Full-depth generative model — graded per-subsystem viability + closed world↔body↔mind loops (DESIGN)

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and planning — specs marked as proposed rather than applied, organs designed but not built, plans that were later superseded — and each page states which it is.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

Design-only. FE-touching ⇒ every cure needs a /lab-team-review MERGED VERDICT + typed spec + paired pre-registered RED + ship-gate checklist + owner go-ahead (live-stream guard) before code. Adversarially designed + hardened by the full-depth design workflow (wf_4cf9ba90-5ce, 20 agents: 3 grounding → 4 depth domains × math-breaker/embodiment/architect → architect merge). MERGED VERDICT = SIGN-WITH-CHANGES.

Why this exists (owner correction, 2026-07-11)

Fitting into a Minecraft body is not license to reduce the models or their hierarchical depth — it is the opportunity to make them real and really digital. A human arm, down to the cells, gets tired, has extra energy, is nominal, is critical. The model must carry graded per-subsystem viability (cell→tissue→limb→ organism) and close all the loops world↔body↔mind↔body↔world at full depth. See [[feedback_model_depth]].

Measured evidence licensing this: the current metabolism is a single global energy+satiety scalar, each 4 coarse bins, setpoint-peaked C {2:+3, 3:0} (curriculum.ex:33) — flat at "full", no reward for a reserve. The v2 regulation gate (metabolism_regulation_gate_v2.md, FALSIFIES) measured this flat drive dying 6/12 worlds (genuine self-drains, food reachable) and looser in dispersion than a hoarding foil. The shallowness fails, measured. This design is the honest, staged path to the depth — it does not claim to deliver the full per-subsystem/full-loop mandate in one shot.

The unified model — a 3-tier categorical AIF stack, one closed loop, one opt-in organ (`:homeostat`)

All four reviewed depth domains collapse to ONE structure. :metabolism stays frozen (its gates hold); new depth lives behind an opt-in :homeostat organ absent from default/0 ⇒ byte-identical default.

L0 — process substrate (body; NOT a belief — the Markov-blanket process side)

Continuous per-subsystem reserve scalars on the bridge: ATP energy, gut buffer, per-limb fatigue (arm/leg), soma/health. Metabolism.step advances each by wall-clock dt with acted-subsystem attribution (mine/attack → arm, forward/turn/jump → legs; @upkeep on the core every tick incl. :noop = the load-bearing "no free hold" that makes the death edge bite). Drain/refill/death are real world-consequence.

L1 — subsystem interoceptive factors (mind; the graded depth)

Per-subsystem reserve factors, each a standard per-factor A[no×ns] / B[nu×ns×ns] / C[no] / D[ns] Dirichlet factor, ns=no=6 with the owner gradient {0 critical, 1 depleted, 2 tired, 3 nominal, 4 sated, 5 surplus}, init_a :diagonal (self-sensing), b_init :emptying/:filling/:fatiguing durably seeded (pb_seed≈50 ⇒ W_b→0 faster). Members: energy_reserve, gut_satiety (dissociated from energy via a body-side gut→energy digestion transfer, so "full gut / low energy just after eating" is representable), muscle_fatigue (work-accruing / rest-recovering — a fatigue signal, NOT an appetite brake), soma_integrity (health channel).

L2 — organism viability parent (deepest; Cure 5, gated hardest)

One Sg=4 SlowContext/Hierarchy2 marginal filter {thriving, nominal, strained, critical}, NEUTRAL C (never a second reward). Acts ONLY through the DOWN empirical prior. Its transition B^G is DIRECTIONAL (seeded off-diagonal decline drift, OR a new pB^G Dirichlet + learn rule) — not the stock sticky diagonal (a low-pass smoother, incapable of anticipation). Informative W_c maps subsystem-reserve(6) → organism- viability(4), keyed on homeostat factor names only (the 12 default factors keep uniform W_c ⇒ inert even at δ>0).

The closed loop (every segment on a named seam — no dangling half)

world→body: MC food/health/action stream → bridge store drain/refill → real death edge.
body→mind: Metabolism.inject → felt_* observations → subsystem beliefs (mind reads only injected obs).
mind UP: child extrinsic likelihood (not posterior) folds into the parent via W_cᵀ predict/correct; cavity honoured by timescale separation (down-prior = last-tick parent belief, injected before this tick's up-fold — no child sees its own current posterior).
mind DOWN: parent viability → each child's empirical prior, added-alongside / precision-weighted the child's own predictive emptying-B forward prior (NOT the replacement blend that dilutes the one predictive term). δ=0 short-circuits to the exact forward prior (byte-identical).
mind→body execution: muscle_fatigue → Motor.pi loop-gain (fresh 1.0 → spent ~0.35) → weaker servo → degraded aim reafference (genuine consequence for MINING; :motor_cortex-gated, scoped honestly).
mind→body precision/affect: 2-axis interoceptive affect {urgency u = believed critical mass, comfort c = believed high-satiety mass} + organism-viability posterior → Hormones γ/lr dial (gated on the ORGAN, not on l2==nil). Criticality sharpens γ; comfort broadens (u-gated); lr damps with a hard floor > 0. Transient: re-based each tick, stripped before persist.
body→world: emit → MC actuates → world changes → re-sense.

Graded viability + allostasis (the fix for the measured death)

Replace BOTH the drain-setpoint C {2:+3,3:0} AND the reward-smuggling monotone foil {2:+2,3:+4} with a reserve-holding INTERIOR-PEAK C: reserve_ramp(6) = [-8, -3, -1, +1, +2.5, +2] — positive gradient nominal→sated (refill pressure returns the instant belief slips below sated) with surplus(2.0) < sated(2.5) so the argmax is an interior buffer bin, never the ceiling ⇒ bounded, non-hoarding, non-saturable-at-the- edge. Anticipation needs no new EFE term: the existing depth-5 Plan rollout applies emptying-B forward, so a non-eat policy predicts a penalised low-reserve future ⇒ eat-before-edge for free. The L2 parent adds pre-edge demand-shift on top, if it earns it.

Timescales (yuga 4:3:2:1) & work/fatigue

Kali ts=1: felt A (fast). Dvapara ts=2: reserve B. Fatigue on a distinct faster clock (~3 s) vs energy (~8 s) — the first per-factor timescale differentiation. Satya ts=4: the viability parent (slow ⇒ its down-prior is anticipatory). Work/fatigue: graded effort cost in the BODY store (never policy logits), inefficiency rises as the limb tires, recovery energy-rate-limited, fatigue lowers Motor.pi ⇒ work-rest pacing = the allostatic brake the flat scalar lacked. The parent's period does NOT re-derive the default @l2_period (byte-identity).

Staged gate ladder (one cure at a time — each a recorded verdict before the next)

CURE 1 — graded-reserve interior-peak setpoint (ships first). Single organism energy factor, cardinality 4→6 (Designer.transition(:emptying,…) is already ns-generic — no engine change) + a new drive_c(:reserve, no) interior-peak shape keyed by the drive_shape gene (default :setpoint ⇒ byte-identical). Parent OFF, no factoring, no affect, no fatigue — a win is attributable to exactly one change. PASS iff: (a) N≥12 survival ≥11/12 (vs measured ~6/12); (b) allostasis_index = believed reserve at eat-onset, (:reserve − :setpoint) CI-excludes-0 positive (eats earlier); (c) two-ended satiation — fights harder near critical in a scarce world AND stops eating / does not hoard in a rich world; (d) beats BOTH :setpoint and :saturable, survival-count CI excluding each. FALSIFIES/REJECT if survival not improved, OR indistinguishable from :saturable (then it is just "eat more"), OR rests/hoards into starvation.
CURE 2 — per-subsystem factoring (energy_reserve + gut_satiety via gut→energy transfer + soma_integrity). PASS iff a cross-subsystem dissociation Δ CI-excludes-0 (beliefs measurably decouple), with a per-subsystem severed-limb falsifier for EACH factor.
CURE 3 — work↔fatigue limb tier (muscle_fatigue → Motor.pi). Blocker prereq: reindex motor_config by NAME not Enum.take(-5) (mc.ex:152). Per-mechanism ablation arms (C-only / inefficiency- only / pi-only) so any win is attributable.
CURE 4 — affect→precision loop (2-axis urgency/comfort → γ/lr + interoceptive γ_m). Un-gate from L2 by gating on the ORGAN; reset_baseline before apply, strip γ+lr+γ_m before persist; lr floor fenced > 0.
CURE 5 — anticipatory organism-viability parent (directional B^G, an explicit engine addition, NOT the stock smoother). Decisive discriminator: the δ>0 arm must beat δ→0 on a cross-subsystem allostasis_index; if δ→0 matches, keep Cures 1–4 and REJECT the parent.

Invariant guarantees (each a REJECT-on-fail gate, proven per-seam)

Categorical AIF only — standard A/B/C/D/E Dirichlet factors + exact SlowContext marginal filter; q=softmax(prior+Σγ_m·lnA); additive G=ΣG_f = epistemic + pragmatic. No Nx/Rust/NIF/backprop/RL.
Additive + gated + byte-identical — depth behind :homeostat absent from default/0; new C shapes queried only for homeostat factors; new knobs default inert, drawn LAST. Binding fences: do NOT lift @factor_cap globally, do NOT re-derive @l2_period, gate step-path edits on :homeostat in active_organs (NOT l2==nil, which metabolism_l1_phase0 also traverses). Per-seam mad<1e-12 over default AND every touched lineage.
No scalar-per-action — every cost enters via that action's own B column → felt obs → belief, NEVER policy logits; affect γ multiplies ALL policies uniformly (action-independent). Guarded by action-clone-invariance (extended) + a NEW action-history-invariance test (any demand estimate is a fn of a hidden-state posterior only — the "recent-costly-fraction" driver is DELETED).
Monotonic decay / no smuggled reward — info terms use the floored wnorm kernel ⇒ W→0 as counts→∞, C-independent; C is a bounded log-preference (interior-peak, surplus<sated), not an info term; the parent carries NEUTRAL C. lr floor > 0 so criticality can't freeze Hebbian learning.

Claim fence + what is deliberately still simplified (named, not smuggled)

Fence: every reserve/felt/viability/urgency/comfort/fatigue float is a MODEL VARIABLE, never a felt state. Passing a gate demonstrates graded self-maintenance / allostatic pacing as behaviour only — a necessary-not-sufficient substrate with ZERO weight for awareness/consciousness/life. "The arm gets tired" is a limb-ATP/soreness proxy, not sentience; it must NOT be narrated as felt hunger/pain in any stream overlay. Still simplified (honest ceilings): (1) mean-field, not exact joint — the planner can't yet anticipate "a tired arm wastes ATP" (deferred joint conditioning); (2) cavity approximate (timescale separation, not exact leave-one-out); (3) subsystem depth staged across Cures 2–5, parent-of-parents deferred; (4) fatigue→Motor.pi is a world consequence for MINING only (leg fatigue via cost+C until a locomotion servo exists; soma_integrity gated on measured health-channel variance); (5) short horizon may not reach the penalty region (remedy: L2 parent); (6) affect is 2-axis, full emotion vector stays read-only.

Top risks

Anticipation may not separate (structural): Cures 1–4 give raised setpoint + reactive rollout-anticipation; genuine pre-edge allostasis needs Cure 5's directional parent, pre-registered to FAIL its δ→0 discriminator if built as the stock smoother.
Reward-smuggling via C shape or a demand estimate (the one corpus REJECT): a monotone ramp is the saturable foil; a history-conditioned demand estimate is a cross-tick action→preference loop the within-tick clone test is blind to. Mitigations binding (interior-peak surplus<sated + discrimination arm; delete the history driver + action-history-invariance guard).
Fake/thin world loops: a subsystem whose belief drains with no world consequence is a preference-hack — per-subsystem severed-limb falsifiers required (muscle_fatigue, soma_integrity are the thin links).
Byte-identity / cross-lineage perturbation (motor_config positional index, global @factor_cap, @l2_period, l2==nil-gated edits) — organ-gated edits + per-seam mad-check over every touched lineage.
Affect-loop instability / compounding — reset_baseline before apply, strip γ_m before persist, drop :direct per-sub salience, u-gate satiety_attenuate, contraction-bound κ_s, tune interacting brakes jointly.

sha256 87bf9a129d405ecd — 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 is a design document, not code, and it says so twice over: every change here needs a review verdict, a typed specification, a registered paired test and the owner's go-ahead first.

It exists because of a measured failure. The current internal state is a single coarse scalar whose preference is flat once full, so there is no pull toward keeping a reserve. A test measured that shape dying in half of its worlds with food reachable. The shallowness fails, measured, and that is what licenses the deeper design.

The proposal is a three-tier stack behind one opt-in organ: continuous stores in the body, graded per-subsystem beliefs above them, and one slow organism-level parent above those. Every segment of the loop from world to body to mind and back is named.

It is staged into five cures, one at a time, and the last one is committed in advance to be rejected if it cannot beat its own switched-off version.

Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 87bf9a129d405ecd

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 design document with a review verdict of signed-with-changes attached, and it repeats at the top that nothing here is code. Every cure needs a merged verdict, a typed specification, a registered paired test, a ship-gate checklist and the owner's go-ahead first.

The motivation is an owner correction and a measurement. The correction is that fitting a model into a game body is not a licence to reduce its depth. The measurement is more concrete: the current internal state is a single global scalar in a few coarse bins, whose preference is flat once full, so there is no pull toward holding a reserve. A registered test measured that shape dying in half of its worlds with food reachable, and being more scattered than a hoarding foil. The page draws the conclusion carefully: the shallowness fails, measured, and this design is a staged path to depth which does not claim to deliver the whole mandate at once.

The proposal collapses four reviewed areas into one structure behind a single opt-in organ, leaving the existing organ frozen and the default configuration byte-identical. It has three tiers. At the bottom, continuous per-subsystem stores in the body, advanced by real elapsed time, with attribution to whichever subsystem acted. There is an upkeep charge on every tick, so that holding still is not free, which is what makes the death edge bite. In the middle, one interoceptive factor per subsystem, each a standard factor over a six-step gradient from critical to surplus, seeded so it can sense itself. At the top, one slow organism-level parent with deliberately neutral preference, so it can never become a second reward. It acts only through the prior it sends downward, and has a directional transition rather than the stock smoother, because a smoother cannot anticipate.

Every segment of the loop is named and none is left dangling. World to body, then body to mind. Then upward folding with a timescale separation, so no child sees its own current posterior, and downward as a prior added alongside rather than replacing the child's own prediction. Then mind to body through a loop gain that weakens as a limb tires, mind to body again through an affect-to-precision dial, and body back to world.

The preference fix is stated precisely. Both the flat setpoint and the monotone foil are rejected, and replaced with an interior peak in which the most preferred bin is a buffer rather than the ceiling, so the pull is bounded and non-hoarding. The page notes that anticipation needs no new term, because the existing deep rollout already predicts a penalised low-reserve future.

A staged ladder of five cures follows, each with its own registered conditions and each requiring a recorded verdict before the next. The first ships alone so a win is attributable to exactly one change. The last is committed in advance to be rejected if its switched-off version matches it.

The closing sections list invariants that reject on failure, and a stated limit saying every one of these values is a model variable and never a felt state. Then comes an explicit list of what remains simplified, and a ranked list of risks. That list includes the one the reviewers rejected before: smuggling reward back in through the shape of a preference or a history-conditioned estimate.

Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 87bf9a129d405ecd