Scientific posture. This paper reports empirical results from a controlled ablation study of a single system under synthetic task conditions. It does not claim validation of broader theoretical frameworks, biological relevance, or computational necessity of any particular mathematical structure.

The Question

Multiple structural hypotheses exist for hierarchical reasoning: specific graph topologies (600-cell), golden-ratio frequency scaling, spectral preconditioning, cross-layer coupling. Which ones actually drive performance?

This study systematically isolates each factor through seven controlled experiments.


Seven Experiments

  1. Frequency hierarchy analysis — does hierarchical frequency structure exist?
  2. Emergent ratio analysis — does φ-like scaling emerge from the geometry?
  3. Regime performance comparison — polytope regime vs dyadic baseline
  4. Coupling suppression dose-response — is coupling causally linked to performance?
  5. Topology comparison — does graph structure matter under matched dynamics?
  6. Full-stack feature ablation — which individual features drive performance?
  7. Spectrum sensitivity — does the specific distribution of learning rates matter?

The Dominant Factors

Key Result

Learning-rate scale is the dominant factor. Removing eigenvalue-scaled rates degraded quality by 15.45% ± 1.2% — approximately 50× larger than any other individually ablated feature.

Cross-layer coupling is causally beneficial (r = 0.96, p < 0.01, monotonic dose-response across seven suppression levels).


What Didn't Matter

FeatureAblation EffectVerdict
Learning-rate scale15.45%Dominant
Cross-layer coupling0.29% (but r=0.96 in dose-response)Causally important
Shell structure0.02%Negligible
One-way crystallisation0.00%No effect
Ico/dod populations−2.25%Negative

Topology comparison: the complete graph performed comparably or better than the 600-cell under matched dynamics. Six different spectral distributions produced performance within 0.13% of each other.


The Role of the 600-Cell

The 600-cell provided one principled route into a productive rate regime — its eigenvalue spectrum yields learning rates in the right range — but other rate distributions in the same range performed comparably.

The benefit comes from operating in the correct regime, not from the specific geometry.


Scope of Claims

  • No claim about biological neural systems
  • No claim that φ is necessary for intelligence or reasoning
  • No claim that the 600-cell is a required computational substrate
  • No validation of broader physics theory
  • System-specific conclusions only — one system, synthetic tasks

Performance was dominated by learning-rate scale and cross-layer coupling. What the system is built from mattered less than how it was tuned.

Code & Reproducibility. Seven Python scripts, structured JSON results, and full LaTeX source available at github.com/vfd-org/hierarchical-reasoning-ablation-study.