# Эксперимент: Pole-Certified SSM Initialization (#1290)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built a Hankel-pencil pole estimator with shift persistence, conjugate-pair rejection, contour winding, margin evaluation, and a deterministic extrapolation comparison. The noiseless generalized-eigenvalue identity was verified to about 8.3e-14, and the contour winding count matched the expected zero count with margin 0.0988. However, noisy certification almost always rejected all modes, so no reliable initialization benefit was observed; the proposed effect is not supported by this MVP.",
  "metrics": {
    "baseline": "Mean extrapolation MSE: 1.07e-4, 1.15e-3, and 9.30e-3 at noise 0.01, 0.035, and 0.1 respectively.",
    "idea": "Mean certified pairs: 0.2, 0.0, and 0.0 at noise 0.01, 0.035, and 0.1 respectively; MSE was infinite when no modes were certified, and 1.058e-4 in the sole successful low-noise trial."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py",
    "results.json"
  ],
  "limitations": "This is a small synthetic extrapolation test, not a trained GRU/SSM or language/long-range benchmark. The estimator uses fixed order and simple proximity-based persistence rather than a fully tuned SVD/conditioning/contour certification procedure; no GPU, learned dynamics, FLOP-matched training, or multi-dimensional probes were tested."
}