L5 — Sensorimotor / motor hierarchy (embodied action)

The Cookbook · cookbook/recipes/L5-sensorimotor.md @ 575fc93d9d31 (main) — opens the published snapshot e850f872196d

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.

PlainClearPrecise

Your browser cannot switch reading levels, so the document itself is shown.

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. The "mind-body as one" rung: a no-backprop active-inference agent that moves — a proprioceptive motor hierarchy where the body senses its own configuration, a continuous servo fulfils goals projected down the hierarchy, and reafference learns the transition model. One proven live craft chain, one proven synthetic body-world coupling PASS, and its symmetric NEGATIVE, all carried at their real ledger class. A developmental SIMULATION of embodied action, never a person.

This recipe sits on the same single engine as every other rung (core.py, the discrete POMDP loop, exact conjugate-Dirichlet learning, AST-guarded no-backprop). Nothing here is "active inference demonstrated"; "active inference" is the framing LENS only (textbook level: Parr/Pezzulo/Friston, Active Inference, MIT Press 2022). The composed appliance is variationally-controlled (audited), module-exact only at the single-step categorical body→mind interface, never globally exact.

Ingredients (which pantry engines/primitives this calls by name)

The motor hierarchy (from strings): the proprioceptive diagonal-A prior, the continuous signed-error servo (Class-B predictive coding), reafference learning the transition model. Shipped as the additive, genome-gated :motor_cortex organ — default colony stays byte-identical (mad < 1e-12 golden fixture).
The live RCON craft-chain bridge (from strings): a kin-9 lineage on the real PaperMC server, server-authoritative via RCON; the motor-ablation discriminator that collapses harvest ~700×.
The Embodiment A3 synthetic protocol (from uni-mind): the held-once body↔world coupling bound — a PASS design (#1) and a symmetric NEGATIVE design (#2), each measured as held Δ (agent − policy-shuffle).
The 277 offline-test suite (from strings): the motor-hierarchy regression floor.
The JAX POMDP + EFE + Dirichlet engine (core.py): the loop everything reuses verbatim; float32 host (anchors hold to ~6e-8, NOT the f64 tier).

Method (numbered build steps)

Break the non-identifiable factor. A uniform-A single-modality proprioceptive factor is non-identifiable — the posterior sits stuck at 0.0. Install the proprioceptive diagonal-A prior to break the symmetry; the posterior shifts 0.0 → 0.75 (the body can now infer its own configuration). This is the Class-A anchor of the rung.
Model the continuous servo + reafference on the same loop. A continuous signed-error servo (Class-B predictive coding) fulfils goals projected DOWN the hierarchy; reafference (the sensory consequence of the agent's own action) trains the transition model. No new optimizer, no backprop — the only learning rule remains counts + lr * sufficient_stat over A/B/D/E Dirichlet tensors, AST-guard live.
Run the live mechanism gate (server-authoritative). A kin-9 lineage bootstraps the craft chain wood → planks → sticks → wooden_pickaxe + sword, confirmed server-authoritative via RCON (the brain never reads coordinates; it perceives only the symbolic/visual senses across a fixed Markov blanket). Register the motor-ablation discriminator: with the motor organ ablated, harvest must collapse ~700×. This is the live RCON mechanism gate (live Class C), separate from the synthetic A3 protocol of step 4.
Separately run the Embodiment A3 synthetic protocol (held-once). Pre-register the bar, seal the held set, touch it ONCE, read the verdict off the CI bound that excludes the threshold (M2). Score held Δ (agent − policy-shuffle) over ≥5 disjoint seeds. Run both the PASS design (#1, fast immediate-reward axis) and the NEGATIVE design (#2, slow Z-bottleneck), each a synthetic-process design (NOT live-appliance, NOT recorded-hardware). Mind the length-confound: a Σ_t log p score penalises survival, so report survival + per-step (length-normalised) loglik separately and hard-fail premature death, or a fast-dying control looks like a false NEGATIVE.
Keep the offline floor green. The 277 offline motor tests stay GREEN; the genome-gated organ keeps the default colony byte-identical.

Gate (exact ledger figures)

L5.1 — Embodiment A3 Design #1 (fast immediate-reward axis) HELD PASS. Held Δ (agent − policy-shuffle) = +0.092, CI [+0.038, +0.157], UNI-signed. Class C. (Falsifier below.)
L5.3 — motor hierarchy live + mechanism PASS. Diagonal-A posterior shifts 0.0 → 0.75; the kin-9 craft chain reproduces server-authoritative via RCON; motor-ablation collapses harvest ~700×; 277 offline tests green. Class A (posterior shift) / C (live RCON gate) / E (277 tests).

A clean L5 is: live mechanism gate PASS (craft chain reproduces server-authoritative; ablation collapses harvest) AND A3 Design #1 held Δ CI excludes 0 AND the diagonal-A posterior is off 0.0 AND 277 offline tests green.

Falsifier (operable)

Any one of these falsifies the rung:

Harvest does not collapse under motor ablation; OR
the kin-9 craft chain does not reproduce server-authoritative (RCON disagrees); OR
the diagonal-A posterior stays stuck at 0.0; OR
a re-registered held-once synthetic A3 run with ≥5 seeds drops the CI lower bound to ≤ 0 (overturns Design #1).

Recorded NEGATIVE(s) (first-class, inline)

L5.2 — Embodiment A3 Design #2 (slow Z-bottleneck, compressed 2-modality EMA, delayed-reward) HELD NEGATIVE. Held Δ = −0.091, CI [−0.134, −0.055]. Class C. This is a synthetic protocol, and it is the only sealed motor NEGATIVE today: K-negative = 1, so no Section 0.6(B) bound is owed (a bound requires K≥3 structurally-distinct held NEGATIVEs, each changing ≥2 of {coupling topology, timescale source, information bottleneck, control path}; M5). Its own falsifier: a structurally-distinct corrected slow-Z-bottleneck re-run that pushes the CI above 0 overturns it.

The shape of the result is the load-bearing finding: under the registered synthetic protocol the active channel mattered — it was load-bearing on the fast immediate-reward axis (#1, vs a same-perception / same-learning policy-shuffle control) but did not matter on the slow delayed-reward compressed-interoceptive-bottleneck axis (#2). This is the framing LENS only, NOT active inference demonstrated and NOT a capability win. The NEGATIVE delimits the PASS; it is not a failure to hide. The live behavioural / ablation tallies for motor (K-of-6 over 6 h, paired MOTOR_SHUFFLE=1 control) are PARKED — accruing, not a sealed hold.

SIGNED consult — Design #3 `Proprioceptive Servo Bridge` (DESIGNED / not-yet-run)

UNI-GPT consult 2026-06-27 (SIGN, Q6) folds in the next design to accrue, aimed at a live PASS (not bound-seeking — no strawman built to farm negatives). It is a Class-C UNI design; the standard part is only the AIF control vocabulary. It changes 4 dimensions vs #1/#2: (1) coupling topology — scalar reward/Z coupling → closed-loop triadic (task policy → motor setpoint → proprioceptive error → corrective action); (2) timescale source — immediate-reward / slow-Z → event-triggered fast motor correction (update when proprioceptive PE crosses threshold, else hold the current primitive); (3) information bottleneck — global Z → low-dim proprioceptive residual r_motor = desired_pose/velocity − inferred_pose/velocity; (4) control path — direct discrete action bias → a motor shim turning policy intent into continuous/local corrective commands.

Protocol L5_D3_PROPRIO_SERVO_BRIDGE_HELD_v1 (designed): paired kin seeds, one cure at a time (no Z-bottleneck cure, no new learning rules, no exploration bonus, no reward edits). Held bar: held_motor_delta = treatment − control, paired bootstrap 95% CI excludes 0 positively, point estimate ≥ +0.05 (smaller than #1's +0.092 but nontrivial). Live bar: paired CI for live motor-success delta excludes 0 positively AND no earned survival/allostasis/task metric regresses beyond non-inferiority. motor_score = task-relevant successful actions − collision/overshoot/oscillation − stalled-control − energy/recovery penalties ("more movement" must NOT count as success). Discriminator = delayed/slipped embodiment trials (actuator noise / one-step delay / friction change / positional perturbation / contact-state ambiguity). Ablations that MUST collapse the gain: (A) ε_prop := 0 (shim present, compute runs); (B) ε_prop shuffled across time/kin; (C) open-loop setpoint only (mapping kept, no feedback correction).

K-discipline (held DOWN): if Design #3 fails cleanly and is truly structurally distinct with all pre-registered bars/controls valid, the ledger would have at most K-negative = 2 — still no Section 0.6(B) motor bound owed until K≥3. It does not count toward K≥3 if <2 structural dimensions changed, if undertuned vs control, if it silently combines cures, if the protocol drifts so deltas aren't comparable to #1/#2, if the discriminator is absent/too easy, if the ablation is hardcoded (disables the whole motor stack) rather than causal (only ε_prop), or if a bug/seed-imbalance/env-drift/logging-failure breaks paired inference. Status: DESIGNED / not-yet-run — no result, no rung raised. Design #3 RAISES NOTHING and the rung's status is UNCHANGED; the Design #2 negative is not erased by it.

HONEST FENCE — proven (PASS) + NEGATIVE (symmetric bound)

proven — a held, sealed, UNI-signed PASS exists with its falsifier still live: the diagonal-A posterior shift (Class A), the live RCON craft-chain mechanism gate (Class C), the A3 Design #1 synthetic PASS (Class C, +0.092 [+0.038, +0.157]), 277 offline tests (Class E). Carried inseparably with its symmetric NEGATIVE (A3 Design #2, −0.091 [−0.134, −0.055], K-negative = 1, no Section 0.6(B) bound owed).

The A3 PASS allows exactly one claim and nothing stronger: "variationally-controlled active-inference evidence on body↔world coupling under the registered SYNTHETIC protocol." It is synthetic-process only (NOT live-appliance, NOT recorded-hardware) and NOT near-optimal control (the agent plateaus at 0.222 versus an oracle's 1.0 — the active channel mattering is the entire allowed claim).

Not claimed at L5: not general motor intelligence, not human-like embodiment, not AGI, not consciousness, not "created life", not "active inference demonstrated". Design #3 is designed, not run: it RAISES NOTHING. Live behavioural K-of-6 motor tallies are PARKED. Honest program position: ~2 of 11+ developmental rungs earned — a bounded peek, a toy world, never a person.

sha256 a5eccba95bc70272 — of the original file, so what was ingested stays checkable.

Written for this website — not the document. This is a plain-language retelling, written to help you meet the document. It is not the source, and it is not evidence. It has not yet been checked by a person. Read the document itself → (or choose Precise in the reading-level control above)

This recipe is the rung where the simulated agent moves. The body senses its own configuration, a servo carries out goals handed down from above, and the sensory consequences of its own actions teach it how the world responds. It is a simulation of embodied action, never a person.

The one thing this page says is that a positive result and a negative result are carried together and mean different things. On a fast axis where reward arrives immediately, the active channel mattered. On a slow axis where reward is delayed and the body's own signals are squeezed through a narrow bottleneck, it did not. Both were sealed, both are reported, and the negative delimits the positive rather than being hidden by it.

The page is precise about what the positive allows. It is a synthetic protocol rather than a live appliance or recorded hardware, and it is not near-optimal control: the agent plateaus far below an oracle. That the active channel mattered at all is the entire permitted claim.

A further design has been signed off to run next. It has not been run, it raises nothing, and it does not erase the negative.

Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is a5eccba95bc70272

Written for this website — not the document. This is a clearer retelling, written to help you meet the document. It is not the source, and it is not evidence. It has not yet been checked by a person. Read the document itself → (or choose Precise in the reading-level control above)

This chapter is the mind-and-body rung. It builds a motor hierarchy on the same engine as every other rung. The body infers its own configuration, a continuous error-correcting servo fulfils goals projected down the hierarchy, and the sensory consequences of the agent's own movement train the model of how things change. The page opens by disclaiming the obvious reading — nothing here is a demonstration of the underlying theory, and the composed system is exact only at one narrow interface, never globally.

The first build step is the most interesting, because it is a fix for a genuine defect. A single uniform sensory channel for proprioception is not identifiable at all: the posterior simply sits at nothing. Installing a structured prior breaks that symmetry and the posterior moves, which is what lets the body infer its own configuration at all. From there the servo and the reafference are modelled on the same loop, with no new optimiser and no gradient methods, and the organ ships as an opt-in addition so the default colony stays byte-identical.

Two quite different gates are then run and the chapter refuses to blur them. One is a live mechanism gate on a real server, where a lineage bootstraps a crafting chain and the result is confirmed by the server rather than by the agent's own report. Its registered discriminator is simple: remove the motor organ and the harvest must collapse. The other is a synthetic held-once protocol, where a bar is registered, the held set is sealed and touched once, and the verdict is read off the interval that excludes the threshold across several disjoint seeds. The chapter even flags a scoring trap in this protocol — a score that sums over time penalises survival, so a fast-dying control could masquerade as a negative unless survival and per-step scores are reported separately.

The results are symmetric and both are sealed. One design, on a fast immediate-reward axis, passed with its interval clear of no-difference. The other, on a slow delayed-reward axis with the body's signals compressed through a narrow bottleneck, held negative. The chapter says the shape of that pairing is the finding: the active channel mattered on one axis and did not matter on the other. It also counts carefully — this is the only sealed motor negative so far, so no formal negative bound is owed yet, since a bound requires several structurally distinct negatives.

A signed consultation design is folded in for what to run next, aimed at a positive rather than at farming further negatives. It changes several structural dimensions at once relative to the earlier designs. The coupling becomes a closed loop, the trigger becomes an error crossing a threshold, the bottleneck becomes a small local residual, and the control path becomes a shim producing continuous corrections. It specifies its bars, its discriminator of perturbed trials, and the ablations that must collapse the gain. It also spells out the conditions under which it would not even count toward a future bound. Its status is designed and not run: it raises nothing.

The closing verdict carries the pass and the negative inseparably, and lists what is not claimed — not general motor intelligence, not human-like embodiment, not created life, not a demonstration of the theory. The live behavioural tallies are parked and still accruing rather than sealed.

Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is a5eccba95bc70272

L5 — Sensorimotor / motor hierarchy (embodied action)

How to read this page

Precise — the source document