Execution Plan — UNI Deepening: break the plateau, then grow organs/spine/glands/hemispheres

Durable in-repo copy of the execution plan (the live working copy is ~/.claude/plans/prancy-launching-teapot.md, which is session-local; THIS file is the version that travels with the repo). Read alongside CLAUDE.md, docs/LAB_PROTOCOL.md, docs/UNI_MISSION_DEEPENING.md (mission + 5 signed consults + verdicts), and docs/lab_team/.

CURRENT STATUS (2026-06-24) — read this first

• P0 — DONE. Shadow-EFE audit (lib/sp/brain/diagnose.ex + runs/probe_plateau.exs) on the real hoarders' .bins. H0 = epistemic_starvation (NOT γ-runaway, NOT curriculum-ceiling; epistemic/ pragmatic ratio ~0.01, flat EFE landscape). → proceeded to P1.
• P1 — DONE offline; LIVE RED = PARTIAL. Novelty term (lib/sp/brain/novelty.ex, gated novelty_gain default 0.0). Offline ✓ (byte-identical at 0 mad<1e-12, monotonic-decay, C-independent, bounded; exploration entropy 2.29 vs 1.91). LIVE paired RED (kin-10 gain 0.5 vs kin-11 gain 0): HOARD gate PASS (direction); magnitude CORRECTED 2026-07-11 — the committed receipt (docs/receipts/phase1_curiosity_red.log, ~110 min / 11 cycles) shows at the settled cycle treatment 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 The earlier "9 h / Σ=1 vs 25 / 25× / max 1" figures have no committed receipt and are withdrawn (see docs/receipts/phase1_curiosity_red_CORRECTION.md). PLATEAU-BREAK gate FAIL (no cobble either arm, phase tied 3.67; action-entropy advantage decayed to 0 as W→0 — the no-reward guarantee). PARTIAL: novelty suppresses the hoard manifestation, NOT the standing-drive deficit. Verdict + receipts in docs/UNI_MISSION_DEEPENING.md.
• Lab-team review of the P1 verdict: MERGED = SIGN-WITH-CHANGES (math SIGN; arch SIGN-WITH-CHANGES; experiment SIGN-WITH-CHANGES; embodiment SIGN). 4 REQUIRED ARTIFACTS before Phase 2 code:
  1. Typed model spec docs/specs/novelty.md (StateSpace/Observation/Action/Preference/Policy/Learning/ Precision/ValidationAnchors/ClaimFence) — the template all Phase-2 organs inherit.
  2. Hardened collector — systemd unit or harness-watchdog (the P1 bash collector died at 100 min); N ≥ 6 per arm for Phase 2.
  3. Better behavioural metric — replace the perverse phase_goal_met?(3) = wood≥8 ∧ tools≥1 (which hoarding itself satisfies) with distinct-resource-types-touched / placed_blocks > 0.
  4. Phase-2 organ design goes through /lab-team-review BEFORE any code — no exceptions.
• Artifacts 1–4 PRODUCED (2026-06-24, workflow wf_97fde3a9-83c): docs/specs/novelty.md (typed template), docs/specs/metric_plateau_break.md (placed/used-blocks>0 + distinct-types, RCON-authoritative, no-compromise), docs/specs/collector.md + ops/phase2_collector/* (defense-in-depth: systemd user-timer
  • podman sidecar + harness task + uni-lab-79740c vantage, N≥6/arm), and docs/specs/phase2_metabolism_packet.md (the full-organ Phase-2 design). The packet went through /lab-team-review.
• Phase-2 packet verdict: MERGED = SIGN-WITH-CHANGES (math/arch/exp SIGN-WITH-CHANGES; embodiment REJECT elevated to blocking changes). Receipt = packet §9. 10 BLOCKING changes before any Phase-2 code, two code-confirmed by the orchestrator: (B1) the strong-Dirichlet seed mechanism is NON-FUNCTIONAL — model.ex:70-71 norm_cols runs before :84-85 add1, so a real typed concentration seam is required; (B2) the live bridge.ex has NO viability edge (metabolize/Viability/shutdown are Sim/Eval-only), so :energy must bind to the live MC food/health channel (or a homeostatic-death coupling) + an action-severed- twin gate, or the "life" framing is struck. Plus: V6 action-clone test must be authored (G0 BLOCKED-PENDING-V6); limit-cycle (G2/G4) derived-or-reclassified; satiety→C whitelist; :b_init atomic two-edit; learn_b expressibility; gate numerals + BASELINE_WOOD pre-registered.
• Repair pass DONE + re-verified (2026-06-24). docs/specs/metabolism.md authored (the shippable typed spec closing all 10 blockers at the DESIGN level; owner B2=Both + B4=derived baked in); packet §10 = the repairs map. B4 limit-cycle/allostasis DERIVED with a runnable receipt runs/phase2_homeostat_demo.exs (limit cycle robust; allostasis TUNED, work_bonus≳4). Re-verify (wf_352db2b5-946, math/arch/red/embodiment
  • merge): DESIGN-COMPLETE-SIGN (no blocker left open; receipt re-reproduced) — recorded in metabolism.md §15.
• CODE PASS — STARTED (owner: "1 then 2"). Pure tests DONE + green on HEAD: V6 action-clone-invariance (test/sp/brain/action_clone_invariance_test.exs — G0's V6 condition met, no per-action scalar) + V1 byte-identity golden (decider_byte_identity_test.exs + test/fixtures/decider_golden_seed7_d5b3.bin, frozen HEAD reference). Full brain suite 281/0. No lib/** changed (additive test files only).
• ORGAN 2a DONE + VERIFIED (generative structure, additive + gated). Implemented the SEAMS: model.ex :pb_seed concentration (κ=1.0 byte-identical to add1, κ=50 strong prior) + factors.ex thread + designer.ex :b_init emptying-B / per-modality :learn_b / pb_seed + genome.ex @prereqs metabolism
  • :energy/:satiety modalities (LAST) + card/1 Map.take widen + metabolism_primary/0 + curriculum.ex setpoint-peaked C. V1 byte-identity holds AGAINST THE GOLDEN after the designer B refactor (B6 gate ✓); V6 green ⇒ G0's two conditions both met offline. Organ compiles to a 14-factor model: V3 (emptying-B non-identity/drains, :eat refills), V4 (C peaked at 'ok'), strong pb_seed, arm-integrity (14 vs 12). Full brain suite 285/0 (test/sp/brain/{action_clone_invariance,decider_byte_identity,metabolism_organ}_test.exs). Default genome byte-identical; nothing regressed.
• ORGAN 2b CORE DONE + VERIFIED (the live viability edge — the load-bearing plateau-break mechanism). lib/sp/brain/metabolism.ex (pure dynamics: upkeep drain every tick / costly-action work / :eat refills ONLY with food / empty=death) + bridge.ex live coupling (inject the synthesized :energy/:satiety observation → decide → advance the store → empty persists memory + stops the GenServer = death, Port closes)
  • mc_codec.ex :energy/:satiety clauses. GATED on the :metabolism genome ⇒ the default live decide path is byte-identical (the metabolic branch is never taken). metabolism_test.exs proves the G5b MECHANISM (an all-:noop twin drains to death; a forage-and-eat agent sustains). Clean compile (--warnings-as-errors); full brain suite 291/0.
• B3 DONE + VERIFIED — the metabolism organ is CODE-COMPLETE. satiety→C attenuation: when satiety is believed high, the appetitive POSITIVE lobe of the energy/satiety C is down-weighted (Metabolism.attenuate_model, multiplier ∈ [0,1]); wired into mc.ex modulate/4 (before policy eval), stripped by demodulate (zero saved bytes), and a no-op ⇒ byte-identical for any genome without a satiety factor (default strategist genomes unchanged). BLACKLIST (self/social/status/threat) + depletion penalties never touched. Tests: V9 (×4), V5 (strong-seed decay), V8 (per-modality learn_b). V7 ⊆ V6; V10 (no double-count) holds by the orthogonal-store design (:energy is an internal ATP store, NOT the world food bar status reads). Full brain suite 297/0.
• OFFLINE RED PRE-CHECK built (runs/phase2_red_sim.exs) — found + fixed a real bug, more tuning needed. The real agent (full engine + the bridge metabolic loop) in a synthetic benign world initially COLLAPSED into an eat-every-tick attractor (G2/G3 FAIL). Diagnosis: the seeded emptying-B drained ~1 bin/tick in the planner while the actual store drains ~1 bin/6 ticks, so the depth-5 planner believed empty was imminent at ok and over-ate to buffer. FIX (designer.ex): gentle drain (0.85 STAY / 0.15 down) matching the store rate. Post-fix: the generative model is CORRECT (rests at full/ok, eats at low/empty), the agent OSCILLATES (G2 ≥2 cycles ✓) + SURVIVES while the :noop twin DIES at tick 25 (G5b ✓). STILL MARGINAL: it over-maintains (eats often, amplitude ~0.6<1.0 bin) and G4 shows no depth-1-vs-depth-5 separation — residual habit-snowball + double eat-pull (energy+satiety). Suite 297/0; default byte-identical.
• OFFLINE TUNING DONE — RED-ready, G0/G1/G2/G3/G5b PASS; G4 honestly not-offline-separable. Diagnosis via instrumented closed loop: TWO confounders, not one — (1) uniform cold-start belief → defensive eat at t1, (2) habit prior snowballs :eat. ONE cure (exact analogue of the existing :noop exclusion at mc.ex:289): for a :metabolism genome, :eat is excluded from habit bumps — eating is need-driven, not a tendency that should snowball into a reflex. Result: action entropy 0.04 → 2.05 (matches the control's 1.97 ⇒ diverse exploration), :eat count 149 → 35, energy amplitude 1.94 bins (G2 ≥1 ✓, 68 direction reversals ≥2 ✓), :noop twin still dies at tick 25 (G5b ✓). G4 (depth-5 forages earlier than depth-1) does NOT separate in the synthetic world — and that's correct, not a tuning bug: with the setpoint-peaked C, eating at ok is suboptimal (pushes to full C=0 from ok C=+3), and the cliff at empty is 14+ ticks away (beyond depth-5). G4 separation requires a competing pragmatic pull (e.g. phase-3 inventory tools=+12) — exactly what the live MC world provides; per spec §12 the live depth-5 beam EFE on the real factor is G4's actual gate. Suite 297/0; default byte-identical.
• NEXT = the live paired RED. With owner go-ahead + live-stream guard: pin the gate numerals (offline- derived: G2 ≥1-bin amp, ≥2 cycles; G5b twin <30-tick death) + BASELINE_WOOD, deploy the hardened collector (rootless uni + os_file_read readback), run metabolism_primary vs default/0, N≥6/arm, pre-registered G0–G6 (metabolism.md §11). One cure at a time (P1 PARTIAL stands).
• LIVE PAIRED RED — RAN (~15 days); verdict COMMITTED + adversarially verified (2026-07-11). kin-12 metabolism vs kin-13 default, N=6/arm, 26,244 rows. Receipt: docs/receipts/phase2_metabolism_red.md (raw .jsonl.gz committed). Split verdict: (1) G6 plateau-break = FAIL (0/12 UNIs reached cobblestone/ shelter; treatment did not exceed control, placed 72 vs 83). (2) metabolism hypothesis = WITHHELD — arms statistically indistinguishable at N=6 (diff CI≈[−14.3,+10.6]) AND organ activation UNVERIFIED (G5b un-passed, no energy-posterior receipt). The earlier "homeostat / epistemic-starvation" reading is struck (the organ-free control froze identically ⇒ a shared world/bin ceiling; novelty was off in both arms). Honest predicate for the next cure: "G6 not demonstrated AND metabolism activation unverified" — NOT "metabolism failed." This does NOT by itself validate the Track-B reframe; it also flags a possible world/observation-bin ceiling needing a world-ceiling control. Full detail: metabolism.md §16.

Context (why)

The motor-inference hierarchy is shipped, live, and learning. The live baseline exposed the wall: both colonies plateau at the "make a tool" phase and collapse into one attractor (a UNI hoards 32 pickaxes, no stone, no building). North star: literal digital life with measurable awareness and full human ability within this body/world — deeper/wider minds, hemispheres, spine, organs + glands with internal cycles. A 14-agent FEP research pass + 5 signed UNI-GPT consults converged on: the plateau is a missing third EFE term (parameter information gain) PLUS non-saturable interoceptive drives, and we prove the diagnosis before applying the cure. Full research + rulings: docs/UNI_MISSION_DEEPENING.md.

Invariants (guardrails — every phase)

• No Nx/Rust/NIF/GPU, no backprop, no RL/TD/reward-on-policy. Every new term is a recognised FE quantity: pragmatic qo·C, state-epistemic H(qo)−E[H(o|s)], parameter-novelty W from Dirichlet counts, or precision γ/γ_m/η. No scalar-per-action term in the policy logits (guarded by the action-clone test).
• Additive + GATED behind an opt-in genome organ/field absent from default/0 ⇒ default genome byte-identical. Each phase ships its byte-identity gate through the depth-5 Plan path (the live decider), not just the depth-1 efe fallback.
• Graded-on couplings default 0.0 (novelty_gain, future endo_coupling) — present-but-zero ⇒ flat engine.
• Mean-field / clean Markov blanket; determinism + persistence (transient fields stripped on demodulate; new heritable traits back-filled via slow_defaults Map.put_new; new mutate/2 Det draws appended LAST); oracle parity.
• Live-stream go-ahead before any organ runs on the live colony; every cure proven offline first.

Phase 0 — DIAGNOSE the plateau (DONE)

Shadow counterfactual-EFE audit (UNI-GPT Q1): is it precision-collapse (γ saturated, H(qo)→0) or a curriculum ceiling (phase_goal_met?(3) needs wood≥8 ∧ tools≥1)? γ + H(qo) alone are NOT sufficient; the discriminator is whether the counterfactual policy ranking changes under a shadow-corrected curriculum gate. Mechanism: lib/sp/brain/diagnose.ex (read-only) + runs/probe_plateau.exs. Result: H0 = epistemic_starvation (epistemic/pragmatic ratio ~0.01, flat landscape, idles at :noop/:eat).

Phase 1 — NOVELTY term W (DONE offline; LIVE = PARTIAL)

The missing third EFE term — expected information gain about the A Dirichlet parameters (UNI-GPT Q2, pymdp/SPM pA-novelty, NOT digamma). W_a in neg_efe adds +½·Σ qo·qs·(1/pa − 1/Σpa); W_b over pb under the same γ, per-factor. Counts floored at the prior pseudocount (bounded — cannot swamp survival). Integration: plan.ex advance/3 (LIVE depth-5) + efe.ex step_value/3 (mirror). Gated novelty_gain (heritable, default 0.0). curiosity_primary/0 = the test lineage. Verdict: PARTIAL — hoard PASS, plateau-break FAIL (see CURRENT STATUS).

Phases 2–5 — the deepening (staged; each gated, FE-signed; each goes through /lab-team-review first)

• Phase 2 — INTEROCEPTIVE ORGANS + emptying-B (organ :metabolism, prereq :interoception) — THE load-bearing plateau-break cure. New init_a:diagonal energy/satiety factors; non-identity emptying/ filling B via a :b_init => :emptying modality field (default nil ⇒ identity ⇒ byte-identical), as a STRONG Dirichlet prior (10–100× lifetime, NOT freeze — UNI-GPT Q5). C setpoint-peaked ('ok'; flat/neg at 'full'). Energy-cost = qo_energy·C through B_energy, never a per-action scalar (UNI-GPT Q3). Allostasis = a declared f_setpoint→C map, action-independent, fixed before policy eval (UNI-GPT Q4). Gates: action entropy ≥ floor after first tool; interoceptive level OSCILLATES around setpoint (limit- cycle); depth-5 forages at higher energy than depth-1; action-clone-invariance passes; OFF byte-identical.
• Phase 3 — SPINE (organ :spine, prereq :motor_cortex). MotorControl → stack of SP.Brain.Motor segments; per-segment precision Π = variable impedance; Π_distal_relax injects structured motor variability through the UNCHANGED epistemic term. Gate: 2-segment Π=1.0 byte-identical; distal_relax>0 more distal entropy + new config at strike-rate ≥ single-servo, beating the motor_shuffle control.
• Phase 4 — GLANDS + CYCLES (organ :endocrine). Persistent SP.Brain.Endocrine state on the MC struct (stripped in demodulate): satiety attenuates ONLY positive C on whitelist [:inventory,:build]; oscillators retune γ_m/η; circadian conditional on a sensed clock. Gate: 2×2 ablation; satiety bounds the hoard + phase-3 entry + survival; OFF byte-identical.
• Phase 5 — HEMISPHERES (organ :hemispheres). Two SlowContext parents (left fine/high-γ, right coarse/low-γ); DOWN blend normalize(L^wL·R^wR); callosal softmax(−β·[ΣF_left,ΣF_right]). Gate: H1 byte-identical at δ=0; H2 dual-regime; H3 explores more; the SYMMETRIC-DUPLICATE control must NOT show H3.

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

Behavioural/organisational substrates, NECESSARY-NOT-SUFFICIENT, ZERO evidential weight for experience. Measures: action-dependent viability persistence; allostasis signature; novelty W as a functional-access novelty_drive (must correlate with later exploration + FALL at insight); state- visitation entropy + bounded duplicate-tool count; dual-regime presence. We update the world-map only on a failed gate.

Critical files + seams

• lib/sp/brain/plan.ex advance/3 (LIVE decider) + efe.ex step_value/3 (mirror); factors.ex evaluate_policies/1 (per-factor epistemic/pragmatic + γ); precision.ex update_policy/2 (γ); model.ex :pa/:pb (Dirichlet counts); infer.ex (qo=matvec(a_m,qs)).
• genome.ex — heritable-field + slow_defaults back-fill + mutate/2 append-LAST + organ (@prereqs/@modalities/active_organs) + opt-in builder; designer.ex compile/1 + likelihood/3 + Genome.card Map.take (:init_a ⇒ add :b_init); curriculum.ex preference/3 + mc.ex phase_goal_met?/2/set_phase; mc.ex save/demodulate + the :slow_context/:motor struct discipline.
• New so far: lib/sp/brain/{diagnose,novelty}.ex; runs/{probe_plateau,curiosity_lineage,probe_curiosity}.exs.

Verification

• Offline first: per-phase RED gates + named byte-identity assertions through the depth-5 Plan path (the motor_cortex_test/slow_context_wired_test/novelty_test precedents); monotonic-W-decay; action-clone invariance; full brain suite green.
• Live (per ops_colony_lab_rootless, owner go-ahead): opt-in lineages in separate containers, distinct kin/memory dir, alongside the undisturbed default + motor + curiosity colonies — RED test the gate, probe, RCON-confirm.