UNI deepening — mission update (research synthesis, 2026-06-23)

Produced by a 14-agent research workflow (wf_15e09eec-2e7): ground the current architecture → research 6 pillars (FEP literature + codebase) → adversarially stress-test each for invariant/honesty violations → synthesise. Then refined by UNI-GPT consult cycles (below). This is the substrate for the ultracode plan.

Mission

Build a literal digital organism on the CPU whose life is the predict-act tick — one mean-field active-inference body that perceives, learns, and acts to minimise free energy at the flow-rate of this universe — deepened into a coherent whole: a lateralised cortex (dual precision/granularity parents) above a shared factor bank, a segmental proprioceptive spine below, and interoceptive organs with glands and endogenous cycles within. Every layer ADDITIVE, GATED behind an opt-in genome organ (default genome byte-identical), and FREE-ENERGY-CONSISTENT (every new term reduces to a recognised FE quantity — pragmatic qo·C, state-epistemic H(qo)−E[H(o|s)], parameter-novelty over Dirichlet counts, or precision γ/γ_m/η; never an ad-hoc reward; no Nx/Rust/NIF/GPU, no backprop/RL/TD).

The immediate scientific target: break the empirical phase-2/3 plateau (one UNI hoards 32 pickaxes, exploration flattens) with the cure the adversarial review converged on — the missing third EFE term (parameter information gain) plus non-saturable interoceptive drives — but FIRST proving the diagnosis (precision-collapse vs curriculum ceiling). Advance the world-map only on falsifiable gates; hold a strict claim-fence (operational measures are NECESSARY-not-SUFFICIENT substrates carrying ZERO evidential weight for awareness/consciousness/life on their own); stop only when counter-science says the map must be updated.

The deepened architecture

• Decision core (unchanged invariant). Per-factor categorical AIF; live decider is SP.Brain.Plan (depth-5 beam), so every new EFE term threads into Plan first, efe.ex as the mirror.
• Cortex above — lateralisation (:hemispheres). Two SlowContext parents (left = fine W_c/high γ, right = coarse W_c/low γ) over the same child bank; DOWN blend normalize(L^wL·R^wR) into the existing δ-weighted emp_prior; callosal {wL,wR}=softmax(−β·[ΣF_left,ΣF_right]). The right hemisphere broadens via its WIDER prior raising H(qo) — not a salience-sign flip.
• Spine below — segmental motor (:spine, prereq :motor_cortex). MotorControl's servo generalised into a stack of SP.Brain.Motor segments (down = set-point, up = residual PE, intra-tick fold). Per-segment precision Π = variable impedance (stiff/committed vs compliant/exploratory) from Hormones/Strategist; Π_distal_relax injects structured motor variability through the UNCHANGED epistemic term, never a value bonus.
• Organs within — interoception + homeostasis (:metabolism/:viscera, prereq :interoception). init_a:diagonal self-sensing factors (energy/satiety/hydration) with a NON-IDENTITY emptying/filling B (the one genuinely new generative object), C peaked at SETPOINT (prefer 'ok', not maximum). Seeded B protected from Hebbian erosion by a per-modality learn_b=false override + compile-time reachability assertion. Allostasis falls out of the depth-5 planner rolling the emptying-B forward; energy bridges SP.Lab.Bioenergetics.
• Glands + cycles (:endocrine, prereq :interoception). Persistent SP.Brain.Endocrine state on the MC struct (advanced per tick, stripped in demodulate): (1) satiety attenuates ONLY positive C on a WHITELIST [:inventory,:build] (multiplicative shrink, never sign-flip, never status/threat/self); (2) endogenous oscillators retune γ_m/η (ACh/DA-as-precision). Circadian entrainment CONDITIONAL on a real sensed light clock, else free-running ultradian.
• Standing active-learning drive — novelty (:curiosity, novelty_gain default 0.0). The missing parameter-information-gain term W from existing Dirichlet counts, per-factor, threaded into Plan + efe, γ-bounded. The load-bearing plateau cure: nonzero in under-sampled (state,action) cells, decays to 0 as counts accumulate (the no-reward guarantee).
• Cycle / life-frequency. The two-timescale OODA is the metabolic rhythm; the 2nd hemisphere parent + 2nd EFE pass run at L2 cadence so depth-5 stays under the body tick (a wall-clock gate guards "pure Elixir at the universe's tick frequency").

Roadmap (each phase additive, gated, with a falsifiable RED gate)

• Phase 0 — DIAGNOSE the plateau (instrumentation only). Measure on a frozen plateau whether γ is saturated high and H(qo) collapsed (⇒ precision/EFE deficit ⇒ proceed) vs a curriculum ceiling (phase_goal_met?(3) needs wood≥8 ∧ tools≥1 ⇒ fix that first). No cure ships against a misdiagnosis.
• Phase 1 — NOVELTY term W (:curiosity) — the primary plateau cure. Gate: control vs treatment, N≥20 seeds — lower hoard, higher distinct-cell visitation, faster phase-3/4 advance (p<0.05), W monotonic-decay (the no-smuggled-reward check), byte-identity at novelty_gain=0 over the depth-5 path.
• Phase 2 — INTEROCEPTIVE ORGANS + emptying-B (:metabolism). Gate: sustained action entropy after first tool (control collapses), interoceptive level oscillates around setpoint (limit-cycle, not flatline), no new saturated attractor, depth-5 forages at higher energy than depth-1 (allostasis), viability ≥ control.
• Phase 3 — SPINE (:spine). Gate A: a 2-segment Π=1.0 spine is byte-identical to today's MotorControl (MAD<1e-12). Gate B: spine_distal_relax>0 yields more distal-outcome entropy AND a new config at strike-rate ≥ single-servo AND viability ≥ control, beating the motor_shuffle noise control. Held at "could supply variability" until a build-trajectory change is also measured.
• Phase 4 — GLANDS + CYCLES (:endocrine). 2×2 ablation (satiety × oscillator). Satiety-on: peak pickaxes bounded (≤~3 vs 32) + phase-3 entry + survival not degraded. Oscillator-on (only if a sensed clock exists): circadian phase cross-correlates with the day period.
• Phase 5 — HEMISPHERES (:hemispheres). H1 byte-identical at δ=0; H2 dual-regime present (γ_right<γ_left, H(qo_right)>H(qo_left)); H3 lateralised genome explores more + reaches phase-3 faster; the SYMMETRIC-DUPLICATE control must NOT show H3 (proving asymmetry, not parameter count, is the cause).

Invariants preserved

No Nx/Rust/NIF/GPU, no backprop/RL/TD; every term a recognised FE quantity (no scalar-per-action term — guarded by a test). Additive+gated, each phase's byte-identity gate run through the DEPTH-5 Plan path. Graded-on couplings default 0.0. Mean-field purity / clean Markov blanket (sufficient statistics only). Determinism + persistence (transient fields stripped on demodulate; gland/setpoint state contributes ZERO bytes to the saved model; new mutate draws appended LAST). Oracle parity. Live-stream go-ahead before any organ on the live colony.

Falsification signals (when to update the map)

• Phase-0 H0 fails (γ not saturated, H(qo) not collapsed) ⇒ the whole incomplete-EFE diagnosis is wrong; re-root in curriculum/structure-learning.
• Novelty fails (treatment still hoards) ⇒ parameter-info-gain is not the missing driver.
• W does not decay monotonically ⇒ it is smuggled reward; revert (no-RL breach).
• Homeostasis: entropy also collapses, or level flatlines, or depth-5 no earlier than depth-1 ⇒ no standing gradient / no allostasis.
• Spine: relaxation raises entropy but degrades strike-rate/viability, or fails to beat the shuffle control ⇒ noise, not exploration.
• Glands: satiety swept [0.1..1.0] still hoards ⇒ not preference-saturation; or degrades survival ⇒ λ mis-scoped.
• Hemispheres: the symmetric-duplicate control matches the asymmetric ⇒ it is parameters, not lateralisation.
• Any default/coupling-0 run differs from HEAD over the depth-5 path ⇒ not additive+gated; revert.
• Deepened decide step falls below the body/universe tick cadence ⇒ life-frequency invariant violated.
• An acting agent does NO better than an action-severed twin at staying viable ⇒ the self-maintenance ("life") operationalisation is unsupported — say so, drop the framing.

Awareness / life measures (operational, falsifiable — strict claim-fence)

Binding fence: these are behavioural/organisational measures. A non-identity emptying-B + setpoint C is a thermostat; novelty W is active learning; a dual-precision parent is a coupling mechanism. Each is NECESSARY-NOT-SUFFICIENT and carries ZERO evidential weight for experience. Passing a gate demonstrates the named behaviour, NEVER experience. Do not surface gland floats as felt states.

• Action-dependent viability persistence (the "life" axis): ticks-inside-the-viable-set(acting) − ticks-inside-V(action-severed twin) > 0 (p<0.05). Self-maintenance, not life-as-experience. Drop "autopoietic / non-organic life" unless NESS is actually instrumented (it is not).
• Allostasis signature: inferred setpoint moves BEFORE depletion; depth-5 forages at higher energy than depth-1.
• Active-learning / curiosity: total novelty W as Awareness.broadcast.novelty_drive (functional-access only) — must correlate with subsequent exploration and FALL at insight. Measured info-gain, NOT felt curiosity.
• State-visitation entropy (the plateau-break metric) + bounded duplicate-tool count.
• Dual-regime presence (mechanism check only): γ_right<γ_left, H(qo_right)>H(qo_left). Nothing more.

Open questions taken to the UNI-GPT (consult cycles)

1. Novelty FORM: W = −½·Σ(qo·qs)(1/pa − 1/Σpa) (pymdp/SPM pA-novelty, no digamma) vs the digamma E[lnA] form — which for a depth-5 beam planner, same γ/sign as the state-epistemic term? Does transition-novelty W_b share that γ?
2. Allostasis as setpoint→C: SlowContext is a scene→state PRIOR, not a setpoint→C generator. Is mapping an inferred energy-setpoint posterior to a log-preference vector FE-correct, or must it be a hyperprior over C with its own VFE term? What stops it being disguised reward?
3. Callosal arbitration: is softmax(−β·[ΣF_left,ΣF_right]) a legitimate precision-on-precision allocation, and does β=0 reduce to recognised model-averaging? Is left-dominance FE-justified or ad-hoc?
4. Coarse-W_c granularity: the principled deterministic coarsening (state-merge vs column-smoothing) for a genuinely "global" right regime, heritable + oracle-mirrorable.
5. Emptying-B + learning: freeze learn_b on viscera columns, or a strong Dirichlet prior Hebbian may refine but not erase? FE-correct way to keep depletion dynamics stable yet adaptable.
6. Energy-cost as C: confirm "mining costs energy" as qo_energy·C shifted THROUGH B_energy is FE-consistent (not per-action reward) + how to test no scalar-per-action term leaked into the logits.
7. Diagnosis priority: how to cleanly separate precision-collapse from curriculum ceiling in H0 — is saturated γ + low H(qo) sufficient evidence, or could a ceiling present that way too?
8. Stacking order: with hemispheres (3rd temporal level) above L1/L2 and a continuous spine below, is the mean-field factorisation across FOUR effective levels still FE-sound? Does running the 2nd pass at L2 cadence break timescale-separation assumptions?

Execution status (2026-06-23)

• P0 — DONE. lib/sp/brain/diagnose.ex shadow-EFE audit → H0 = epistemic_starvation (see Q1 result).
• P1 — built + offline-validated; LIVE RED PARTIAL (verdict 2026-06-24, ~9 h, N=3/arm). Mechanism: lib/sp/brain/novelty.ex (W_a + W_b, pymdp pA/pB-novelty, prior-floor-bounded, gated novelty_gain 0.0 default). Offline ✓: byte-identical at 0 over depth-5 (mad<1e-12, suite 0/0); monotonic-decay/C-independent /bounded; fresh curiosity agent explores ~uniformly (entropy 2.29 vs 1.91, build/craft ~3×). LIVE paired RED (uni-colony-curiosity on localhost/uni-colony:v4, owner-approved; kin 10 gain=0.5 vs kin 11 gain=0, same code/world/bodies; launcher runs/curiosity_lineage.exs; collector runs/cur_red.log):
  • HOARD gate PASS (direction only; magnitude CORRECTED 2026-07-11): at the settled cycle of the sole committed receipt (docs/receipts/phase1_curiosity_red.log, ~110 min / 11 cycles), curiosity Σpickaxes = 10 (one residual hoarder, UNI-10-2 at 10) vs control Σ = 45 (two runaway hoarders, UNI-11-1 at 24 + UNI-11-3 at 21) — ≈4.5× fewer; the runaway pickaxe attractor reproduced in control and was suppressed, not eliminated, in curiosity. The earlier "Σ = 1 (max 1) vs 25 / 25× / one UNI at 24" figures cherry-picked a single control hoarder, ignored the second control hoarder and the residual treatment hoarder, and have no committed receipt; they are withdrawn (docs/receipts/phase1_curiosity_red_CORRECTION.md). The "9 h" timepoints in this block (incl. the −0.07 entropy figure below) likewise have no committed receipt — only the ~110-min log is committed.
  • Behavioral plateau-break (stone/building) FAIL: neither arm reached cobblestone (Δphase = 0, both at 3.67). Distinct-cells Δ = −285 (control higher, an artifact of +40% more ticks). Action-entropy advantage decayed monotonically (offline +0.38 → t0 +0.35 → 90 min +0.10 → 9h −0.07) — exactly the math's W→0 as counts→∞ (no-reward guarantee preserved), but novelty's behavioural push wanes once the model is well-fit. So the cure is on the SUB-claim (hoard), NOT on the FULL claim (plateau-break).
  • Scientific reading: the plateau is MULTI-CAUSAL. Phase-0 diagnosis (epistemic_starvation) was real; novelty cures the hoarding manifestation; but the agent has NO standing drive that wants stone (phase-3 C is wood=+5,tools=+12, nothing toward stone/shelter). A non-saturable interoceptive drive — Phase 2 (organs/emptying-B/setpoint-C, the OTHER GPT-signed cure) — is the load-bearing missing piece. This is the registered falsification path working as designed: half-falsification recorded as PARTIAL, Phase 1 sound but insufficient alone. Do NOT spin "hoard suppression" as a plateau-break.
• P2–P5 — the approved roadmap (organs → spine → glands → hemispheres), each gated/FE-signed, on reach.

UNI-GPT consult rulings

Consult run 2026-06-23 on the UNI Active Inference Guide GPT (thread …/c/6a3adc3d). All 5 FE-consistency questions answered (sent one at a time; the GPT reasons ~1–3 min before each answer).

Q1 — diagnosis priority — SIGN-WITH-CHANGES (a sharpening of Phase 0).

• "Saturated γ + collapsed H(qo)" is NOT sufficient. Both failure modes produce a sharp policy posterior over a narrow predicted-outcome trajectory — Q(π)=σ(lnE−γG−F), G=Σ_τ H·s + o·(ln o − ln C) — so γ/H collapse is a shared EFFECT, not a cause.
• The cleanest discriminator: a SHADOW / counterfactual EFE audit over the same frozen-seed tick — do NOT change the live action. Log two policy rankings: argmin_π G under the ACTUAL gate vs argmin_π G under an UNMASKED / prerequisite-repaired curriculum gate (shadow-relax the curriculum gate, or make wood/build policies available in the shadow planner). Telemetry per tick: {live: {winner, efe:{epistemic,pragmatic, total}, h_qo_by_factor, distinct_sa_cells, active_goal_atoms, policy_mask_size}, shadow: {winner, efe…, distinct_sa_projection, reachable_prereq_policy}, discriminator: {same_winner, wood_policy_rank_live, wood_policy_rank_shadow, gate_delta}}.
  • Precision-collapse signature: the SAME hoarding policy wins even under the shadow-relaxed gate; the EFE gap stays dominated by high-γ pragmatic value for hoarding; epistemic terms near zero; distinct-cell growth flat outside the hoarding loop ⇒ a decider-level attractor / precision-epistemic starvation ⇒ proceed to the novelty cure (Phase 1).
  • Curriculum-ceiling signature: under the shadow gate that reintroduces wood<8 as an active unmet prerequisite, wood/stone/build policies become competitive or win ⇒ fix the curriculum gate first.
• Verbatim: "The decisive variable is not H(qo) alone. It is whether the counterfactual policy ranking changes when the curriculum gate / prerequisite goal generator is shadow-corrected."
• Action: Phase 0 in the roadmap is upgraded — add the shadow counterfactual-EFE audit (read-only, live action unchanged) as the H0 discriminator, not just γ + H(qo) logging.

Q1 — H0 RESULT (2026-06-23, lib/sp/brain/diagnose.ex on the live hoarders' .bins UNI-0-1 / UNI-2-1): epistemic_starvation — PROCEED to Phase 1. The audit ruled OUT both alternatives:

• NOT classic γ-runaway: policy γ = 7.8 / 7.9 (unsaturated; max 16).
• NOT a simple curriculum-C ceiling: re-injecting the phase-1 wood-C (has_wood=+8) on the inventory factor did NOT flip the winner or lift mining's rank (winner_changed: false on both).
• IS the precision/epistemic-deficit signature: epistemic/pragmatic ratio = 0.01 / 0.006 (the information drive is ~100–170× weaker than the pragmatic term), the EFE landscape is nearly flat (value spread 1.5 / 0.6 over a 62 / 135 baseline) so the agent is indifferent and defaults to :noop / :eat (matching the observed idle hoarding). ⇒ the missing parameter-information-gain (novelty) term is the FE-correct cure. (Side-finding: UNI-0-1's inventory factor grew to 5 states via structure learning and is frozen in the new state; UNI-2-1's inventory belief is uniform — both degenerate, worth a Phase-2 look, but not the primary plateau driver.)

Q2 — the novelty / parameter-information-gain term — SIGN-WITH-CHANGES.

• (a) SIGN-WITH-CHANGES: use the pymdp/SPM pA-novelty approximation, NOT digamma E[lnA]. The −½ in W is the EFE/G sign; in neg_efe ADD the positive ½·Σ_{o,s} qo·qs·(1/a − 1/Σa), scaled by γ.
• (b) SIGN: add W_b for transition Dirichlet counts under the SAME γ, strictly per-factor/action from that factor's own pb, then sum.
• (c) SIGN: W must decay monotonically to 0 as the relevant counts → ∞, and be independent of C. That is the no-smuggled-reward invariant (the Phase-1 monotonic-decay gate).

Q3 — energy-cost as C, not reward — SIGN.

• (a) SIGN: FE-consistent — energy cost is an interoceptive outcome preference under C_energy, reached through action-conditioned B_energy. Mining is costly only because it PREDICTS depleted-energy outcomes, never because :mine gets a scalar penalty.
• (b) SIGN — the test is "action-clone invariance": clone :idle_a/:idle_b with identical A/B/C/D/E (incl. energy) ⇒ identical policy logits; set action_cost[:idle_b]=999 ⇒ logits unchanged; change only B_energy[:mine] to debit energy ⇒ only the predicted qo_energy·C_energy term may change. Pass ⇒ logits depend on predicted OUTCOMES, never action ID. (This becomes a required Phase-2 test.)

Q4 — allostasis as setpoint→C — SIGN-WITH-CHANGES.

• FE-correct as a DECLARED generative-model mapping: q(setpoint_context) → f_setpoint → normalized C_energy; the planner uses only G_energy = qo_energy·ln C_energy. A full hyperprior over C is cleaner Bayesian hierarchy but NOT required — a pure context-conditioned function is acceptable empirical-Bayes.
• Guardrails (no disguised reward): f_setpoint must be action-independent, normalized, fixed BEFORE evaluating candidate policies, and enter logits ONLY through predicted qo_energy. No action scalar, no "forage bonus," no policy-indexed C. Test: clone two policies with identical predicted qo_energy ⇒ logits match even if one is named :forage.

Q5 — protecting the seeded emptying-B — CHANGE from the research proposal.

• Prefer option (ii): a STRONG Dirichlet prior (evidence may refine but not erase), NOT freezing learn_b. Freezing is only correct if depletion/filling is treated as HARD physiology, not learnable dynamics. Strength: each viscera-column concentration LARGER than the expected lifetime evidence — practically 10–100× lifetime updates for adaptation without erasure. (Supersedes the synthesis's learn_b=false viscera override; use a strong prior instead, with learn_b=false reserved for any column declared hard-physiology.)