# Эксперимент: Spectral pinning of neural modules (#1125)

{
  "worked": true,
  "confidence": 8,
  "verdict": "Built a deterministic spectral-pinning experiment for weighted graph modules with greedy grounded-Laplacian pin selection and exact Kronecker disagreement dynamics. Measured late-time energy-decay slopes matched -2c lambda_min(L+P)lambda_min(Gamma) within 3.2% across coupling values and within 0.02% across pin strengths, satisfying the 25% criterion. Greedy pins produced a larger gap than random selection, although highest-degree selection tied greedy on this graph and diminishing returns were not strictly monotonic over the full sweep.",
  "metrics": {
    "baseline": "Random two-pin sets: mean grounded gap 0.146681; mean observed decay slope -0.284277",
    "idea": "Greedy pins [3, 6]: grounded gap 0.176882; observed decay slope -0.342809; theoretical slope -0.342808; maximum rate-prediction relative error 3.20%"
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 spectral_pinning_experiment.py",
  "files": [
    "spectral_pinning_experiment.py",
    "results.json",
    "run_output.txt"
  ],
  "limitations": "This verifies the linear spectral mechanism rather than training real MLPs, MoE experts, or distributed workers; no accuracy, throughput, communication cost, asynchronous noise, or learned graph-weight estimation was tested. The graph is small and synthetic, and greedy did not outperform highest-degree selection in this instance."
}