TWO GOALS COMBINED. (1) Replication of paper6 step 6 perturbation-recovery (pilot N=2 showed 79% kill / 74% reset attractor recovery; numbers UNVERIFIED). (2) Boundary-maintenance discriminator (gap.priority=0.8, SELF.boundary-maintenance): mute coupling vs verbal coupling under perturbation. Mute strips text 'current_concept' from broadcasts (numeric progress only). Tests whether self-repair ability is mediated by language content or by substrate-level structural coupling.
PRIMARY (boundary-maintenance, registered):
- LANGUAGE-MEDIATED: mute coupling produces lower attractor recovery than verbal coupling under same perturbation (Δ recovery ≥10% on max-depth ratio). Boundary maintained via verbal/symbolic content.
- SUBSTRATE-LEVEL: similar recovery (<5% Δ). Boundary-maintenance is not language-mediated; substrate-level mechanisms (numeric coupling, role structure) carry the recovery.
- GRADED: 5-10% Δ — language matters but not exclusively.
SECONDARY (step 6 replication, registered):
- SELF-REPAIR-POSITIVE: post-perturbation max_depth ratio ≥70% of baseline (matched mute or verbal arm) (Mann-Whitney p<0.05 vs random matching baseline).
- PROGRAM-FRAGILE: ratio <40% (collapse).
- PARTIAL: 40-70% (functional substitution but reduced).
- REPLICATION outcome: report mean ± SD vs pilot 79% / 74% point estimates.
KILL: 0/30 runs complete cleanly → engine surface bug, void. Round 30 = mid-run (gen 1 of 5). Allows 3 generations of post-perturbation recovery time. Same as pilot.