Metabolism regulation gate — PRE-REGISTRATION (2026-07-11, before T0)

Evidence & Verdicts · docs/receipts/metabolism_regulation_gate.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-seven dated pages: receipts, pre-registrations, handoffs, validation records and review verdicts. A receipt is written at the moment a piece of work was checked. It names what was claimed, the commit and the seed, what was actually run, and the outcome in one of a small set of controlled words. Then it names what the work did not achieve. That last part is what makes it a receipt rather than an announcement. A pre-registration is the same discipline run in advance: the conditions that would count as a pass and the conditions that would falsify the claim are written down before the run, so neither can be adjusted once the numbers arrive.

That is why so many small dated stubs are an audit trail rather than noise. No one of them is meant to be a good read. The value is in the sequence and in the dates, because you can watch a prediction be registered, then the run happen, then the verdict land — sometimes against the prediction. Pages here record a falsified result, a rejected fix, a retracted overclaim, and a green receipt that turned out not to be reproducible from the commit that carried it. A record that carried only successes would be worth a good deal less than this one.

A gentle way in is to read a pre-registration first, so the shape becomes familiar, then a result page, then one of the corrections. This section sits off the main navigation on purpose: it is the record you check the rest of the site against, not the place to begin.

What it is not: documentation, and not a summary. Nothing here has been tidied in hindsight. Every entry reads as of its date, a later entry may overturn an earlier one, and the presence of a page is not a claim that its result stood.

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.

Registered BEFORE any run (Lab Protocol: pre-register PASS + FALSIFIES + numerals before T0; no post-hoc retuning). Adversarially designed + hardened by the 5-persona pre-registration panel (workflow wf_deef92cb-8b9: experimentalist + math-breaker + embodiment + merge). The verdict will be recorded beside this registration.

Question (one line)

Does the setpoint-PEAKED energy/satiety drive-C regulate the raw metabolic store tighter around the "ok" setpoint than a still-eating, no-peak monotone-appetite control — measured as paired MAD on the ground-truth store across ≥6 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 + the edge real; this asks whether the setpoint SHAPE does regulatory work.

Arms (GENOME-LEVEL — the C shape is a genome property emitted at EVERY curriculum phase, not a launcher one-shot)

ARM T (treatment): setpoint-peaked energy & satiety C = %{0=>-8.0, 1=>-2.0, 2=>3.0, 3=>0.0} (the live shape, curriculum.ex:33-34; reachability receipted — metabolism_activation_gate_LIVE.md).
ARM C (control — the endorsed harder foil): flattened/saturable monotone "more-is-better" energy & satiety C = %{0=>-8.0, 1=>-2.0, 2=>2.0, 3=>4.0}. Bin-0 stays steeply dispreferred so it STILL eats and stays alive, but there is no homeostatic peak — reward keeps rising toward "full". Isolates the setpoint PEAK, not "all reason to eat."
ARM Z (optional sanity, EXCLUDED from PASS): fully drive-severed C = all-zero. Death is construction-forced (C=0 + novelty_gain=0 + :eat habit-excluded ⇒ never eats by wiring) → labelled construction-forced, excluded from the CI/PASS discriminator.

A/B/D/E/policies/precision 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"; satiety→C attenuation stays partly live in ARM C (metabolism.ex:79 no-ops only on non-positive C; ARM C's bin2/3 are positive) — no smuggled advantage.

Primary metric (SINGLE)

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 MADs). Per matched seed: D = MAD(ARM C) − MAD(ARM T); D>0 ⇒ treatment regulates tighter. Secondary/tertiary (descriptive only, NOT the gate): time-in-band (raw st.energy > 0.25) and survival fraction.

PASS — pinned NOW, no post-hoc retuning. PASS requires ALL of:

median-over-seeds D ≥ 0.10 store-energy units.
paired 95% bootstrap CI of mean D excludes 0 (lower bound > 0); 10,000 resamples, WORLD/seed as the resampled unit, N ≥ 6 paired seeds.
SANITY FLOOR (uncontested, no CI): treatment survival fraction ≥ control survival fraction.
ATTRIBUTION GREEN: per-arm :eat count logged; ARM C mean :eat count strictly > 0 (it genuinely eats — else VOID, not FAIL); treatment tightness not achieved by never-eating.

Locked constants: setpoint center 0.625, band threshold 0.25, effect size 0.10, N≥6, 900s run / 15s probe / 120s warm-up / 45s food.

FALSIFIES (setpoint-peaked C does NO regulatory work; strike the framing)

Paired 95% CI of mean D includes 0, OR median D < 0.10 — the still-eating monotone control regulates within 0.10 of (or tighter than) the setpoint treatment ⇒ a saturable/reflex appetite suffices, the peak adds nothing.

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

(a) any arm's live energy/satiety C deviates from its registered map at any probe (phase-advance C-restoration leak fired, mc.ex:487); OR (b) ARM C mean :eat ~0 (degenerated to the vacuous drive-severed case); OR (c) the offline invariant gate is not green before deploy.

Replication unit

One independent mc-server WORLD/SESSION = separate container, distinct kin, distinct memory dir (CLAUDE.md live-stream guard), exactly ONE arm per world. N≥6 distinct world-seeds per arm, PAIRED by seed. Bootstrap resampled at the WORLD level. 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; both arms NEVER share a world.

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, IDENTICAL schedule both arms, per-world, feed timestamps logged.
Each probe records per agent: RAW st.energy + st.satiety (authoritative); the live energy+satiety C vector (leak detector); :eat selection count; RCON list presence (survival) + cause-of-death.
Fresh minds + a kin group unused by any lineage (NOT the exploratory activation-gate run). Factor indices resolved BY NAME via Genome.active_modalities/1 (as MC.satiety_attenuate, mc.ex:281-283) — never a hardcoded Enum.at(-2). novelty_gain=0 in ALL arms (the state-epistemic term efe.ex:97 is NOT gated and remains live in every arm — pre-registered as why the control is not a strawman).

Blocking changes to satisfy before T0 (all required)

Close the phase-advance C-restoration leak — genome-level :drive_shape field (energy/satiety C emitted at EVERY phase) AND/OR pin phase. Launcher one-shot is NOT genome-equivalent.
World-level replication (≥6 seeds/arm, one arm per world, bootstrap at world level).
Control = saturable map {0=>-8,1=>-2,2=>2,3=>4} (still eats), not fully-zeroed.
Metric = MAD of raw store from 0.625 (not ebin, not survival/time-in-band).
Factor indices BY NAME (Genome.active_modalities/1).
Offline invariant gate GREEN before deploy: T/C byte-identical on A/B/D/E/policies (only energy+satiety C differ), action_clone A1/A2/A3 pass, novelty_gain=0 all arms, AND a multi-tick offline roll shows each arm's energy+satiety C stays at its registered map THROUGH a forced phase advance (proves #1).
Instrumentation/attribution (eat counts, cause-of-death, live C vector each probe; VOID if ARM C mean eat ~0 or C deviates).
This registration committed before T0 with a harness-managed continuous collector.

Claim fence (verbatim, binding)

The store-energy MAD is a model-variable regulation measure, never felt hunger/comfort/experience; a PASS demonstrates setpoint-regulation BEHAVIOUR only. G4 allostasis stays an independent FAIL (g4_allostasis_horizon_limited.md) and is NOT folded into this gate.

VERDICT (recorded 2026-07-11, after the run) — FALSIFIES (confirmed by adversarial verification, not overturned)

Run (T0 2026-07-11): 6 setpoint (ARM T) + 6 saturable (ARM C) sessions, one arm per world, across 6 distinct-seed Paper 1.16.5 worlds (mc-gate-1..6, seeds 1..6), fresh minds, kin 77, 900 s / 15 s probe / 120 s warm-up (53 scored ticks/agent), food give @a cooked_beef 64 every 45 s identical both arms. All 12 sessions embodied (RCON-confirmed exactly one UNI per world), all c_ok=true (no C-restoration leak), ARM C mean :eat = 257.2 (attribution green) ⇒ NOT VOID. Raw RESULT lines: metabolism_regulation_gate_results.txt. Launcher runs/regulation_gate.exs, analyzer runs/analyze_regulation.py (random.seed(20260711)).

Data (primary metric = MAD of raw store energy from setpoint 0.625):

seed/world	MAD(T setpoint)	surv T	eat T	MAD(C saturable)	surv C	eat C	D = C−T
1	0.1663	alive	49	0.2728	alive	174	+0.1065
2	0.1467	alive	44	0.3019	alive	358	+0.1552
3	0.1845	alive	32	0.2578	alive	183	+0.0733
4	0.2688	died	–	0.3422	alive	368	+0.0734
5	0.2303	alive	44	0.2800	alive	237	+0.0497
6	0.2454	died	–	0.2864	alive	223	+0.0410

median D = +0.0733 < locked +0.10 ⇒ condition 1 FAILS (the pre-registered FALSIFIES trigger fires verbatim).
mean-D 95 % bootstrap CI = [+0.0562, +0.1144] excludes 0 ⇒ condition 2 passes (the direction is real).
survival floor T = 0.67 < C = 1.00 ⇒ condition 3 FAILS (independent second failure).
attribution ARM C mean eat = 257.2 > 0 ⇒ condition 4 passes.

VERDICT = FALSIFIES — over-determined on two independent locked conditions (median D < 0.10 AND survival floor T < C).

Confirmed by a 4-lens adversarial verification (workflow wf_66947cb9-c31, 5 agents, unanimous high-confidence, not overturned): the arithmetic is exact (independent recompute); the negative is NOT a false-negative artifact — the magnitude shortfall holds even among the 4 non-death worlds (survivor-only median +0.090 < 0.10), and the 2 treatment deaths (seeds 4, 6) are terminal self-drains to store = 0 under the arm's own policy while world-food was available (the ARM-C agent ate 223–368× and survived on the very same seeds), NOT a provisioning VOID; the rescue lens found no honest reading that yields PASS.

Honest reading (behaviour-only, claim-fenced): the setpoint-PEAKED drive-C does a real but sub-committed amount of store-regulation work — it held the raw store closer to 0.625 in 6/6 worlds while eating ~7× less (32–49 vs 174–368 feeds) — but the effect (~0.073–0.083 store-units) falls below the +0.10 magnitude bar committed before T0, so the strong claim "the peak does large regulatory work" is struck. A second, largely-independent finding: the tighter-regulating lean eater carries a viability cost — it persisted in 4/6 worlds vs the over-eater's 6/6 — i.e. in this world the buffer-hoarding monotone appetite is operationally more survivable. Entanglement noted (the same 2 deaths also depress D on seeds 4, 6 via right-censoring), though the magnitude failure stands alone. Mechanism of the two deaths (adaptive thin-buffer vs degenerate forage-cessation/stuck-state) is UNDETERMINED from these receipts and is not asserted.

Claim fence (binding): store-MAD, feed counts, and in-world persistence are necessary-not-sufficient operational substrates with ZERO evidential weight for awareness / preference / life. 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. This FALSIFIES is not a pass, not a PARTIAL, not a rescue — it is logged as evidence. Scope honesty: the arms swap the whole interoceptive drive SHAPE (peak vs monotone), not a C-only isolate (satiety→C attenuation stays partly live in ARM C), so the finding is the peaked-vs-monotone contrast.

Go-live implication: this gate does NOT clear, so the standing "go full live on a cleared gate" is NOT triggered by this run — nothing here authorizes deploying the peaked drive-C lineage to the public-streamed colony. The separately-recorded metabolism activation gate (organ live + death edge real, pos/neg/neg/pos) remains PASSED in production (metabolism_activation_gate_LIVE.md); that is the cleared receipt — this regulation gate is a falsified one.

Carry-forward (evidence, not clearance): (a) the small-but-consistent directional effect + the viability-cost finding are real inputs for the next design — they indicate the setpoint peak needs either a wider preferred band, an allostatic buffer margin, or the epistemic/foraging layer to close the survival gap before the SHAPE can do ≥0.10 units of work; (b) fix the regulation_gate.exs live-nil→0.0 coercion (drop a transient live-nil read rather than score store = 0; arm-symmetric, did not change this verdict) before any re-run; (c) N = 6 is thin — a re-run would widen N and resolve the death mechanism.

sha256 998fb4161d838c32 — 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)

A pre-registration written before any run, with the verdict recorded beside it afterwards, and the verdict is that the gate does not clear. The question is whether a preference shaped around a set point holds an internal store more tightly than a still-eating control with no peak. The bars, the effect size, the number of worlds and every constant were pinned in advance, including the conditions that would make a run void rather than either a pass or a failure. A binding limit sits alongside it: the measure is a model variable, never a felt state. Nothing here authorises deploying that shape.

Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 998fb4161d838c32

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)

A pre-registration written and hardened before anything ran, with the verdict added beside it rather than in place of it. The question is stated in one line and is narrower than it first looks: not whether the drive exists, which an earlier gate already settled, but whether the shape of it does regulatory work.

The arms are defined at the level of the genome rather than as a setting applied once at launch, and the reason is given, because a one-shot setting is not equivalent to a property emitted at every phase. The control is chosen deliberately as the harder foil: it still strongly avoids emptiness, so it keeps eating and staying alive, but it has no peak, which isolates the peak rather than the reason to eat at all. A third arm is included as a sanity check and explicitly excluded from the pass, because its death is forced by construction.

The scope is stated honestly. This swaps the whole shape rather than one term, and one related mechanism stays partly live in the control, so the finding is a contrast between two shapes and is described as such.

There is a single primary metric, defined on the raw store rather than on the belief about it, with the difference spelled out. Bodies sharing a world collapse to one datum, and the paired difference is taken by seed. Secondary measures are named as descriptive only and explicitly not part of the gate.

The pass conditions are pinned with numbers. There is an effect size, an interval that must exclude zero, and a resampling unit that is the world rather than the body. There is a floor on survival, and an attribution check that the control genuinely ate, so tightness cannot be achieved by never eating. What would show it wrong is the mirror of that. A separate void category then covers three ways a run could be invalid rather than negative, including a leak that would restore the preference behind the experiment's back.

The replication unit gets its own section, rejecting many bodies in one world by name as pseudo-replication, citing the earlier phase where that lesson came from, and requiring that the two arms never share a world.

Run parameters are given in full: how long, how often to probe, how much warm-up to discard, and an identical feeding schedule in both arms. One detail is worth noting, that indices are resolved by name rather than by position, which is the kind of thing that breaks quietly. A list of blocking changes must be satisfied before the clock starts.

The verdict, recorded afterwards, is that this gate does not clear, so nothing here authorises deploying that shape to the live colony, and the page names which earlier gate remains the cleared one. What is carried forward is called evidence rather than clearance: a small consistent directional effect and a cost to viability, which indicate what the next design would need. One coercion in the harness is named as needing fixing before any re-run, with a note that it was symmetric across arms and did not change this verdict, and the number of worlds is called thin.

Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 998fb4161d838c32