UNI Universal Natural Intelligence

Wiki · The Colony & the Method

Motor-Inference Hierarchy — LIVE RED test (P4), pre-registered

The Colony & the Method · docs/MOTOR_RED_TEST.md @ 44baf03d5041 (gen2-runtime) — opens the published snapshot ac338733bbba

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.

Eighty-four pages about the colony. Each agent is an Elixir process holding a generative model and doing inference, attached to a body that logs into a Minecraft world as an ordinary player. Around that sit the broadcast suite that films them and the runbooks that keep the whole thing running. There are typed specifications for each organ of the model, plus the world and genome specs. There are also the adversarial review personas used to attack a proposed change before it ships.

It is for the reader curious how a running system is put together and how it is held to account. The accountability half is the more distinctive. There is a lab protocol governing evidence and attribution, and a claim fence that restricts the vocabulary a claim is allowed to use. There is a public gate log. And there is a standing invitation to reproduce any verdict from the commit and the seed named in its receipt.

Start with the public read, then the lab protocol, then the falsification invitation. If you want the mathematics rather than the operations, go straight to the typed organ specs.

What it is not: a description of a mind, and not all one kind of document. A large part of this corpus is design and planning — specs marked as proposed rather than applied, organs designed but not built, plans that were later superseded — and each page states which it is. A specification is not a running system, and these pages are careful about the difference; the reader should be too. Eight documents were withheld from publication because they describe private infrastructure.

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.

Registered before the run (RED-first discipline, same as HARVEST_FIX_PLAN.md). The claim under test and the falsification gates are fixed here so the result cannot be moved after the fact.

Claim under test (and the claim FENCE)

A categorical motor-inference hierarchy can learn a Minecraft harvest motor chain under the registered conditions — the parent option projects a desired proprioceptive configuration (C_motor) DOWN, the motor inner loop (SP.Brain.MotorControl + SP.Brain.Motor) fulfils it by descending proprioceptive prediction error while inferring the control sign from reafference, and the proprioceptive reafference folds UP to learn B_motor (muscle memory).

NOT claimed (until every bar below is met): "the agent learned harvesting", "human-like motor control", "full human motor inference". If a gate fails, the result is WITHHELD (recorded as a partial), never percent-scored or spun.

What is already validated OFFLINE (committed, deterministic)

  • P1 motor_cortex_test.exs — proprioceptive pipeline additive + :motor_cortex-absent byte-identical (action-sequence + posterior mad < 1e-12); 272-test brain suite green.
  • P2 motor_cortex_test.exs — the motor factors learn A_motor/B_motor, the config posterior becomes informative (peak 0.75), the habit E accumulates, and learned A/B persist across MC.save/load.
  • P3 motor_control_test.exs — the inner loop converges closed-loop against a simulated body, inferring the control sign + axis from reafference, nulling the target-relative prediction error; motor_cortex_test P3 — the live MC.step mine_log option emits fine primitives and self-terminates on dig=broke; the default genome never engages it.

The live RED test below adds the only thing offline tests cannot: that the loop closes through a real mineflayer body in a real world and the chain is learned there.

Conditions (fixed)

  • Lineage: 6 UNIs, Genome.motor_primary() (the 12 default factors + the 5 motor factors), kin 9, usernames UNI-9-1..6, separate memory dir runs/colony_motor/ (a motor brain never loads into a default UNI). Launch: runs/motor_lineage.exs inside the colony BEAM (rootless on the lab, per ops_colony_lab_rootless), MC_HOST=mc-server.
  • Body-assist OFF: :mine routes through the motor inner loop (fine primitives), NOT the mineTree approach-script. (The inner loop's step_forward/hold_mine close range + strike; the brain aims.)
  • Optimistic-B ablated: the coarse harvest bridge is not enabled for this lineage (mine_prior off), so any harvest is attributable to the motor hierarchy, not the task-level prior.
  • Forest seed 8675309 (trees present); phase 1 (the wood-seeking curriculum).

The 3-part gate (pre-stated; PASS requires ALL three)

  1. Behavioral. ≥ K = 3 of 6 motor UNIs reach wood ≥ 3 within N = 6 h, with body-assist off and optimistic-B ablated. (Withheld if < 3.)
  2. Mechanism. For ≥1 passing UNI, the trace shows the full chain in order: :mine committed → mine_log option active (brain.motor ≠ nil) → fine-primitive sequence (turn/pitch/step) → inner-loop target-relative error falls → body logs "motor strike block_broken wood_delta>0" → the motor factors' B updates on the relevant transitions (probe). Evidence: body stderr + runs/probe_motor.exs.
  3. Ablation. A paired control performs worse (paired CI excludes a registered margin):
    • control A: motor-learning off (learn_b=false on the motor lineage), and
    • control B: motor-policy shuffled (the inner loop emits a random fine primitive each tick). If neither control is worse than the live hierarchy, the hierarchy is not the cause ⇒ WITHHELD.

Probes / evidence

  • runs/probe_motor.exs — reads a live motor UNI: option state (brain.motor), the 5 proprioceptive posteriors, and B_motor (off-identity mass) for the motor factors.
  • Body stderr — the fine-primitive strikes (motor strike dig_started / block_broken wood_delta).
  • RCON data get entity UNI-9-* Inventory — independent wood confirmation (as in the harvest fix).

Simulation results (offline, deterministic — runs/motor_sim.exs)

The end-to-end sim closes the loop through the REAL MC.step + inner loop with a simulated body (one tree in front). It de-risks the live run by validating the whole mechanism + the ablation logic short of a real world. Optimistic-B OFF, body-assist OFF throughout.

run wood strikes mine_log options B^mine off-identity result
live hierarchy (depth 5, 4000 steps) 2660 2660 1330 0.0 → 5.99 wood≥3 ✓
live hierarchy (depth 1, 2000 steps) 1475 1475 738 0.0 → 5.98 wood≥3 ✓
ablation B — inner loop SHUFFLED (depth 1, 2000) 2 2 79 learns :mine→no-change partial ✗

Reading: the brain EXPLORES :mine (738–1330 options) with no optimistic-B; the continuous inner loop FULFILS the proprioceptive target (aim→approach→strike); wood enters (1475–2660); the high-level B^mine learns empty→has_wood from REAL success, bootstrapping harvest. Zero raw :mine ever reaches the body (the option always emits fine primitives). The shuffled control collapses harvest ~700× (1475 → 2), isolating the inner-loop SERVO POLICY as the cause — not chance, not the high-level prior. This is the behavioral + mechanism + ablation gate validated in simulation.

LIVE results (lab, uni-colony-motor on localhost/uni-colony:v3, kin-9, mc-server seed 8675309)

Deployed as a SEPARATE container (UNI_AUTOSTART=0) — the default streamed colony (kin 0–3) stays online + undisturbed. All 6 motor UNIs (UNI-9-1..6) connected (RCON list).

  • Mechanism — PASS (live). Body stderr shows the inner-loop chain on REAL logs: motor strike dig_started=birch_log → motor strike block_broken wood_delta=1 (repeatedly), via the fine primitives (NOT the old mine_tree body-assist — body-assist is off for :mine). The brain selects the option, the continuous servo aims+approaches+strikes, real wood enters.
  • Behavioral — passing early, multi-hour tally running. Within minutes, RCON (the server's authoritative inventory) shows UNI-9-2 holding birch_planks + stick + wooden_pickaxe + wooden_sword — it harvested real wood and bootstrapped wood→planks→sticks→TOOLS with optimistic-B OFF and body-assist OFF. The formal K=3-of-6 / wood≥3 / N=6 h tally accrues over the running window.
  • Ablation — validated in sim (~700× collapse); the live paired control (shuffle) is the next run.

Status

  • Offline P1–P3 (unit + integration): PASS (committed; full brain suite 277 / 0).
  • End-to-end SIM (behavioral + mechanism + ablation): PASS (table above).
  • LIVE Minecraft mechanism: PASS (real strikes land real wood; RCON-confirmed harvest→tools).
  • LIVE behavioral K-of-6 over N h: RUNNING (UNI-9-2 already reached tools; tally accrues).
  • LIVE ablation paired control: pending the shuffle run (sim already shows the ~700× collapse).

sha256 9f5ad4b7e2b62ec6 — of the original file, so what was ingested stays checkable.

Plain — written for this website, not the source document

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. (or choose Precise in the reading-level control above)

This page registers a test before the run, so the result cannot be moved afterwards. It states exactly what is being claimed, what would count as passing, and what would count as failing.

The claim is narrow. A layered motor system can learn a chain of movements in the live game. A higher level projects a desired body configuration downward, and an inner loop works out which way to move by watching what comes back. The results fold upward into something like muscle memory.

What is not claimed is written just as plainly, and includes any claim that the agent learned to harvest, or anything about human-like motor control. If a gate fails, the result is withheld rather than scored.

The results section keeps offline work, simulation and the live run separate. Two parts pass live, one is still running, and the paired control is named as not yet done, with the simulated version standing in for now.

Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 9f5ad4b7e2b62ec6

Clear — written for this website, not the source document

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. (or choose Precise in the reading-level control above)

This document registers an experiment before it runs, which is the point of it. Both the claim and the conditions that would refute it are fixed in advance so the outcome cannot be reinterpreted later.

The claim under test is stated as a single sentence and is deliberately narrow. A layered motor arrangement can learn a harvesting movement chain under stated conditions. The level above projects a desired body configuration downward. An inner loop reduces the difference between that target and the body's actual configuration, while inferring which direction its controls push by watching what comes back. That returning signal folds upward, so the transitions are learned.

Immediately beneath it is a list of what is not claimed, including that the agent learned to harvest, human-like motor control, or full human motor inference. If any gate fails, the result is withheld and recorded as partial, never scored as a percentage.

A section then lists what had already been shown away from the live world. The new pipeline is additive and leaves the standard configuration behaving identically. The motor parts do learn, and the learning survives being saved and reloaded. And the inner loop converges against a simulated body while working out its own control direction. The page is explicit about what these cannot do, which is show the loop closing through a real body in a real world.

The conditions are then fixed in detail: the size of the group, which configuration it runs, a separate memory location so a motor mind never loads into a standard one, and two deliberate removals. A helper script that used to do the approaching is switched off, and a coarse shortcut prior is not enabled, so that any success is attributable to the thing under test rather than to either crutch.

The gate has three parts and all must pass. A behavioural part requiring a minimum number of agents to reach a minimum amount of wood within a time window. A mechanism part requiring the trace to show the whole chain in order, from the choice of action through the fine movements to a real block breaking and the transitions updating. And an ablation part requiring two paired controls to do worse, with the honest addition that if neither control is worse then the arrangement is not the cause and the result is withheld.

Simulation results come next in a table, run through the same real decision path with a simulated body. The reading given is that the shuffled control collapses harvesting by a very large factor, which is what isolates the inner loop as the cause rather than chance or the higher-level prior. The page labels all of this as validated in simulation.

The live results are then reported separately and honestly. The mechanism passed, with the body's own logs showing real strikes on real logs and real wood entering. The behavioural part is described as passing early with the formal tally still accruing over a multi-hour window. And the paired control has not been run live at all, which the status list states plainly rather than letting the simulated collapse stand in for it.

Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 9f5ad4b7e2b62ec6