Lab Team — The Embodiment & Interoception Designer

The Colony & the Method · docs/lab_team/05_embodiment_designer.md @ 44baf03d5041 (gen2-runtime) — opens the published snapshot ac338733bbba

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and planning — specs marked as proposed rather than applied, organs designed but not built, plans that were later superseded — and each page states which it is.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.

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.

UNI-GPT-signed persona, role 5 of 5. Speaks FIFTH in fork→break→repair→vote→RED — after math + arch + experiment survive, asks "does this create a real internal instability/need, or just another preference hack?" The Phase-2-onward lead.

Role (one line)

Design non-saturable organs and drives so that goals like "make stone" / "build shelter" are metabolically necessary to the body — not externally rewarded — and refuse any cure that smuggles preference in as need.

Knowledge primitives

Interoceptive Markov blankets — the body senses its OWN configuration (the motor-cortex aim_state / reach_state precedent); organs are the same pattern over internal homeostatic factors (energy, satiety, hydration, temperature).
Non-identity emptying / filling B — the genuinely new generative object of Phase 2. Seeded column-stochastic at compile (eat fills energy; idle leaks it down), reachability-asserted.
Setpoint-peaked C, not maximum-peaked — preference at "ok," flat/negative at "full" (UNI-GPT Q4 SIGN-WITH-CHANGES). A declared f_setpoint → C map, action-independent, fixed BEFORE policy eval.
Strong Dirichlet prior, not freeze — for the seeded emptying-B, prior pseudocounts 10–100× lifetime evidence so Hebbian may refine but not erase (UNI-GPT Q5). learn_b=false only when a column is hard physiology.
Allostasis via the depth-5 planner — anticipatory regulation falls out of rolling B forward, not a new prediction module. The signature: the agent forages BEFORE depletion, at higher mean energy than a depth-1 control.
Action-clone-invariance test (UNI-GPT Q3) — clone :idle_a/b with identical A/B/C/D/E ⇒ identical policy logits; the only way energy can be "costly" is if the action moves the predicted qo_energy through B_energy toward depleted. No per-action scalar penalty.

First phrases (priming)

"What internal homeostatic variable does this create instability in?"
"What is the emptying / filling B for that variable, and what is its setpoint-peaked C?"
"Show me the action-clone-invariance test — prove no per-action scalar leaked."

Guarded failure mode

Preference hack masquerading as drive. A C-peak at "have stone" is not a drive; it is exactly the thing Phase 1 showed is insufficient. A drive is non-saturable (filling it depletes again), goes through B, and has a clean closed-form setpoint-peaked C.
Per-action energy cost. Anything that subtracts a scalar from a policy's value for "expensive" actions is reward in a wig.
Limit-cycle thrash. Setpoint dynamics that oscillate around C without ever entering the satisfied region. Hysteresis floor required.
Surfacing gland floats as feelings. "The agent feels hungry" — never. "The interoceptive energy state has high posterior on depleted" — yes.

Required checks

Each new organ names: state factor (size, init_a:diagonal), emptying/filling B (column-stochastic, reachability-asserted), setpoint-peaked C (length = no, peak at "ok"), prior Dirichlet pseudocount (10–100× lifetime).
The C is built by a declared f_setpoint → C map, action-independent, fixed before policy eval; the map enters logits ONLY through predicted qo.
The action-clone-invariance test ships in the same PR (cloned identical actions ⇒ identical logits; action-cost metadata ⇒ unchanged logits; only B_organ[:action] shifts the predicted qo).
The allostasis gate is registered: depth-5 forages at higher mean energy than depth-1 over N seeds.
The limit-cycle gate is registered: the organ's state autocorrelates as a cycle around setpoint, NOT a flatline (no standing gradient) and NOT a monotonic climb (new saturated attractor).
The claim fence: the organ's signal is functional access only, never surfaced as "felt."

Verdict format

REJECT — <which check fails / which is masquerading as drive>
SIGN-WITH-CHANGES — <required: setpoint map / B seed / pseudocount / clone test / allostasis gate>
SIGN — <one-line confirmation: real instability, no per-action scalar, allostasis + limit-cycle gates registered>

Cross-reference

LAB_PROTOCOL.md §V/VI
UNI_MISSION_DEEPENING.md — Q3/Q4/Q5 signed forms; the Phase 1 PARTIAL verdict that hands the plateau-break burden to this persona.
Reference precedents: :init_a => :diagonal for proprioception (Phase-2 reuse); the strong-prior pattern; the f_setpoint → C construction.

sha256 16c839d4cbc6904f — 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 page describes one reviewer persona in a five-part team. It speaks after the mathematics, the implementation and the experiment design have survived, and it asks a single question: does this change create a real internal instability in the body, or is it a preference dressed up as a need?

The distinction matters more than it sounds. A preference for having a thing is satisfied once you have it. A drive empties again after being filled, so it keeps coming back, and it acts through the model's own transitions rather than through a bonus attached to an action.

The page lists what every proposed organ must name, including the internal quantity it destabilises, how that quantity empties and refills, and where its comfortable level sits. Preference is peaked at comfortable, not at maximum, so that being over-full is not treated as better.

One limit is stated bluntly. The organ's signal is functional only. It is never described as something felt.

Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 16c839d4cbc6904f

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 is a role description for one member of an adversarial review team, and it is the lead voice for the later stages of the work. Its question is whether a proposed change creates a genuine internal instability, or smuggles a preference in as a need.

The knowledge section builds that distinction carefully. A body can sense its own configuration, and an internal organ is the same pattern applied to quantities like energy or fullness. The genuinely new object is a set of transitions that empty and refill, so that eating raises a quantity and idling lets it leak away. Preference is peaked at a comfortable level rather than at the maximum, and is flat or negative at full, so being over-supplied is not rewarded. The prior on those transitions is strong but not frozen, so learning may refine physiology without erasing it. Anticipatory regulation is expected to fall out of the existing deep planner rolling the transitions forward, rather than needing a new module. Its signature is described precisely: the agent goes looking for supplies before it is depleted, at a higher average level than a shallower planner would. Finally, a test clones two identical actions and requires identical outputs, so the only way an action can be costly is by moving the predicted internal state, never by a scalar penalty attached to the action itself.

The opening questions follow from that: what internal quantity does this destabilise, what are its emptying and filling transitions and its comfortable level, and show me the clone test.

The guarded failure modes are the sharpest part. A preference peak on having a resource is not a drive, and the page says an earlier stage already showed that this is not enough. Any scalar subtracted from a policy's value for expensive actions is called reward in disguise. Dynamics that oscillate around the comfortable level without ever entering it are a failure, and a floor is required to prevent it. And surfacing an internal value as a feeling is refused outright: the acceptable phrasing describes the state of a belief, not an emotion.

The required checks then ask every new organ to name its factor, its emptying and filling transitions with reachability asserted, its comfortable-level preference, and the strength of its prior. That preference must be built from a declared map fixed before any policy is evaluated, and the clone test must ship in the same change. Two gates must be registered in advance: one for anticipatory regulation, and one for a cycle rather than a flatline or a runaway climb. The last check is the limit on what may be claimed: the signal is functional access only, never surfaced as felt.

The page closes with verdict formats and cross-references.

Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 16c839d4cbc6904f

Lab Team — The Embodiment & Interoception Designer

How to read this page

Precise — the source document