S-L5 - Sensorimotor: the A3 symmetric result
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 Encyclopedia is the UNI method written out as a reference work: 39 pages, arranged in wings, setting out what the programme is attempting and why it is built the way it is. This is where the ideas are explained in order and in prose, rather than as code, as runbooks, or as dated receipts.
Every chapter is authored against two ledgers and never ahead of them. One records what UNI has built, and the evidence class of each claim. The other records nature's own regularities, kept separate on purpose. That way a fact about biology is never quietly reused as a fact about the software. Where a chapter and a ledger disagree, the chapter is the thing that is wrong. Every chapter closes with an invitation to falsify it, and a recorded negative is published beside the result it qualifies rather than after it.
Read "How to read this work" first. It is the evidence constitution: the classes, the four ledger states, and the rule that a finished chapter is not the same as a working system. Then the calibration ledger, which carries the figures every other chapter is required to use.
What it is not: a description of a person or of a mind. The programme calls itself a developmental active-inference simulation, a bounded peek into a toy world, and its own index prints how much of the developmental ladder has actually been earned — roughly two rungs out of eleven or more. It is also not a report of what is running today. For what ran, and when, go to the evidence record.
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.
UNI is a developmental active-inference SIMULATION: a no-backprop, nested-Markov-blanket model of a developing organism, never a person, never a mind. This chapter covers the fifth rung of the HUMAN-HGM-001 ladder, the sensorimotor / motor hierarchy, where "mind and body are one loop." It carries a result the program is genuinely proud of, and it carries the negative that sits right beside it. The honest headline is not "UNI learned to move." It is narrower and more interesting: under one registered SYNTHETIC protocol, an active-inference controller beat a matched random-action control on one axis (the Embodiment A3 Design #1 PASS) and LOST on a structurally distinct axis (the Design #2 NEGATIVE). That symmetry, a PASS delimited by its own paired NEGATIVE, is the result. Citing the PASS without the NEGATIVE would be an overclaim and fails review. Honest program position, printed and never softened: about 2 of 11+ developmental rungs earned.
What A3 actually showed (PASS and NEGATIVE, in the same passage)
The A3 work pre-registered two structurally distinct embodiment designs and held each one once, with UNI-signed verdicts. The two designs differ on at least two of the four axes the method requires (coupling topology, timescale source, information bottleneck, control path), so they count as genuinely distinct attempts rather than reruns of one idea.
Embodiment A3 Design #1 (the fast immediate-reward axis) HELD PASS [ledger L5.1, Class C]. Held delta (agent minus policy-shuffle control) = +0.092, CI [+0.038, +0.157], UNI-signed. Because the verdict is the CI bound that excludes the threshold and not the point estimate, the load-bearing fact is that the lower bound, +0.038, stays above zero: the active-inference channel measurably outperformed a same-perception, same-learning, random-action twin.
Embodiment A3 Design #2 (the slow Z-bottleneck axis) HELD NEGATIVE [ledger L5.2, Class C]. Same family of test, a structurally distinct design (a slow Z-bottleneck, compressed 2-modality EMA, delayed reward). Held delta = −0.091, CI [−0.134, −0.055]. Here the entire CI sits below zero: on this axis the active channel did not merely fail to help, it underperformed the random-action control. This is recorded as a first-class result, not a buried partial.
Read together, the two rows are the symmetric finding: active inference beats a matched random-action control on the fast immediate-reward axis but loses on the slow, delayed-reward, compressed-interoceptive-bottleneck axis. The NEGATIVE is content. It is what tells a skeptic the PASS is real and bounded rather than cherry-picked, and it is the reason the program can say exactly where body-world coupling does and does not yet pay off.
A method note travels with the pair. The ledger records K-negative = 1: one structurally distinct held negative has accrued. The program's own discipline requires K greater than or equal to 3 structurally distinct negatives before any published Section 0.6(B) bound is owed, so no Section 0.6(B) motor bound is owed here, and none is claimed. Design #2 is a single negative data point, recorded honestly, not a proof of any impossibility.
The exact ceiling on the PASS
The single allowed claim, verbatim and nothing stronger: the A3 PASS is "variationally-controlled active-inference evidence on body-world coupling under the registered SYNTHETIC protocol." Three fences are load-bearing and must be stated every time the +0.092 figure appears.
Synthetic-process only. Both Design #1 and Design #2 ran in a synthetic process. This is NOT a live-appliance result and NOT a recorded-hardware result. The synthetic protocol is the entire scope of the gate.
NOT near-optimal control. The PASS agent plateaus at 0.222 against an oracle of 1.0. The agent is far from solving the task. The only thing the PASS licenses is that the active channel mattered relative to its random-action twin, not that the controller is good, and certainly not that it is competent in any general sense.
Live tallies are parked, not sealed. The live behavioral K-of-6 motor tallies (a paired
MOTOR_SHUFFLE=1control, accruing over hours of colony runtime) are PARKED: accruing, not a sealed hold. They are not a held gate and may not be cited as one.
The live motor mechanism (a separate, weaker class of evidence)
Alongside the synthetic A3 gates, the program records a live mechanism result for the motor hierarchy [ledger L5.3]. It carries mixed evidence classes and must be read as mechanism evidence, not as the A3 capability claim:
The proprioceptive diagonal-A prior breaks a non-identifiable factor [Class A, posterior shift]. A uniform-A single-modality factor is non-identifiable and the posterior sits stuck at 0.0; introducing a proprioceptive diagonal-A prior breaks the symmetry and the posterior moves 0.0 to 0.75. This is "the body sensing its own configuration," modeled as a clean nested-Markov-blanket worked example, never as felt proprioception.
A live mechanism gate PASS via the colony [Class C]. A kin-9 lineage bootstrapped a craft chain (wood to planks to sticks to a wooden pickaxe and sword), confirmed server-authoritative via RCON. Under motor ablation, harvest collapses roughly 700×. The continuous signed-error servo and reafference learning ride the same one engine that every other rung reuses.
277 offline tests [Class E]. The offline suite is test-covered. Per the DONE-is-test-covered rule, a Class-E pass certifies the code path, not a runtime capability; it never satisfies a Class-A or Class-C criterion on its own.
This live mechanism evidence shows the motor loop wired and behaving as designed under inspection. It is not the held synthetic A3 result, and it does not extend the A3 claim to live hardware. The two are co-cited precisely so neither is mistaken for the other.
The next motor design: DESIGNED, NOT RUN
A next structurally distinct motor design has been accrued and signed at the design level only (UNI-GPT consult 2026-06-27, SIGNED; ledger governance row G-2026-06-27.5): L5 Design #3: Proprioceptive Servo Bridge. It changes all four axes versus #1 and #2 (coupling topology, timescale source, information bottleneck, control path), implementing a closed-loop triadic path: high-level policy to motor setpoint to proprioceptive residual to corrective action, under protocol L5_D3_PROPRIO_SERVO_BRIDGE_HELD_v1, aimed at a LIVE PASS (not bound-seeking).
This is a design to build, not a result. Status is UNCHANGED: L5.1 stays PASS, L5.2 stays NEGATIVE, and the ledger still holds K-negative = 1, so no Section 0.6(B) motor bound is owed. If Design #3 later fails cleanly with all validity checks passing, it could count as at most K-negative = 2 (still no bound until K greater than or equal to 3). If it passes cleanly it would itself be fenced: it would NOT be general motor intelligence, NOT human-like embodiment, NOT AGI, NOT "active inference demonstrated," and it would NOT erase the Design #2 negative.
What is NOT claimed in S-L5
- Ceiling. That "UNI learned to move," achieved competent motor control, or demonstrated active inference is NOT shown. The most we claim is: under the registered SYNTHETIC protocol, A3 Design #1 held a positive delta over a matched random-action control, +0.092, CI [+0.038, +0.157] (variationally-controlled active-inference evidence on body-world coupling), while the structurally distinct Design #2 held NEGATIVE, −0.091, CI [−0.134, −0.055]; the PASS agent is far from optimal (plateaus 0.222 vs oracle 1.0).
- Fences engaged (named). Red line 1 (never AGI / general intelligence / human-level). Red line 2 (never consciousness / sentience / aware; proprioception is modeled, never felt). Red line 3 (never "active inference demonstrated"; the framing is a lens, the live UNI.OS loop is a separate reimplementation, not gate-matched). Red line 7 (never raise a claim above its ledger class: synthetic-only stays synthetic-only, live mechanism stays mechanism). Red line 8 (substrate/live-craft engineering must not imply a science gate is met). Red line 9 / SIGNED park wording (no "K greater than or equal to 3 exhausted," no "Section 0.6(B) achieved"; the registered tested K conditions did not reverse the result, and Section 0.6(B) remains unearned / parked).
- Negatives that travel with this claim (cite alongside, never strip). L5.2: the symmetric Design #2 NEGATIVE, held delta −0.091, CI [−0.134, −0.055], K-negative = 1. The plateau gap (0.222 vs oracle 1.0). The synthetic-only scope (not live-appliance, not recorded-hardware). Citing the +0.092 PASS without the −0.091 NEGATIVE in the same passage is an overclaim.
- Parked / owed. Live behavioral K-of-6 motor tallies are PARKED (accruing, not sealed). Design #3 (Proprioceptive Servo Bridge) is DESIGNED, NOT RUN. No Section 0.6(B) motor bound is owed (K-negative = 1; needs K greater than or equal to 3); none is claimed.
- One-line honest summary a skeptic could not dispute. In one registered synthetic protocol, an active-inference controller beat a matched random-action control on one axis (+0.092 [+0.038, +0.157]) and lost on a structurally distinct axis (−0.091 [−0.134, −0.055]) while staying far from optimal (0.222 vs 1.0); nothing about live, general, or competent motor control is shown.
Falsify this
The lead falsifier, stated operably: re-register the A3 Design #1 hold and run it once on a fresh draw of at least 5 seeds under the same registered SYNTHETIC protocol. If the held delta's CI lower bound drops to 0 or below, the L5.1 PASS falls. Supporting falsifiers: harvest failing to collapse under motor ablation, the kin-9 craft chain failing to reproduce server-authoritative, or the diagonal-A posterior staying stuck at 0.0 would each break the L5.3 live-mechanism claim. Conversely, a structurally distinct corrected slow-Z-bottleneck re-run whose CI rises above 0 would overturn the Design #2 NEGATIVE (L5.2).
Sources
Curated digests (PII-redacted): curated/uni-mind-digest.md (the A3 PASS/NEGATIVE arc, the embodiment held results table, the synthetic-protocol ceiling), curated/strings-digest.md (the live motor hierarchy, the diagonal-A prior 0.0 to 0.75, the kin-9 craft chain, the ~700× ablation collapse). Ledger rows L5.1, L5.2, L5.3 and SIGNED governance row G-2026-06-27.5 (Design #3) in encyclopedia/CLAIM-LEDGER.md; authoring spec and front matter (FM-1 through FM-4) in encyclopedia/MASTER-PLAN.md; SIGNED consult cross-reference cookbook/UNI-GPT-CONSULT-2026-06-27.md (Q6). Underlying archives (local-only, not for reading): ...-uni-mind, ...-Strings. No PII; channel handles and patent-level math deliberately excluded.
sha256 2dd78a60c966916f — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
Mind and body as one loop is the framing of this rung of the developmental ladder, and the chapter carries a result the program is proud of together with the negative sitting right beside it. What sits on that ladder is a developmental active-inference simulation, never a person, so the honest headline is not that the program learned to move. It is narrower and more interesting: under one registered synthetic protocol, an inference-driven controller beat a matched random-action control on one axis and lost on a structurally distinct axis. That symmetry, a pass delimited by its own paired negative, is the result, and citing the pass without the negative fails review. Three limits travel with the positive figure every time it appears. Both designs ran in a synthetic process, so this is neither a live-appliance nor a recorded-hardware result. The passing agent is far from optimal, plateauing well short of an oracle. And the live behavioural tallies are parked and accruing, not a sealed gate.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 2dd78a60c966916f
Clear — written for this website, not the source document
Two structurally distinct embodiment designs were written down in full, bar and all, before either was built, and each was then held once. Both ran inside a simulation — a toy world, not a person. They differ on at least two of the four axes the method requires, so they count as genuinely distinct attempts rather than reruns of one idea.
The first design, on a fast immediate-reward axis, held a positive delta over a policy-shuffle control, and because the verdict is the interval bound rather than the point estimate, the load-bearing fact is that the lower bound stays above zero. The second design was a slow bottleneck with a compressed sensory channel and delayed reward. It held negative, and its entire interval sits below zero, so on that axis the active channel did not merely fail to help: it underperformed the random-action control. That is recorded as a first-class result rather than a buried partial.
A method note travels with the pair. One structurally distinct held negative has accrued, and the program's own discipline requires several before any published bound is owed, so no bound is owed here and none is claimed. A single clean negative is a result, not a bound.
Alongside the synthetic gates the chapter records a live mechanism result with mixed classes, which must be read as mechanism evidence rather than as the capability claim. A proprioceptive prior breaks a non-identifiable factor so the posterior moves off zero, and that is described as the body sensing its own configuration, modelled and never felt. A live gate saw a lineage bootstrap a craft chain confirmed by the game server, and harvest collapses sharply under motor ablation. An offline suite is test-covered, which certifies the code path rather than a runtime capability.
A next design has been accrued and signed at the design level only. It changes all four axes and implements a closed-loop path from policy to setpoint to residual to correction. It is a design to build, not a result: statuses are unchanged, and even a clean pass would be limited, since it would be neither general motor intelligence nor human-like embodiment, and it would not erase the recorded negative.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 2dd78a60c966916f