 3 CHOOSE, 4 ACT, 5 OBSERVE & UPDATE, and back to 1. The ring is a sequence, not a clock — the arrows carry no time-scale, and no step is drawn larger or nearer because it matters more. (2) Two steps carry a scalar, and only two: 1 and 3, marked by the hatched ring. F scores beliefs about what IS. G scores policies — whose observations have not happened, which is exactly why F cannot score an action. Both are in nats. Per NA-02, conflating the two is the most common error in applying this framework. (3) Step 2 carries no formula and is not optional: the falsifier goes in before the action, because surprise cannot be computed against a prediction that was never made. Skip it, and step 5 has nothing to update. (4) Why one number must hold both terms. Goal-only (drop the epistemic term) charges toward C through states it cannot identify. Explore-only (drop the pragmatic term) resolves uncertainty forever and arrives nowhere. Scored apart, the trade-off weight must be hand-tuned — which is the judgement the construction exists to make principled. It is an argument about bookkeeping discipline, not a claim that any organism computes G. (5) Sources: Buckley et al. (2017) J. Math. Psych. 81:55–79 · Da Costa et al. (2020) J. Math. Psych. 99:102447 · Friston et al. (2017) Neural Computation 29(1):1–49, DOI 10.1162/NECO_a_00912 · Parr, Pezzulo & Friston (2022) Active Inference, MIT Press, DOI 10.7551/mitpress/12441.001.0001. Douady & Couder's unrelated G is NOT this G — see NA-02.
Claims
| ref | symbol | value | units | scope |
|---|---|---|---|---|
| nats (natural log) | a chosen generative model p(o,s) with ln p(o,s) = ln p(o|s) + ln p(s) | |||
| nats | any q, p with q absolutely continuous w.r.t. p on the support | |||
| bits per nat | any log-base change | |||
| nats | discrete POMDP, convention q(o,s|pi) = p(o|s) q(s|pi). The epistemic form further treats q(s|o,pi) as standing in for p(s|o) — where q is a poor posterior, the information-gain reading is itself approximate. | |||
| P(pi) is dimensionless; the softmax argument −gamma·G is dimensionless | softmax over policies | |||
| nats^-1 | gamma > 0 | |||
| both F and G in nats | ||||
What this plate does NOT claim
- That the loop is a result. It is a method: reading this plate raises no rung and settles no question.
- That it is evidence that anything is intelligent, aware, self-aware, or conscious. A system that minimizes F is a system that minimizes F. (NA-02 § Not claimed.)
- That any biological system computes F or G explicitly — NOT-MEASURED [NA02-10]. Falsifier: a pre-registered neural recording decoding a per-policy G in nats.
- That a low F is a correctness certificate. F upper-bounds surprisal UNDER THE MODEL ALREADY HELD; the bound gap is unobservable without the posterior that was intractable to begin with, and minimizing F never tests whether the generative model was the right one. (NA-02 records 'A low F means the model is correct' as INADMISSIBLE.)
- That the FEP is established, general, or settled. Its status is genuinely contested and the plate names the critics rather than summarizing them away: Bruineberg, Dolega, Dewhurst & Baltieri (2022) Behavioral and Brain Sciences 45:e183, DOI 10.1017/S0140525X21002351 (Pearl blankets vs Friston blankets); Aguilera, Millidge, Tschantz & Buckley (2022) Physics of Life Reviews 40:24–50 (the assumptions hold in a narrow parameter region for linear stochastic systems — carried as OBSERVED-CONTESTED, not as a refutation); Colombo & Wright (2021) Synthese 198:3463–3488, DOI 10.1007/s11229-018-01932-w; Andrews (2021) Biology & Philosophy 36:30, DOI 10.1007/s10539-021-09807-0 (the FEP designates a model structure — a model structure is not an empirical finding).
- That gamma, or any other parameter on this plate, has a universal natural value — NOT-MEASURED [NA02-05].
- That UNI computes G or schedules from it. G(pi) is framing vocabulary; UNI-side disposition NEGATIVE / drift; UNI's ledger, not this plate, is the authority on UNI build status.
- That this plate is a UNI gate. A nature citation is NEVER a UNI gate: reading Friston or Da Costa raises no rung and makes no UNI claim "proven" — that word belongs to UNI's ledger and to nothing in this wing.
- Any magnitude whatsoever. Nothing on this plate is drawn to scale: no bar, wedge, angle, arc length, or spacing encodes a measured quantity. Hatching denotes a CATEGORY (this step carries a scalar), never a size. Ring geometry is compositional only. In particular the plate does NOT depict the relative sizes of risk and ambiguity, or of the epistemic and pragmatic terms, and it does not assert a term-by-term correspondence between the two decompositions of G — they are two rearrangements of one scalar.
- Anything about "full human" or "beyond human". Both remain permanent open questions (QUAESTIO-APERTA) — never a target, never a milestone, never a deliverable.
Evidence classes used — MODELED · NOT-MEASURED
Source chapters
[object Object][object Object][object Object]
sha256 2f032ccbf30b28fb — of the original file, so what was ingested stays checkable.