SkewedBell GearboxX feedforward control failed to achieve x-axis restoration (x_dev_max=0.0402 vs baseline 0.0205) due to a positive-feedback sign-convention mismatch.
A predictor minimizing observational error does not necessarily minimize closed-loop control error if the feedback sign convention is incorrect, leading to destabilizing positive feedback.
Absolute Event Rate: 0.0402% vs 0.0205%
We present PHILIA-EcoSensory Swarm v56 Phase 3, a pre-registered 3-cell head-to-head ablation study (500 runs, N=20/40, seeds 482–521, T=80, 000 steps) designed to determine whether SkewedBell (SB₃param) GearboxX feedforward control achieves x-axis RESTORATION (xdev < 0. 010) in the PHILIA Phase Vector Φ= (x, y, z). Building on v55 Phase 2B (DOI: 10. 5281/zenodo. 20025150), which established Q3=INTRINSIC (x-drift as emergent property) and Q4=PARTIAL (linear GearboxX, 50% reduction), v56 pre-registered the hypothesis that a SkewedBell-shaped correction (LOOCV RMSE=0. 000436, R²=0. 9995, approximately 76% improvement over Linear) would achieve full RESTORATION. Primary finding: Q6 = FAIL → §Conclusions = CASEK. F3 (SB₃param) yielded xdevₘax=0. 0402, more than double the baseline (F1=0. 0205), while F2 (Linear, v55 E5 replication) achieved xdev=0. 0100. Post-hoc code analysis confirmed a control-sign convention mismatch: gearboxₓₛb3correction (w) applied +sbₓ (w) (positive, drift-amplifying) instead of −sbₓ (w) (negative, drift-suppressing), constituting a feedback-direction inversion rather than a simple coding error. The quantitative signature is highly consistent: |F3−F1| ≈ 0. 020 ≈ SB-predicted drift magnitude, consistent with drift doubling by positive feedback. Key finding: A predictor minimizing observational error does not necessarily minimize closed-loop control error. Descriptor-level fit quality can remain excellent (R²=0. 9995) while feedback control becomes destabilizing under incorrect sign convention. This ambiguity was not detected during pre-registration review (크리스 0번 관문 16/16 PASS, 4-AI sequential review) — recorded as a methodological limitation; v57 will add uₓ sign sanity check to §7 pre-registration items. All results executed locally (AMD Ryzen 9 9800X3D + RTX 4080 SUPER 16GB). Pre-registration: philiaᵥ56ₚreregistrationfinal. md. Sequential review: 판도라→루카스→크리스 PASS. v55 DOI: 10. 5281/zenodo. 20025150v54 DOI: 10. 5281/zenodo. 20003245v53 DOI: 10. 5281/zenodo. 19974690v50 DOI: 10. 5281/zenodo. 19887194v48 DOI: 10. 5281/zenodo. 19815222 실측이 생명. 증명이 아닌 기술. Measurement is life. Description, not proof. — Trinity AI Research Team | May 2026 | PHILIA OS — 0∞1∞0. 5∞
신두섭 (2026) studied this question. SkewedBell (SB_3param) GearboxX feedforward control vs. Baseline (F1) and Linear GearboxX (F2) was evaluated on x-axis RESTORATION (x_dev < 0.010). SkewedBell GearboxX feedforward control failed to achieve x-axis restoration (x_dev_max=0.0402 vs baseline 0.0205) due to a positive-feedback sign-convention mismatch.