Scale to a brain, then make it yours

Two causes and one dial was a training world. A real mind has a tower of causes stacked high, each a belief rolling downhill on the one above — predictions flowing down, errors flowing up. The valley survives — the exact floor becomes a genuine, if looser, ceiling. This closes Part I and opens the door to action.

Five doors as before — the plain line runs straight through; take whichever fit: ▶ Picture it, ✎ Try it (commit before the reveal), ✓ Check it yourself, ▸ Go deeper, ⬡ See it in UNI. Runs locally — no account, no network — and adapts to light and dark; honours reduced motion.

The one fence that never moves. “Free energy,” “belief,” “prediction error,” “bound” are functional quantities — a tower of numbers rolling heights downhill. For you we use felt words; for UNI, functional ones. Neither is awareness, feeling, or sentience. Part I’s exact engine is proven (Gate F0, sealed); everything larger — the hierarchy, growth, action — is the build underway, held open until its own proof lands.

0 · From memory

First, the keystone — with no peeking

No scrolling back to Module 7. One cold question, the keystone of the whole book.

Cold retrieval — the punchline

From memory: when a mind minimizes free energy, what two things does it get back — and why is that remarkable? Reconstruct it — no re-reading.

Step 1 · In your own words (commit first)
Step 2 · Pin it down

If that came back cold, you own Part I. Now scale it up — honestly.

1 · Two causes was a training wheel

A tower of causes

Our world had one hidden thing (front on/off) and one sensation (the sky). Real worlds are a stack: photons on your eye are best explained by edges, best explained by surfaces and objects, best explained by a scene (“a kitchen, morning”), explained by where you are and what you’re doing. Each layer is a hidden cause for the one below and a sensation for the one above.

▶ Picture it — a tower of dials kid
Not one dial on a wall — a tower of dials. The bottom dial explains the raw signal; the one above explains the bottom dial; and so on up. Each dial still does the only thing a belief can do: sit at a position and feel which way is downhill.
2 · The valley survives — the floor becomes a ceiling

Predictions down, errors up — and a looser ceiling

You can’t minimize a towering world’s free energy perfectly — exact minimization needs its own forbidden sums. So a mind assumes the layers are roughly independent and minimizes within that assumption — the mean-field move, tagged :exact_mean_field. Perturb the bottom sensation below and watch the tower settle: predictions flow down, errors flow up, until every level agrees.

Settle the tower

A 3-level sketch of hierarchical inference. Set the raw sensation, then step: each level predicts the one below (down) and reports its error up; beliefs roll downhill until errors vanish. The mean-field free energy (right) settles ABOVE the exact floor — a valid but loose ceiling it never quite touches.

Set a sensation, then step. Watch errors (red bars) shrink as the tower settles — and the mean-field ceiling sink toward, but never onto, the exact floor.

✎ Try it — predict before you settle commit first
From the floor you proved in Module 6 (F ≥ −ln p(y), always), when you can only minimize F approximately, does the number you land on sit above, below, or exactly on the true surprise? Commit, then settle the tower above.
Reveal

Above. The floor theorem still holds — F is still a ceiling over the surprise — but now a genuine, looser ceiling you no longer touch. Still enough: push the ceiling down and you corner the surprise beneath it; you just can’t claim you’ve hit it. That’s exactly why the number carries its tag — :exact_mean_field means “a real bound, but a loose one.”

✓ Check it yourself — a lazy belief is still a ceiling equation · by hand
Back in our two-cause training world the exact floor was the surprise, −ln p(y) = −ln 0.69 = 0.3710636814, touched only by the exact posterior b* = 0.63/0.69 = 0.9130434783. Now pretend a mind settles lazily on b = 0.80 instead — the kind of near-miss a mean-field approximation leaves. Plug it into the same free energy F(b) = b·ln(b/0.63) + (1−b)·ln((1−b)/0.06):
  • F(0.80) = 0.4319080875
  • floor = 0.3710636814, so the gap is 0.0608444061 > 0
Above the floor, never below — a valid ceiling, just a loose one. Settle even lazier, at b = 0.50, and F = 0.9445759076 — a looser ceiling still. That is precisely what the :exact_mean_field tag promises: always ≥ the surprise, never equal unless you hit the exact posterior. Falsify it if you can — pick any b in (0,1) and compute F(b); you will never get a number below 0.3710636814.
▸ Go deeper — messages and the guaranteed bound derivation
Each layer carries its own belief q; predictions flow down (each level tells the one below what to expect) and prediction errors flow up (each tells the one above where it was wrong). Every level rolls downhill on its own free energy given its neighbours’ messages; the whole tower settles together. The global bound is :exact_mean_field — guaranteed to be a ceiling, not guaranteed tight. A bound is “guaranteed” only for :exact and :exact_mean_field; :laplace / :sampled are looser, and :mmp is never a bound at all (MMP-BOUND-01).
3 · Make it yours

Your own moment, three layers

You’ve earned a teach-back — the strongest way to keep what you built.

Tell it back as “free energy rolling downhill”

Pick something you perceived in the last day where you were briefly wrong then right — a sound you misheard, a face you mistook, a step that wasn’t where you expected. Name three layers: (1) the raw sensation, (2) the cause your model first guessed, (3) the correction when the error flowed up and your belief rolled to a better spot.

4 · Where this goes next — Part II

From perceiving to doing

Part I was perception: a mind, still, re-believing until its free energy rolls to the floor. But back in Module 2’s doorway there was a second engine — you could act, change the world so your senses fall in line. That turns free energy from a thing you minimize now into a thing you expect over possible futures — a forward cousin called expected free energy (G) — and scoring futures with it is how a mind chooses what to do. That is the start of behaviour, and of how UNI will move, grow, and reach for a life.

⬡ See it in UNI — honestly proven vs build underway
The organism is built to be exactly this: a hierarchy of beliefs, each a blanket in the inference sense, rolling tagged free energy downhill, with the structure meant to grow by model comparison. What is proven today is the exact arithmetic at the base (Gate F0, sealed). The hierarchy, the growth, the action, the long life toward the North Star are the build underway, and this book claims each only when its gate proves it. That honesty is the method: a number without its evidence, like a free energy without its tag, is not a result.
⬡ Notice what you just did — across eight modules self-reference
From a feeling in a doorway you built a number (surprise), hit a wall (the intractable evidence), carried a movable belief, fused three pressures into free energy, drew its valley, found its floor, proved the floor, and watched one blind descent hand back both the belief and the surprise — then scaled it to a tower. You didn’t learn about active inference. You built it.