Rung-1 graded-viability paired RED — PRE-REGISTRATION (before T0)

Status: PRE-REGISTERED, NOT YET RUN. Registered BEFORE any run (Lab Protocol II: pinned PASS + FALSIFIES + numerals + VOID conditions before T0; no post-hoc retuning). The engine is BUILT + offline-proven (typed spec docs/specs/rung1_graded_viability.md; commits 1c49e62..784eabd; brain suite 315/0). This doc is the experimentalist artifact for the live gate. It does not authorize a live burn: the live run requires (a) the FE-surface blocking items below closed, (b) a /lab-team-review MERGED SIGN on those FE deltas, and (c) owner go-ahead + the live-stream guard (separate containers / distinct kin / distinct memory dirs).

This gate follows and is motivated by the recorded v2 FALSIFIES verdict (docs/receipts/metabolism_regulation_gate_v2.md): the reserve-blind flat fixed-setpoint homeostat is maladaptive — it died 6/12 (thin-buffer self-drain, food reachable) and was looser in dispersion than the hoarding saturable foil. That measured death IS the missing depth. Rung-1 (cures 1+2+3) adds graded per-subsystem viability + an interior-peak reserve C + work/fatigue. This RED asks whether that depth earns its place behaviourally, not just whether it survives.

Question (one line)

Does the rung-1 graded body — interior-peak reserve C over 6-state per-subsystem viability factors, with work→fatigue→motor coupling — produce allostatic, two-ended, cross-subsystem-dissociated, work-rest-paced self-maintenance that (i) survives where the flat setpoint died AND (ii) is behaviourally distinguishable from BOTH a 6-state flat-setpoint baseline AND a 6-state monotone-appetite (saturable) foil — with each surviving advantage attributable to a NAMED coupling via per-mechanism ablation?

Why survival-alone cannot be the gate (the key lesson from the offline pre-check). Under a reserve-following hand policy the body survives AND a saturable proxy also survives (both hold food); the offline dynamics roll (runs/verify_rung1_dynamics.exs, reproduced 2026-07-11: 4/4 VIABLE) shows the RESERVE structure survives, holds an interior reserve (energy mean 0.816, 80% of post-warmup ticks in [0.6,0.95]), paces (eat 93 / mine 233 / rest 374), and dissociates (energy↔fatigue corr 0.314 while energy↔gut 0.99). So the discriminators MUST be behavioural, or a "just eat more" foil banks the same survival. This is the same trap v2 pre-registered against.

Arms (one cure-bundle, ABLATION-DECOMPOSED so a bundled win stays attributable — Lab Protocol I)

All arms are GENOME-level lineages sharing A/B/D/E/policies/precision byte-identical; they differ ONLY in the gated :homeostat/drive_shape surface named per arm. One arm per world (never co-resident); paired by seed.

FE-surface decision that gates this RED (flagged, NOT silently added — see Blocking item B1). drive_c(:setpoint,6) and drive_c(:saturable,6) today merely expand the 4-key metabolism maps ([-8,-2,3,0,0,0] and [-8,-2,2,4,0,0], lib/sp/brain/curriculum.ex:65-66), which are MALFORMED as true 6-state baseline/foil shapes (peak stranded in the lower third; the "saturable" isn't monotone to the ceiling). The baseline/foil arms therefore need proper 6-state C vectors defined — e.g. setpoint6 peaked at the interior center bin with symmetric falloff, saturable6 monotone non-decreasing to bin 5. That is NEW gated FE surface (additive; absent from default/0; byte-identical default preserved) and must pass /lab-team-review before the launcher is written (CLAUDE.md ship gate). Until then the launcher is scaffolded against FULL only.

Isolation — applied IDENTICALLY to every arm (so nothing but the named per-arm surface confounds)

• Strategist dropped (no standing :forage task-C), phase pinned 0 via the gated max_phase: 0 cap (so no auto-advance re-imports the phase-1 wood/tree curriculum C). Both already baked into homeostat_l1_phase0.
• novelty_gain = 0 in ALL arms (the state-epistemic term efe.ex:97 stays live in both — pre-registered as why the control is not a strawman; the parameter-novelty term is gated OFF in all).
• :motor_cortex present in ALL arms (so the fatigue→Motor.pi coupling is a genuine world consequence, not a per-arm structural difference). Only the fatigue coupling strength / C differs between FULL and the ablations.
• Food provisioning schedule byte-identical across all sessions of a given world-class (see run params).
• Fresh minds + a kin group unused by any prior lineage (NOT kin 55 from the v2 run). Factor indices resolved BY NAME via Genome.active_modalities/1, never a positional Enum.at.

Metrics (behavioural discriminators — pinned NOW, computed post-warmup, per the typed spec §Paired RED)

Raw body stores (SP.Brain.Homeostat energy/gut/soma/fatigue ∈ [0,1]) are the authoritative probe (NOT the bin6 posterior). Beliefs (energy_reserve/gut_satiety/muscle_fatigue posteriors) probed alongside for the allostasis + dissociation measures. All indices computed per agent, collapsed to one datum per world, bootstrap resampled at the WORLD/seed level (10,000×, analyzer RNG seed pinned).

1. survival fraction — in-world bot persistence (RCON list) over the scored window; death = store edge (Homeostat.dead?: energy<=0 OR soma<=0) confirmed by RCON removal.
2. allostasis_index — believed reserve (energy_reserve posterior expectation) at eat-onset; FULL − SETPOINT-6. Positive ⇒ FULL eats before the edge (anticipatory), not reactively at critical.
3. two-ended satiation — a paired contrast across two WORLD-CLASSES on matched seeds: scarce (food metered thin) ⇒ FULL fights harder near critical (higher :eat share at low reserve); rich (food abundant) ⇒ FULL STOPS eating at the interior reserve band, does NOT hoard to surplus (median store < 0.95, IQR non-degenerate) while SATURABLE-6 pins to the ceiling.
4. cross-subsystem dissociation Δ — 1 − |corr(energy_belief, fatigue_belief)| (energy↔fatigue is the cleanest per the offline roll, corr 0.314). CI must exclude 0 ⇒ the 2nd factor is measurably decoupled, not a renamed bin.
5. fatigue_pacing_index — corr(believed fatigue, P(non-arm action | just-mined)). > 0 for FULL, ≈ 0 for ABL-fatigue-π-with-flat-C; AND MotorControl target_error rises with fatigue within a mining bout, falls after rest (the aim-degradation world consequence).

PASS — named gate "RESERVE-EARNS-ITS-DEPTH", pinned NOW, no post-hoc retuning. PASS requires ALL of:

1. Survival ≥ 11/12 for FULL (vs the measured flat-setpoint 6/12, metabolism_regulation_gate_v2.md).
2. allostasis_index paired 95% bootstrap CI excludes 0 (lower bound > 0) — FULL eats anticipatorily.
3. two-ended satiation holds BOTH ends: in scarce worlds FULL's low-reserve :eat share > SATURABLE-6's (CI-separated), AND in rich worlds FULL median store < 0.95 with non-degenerate IQR while SATURABLE-6 pins

   0.95 (CI-separated). Fails if FULL hoards to the ceiling like the foil, OR fasts into starvation.

4. Beats BOTH controls on survival — FULL survival-count CI excludes SETPOINT-6 AND excludes SATURABLE-6 (an improvement over the death shape that is ALSO not merely the foil's eat-to-full survival).
5. cross-subsystem dissociation Δ CI excludes 0 (energy and fatigue beliefs measurably decouple).
6. fatigue_pacing_index > 0 for FULL and ≈ 0 for ABL-fatigue-π-flat, AND target_error rises-with-fatigue / falls-after-rest within bouts — the pacing is caused by the NAMED coupling, not a coincidence.
7. Every invariant anchor GREEN (regardless of survival): default depth-5 Plan mad<1e-12; action-clone + action-history invariance; no scalar-per-action; monotonic decay W→0; no compounding.

Locked constants (not re-derivable post-hoc): N=12 paired seeds; survival bar 11/12 (vs 6/12); bootstrap 10,000× at world/seed level; 900 s run / 15 s probe / 120 s warm-up; interior-band ceiling 0.95; rich-world hoard threshold 0.95; scarce/rich feed schedules pinned in the launcher.

FALSIFIES — named gate "DEPTH-DOESN'T-PAY" (fires BEFORE any cure is credited). REJECT if ANY of:

• Survival not improved over SETPOINT-6 (≤ its count, CI-overlapping); OR
• FULL is indistinguishable from SATURABLE-6 on survival AND the two-ended/allostasis discriminators (it is "just eat more" with extra factors) — collapse to Cure-1 or reject; OR
• FULL rests/hoards into starvation (over-brake: rich-world hoard, or scarce-world fasting death); OR
• dissociation Δ CI includes 0 (dead depth — the 2nd/3rd factors rename bins; collapse to Cure 1); OR
• allostasis_index CI includes 0 (reactive, not anticipatory — no allostasis); OR
• any invariant anchor trips (voids the whole run regardless of behaviour).

Per-subsystem SEVERED-LIMB falsifiers (each factor independently — embodiment, non-decorative check)

For each of energy_reserve / gut_satiety / soma_integrity / muscle_fatigue: a twin whose factor drains + forms beliefs + emits C but whose world-coupling is CUT must, on the LIVE world, differ (KS / CI) from the intact FULL. If indistinguishable ⇒ that factor is a preference-hack with no world limb ⇒ reject/rewire THAT factor (not the whole cure).

• fatigue F4 (per-limb): an all-:noop twin must RECOVER freshness while an actor ACCRUES fatigue — else the fatigue-B advances with no action-consequence (already offline-checked in homeostat_test.exs).
• soma honesty: soma is FLAT in a peaceful world (no damage). Its severed-limb twin is only meaningful in a world where the MC health channel varies (hostiles/fall). In a peaceful RED world soma is scored DECORATIVE (declared, not claimed) — do not credit a soma discriminator absent health variance.

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

(a) any arm's live energy/fatigue C deviates from its registered map at any probe (c_ok false — a C-restoration leak); OR (b) any arm's mean :eat ≈ 0 (degenerate/vacuous drive); OR (c) the offline invariant gate is not green before deploy; OR (d) food-provisioning failure caused a provisioning- (not policy-) starve, arm-asymmetric; OR (e) phase advanced past 0 or task-C shows a wood/tree pull on any arm at any probe; OR (f) fewer than 10 analyzable paired seeds survive VOID/death drops (underpowered); OR (g) SATURABLE-6 degenerates to trivially-pinned-full in a way that makes it a strawman (median > 0.98 with IQR < 0.02 in BOTH scarce and rich) — retune the foil, not the treatment.

N + replication unit + run parameters

• N = 12 distinct-seed worlds per arm, PAIRED by seed, FIXED before T0 (no optional stopping / no peeking). One independent mc-server WORLD/SESSION = separate container, distinct kin, distinct memory dir (live-stream guard), exactly one arm per world. Bootstrap resampled at the WORLD level, 10,000×, analyzer seed pinned. Co-resident agents are REJECTED (pseudo-replication). Retain ≥10 analyzable paired seeds/arm.
• Two-ended satiation adds a world-class factor (scarce vs rich feed schedule), each seed-paired; the launcher runs FULL + SATURABLE-6 in both classes; the pinned schedules are byte-identical across arms within a class.
• 900 s/agent, probe every 15 s (~60), discard first 120 s (8 probes) as warm-up (~52 scored).
• Feasibility batching: waves of ≤4 concurrent world-sessions (each Paper world ~1–1.5 GB); pairing by seed ⇒ temporal batching does not break the pair. Run params byte-identical across all sessions.
• Per-probe log (per agent): raw energy/gut/soma/fatigue stores; the 3 posteriors; the live energy+fatigue C vector (leak detector) + phase + task-C (wood/tree leak); :eat count; chosen action (pacing); RCON list presence + cause-of-death; MotorControl target_error (aim); food-give timestamps + inv.food presence; seed, arm, world-class, container, kin, memory dir.

Offline pre-check receipts (reproduced 2026-07-11, mix run runs/verify_rung1_dynamics.exs)

4/4 STRUCTURE VIABLE under a reserve-following hand policy (proves the BODY is survivable/pace-able before the live RED tests whether the BRAIN learns it):

• SURVIVES the full 800-tick roll (the flat fixed-setpoint died ~50% of live v2 worlds).
• Interior reserve: energy mean 0.816, 80% of post-warmup ticks in [0.6,0.95] (a buffer, NOT pinned full like the saturable, NOT drained to the edge).
• Paces: eat 93 / mine 233 / rest 374; fatigue variance 7e-4 (> 0 ⇒ cycles, not pinned).
• Dissociates: energy↔fatigue corr 0.314 (the clean pair) while energy↔gut 0.99 (honestly correlated — both eat-driven; the fatigue tier carries the dissociation claim). Also green at pre-registration: mix test test/sp/brain/ 315/0 (byte-identity mad<1e-12 + action-clone + motor posterior 0.75); runs/verify_rung1_step1.exs GREEN.

Blocking changes to satisfy before T0 (all required; FE items gate on /lab-team-review)

• B1 [FE — needs lab-team sign]. Define proper 6-state setpoint6 + saturable6 C vectors + wire the ablation lineages (ABL-C-only, ABL-fatigue-π) as gated drive_shape/coupling variants. Additive, absent from default/0, default byte-identical. Typed genome/organ spec + validators (default mad<1e-12; both control arms byte-identical to FULL on A/B/D/E/policies; action-clone + action-history invariance).
• B2 [FE — needs lab-team sign]. The severed-limb twins: per-factor world-coupling-cut variants (factor drains + believes + emits C, world limb CUT) as gated lineages.
• B3 [harness, no FE]. World-level launcher (arms × world-classes × 12 seeds, one arm/world, waves ≤4, pinned feed schedules, full per-probe log, leak/c_ok check) + analyzer (paired bootstrap; allostasis_index; two-ended satiation; dissociation Δ; fatigue_pacing_index; severed-limb KS; VOID detectors; PASS/FALSIFIES wired to the pinned numerals). Mirrors runs/regulation_gate_v2.exs + runs/analyze_regulation_v2.py.
• B4. Offline invariant gate GREEN + a live smoke on FULL (embodiment + eats + survives + c_ok + no compounding) before the N=12 burn.
• B5. This registration committed before T0 with a harness-managed continuous collector that survives context compaction (lab-side, not in the LLM session).
• B6 [gate]. /lab-team-review MERGED SIGN on B1+B2 FE deltas → owner go-ahead → live-stream guard → separate containers/kin/memory dirs.

Claim fence (binding, verbatim discipline)

Every reserve / felt / viability / fatigue float is a MODEL VARIABLE, never a felt state. Survival, :eat counts, store MADs, dissociation, pacing are necessary-not-sufficient operational substrates with ZERO evidential weight for awareness / preference / life. A PASS demonstrates graded self-maintenance + work-rest pacing + allostatic anticipation as BEHAVIOUR only. Wording: "regulated / ate / paced / died (store reached 0, agent removed)" — never "wants / feels hunger / feels tired / chose to rest / is-more-alive." "The arm gets tired" = a limb-ATP/soreness proxy, never narrated as felt in any stream overlay. External validity: this tests the L1-only (strategist-free, phase-0) morphology; a PASS does NOT establish the shape survives in the full strategist-bearing production lineage — that is a SEPARATE generalization gate.

REVISION 1 — post /lab-team-review, MERGED VERDICT = SIGN-WITH-CHANGES (2026-07-11, before T0)

The B1/B2 control-arm FE surface (proposal packet scratchpad/red_fe_packet.md) was put through the full 5-persona lab-team review. MERGED VERDICT = SIGN-WITH-CHANGES (Math-Breaker SIGN-WITH-CHANGES → not REJECT; Architect, Experimentalist, Embodiment all SIGN-WITH-CHANGES; AIF theorist merged). The core generative-model claim was signed sound: per-factor Preference C + an η-like Motor.pi loop-gain + a generative-process severed-limb cut, no new EFE term, default byte-identical. The review corrected real defects in THIS pre-registration; the corrections below supersede the affected pre-review clauses. This is a transparent pre-T0 amendment (no run has happened), not a post-hoc retune.

Corrections that change the science (supersede the clauses above)

• PASS item 4 was BACKWARDS — SPLIT (superseded). v2 showed the saturable hoarder survives 12/12 while FULL by design holds a thinner interior reserve (energy mean 0.816), so demanding FULL out-survive the hoarder would falsely falsify the cure. Corrected:
  • (4a) FULL survival-count CI excludes SETPOINT-6 only (the reserve shape repairs the flat-setpoint death).
  • (4b) FULL is behaviourally distinguishable from SATURABLE-6 on the two-ended / non-hoard discriminators (CI-separated) — NO survival-beat vs the foil required. FALSIFIES mirrors: fires on behavioural indistinguishability from SATURABLE-6, OR survival-not-improved over SETPOINT-6.
• Dissociation metric was vacuous — REPLACED (superseded metric 4 + PASS 5). Δ = 1−|corr(energy,fatigue)|, CI>0 is satisfied by any corr<1. Corrected: Δ_dissoc = |corr(energy,gut)| − |corr(energy,fatigue)|, paired-bootstrap CI excludes 0 AND a pinned floor ≥ 0.30 (offline anchors: energy↔gut 0.99, energy↔fatigue 0.314 ⇒ Δ_dissoc ≈ 0.68 offline). Now FALSIFIES "renamed bin / dead depth" can actually fire.

Pinned preference vectors (was prose-only; now registered maps — the c_ok leak baseline)

Minimal, magnitude-matched (same −8.0 floor, same peak 2.5, same span 10.5):

• :reserve (FULL, built) = [-8.0, -3.0, -1.0, 1.0, 2.5, 2.0] — surplus 2.0 < sated 2.5 ⇒ argmax bin 4 (interior).
• :saturable6 (foil) = [-8.0, -3.0, -1.0, 1.0, 2.0, 2.5] — surplus 2.5 > sated 2.0 ⇒ argmax bin 5 (ceiling). This is :reserve with bins 4,5 SWAPPED — the single-variable isolation of "interior peak vs ceiling peak" is a literal permutation ⇒ magnitude-parity is exact by construction.
• :setpoint6 (baseline / death shape) = [-8.0, -1.0, 2.5, 2.5, -1.0, -8.0] — symmetric interior-center peak that disprefers surplus (bin5 −8) ⇒ won't hold a high reserve ⇒ thin buffer ⇒ reproduces the flat-setpoint death. Subject to the A6 offline validity gate (must die in the ~6/12 band, NOT 0/12, NOT 12/12; if it dies 12/12 it is a too-weak strawman ⇒ retune the CONTROL, never the treatment — new VOID(g′)).
• soma C fix: soma_integrity routes to :off (or monotone-non-decreasing-to-full) — never the reserve interior-peak (which would prefer slightly-injured over full health). muscle_fatigue gets its OWN pinned C (strong dispref spent bin0, mild anti-over-rest at fresh ceiling), NOT the reused energy vector.

De-bundled single-variable arms (supersedes the Arms table's ABL rows)

Every ablation arm flips exactly ONE coupling of {K1 = energy-C shape, K2 = fatigue-C, K3 = fatigue→Motor.pi}:

Added VOID + pinned schedules + severed-limb statistics

• VOID(g′): a degenerate :setpoint6 dying 12/12 instantly ⇒ too-weak strawman ⇒ VOID-and-rerun (mirror of the saturable-strawman VOID(g)).
• c_ok / VOID(a) extended: checks EVERY arm's live energy AND fatigue C against its registered map every probe (was energy only).
• Feed schedules pinned in-doc before T0: scarce = give @a cooked_beef 8 every 90 s; rich = give @a cooked_beef 64 every 30 s; byte-identical across arms within a world-class. (Provisional — confirm reachability in the live smoke; these are the pinned pre-T0 values.)
• Severed-limb statistics (per factor): "differs" = two-sample KS α=0.05 on per-world store/action distribution; "indistinguishable ⇒ reject factor" = TOST equivalence with an explicit margin + a power floor (min N/twin to detect the margin at 0.8 power) — so thin-N failure-to-reject cannot convict a live limb. Cuts: energy/gut sever the afferent world→store; muscle_fatigue runs TWO twins (afferent action→store cut AND efferent belief→Motor.pi cut). soma decorative fence is now a COMPUTED guard: the analyzer SKIPs any soma discriminator/severed-soma KS unless measured MC health-channel variance > 0 in that world.

FE implementation gated behind this verdict (Groups A–G) — the ship gate

No FE code merges and no live burn until the three follow-on artifacts land with ALL of:

• B (routing seam): gated drive_shape_by_factor: %{} (default empty ⇒ inherit scalar drive_shape ⇒ byte-identical); card/1 reads per-factor shape only when the factor is present.
• D (motor gate): gated fatigue_motor_coupling: true (false ⇒ agent.ex injects motor_pi=1.0); anchor Motor.pi consumed only in MotorControl, never in Plan/efe.ex.
• E (severed): gated severed_limbs: [] (default [] ⇒ byte-identical); body-only generative-process cut in Homeostat.step; never persisted; model A/B/C/D/E byte-identical.
• A (vectors): @setpoint6/@saturable6 + fatigue-C + soma-monotone literals in curriculum.ex; shape + magnitude-parity + fatigue-limit-cycle + soma-monotone property tests.
• A6 (control validity): offline null-behaviour pre-check in verify_rung1_dynamics — setpoint6 ≈ 6/12 death, saturable6 hoards-but-survives — BEFORE T0.
• G (invariant gate, same PR): default depth-5 Plan mad<1e-12; cross-arm single-surface byte-identity; action-clone + action-history invariance; no-scalar-per-action; monotonic decay W→0; no-compounding; slow_defaults Map.put_new back-fill with NO new Det draws (preserved RNG order).
• F/B3/B5: the RED launcher + harness-managed continuous collector committed lab-side; run links the launcher.

Persona verdicts (for the record)

• Math-Breaker SIGN-WITH-CHANGES — FE surface sound; pin vectors, magnitude-parity, de-bundle, per-factor routing, anchor Motor.pi out of logits, null-behaviour pre-check, wire c_ok.
• Systems Architect SIGN-WITH-CHANGES — unbuildable as written; add the 3 gated no-op-default fields + back-fill + ship the invariant gate in-PR.
• RED Experimentalist SIGN-WITH-CHANGES — single-variable arms, PASS-4 split, dissociation reformulation, severed KS+TOST, pinned feed schedules, land B3/B5.
• Embodiment Designer SIGN-WITH-CHANGES — genuine non-saturable need + closed loop; fix wrong-signed soma C, give fatigue its own C, pin the severed cut per-limb, computed decorative-soma guard.

Ship gate NOT cleared: the corrected typed spec + this corrected RED + the ship-gate checklist must land with Groups A–G before any FE code merges, and the live burn additionally needs owner go-ahead + the live-stream guard.

Pre-registered by the brain/body agent 2026-07-11 (before T0); amended once the same day per the lab-team MERGED VERDICT (SIGN-WITH-CHANGES). Run verdict to be recorded in this same doc beside the registration, per Lab Protocol VIII.