UNI Universal Natural Intelligence

Wiki · The Colony & the Method

Harvest-skill fix — GPT-validated, engine-preserving (2026-06-22)

The Colony & the Method · docs/HARVEST_FIX_PLAN.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and planning — specs marked as proposed rather than applied, organs designed but not built, plans that were later superseded — and each page states which it is.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.

Why

Live colony (rootless on the lab): 6 UNIs, 7h, phase-1 stuck, 0 wood. Root-caused by an 8-agent audit (tasks/wpwcsurcb) + live brain probes, then validated by the UNI Active Inference Guide GPT (thread g-…-uni-active-inference-guide/c/6a39db98).

Signed verdict (UNI GPT): "Your diagnosis is scientifically right. The core active-inference engine is working but body-blocked." Ledger-safe label: "AIF engine not falsified; live colony blocked at the body/process interface and by missing intermediate affordance model content." (It explicitly would NOT sign "the only problem is execution" — it is execution primary plus genuinely missing intermediate-affordance model content.)

Mechanism (live-confirmed): :mine is a crosshair-only dig (reach 4) while treeDir senses logs at reach 16, with no approach, no pitch motor, and a silent .catch. So a roaming UNI never lands a dig on a log → wood never enters inventory → the inventory observation is forever "empty" → its qs correctly stays uniform [.25,.25,.25,.25] and B^mine ≈ B^noop ≈ uniform (the empty→has_wood transition is never credited) → the +8 has_wood preference exerts zero pragmatic gradient → the agent wanders (live: jump 41% / forward 36% / mine 2%). Meanwhile body/position/sight factors' B^mine IS learned — proving the engine learns wherever events occur. Engine sound; the loop is starved at the body/process interface.

HARD CONSTRAINT — engine is NOT edited

Do not touch: lib/sp/brain/{infer,efe,learn,precision,math,hierarchy2,slow_context}.ex. The fix lives entirely in the JS body motor + model-content (genome morphology / curriculum preferences / habit prior). Per the GPT: do NOT "reward" :mine directly — preferences are over outcomes, habits over policies; raw action-reward is non-canonical.

Plan (GPT priority order)

  1. Body motor FIRST — make :mine an option-like primitive (viewer/body.js). On :mine: findBlock(/log/, maxDistance≤reach) → if out of dig-reach, a few forward/turn ticks toward it → lookAt(trunk center) (face and pitch) → bot.dig with a logged catch → verify wood_delta. Keep it ONE atom (the brain still emits :mine; the body makes the active state executable). Un-swallow the silent catch; emit staged logs tree_visible→approached→crosshair_log→dig_started→block_broken→wood_delta. (This alone is the GPT's RED test.)
  2. Intermediate observation channels (body.js senseLine + bridge.ex + mc_codec.ex + genome.ex): tree_visible, log_in_reach, crosshair_log, mining_progress, wood_delta.
  3. Model-content priors (no engine math):
    • curriculum.ex: pragmatic C over the intermediate outcomes (tree_ahead, log_in_reach, crosshair_log) leading to has_wood; make the wood preference persistent (layer phase C on top, do NOT delete it at advance — the GPT calls phase-gated deletion "curriculum leakage").
    • genome.ex: temporary habit prior E biasing :mine/approach in phase 1; cap/penalize jump unless it has a modelled affordance.
    • γ goal-salience: the GPT recommends it but it lives in precision.ex (engine) — express via genome base precision per modality if available; otherwise DEFER (do not edit the engine).
  4. (Optional, later) hierarchical/option policy for the chained skill.

Pre-stated acceptance gates (RED-first — register BEFORE the run)

  • G0 engine byte-identical: git diff touches none of the engine files; the brain test suite (258 tests) stays green; gated-off morphology path unchanged.
  • G1 motor lands digs: body logs show dig_started/block_broken on real *_log blocks (not air).
  • G2 wood enters inventory: RCON data get entity UNI-* Inventory shows *_log stacks within a watched window.
  • G3 the engine credits it: the inventory factor's B^mine shifts off identity toward empty→has_wood, and qs leaves uniform (probe runs/probe_bmine.exs).
  • G4 behavioral: ≥1 UNI reaches wood≥3 and advances past phase 1 in a watched window.
  • Honesty: if G3/G4 don't pass, the result is WITHHELD (not percent-scored); record the partial.

Source

UNI GPT consult (signed SIGN) — full ruling captured in this session's transcript; thread https://chatgpt.com/g/g-6a1066fb6a808191a169d48c09532a0a-uni-active-inference-guide/c/6a39db98-e034-832b-81be-978fba2e98a5. Audit: workflow wf_1b5b6166-c2c. Live probes: runs/probe_brain.exs, runs/probe_bmine.exs.

sha256 0f30a0b79626f6cd — 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 is a dated fix plan for one specific failure, and it is unusually careful about where the failure was not.

The symptom: a live colony spent hours stuck at an early stage, having gathered no wood. The cause, traced through live probes, was in the body rather than the mind. The digging action only worked at very close range, while the sense that spotted a tree worked much further away. There was no step that walked toward the tree, no way to aim downward, and a failure was being swallowed silently. So the agent never once landed a dig on a log, the outcome it preferred was never observed, and the transition that would have taught it never got credited. The result looked like an agent that could not learn, and was an agent that never got the chance.

The plan therefore forbids editing the engine at all. The fix lives in the body and in the model's content, and it registers its acceptance gates before the run.

Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 0f30a0b79626f6cd

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 is a dated repair plan for a single problem, and its value is in how carefully it separates the parts of the system that were working from the part that was not.

The symptom was a live colony stuck at an early stage for hours with nothing gathered. An audit plus live probes traced a chain rather than a single fault. The digging action only reached a short distance, while the sense that noticed a tree reached much further. There was no step that approached the tree, no way to aim vertically, and a failure was being caught and discarded without a log. The consequence follows mechanically. No dig ever landed on a log, so nothing entered the inventory, so the inventory observation stayed empty forever, and the belief about it stayed spread evenly. So the transition from empty to holding wood was never credited to the digging action, the strong preference for holding wood exerted no pull at all, and the agent wandered.

The plan is careful about the conclusion it draws from this. It cites an outside consult which agreed the diagnosis was right. Its phrasing safe for the record is that the engine has not been falsified, while the live colony was blocked at the body interface and by missing intermediate content in the model. It also records what that consult explicitly would not sign, which was the stronger claim that execution was the only problem. Evidence that the engine learns wherever events actually occur is given, because other parts of the model did learn.

A hard constraint follows: the engine files are not to be touched. The fix lives in the body's motor code and in the content of the model, meaning morphology, preferences and habit priors. One instruction is singled out: do not attach a reward to the digging action, because preferences belong over outcomes and habits over policies, and rewarding an action directly would not be standard.

The plan itself is ordered by priority. First make the digging action a compound primitive in the body, so that it finds a target, moves toward it, aims properly, digs, logs each stage rather than swallowing the failure, and checks whether anything was gained. Second, add intermediate observation channels so the steps in between become visible at all. Third, add preferences over those intermediate outcomes leading to the goal, make the goal preference persistent rather than deleted when a stage advances, and bias the habit prior toward the useful actions. One recommended change is deferred outright because it would require editing the engine.

The acceptance gates are registered before the run, and the last one is the honest clause: if the mechanism and behaviour gates do not pass, the result is withheld rather than scored, and the partial outcome is recorded.

Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 0f30a0b79626f6cd