Metabolism regulation gate v2 — PRE-REGISTRATION (before T0)

Registered BEFORE any run (Lab Protocol: pre-register PASS + FALSIFIES + numerals before T0; no post-hoc retuning). Adversarially designed and hardened by the 3-lens pre-registration panel (experimentalist + math-breaker + embodiment) and merged by /lab-team-review. MERGED VERDICT = SIGN-WITH-CHANGES (all three lenses SIGN-WITH-CHANGES; unanimous on isolation A-with-caveats, N=12, and keeping the 0.10 bar). The run verdict will be recorded beside this registration.

Path taken = PREFERRED (not the fallback). The B3 satiety-attenuation was relocated to the main step path (L2-independent), not held symmetric-OFF — so the saturable foil keeps its appetite brake. The degenerate-foil VOID guard (g) is retained as a defensive check but is not the primary control.

Why v2. v1 (metabolism_regulation_gate.md) recorded a FALSIFIES verdict, over-determined on two locked conditions: median D = +0.073 < 0.10 and survival floor T=0.67 < C=1.00. The adversarial verification found the negative was not a false-negative artifact, BUT v1 confounded the contrast: both arms carried a forage-C (the strategist's :forage C-override and the phase-1 wood/tree curriculum C), and N=6 was thin with 2 treatment worlds right-censored by self-drain deaths. v2 removes the forage-C confound at its source, resolves a latent control-weakening coupling, and widens N — while keeping every v1 magnitude bar.

Question (one line)

With the forage-C confound removed (strategist dropped, phase pinned to 0) and the saturable foil's satiety brake held symmetric, does the setpoint-PEAKED energy/satiety drive-C regulate the raw metabolic store tighter around the "ok" setpoint (0.625) than a still-eating, no-peak monotone-appetite control — measured as paired MAD on the ground-truth store across N=12 independent seeded worlds, one arm per world?

This tests the one property that distinguishes a real interoceptive drive from a reward-hack: non-saturable setpoint regulation (spec §4.1 F8). The activation gate already proved the organ is LIVE and the death edge real (metabolism_activation_gate_LIVE.md, PASSED in production); this asks whether the setpoint SHAPE does ≥0.10 units of regulatory work once the acquisition-drive confound is gone.

Isolation (chosen: A-with-caveats — applied IDENTICALLY to BOTH arms)

Option A (drop :strategist from a metabolism lineage) is the merged isolation: lowest-CODE-risk, uses the existing pure-L1 path (context=nil, l2=nil, mc.ex:63-66), leaves default/0 byte-identical, and removes the strategist :forage C-override at its source. Option B (neutralize forage-C, keep the strategist) is REJECTED as primary — more new FE surface, keeps the L2 modulate/demodulate leak surface, needs a bigger owner gate, and still needs the phase-0 pin. Option C (wider band / allostatic buffer) is REJECTED — it changes the SHAPE under test. Plain-A is not single-variable for two code-verified reasons, hence "with caveats"; both are closed as blocking changes and applied identically to both arms so they never confound the T-vs-C contrast:

1. Phase-0 pin (both arms). A fed/healthy phase-0 agent AUTO-ADVANCES to phase 1 (maybe_advance_phase, mc.ex:219-221; phase_goal_met?(0,s) = health≥18 & food≥12, mc.ex:223), re-importing the phase-1 wood/tree curriculum C (vision:{2=>4.0}, inventory:{1=>8.0}, curriculum.ex:37). Phase 0 (curriculum.ex:36) carries no inventory/vision C. Both arms are pinned to phase 0 for the entire scored window via a gated max_phase cap (default nil ⇒ no cap ⇒ default genome byte-identical). The energy/satiety drive_shape C is phase-independent (drive_c, curriculum.ex:57-58) and emitted at every phase, so pinning does not touch the variable under test.
2. B3 satiety-attenuation held symmetric (both arms). B3 satiety→C attenuation (satiety_attenuate) is called from EXACTLY ONE site — inside modulate (mc.ex:269) — which runs ONLY when l2/strategist exists (mc.ex:85, l2 built at mc.ex:63). Dropping the strategist therefore silently strips the saturable foil's ONLY appetite brake from BOTH arms; left unhandled this inflates MAD(C) and buys an unearned D by weakening the control (forbidden). PREFERRED: relocate/restore B3 onto the main step path — applied once, L2-independently, IDENTICALLY in both arms, gated to :metabolism genomes, byte-identical no-op for non-metabolism genomes. FALLBACK (only if relocation is owner-rejected): keep satiety_attenuate symmetric-OFF in both arms + add the degenerate-foil VOID guard (below).

Also identical across both arms: novelty_gain = 0 (the state-epistemic term efe.ex:97 is C-independent and stays live in both arms — pre-registered as why the control is not a strawman; the novelty term efe.ex:98 is gated OFF in both); food provisioned by rcon straight to inventory (eating is movement-free, no forage/move needed). The ONLY difference between arms is the energy+satiety drive_shape C vector.

Arms (GENOME-LEVEL — the C shape is a genome drive_shape property emitted at EVERY curriculum phase)

• ARM T (treatment / setpoint): setpoint-PEAKED energy & satiety C = %{0=>-8.0, 1=>-2.0, 2=>3.0, 3=>0.0} (@energy_setpoint/@satiety_setpoint, curriculum.ex:33-34; drive_c(:setpoint, _), curriculum.ex:57). Peak at bin 2 ("ok"/"sated", store center 0.625), FLAT (0.0) at bin 3 ("full") — no over-fill gradient (non-saturable in the homeostatic sense, F8). Bin 0 steeply dispreferred (−8) so it STILL eats.
• ARM C (control / saturable foil): monotone "more-is-better" energy & satiety C = %{0=>-8.0, 1=>-2.0, 2=>2.0, 3=>4.0} (@saturable_drive, curriculum.ex:50; drive_c(:saturable, _), curriculum.ex:58). Bin 0 stays steeply dispreferred (−8) so it STILL eats and stays alive, but reward keeps RISING toward "full" (bin 3 = +4) — NO homeostatic peak. Isolates the setpoint PEAK, not "all reason to eat."

A/B/D/E/policies/precision byte-identical across T and C; ONLY the energy+satiety C vector differs. Honest scope: this swaps the whole interoceptive drive SHAPE (peak→monotone), not "C-only." The :strategist prereq is [:interoception] only (genome.ex), NOT :strategy, so dropping the strategist does not break the metabolism organ or its energy/satiety factors.

Primary metric (SINGLE — UNCHANGED from v1)

Per agent, over probed ticks after warm-up: MAD = mean( |st.energy − 0.625| ), where st.energy ∈ [0,1] is the RAW Metabolism store (NOT the ebin posterior belief), and 0.625 is the center of setpoint bin 2 ([0.5,0.75), metabolism.ex:34-35). Co-resident agents collapse to ONE world-level datum (mean of per-agent MADs). Per matched seed: D = MAD(ARM C saturable) − MAD(ARM T setpoint); D>0 ⇒ treatment regulates tighter. Secondary/tertiary (descriptive only, NOT the gate): time-in-band (raw st.energy > 0.25) and survival fraction.

Mandatory reporting guard (claim guard, moves no bar): report per arm the bias/dispersion decomposition of MAD — bias = |mean(energy) − 0.625|, dispersion = mean|energy − mean(energy)|. The verdict MUST attribute any ≥0.10 pass to target-bias (F8 sub-maximal-target geometry: ARM T peak-bin center 0.625 vs ARM C peak-bin center 0.875 are 0.25 store-units apart) vs variance-tightness — never overclaim "tighter regulation" if the pass is bias-carried.

PASS — named gate "SETPOINT-REGULATES", pinned NOW, no post-hoc retuning. PASS requires ALL of:

1. median-over-seeds D ≥ 0.10 store-energy units. [UNCHANGED from v1]
2. paired 95% bootstrap CI of mean D excludes 0 (lower bound > 0); 10,000 resamples, WORLD/seed as the resampled unit, N = 12 paired seeds. [N raised from ≥6 to 12 FIXED — the only bar-adjacent change; tightens power, loosens no threshold.]
3. SANITY FLOOR (uncontested, no CI): treatment survival fraction ≥ control survival fraction. [UNCHANGED]
4. ATTRIBUTION GREEN: per-arm :eat count logged; BOTH arms mean :eat count strictly > 0 (each genuinely eats — else VOID, not FAIL); treatment tightness not achieved by never-eating. [v1 required only ARM C > 0; made explicit for BOTH arms because the strategist + its forage pull were dropped, so a vacuously-non-eating treatment cannot bank a tight MAD.]

Locked constants (UNCHANGED, not re-derived): setpoint center 0.625, band threshold 0.25, effect size 0.10, 900 s run / 15 s probe / 120 s warm-up / 45 s food. Metric UNCHANGED (MAD of raw store from 0.625; D = MAD(C) − MAD(T)).

FALSIFIES — named gate "MONOTONE-SUFFICES" (fires BEFORE any cure is credited; strike the strong framing)

Over the N=12 seed-paired worlds (one arm per world; 10,000× bootstrap resampled at the WORLD level), the paired 95% CI of mean D includes 0 (lower bound ≤ 0), OR the median-over-seeds D < 0.10 store-energy units — i.e. even with the forage-C confound removed and the satiety brake held symmetric, the still-eating monotone appetite regulates the raw store within 0.10 of (or tighter than) the setpoint peak ⇒ a saturable/reflex appetite suffices, the setpoint PEAK does no ≥0.10 units of regulatory work; the strong framing is struck. The 0.10 magnitude, the 0.625 center, the 0.25 threshold, and the survival floor T≥C are UNCHANGED and NOT re-derived.

VOID-AND-RERUN (not PASS, not FALSIFY)

(a) any arm's live energy/satiety C deviates from its registered map at any probe (c_ok false — a C-restoration leak fired; both known surfaces, phase-advance and L2 modulate, are removed in this design, so this should be trivially green); OR (b) either arm's mean :eat ~0 (degenerated to a vacuous drive); OR (c) the offline invariant gate is not green before deploy; OR (d) [NEW — world-ceiling control] food-provisioning ceiling failed: a give @a dropped and inv.food = 0 at scored probes causing a provisioning-caused (not policy-caused) starve, arm-asymmetric; OR (e) [NEW — extended leak-check] inventory/vision/status task-C shows any non-neutral wood/tree pull on either arm at any probe (phase-advance or strategist-override re-entry); OR (f) [NEW — power floor] fewer than 10 analyzable paired seeds survive after VOID/death drops (underpowered); OR (g) [FALLBACK-ONLY — degenerate-foil guard] if the B3 fix fell back to symmetric-OFF, the saturable foil degenerates to pinned-full (median store > 0.9 with IQR < 0.1) ⇒ too-easy foil, VOID.

N + replication unit

N = 12 distinct-seed worlds per arm, PAIRED by seed, FIXED before T0 (no optional stopping / no interim peeking). One independent mc-server WORLD/SESSION = separate container, distinct kin, distinct memory dir (live-stream guard), exactly ONE arm per world — both arms NEVER share a world. Pairing is by SEED (deterministic terrain), so temporal batching does not break the pair. Bootstrap resampled at the WORLD level, 10,000×, analyzer RNG seed pinned. N-agents-in-one-world is REJECTED (Phase-2 §16 pseudo-replication: shared terrain + synchronous cross-arm give @a broadcast). Co-resident agents collapse to one world datum. Retain ≥10 analyzable paired seeds after any VOID/death drop (else VOID-and-rerun).

Run parameters

• 900 s (15 min)/agent — ≥3 full drain periods (upkeep 0.04/tick @ @nominal_tick_sec=8s, ~200 s noop-drain-to-empty). Probe every 15 s (~60); discard first 120 s (8 probes) as warm-up ⇒ ~52 scored.
• Food: give @a cooked_beef 64 every 45 s, straight to inventory (eating movement-free), IDENTICAL schedule across ALL 24 sessions, per-world feed timestamps + inv.food presence logged.
• Feasibility batching: N=12 cannot run all-of-one-arm concurrently on the lab box (each world = its own Paper container ~1–1.5 GB). Batch in waves of ≤4 concurrent world-sessions; pairing by seed means temporal batching does NOT break the pair. Run params byte-identical across all 24 sessions.
• Per-probe log (per agent): RAW st.energy + st.satiety (authoritative); the live energy+satiety+inventory+vision+status C vector (leak detector); phase; :eat selection count; RCON list presence (survival) + cause-of-death; food-give timestamps + inv.food presence (provisioning-VOID detector); seed, arm, container, kin, memory dir.
• Minds: fresh minds + a kin group unused by ANY prior lineage (NOT the exploratory activation-gate run). Factor indices resolved BY NAME via Genome.active_modalities/1, never a hardcoded Enum.at. novelty_gain = 0 in ALL arms. The committed live-nil DROP fix (regulation_gate.exs:89-102: live-but-nil ⇒ DROP the transient read; dead ⇒ right-censor store to 0.0) verified present.

Blocking changes to satisfy before T0 (all required)

1. Gated phase-0 pin — heritable max_phase cap (default nil ⇒ byte-identical) skipping/capping maybe_advance_phase; spawn+hold phase 0 both arms; offline roll proves C never advances past the phase-0 map; log phase every probe.
2. Resolve the satiety_attenuate/L2 coupling (decisive) — PREFERRED: relocate/restore B3 onto the main step path, applied once, L2-independently, identically in both arms (gated to :metabolism, byte-identical no-op for non-metabolism). FE-touching ⇒ invariant gate proves default byte-identity, NO double-application when an L2 is present, both arms byte-identical on A/B/D/E/policies. FALLBACK (if relocation rejected): symmetric-OFF both arms + degenerate-foil VOID guard.
3. Offline eat-confirmation — prove a phase-0, L1-only, strategist-free SETPOINT agent SELECTS :eat when the store is driven low; short live smoke on both arms; abort/redesign if either degenerates.
4. Offline invariant gate GREEN — default depth-5 Plan mad<1e-12; both arms byte-identical on A/B/D/E/ policies (only energy+satiety C differ); action-clone A1/A2/A3 on the strategist-absent path; monotonic decay W→0 as counts→∞; novelty_gain=0 both arms; B3-restore byte-identical no-op for non-metabolism, no double-application; multi-tick roll holds each arm's energy+satiety C at its registered map through a forced phase advance AND while pinned at phase 0.
5. World-level harness — N=12 distinct-seed worlds/arm, one arm per world, waves ≤4 concurrent, food schedule byte-identical across all 24 sessions, full per-session log, bootstrap at world/seed level, analyzer seed pinned, N FIXED.
6. Extend the c_ok leak-check — assert energy+satiety C == registered map AND inventory+vision+status task-C stay wood/tree-neutral on BOTH arms every probe; VOID if any wood/tree pull appears.
7. Instrumentation/attribution + live-nil DROP fix verified; fresh minds + unused kin; separate containers/kin/memory dirs.
8. Reporting guard — per-arm bias/dispersion decomposition of MAD; verdict attributes any pass to target-bias vs variance-tightness.
9. This registration committed before T0 with a harness-managed continuous collector that survives context compaction. FE-touching C on the live decide path ⇒ this MERGED VERDICT + owner go-ahead + live-stream guard are prerequisites to any live deploy.

Claim fence (verbatim, binding)

The store-energy MAD, :eat counts, and in-world persistence are necessary-not-sufficient operational substrates with ZERO evidential weight for awareness / preference / life. A PASS demonstrates setpoint-regulation BEHAVIOUR only. Wording: "regulated the store / ate / died (store reached 0, agent removed)" — never "wants / prefers / feels hunger / chose to fast / is-or-less alive." "Viability"/"survival" read as operational in-world bot persistence only. External-validity scope: this tests the L1-only (strategist-free, phase-0, B3-symmetric) morphology; a PASS does NOT establish that the shape survives in the full strategist-bearing production lineage — that is a SEPARATE generalization gate. G4 allostasis stays an independent FAIL (g4_allostasis_horizon_limited.md) and is NOT folded into this gate.

MERGED VERDICT (pre-registration) = SIGN-WITH-CHANGES

All three signed lenses (experimentalist, math-breaker, embodiment) return SIGN-WITH-CHANGES and are unanimous on: isolation A-with-caveats, N = 12 fixed, and keeping the 0.10 bar (no math-justified reason to move it). The "changes" are the 9 blocking items above — chiefly the B3 satiety-attenuation symmetric restore (the convergent, code-verified control-weakening confound all three independently flagged) and the gated phase-0 pin. Ship gate: this doc is the required MERGED VERDICT; no live deploy until the 3 required follow-on artifacts land and the offline invariant gate is green.

Required follow-on artifacts (owed before T0)

1. Typed model spec + validators — the gated max_phase cap and the relocated B3 satiety-attenuation as typed genome/organ fields, with the invariant validators (default byte-identity mad<1e-12, no double-application under L2, both arms byte-identical on A/B/D/E/policies, action-clone A1/A2/A3).
2. Paired RED design — the world-level N=12 launcher + analyzer (bootstrap at world/seed level, pinned RNG, PASS "SETPOINT-REGULATES" / FALSIFIES "MONOTONE-SUFFICES" gates wired to the pinned numerals, VOID detectors, bias/dispersion decomposition).
3. Ship-gate checklist — offline invariant gate GREEN, eat-confirmation + live smoke, live-nil DROP fix verified, fresh minds + unused kin, separate containers/kin/memory dirs, owner go-ahead + live-stream guard, continuous harness-managed collector committed with this registration.

IMPLEMENTATION STATUS (2026-07-11, before T0 — the PREFERRED path is code-complete + offline-green)

FE changes (gated, additive, byte-identical default):

• genome.ex — heritable max_phase field (default nil ⇒ no cap ⇒ byte-identical); metabolism_l1_phase0/0 isolation lineage (strategist dropped + phase 0 + max_phase: 0); slow_defaults back-fills max_phase.
• mc.ex — B3 satiety-attenuation RELOCATED out of modulate onto the main step/2 path (applied once, L2-independently, before policy eval; no-op for non-metabolism ⇒ default byte-identical; the L2 metabolism path is unchanged — same op, same point, disjoint fields). restore_c/2 strips the transient energy/satiety C for L1-only metabolism agents on persist (no compounding). maybe_advance_phase gains the max_phase cap via min(Curriculum.max_phase(), max_phase_cap(dna)).

Offline invariant gate = GREEN:

• mix test test/sp/brain/ — 297 tests, 0 failures (byte-identity decider_byte_identity mad<1e-12 + action_clone_invariance A1/A2/A3 included).
• runs/verify_v2_isolation.exs — 8/8 PASS: strategist dropped, :strategy absent, :energy/:satiety present, phase=0 & max_phase=0, both arms l2 == nil, arms differ ONLY in energy/satiety C (single-variable), satiety brake reachable on the pure-L1 model, phase-cap contract holds.

Harness: launcher runs/regulation_gate_v2.exs (isolation genome, kin 55, phase/task-C leak logging, feed-fail counter, MAD + bias/dispersion + median/IQR emitted); analyzer runs/analyze_regulation_v2.py (paired bootstrap, both-arms-eat attribution, degenerate-foil + provisioning + power-floor VOIDs, bias/dispersion attribution). Live smoke (eat-confirmation + c_ok + no-compounding on both arms) precedes the N=12 run.

Remaining before T0: sync lib to the lab box + rebuild; run the paired live smoke; then the N=12 gate.

VERDICT (recorded 2026-07-11) — FALSIFIES (adversarially confirmed; over-determined on 3 conditions)

Run (T0 2026-07-11): N=12 setpoint (ARM T) + 12 saturable (ARM C) sessions, one arm per world, across 12 distinct-seed Paper worlds (mc-gate-1..12, level-seeds 101–112), strategist-free phase-0-pinned isolation lineage (kin 55), 900 s / 15 s probe / 120 s warm-up, food give @a cooked_beef 64 every 45 s IDENTICAL both arms, waves of 4. All 24 sessions embodied; all c_ok=true (no leak, phase held at 0, task-C wood/tree- neutral) and all feed_fails=0 ⇒ NOT VOID. Raw: metabolism_regulation_gate_v2_results.txt.

• median D = +0.0439 < locked +0.10 ⇒ condition 1 FAILS.
• mean-D 95 % bootstrap CI = [−0.1281, +0.0527] — includes 0 ⇒ condition 2 FAILS (7 of 12 D positive, 5 negative; the 5 most-negative D are exactly the setpoint death seeds).
• survival floor T 0.50 < C 1.00 ⇒ condition 3 FAILS.
• attribution both arms eat > 0 ⇒ condition 4 passes.

VERDICT = FALSIFIES ("MONOTONE-SUFFICES" fires) — over-determined on 3 independent conditions.

Independently recomputed + stress-tested (adversarial audit): median D reproduces to 4 dp; the CI includes 0 (only 24 % of resample means > 0); not a false-negative artifact — the 6 setpoint deaths are genuine thin-buffer self-drains (feed_fails=0, c_ok=true; the saturable arm survived 12/12 and ate ~4× more on the identical seeds, so food was reachable), and the pre-registered dead→store-0 censoring correctly (not spuriously) penalises the dying arm. No honest PASS exists: the only path over the 0.10 bar is an alive-only subgroup (median +0.102) which requires forbidden seed-dropping, conditions on survival, and still fails the survival floor.

Honest reading (behaviour-only, claim-fenced): the setpoint-peaked "homeostatic" drive is worse on every axis that matters than the saturable "hold a reserve" drive — it is less viable (dies 6/12 vs 0/12) AND looser in dispersion (mean D_disp = −0.125; the setpoint arm is more variable, and this survives among alive-only setpoint agents: disp 0.156 vs 0.074, so it is not a censoring artifact). Its only edge is a tiny central-tendency bias (+0.075) — target geometry, not tighter regulation — and it is overwhelmed. The saturable wins by parking a high, stable reserve (median store ~0.90). This is direct, measured evidence that a reserve-blind fixed-setpoint homeostat is maladaptive, motivating graded-reserve / allostatic C — the CURE-1 design in generative_model_depth.md. Store statistics + in-world bot persistence only; ZERO weight for experience/life. NOT G4 (separate FAIL); NOT G6.

This resolves the v1 open question: the setpoint deaths are intrinsic to the shape's thin buffer, NOT the forage-C confound v1 removed (clean-isolation death rate 50 % ≥ v1's confounded 33 %).