Ground-Model Brief — the `:metabolism` interoceptive organ (Phase-2 design, NOT applied)

⚠ CORRECTED / SUPERSEDED (2026-07-11). This brief's seeding conclusion is wrong and is kept only as a historical design step (error + refutation both in the evidence chain). §"The Dirichlet seed is automatic and correct … No new seeding code" (lines ~135–139 for A, ~241–243 for B) was refuted by the B1 blocker: Model.new/1 runs norm_cols (model.ex:70–71) before add1 (:84–85), so a column's magnitude is normalised to sum-1 before the +1 seed — you therefore cannot seed a strong Dirichlet concentration by shaping/pre-scaling B; only a mild ~1.x prior survives. A dedicated concentration seam (:pb_seed, κ·norm_col + 1) was required and is the mechanism in the shippable spec docs/specs/metabolism.md §6.1 (κ=1 byte-identical to add1). The shape/direction claims (mass drifts toward empty; :eat refills) remain correct; only "the automatic seed is sufficient / no new code" is withdrawn. Authoritative spec: docs/specs/metabolism.md.

Status: DESIGN ONLY. This is a typed proposal for where and how a :metabolism organ would seam into the existing genome → card → designer → model pipeline. No lib/** is edited by this brief; nothing deploys. It is the structural map the Phase-2 typed-model-diff and /lab-team-review build on.

Scope (owner ruling R3): the FULL metabolism organ = energy + satiety hidden factors, a non-identity emptying/filling B, a setpoint-peaked C, energy-cost-as-C routed through B_energy, and allostasis as a declared setpoint→C map. This brief covers the two structural seams that the rest hangs on: (a) the energy + satiety hidden factors with init_a:diagonal self-sensing A, and (b) the non-identity emptying/filling B via a new per-modality :b_init => :emptying field. The C-shape (setpoint-peaked, allostasis map) and the energy-cost-through-B_energy term are flagged where they attach but are deferred to the typed diff.

0. ClaimFence (binding)

Energy and satiety here are categorical hidden factors over discretised interoceptive bins — Dirichlet- learned A/B tensors and a log-preference C. They are NECESSARY-NOT-SUFFICIENT operational substrate with zero evidential weight for awareness / hunger-as-felt / life. "Satiety", "setpoint", "allostasis" name math objects (a peaked C, a non-identity B), never an experience. No float in this organ (a qs entry, a C weight, a B mass) may be surfaced as a "felt" state. Passing any Phase-2 RED demonstrates the named behaviour (e.g. the agent acts to hold a sensed energy bin near setpoint), never that it feels energy.

1. The pipeline this organ must enter (current, verified)

The expression path is a single funnel; a new organ must add itself at exactly the same four points every prior organ did (:motor_cortex is the most recent precedent and is the model to copy):

Genome.@prereqs / @organs        organ → prereqs ; gate
   │   (genome.ex:19-37)
Genome.@modalities               organ → one-or-more {name,no,factor,ns,[init_a]} rows
   │   (genome.ex:46-102)
Genome.<lineage>/0 growth_plan   the opt-in constructor that lists the organ
   │   (genome.ex:152-191)
Genome.card/1 → Map.take(...)    active modalities → card rows (carries :init_a today)
   │   (genome.ex:207-223, the Map.take at :214)
Designer.compile/1               each card row → one Factors sub (A from init_a, B identity, C, D, γ_m)
   │   (designer.ex:32-58)
Model.new/1                      norm_cols + pA=A*1+1 / pB=B*1+1 Dirichlet seed
       (model.ex:64-101)

MCCodec.encode/2 (mc_codec.ex:23-25) walks Genome.active_modalities/1 in declared order and emits one obs-index list per active modality, so a new modality is fed iff its organ is in the growth plan, and it is fed at the position its @modalities row sits. MC.step/2 (mc.ex:75-119) consumes obs_by_factor positionally — factor i ↔ modality i ↔ obs i. Declared modality order is load-bearing; any new row must respect that.

2. Byte-identity invariant — why the default 12-factor genome is untouched

Genome.default/0 (genome.ex:152-156) lists 8 organs; its prerequisite closure does not include :metabolism. Since active_modalities/1 filters @modalities by organ in plan (genome.ex:226-229), no :metabolism row develops under default/0 — the card, the compiled subs, the Dirichlet seeds, and therefore the live depth-5 Plan path are byte-for-byte unchanged. This is the identical gating :motor_cortex / :sight_cortex already rely on (those organs add 5 / 1 factors only for their opt-in lineages and are absent from default/0).

Two NEW things this organ introduces that did not exist for prior organs, each of which must preserve byte-identity by construction:

1. A new per-modality field :b_init (the non-identity B selector). It MUST be absent from every existing @modalities row and MUST default to identity-B when absent (see §4). Because card/1 uses Map.take(&1, [:name,:no,:ns,:init_a]) (genome.ex:214) — an allow-list — the new key is invisible to the card unless that allow-list is extended; so even the plumbing is inert for existing modalities until deliberately threaded. Flag: extending the Map.take list is the single most byte-sensitive edit in the whole organ; see §4.3.

2. A possible new genome field for an allostasis setpoint knob (deferred to the typed diff). If added, it MUST follow the novelty_gain precedent exactly: a struct default that is the inert value, a Map.put_new back-fill in slow_defaults/1 (genome.ex:339-345) so old serialized DNA never raises, and a mutation draw appended LAST in mutate/2 (genome.ex:288-305) so existing lineages' RNG draw order — and thus their reproducible mutation behaviour — is unchanged. Flag: inserting a draw anywhere but last breaks every existing lineage's determinism.

3. Seam (a) — energy + satiety hidden factors with init_a:diagonal

3.1 @prereqs / @organs (genome.ex:19-37)

Add one row to @prereqs:

metabolism: [:interoception],

:interoception is the always-granted base sense (genome.ex:19 comment; ensure_base/1 at :373 forces it into every plan), so the prereq is always satisfiable and an :metabolism lineage is always developable. @organs = Map.keys(@prereqs) (genome.ex:37) picks it up automatically; depth/1 (genome.ex:386-391) computes depth 1 (one level above interoception) for the repair sort. No change to valid?/1, repair/1, closure/1 — they are organ-agnostic and handle a new @prereqs key for free.

3.2 @modalities rows (genome.ex:46-102)

Add two rows. Place them after the motor-cortex block (after line 101) so they are declared LAST; this keeps every existing factor index — including the motor block's "always-final-5" assumption baked into MC.motor_config/1 (mc.ex:133-138, which does obs |> Enum.take(-5)). Flag (ordering hazard): motor_config/1 takes the last 5 obs as the proprioceptive block. If metabolism rows are appended after the motor rows, then for a genome that has BOTH :motor_cortex and :metabolism, Enum.take(-5) would grab [reach,contact,dig,motion,energy] — wrong. Resolution for the design: Phase-2 lineages are metabolism-primary built from default/0 + :metabolism (no motor cortex), so the two organs never co- occur in a registered Phase-2 lineage and the hazard is dormant. If a later lineage combines them, fix motor_config/1 to select by name/index, not by tail position. This is a documented constraint, recorded here so a future combiner does not trip it.

Proposed rows (energy is self-sensed interoception, so init_a: :diagonal exactly as the motor block):

# METABOLISM (Phase-2, opt-in via :metabolism): interoceptive ENERGY-CHARGE and SATIETY the body
# senses about its OWN internal store. init_a :diagonal — the body senses its own charge level, so
# state k a-priori tends to produce sensed bin k (a weak near-identity A breaking the degenerate
# uniform-A symmetry of a single-modality factor, exactly as the motor block). Absent from default/0
# ⇒ default UNIs stay 12-factor + byte-identical.
# energy: internal charge bin — 0 depleted · 1 low · 2 nominal · 3 charged
%{name: :energy, organ: :metabolism, no: 4, factor: :energy, ns: 4, init_a: :diagonal, b_init: :emptying},
# satiety: digestive fullness bin — 0 empty · 1 peckish · 2 sated · 3 full
%{name: :satiety, organ: :metabolism, no: 4, factor: :satiety, ns: 4, init_a: :diagonal, b_init: :emptying}

Why init_a:diagonal (not uniform). designer.ex:62-82 shows the rule: a single-modality factor with no == ns and uniform A is non-identifiable — q(s) is stuck uniform because every state explains every outcome equally (diagonal_likelihood/1 docstring, designer.ex:68-74). Energy and satiety are proprioceptive/interoceptive (the body senses its OWN store), so the weak 0.6-diagonal prior is exactly right and online learning still refines the true A. This reuses the EXISTING :init_a => :diagonal seam (likelihood(:diagonal, no, ns) at designer.ex:62) with no designer change for the A side.

The Dirichlet seed is automatic and correct. Model.new/1 (model.ex:84) sets pa = Enum.map(a, &add1/1) = A*1.0 + 1.0 per column (model.ex:115), the canonical pA = A*1 + 1 prior. The diagonal A's hi=0.6 / lo=0.133 columns therefore seed pa at 1.6 on the diagonal and 1.133 off — a mild, learnable concentration with the @floor=1.0 pseudocount that the Novelty term's W relies on (novelty.ex @floor). No new seeding code.

3.3 The opt-in lineage constructor (genome.ex:152-191)

Add a constructor mirroring motor_primary/0 (genome.ex:176-180):

@doc """
A METABOLISM-PRIMARY genome (Phase-2, opt-in): the default UNI plus its :metabolism organ — the
interoceptive energy + satiety factors (init_a :diagonal) over which it learns a metabolic generative
model, with a non-identity emptying B (the store drains/fills under action). Develops 2 extra factors;
default UNIs keep the 12-factor shape and are byte-identical. A distinct lineage (its saved brains never
load into a default UNI — reconcile/2 starts them fresh on a factor-count mismatch, mc.ex:546-551).
"""
def metabolism_primary do
  repair(%__MODULE__{
    growth_plan: [:interoception, :chemotaction, :proprioception, :vision, :social_sense,
                  :camera_control, :locomotion, :strategist, :metabolism]
  })
end

reconcile/2 (mc.ex:526-529) + compatible?/2 (mc.ex:546-551) guarantee a metabolism brain (14 factors) and a default brain (12 factors) never cross-load: the factor-count differs ⇒ compatible? is false ⇒ a mismatched load starts fresh. Lineage isolation is automatic, identical to the motor/vision precedent.

3.4 MCCodec.outcome/2 — the body-supplied bins (mc_codec.ex)

The codec needs two new clauses, mirroring the motor block (mc_codec.ex:60-64) which trusts the body to send already-discretised values and just bounds them:

# METABOLISM (Phase-2, opt-in): the body sends its internal energy charge and digestive satiety already
# discretised; the codec bounds each to its cardinality. Reached ONLY for a :metabolism genome (absent
# from active_modalities otherwise), so the default path never hits them.
# energy 0 depleted·1 low·2 nominal·3 charged · satiety 0 empty·1 peckish·2 sated·3 full.
def outcome(:energy, s), do: idx(get(s, "energy", 2), 3)
def outcome(:satiety, s), do: idx(get(s, "satiety", 2), 3)

Default 2 (nominal/sated) so an absent body field reads as a neutral mid-store rather than a crisis. The catch-all outcome(_other, _s), do: 0 (mc_codec.ex:65) already makes this safe even before the clauses are added (an undeveloped metabolism modality is never encoded anyway). Flag (body contract): these energy /satiety keys must be produced by the Node body (the external half of the Markov blanket). The brain-side design is inert until the body emits them; until then the diagonal-A factor would sit at its default bin and learn nothing — correct, fail-safe, but means the body schema is a hard co-requisite for any live RED.

4. Seam (b) — the non-identity emptying/filling B via :b_init => :emptying

4.1 The problem with today's B

Designer.compile/1 builds every factor's B as List.duplicate(identity(mod.ns), nu) (designer.ex:47) — the same identity "states persist" transition for all nu actions. That is correct for exteroceptive factors (the world's state is not changed by the agent's discretised action in a known way), but wrong for a metabolic store: energy must DRAIN over time/effort regardless of action, and FILL on :eat. An identity B means the energy factor can never predict its own depletion, so EFE can never prefer acting to refill — the organ would be a dead sensor. R1's permissibility clause (adding generative STRUCTURE the agent can do EFE over) is exactly satisfied by giving energy a real, non-identity B.

4.2 The new field and its default (byte-identical)

Add an optional per-modality field :b_init. Semantics:

Byte-identity: because :b_init is absent from all current @modalities rows and the builder branches on nil ⇒ identity (the current code path), every existing factor's B is byte-for-byte what it is today. This is the exact :init_a precedent: init_a defaults to nil ⇒ uniform A and existing factors omit it (designer.ex:62-63).

4.3 Where Designer.compile/1 changes (designer.ex:43-58)

Two edits, both additive:

1. Replace the hard-coded B line (designer.ex:47):

   #   b: List.duplicate(identity(mod.ns), nu),          # CURRENT
   b: transition(Map.get(mod, :b_init), mod.ns, nu, card),  # PROPOSED
   

2. Add a private selector beside likelihood/3 (designer.ex:60-63), default-preserving:

   # Transition prior selector. Default (nil) ⇒ today's identity "states persist" B for all nu actions
   # (byte-identical). :emptying ⇒ a non-identity metabolic B: a drain drift on every action, refill on :eat.
   defp transition(nil, ns, nu, _card), do: List.duplicate(identity(ns), nu)
   defp transition(:emptying, ns, nu, card), do: emptying_b(ns, nu, eat_index(card))
   

   eat_index(card) resolves the :eat action's position from card.actions (it is Genome.@actions, genome.ex:109, where :eat is index 4) so the refill column is wired to the right action without a magic constant. emptying_b/3 returns nu column-stochastic ns×ns matrices: the drain action-columns put most mass one bin DOWN (toward depleted) with a sticky remainder; the :eat column puts mass one bin UP (toward charged). Each column is normalised (or fed raw to Model.new, which norm_cols-normalises at model.ex:71 anyway).

   The pB Dirichlet seed is automatic: Model.new/1 sets pb = Enum.map(b, &add1/1) = B*1 + 1 (model.ex:85, :115). So the emptying B seeds a Dirichlet that is concentrated toward draining but still learnable (the +1 floor keeps every transition possible) — energy-cost is encoded as prior dynamics over hidden state, which is exactly the right place for "energy-cost-as-C through B_energy" (R3) to attach: a peaked-low C over the energy factor (§5) + a draining B means staying alive costs expected free energy unless the agent acts to refill. This is the categorical, non-reward formulation of metabolic cost.

3. The Map.take allow-list (genome.ex:214) MUST be extended to carry the new key into the card:

   #   Enum.map(mods, &Map.take(&1, [:name, :no, :ns, :init_a])),                # CURRENT
   Enum.map(mods, &Map.take(&1, [:name, :no, :ns, :init_a, :b_init])),           # PROPOSED
   

   Flag (highest byte-sensitivity edit in the organ): Map.take omits absent keys (genome.ex:213 comment), so adding :b_init to the allow-list is byte-neutral for every modality that does not declare it (the resulting card map is identical — no :b_init key appears). Verified mechanism: Map.take(%{name: :x, no: 4, ns: 4}, [:name,:no,:ns,:init_a,:b_init]) == %{name: :x, no: 4, ns: 4}. So this single line is the only genome.ex change for seam (b) and it is provably inert for the default 12 factors. The byte-identity test (novelty_test.exs anchor 5, mad<1e-12 over Plan. action_values(depth:5,beam:3)) MUST be re-run against default/0 after this edit as the gate.

4.4 Mean-field purity (no cross-factor coupling introduced)

emptying_b is a per-factor, per-action transition exactly like every other B. It does NOT read any other factor's belief, does NOT materialise a joint, and is consumed by the same per-factor Infer/Learn path (Factors.infer_states/2 at factors.ex:83-90, Factors.learn/2 at factors.ex:141-148). The mean-field factorisation q(x)=Π_f q(x_f) (factors.ex:11) is preserved: energy is one more independent sub-engine. Flag: the ONLY way this organ could break mean-field purity is if a later refinement made the energy B conditional on another factor's state (e.g. "drain faster when threat=attacking"). That would couple factors and is OUT of scope — if wanted, it must go through a hierarchical parent (the SlowContext seam, mc.ex:299-389), never a cross-factor B. Recorded as a fence.

5. Where the C-shape and allostasis attach (deferred to the typed diff — located here)

Not built in this brief, but the seams are:

• Setpoint-peaked C. card/1 sets preferences via Curriculum.preference(dna.phase, m.name, m.no) (genome.ex:216 → curriculum.ex:42-45). A new entry in Curriculum.@phase_weights (curriculum.ex:29-39) keyed energy: %{...} / satiety: %{...} would give a peaked C — e.g. energy: %{2 => 3.0, 3 => 1.0, 1 => -2.0, 0 => -8.0} prefers the nominal bin (the setpoint), not monotonically "more". A peaked C over an interoceptive factor IS the allostatic setpoint, expressed in the formalism (a preference, not a script). Flag: the C must be PEAKED, not monotone-increasing — a monotone "charged is always best" C is a preference-hack pseudo-drive (the embodiment persona's standing objection); the setpoint peak is what makes it non-saturable and honest.
• Allostasis as a setpoint→C map. The L2 strategist already rewrites factor C per strategic option via strategist_config/1 (mc.ex:395-438, the by_name map). An allostatic shift ("raise the energy setpoint when foraging is cheap / threat is low") is a per-option C override entry there — energy: energy_setpoint_high vs energy_setpoint_nominal — resolved by modality NAME (mc.ex:420-429), so it is morphology-safe and only fires for genomes that express :energy. This is the declared setpoint→C map R3 asks for, and it reuses an existing, tested seam with no new mechanism.
• Energy-cost-as-C through B_energy. As §4.3 notes, the draining B_energy + a peaked-low-disprefer C means inaction drifts the agent into disprefered energy bins, so EFE's pragmatic term qo·C (the only C-channel; see the math fence) penalises doing nothing — metabolic cost without any scalar-per- action reward term. This is consistent with the action-clone-invariance guard (no per-action scalar enters the policy logits; the cost lives entirely in hidden-state dynamics + outcome preference).

6. Persistence / save byte-contribution (verified clean)

MC.save/2 writes :erlang.term_to_binary({brain.dna, brain.model}) (mc.ex:474-476). The metabolism organ adds:

• DNA: one organ atom in growth_plan (and, if added, the deferred allostasis field). The organ atom is already serialized for every lineage; no new transient state.
• Model: two extra Model subs (energy, satiety) in brain.model.subs. These are ordinary factors — their pa/pb/qs/c round-trip through term_to_binary exactly like every other sub. There is no new gland/setpoint TRANSIENT field to strip: unlike the slow-context (emp_prior/emp_delta/last_lik, stripped in demodulate/2 at mc.ex:278-283 and excluded in adopt/2 at mc.ex:557-564), metabolism state IS the persisted learning (the store's learned A/B and the live qs), so it SHOULD persist. Flag: if the typed diff later adds a transient gland readout (e.g. a per-tick allostatic-load scalar derived from qs), it MUST be added to the demodulate/2 strip list AND the adopt/2 exclusion list (mc.ex:562, the -- [:emp_prior, :emp_delta, :last_lik]) so it contributes ZERO save bytes and never resurrects onto a fresh sub — the exact discipline the slow-context fields follow. The R3 metabolism state itself is NOT transient and correctly persists.

set_phase/2 (mc.ex:450-471) re-expresses only C on a phase change when same_shape? holds; the two new factors are shape-stable across phases (their ns is fixed at 4), so a metabolism brain refreshes its energy/satiety C cleanly on phase advance with no structural reset.

7. Edit-list summary (for the Phase-2 typed diff — NOT applied here)

Mean-field / math-fence audit of the design: no scalar-per-action policy term (cost lives in B_energy dynamics + peaked C, both per-factor); no cross-factor coupling (energy/satiety are independent subs); the Novelty W term still decays monotonically over the new factors' Dirichlet counts (W is C-independent and @floor-bounded — novelty.ex — and the new factors are ordinary Dirichlet subs); default genome byte-identical by organ-gating + Map.take omission + nil-default B. Gate before any Phase-2 code: /lab-team-review MERGED VERDICT + re-run of novelty_test.exs anchor 5 (mad<1e-12 over Plan.action_values(depth:5,beam:3) on default/0) as the byte-identity receipt.