PHILIA-EcoSensory Swarm v54 Phase 2A: Functional-Form Robustness of GearboxY within Tested Function Classes under SkewedBell GearboxZ — A Pre-registered 5-Cell Ablation Study

PHILIA-EcoSensory Swarm v54 Phase 2A presents a pre-registered 5-cell ablation study of GearboxY functional forms under SkewedBell GearboxZ (600 runs, N=20/40, seeds 382–441). Building on v53's CONDITIONAL verdict for y(w) independence and v52+'s SkewedBell geometry (ΔAIC=65.5), four GearboxY functional forms were tested: linear (C2), SkewedBell_bounded (C3), Cubic hybrid α=0.7 (C4), and Cubic pure α=1.0 (C5). Primary finding (Q1): NULL — SkewedBell GearboxY produces no meaningful improvement over linear probe (Δ=−0.0006, Cohen's d=−0.134). All four forms show operationally equivalent y-restoration under pre-registered thresholds (Y_res = 0.0073–0.0082, |d| < 0.15). Linear probe is sufficient and parsimonious. Model verification (Q2b): EQUIVALENT — SB is a valid representation but not superior to Cubic pure (Δ=−0.0003, d=−0.064). Mechanistic interpretation: GearboxY operates as an attraction mechanism toward PHI_NATURAL[y]=0.19, not a target-tracking mechanism. Functional form determines aim direction only; convergence to 0.19 is architecture-driven. This distinguishes structural nonlinearity from operational necessity. Exploratory finding: x-axis drift (max |x−0.34| ≈ 0.026) confirmed GearboxY-independent, replicated across all GearboxZ ON cells (v53 Appendix A robust replication). phi_actual_x slope (+0.030~+0.0308) replicates v52+ retroactive finding (+0.033). Phase 2B (COUPLE_XZ ablation) justified. All thresholds pre-registered (philia_phase2a_config.py). Adversarial review: 크리스 0번 관문 11/11 PASS. Independent verification: compare_report.py (JSONL-based recomputation). Execution: AMD Ryzen 9 9800X3D + RTX 4080 SUPER 16GB, Python 3.12 + numba, Windows 11 Home. 실측이 생명. 증명이 아닌 기술. — Trinity AI Research Team | 2026.05.03