# Эксперимент: Shared Symbolic Mechanism Bottleneck (#567)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built the differentiable shared symbolic bottleneck with gated feature primitives, soft operator selection, sparse readouts, temperature annealing, and an MLP comparison. Three quantitative sanity predictions were confirmed: softmax entropy followed the temperature law, exact shared-latent reconstruction residual was 3.13e-32, and noise MSE tracked sigma^2. However, the trained symbolic model had far worse in-range MSE than the MLP (0.0801 vs 0.000285), did not recover a common denominator, and therefore the proposed learned mechanism did not demonstrate the claimed overall effect despite better extrapolation.",
  "metrics": {
    "baseline": "Shared MLP: in-range MSE 0.000284852; extrapolation MSE 0.00317361",
    "idea": "Shared symbolic bottleneck: in-range MSE 0.0800653; extrapolation MSE 0.0449622; 5 active units; no common-denominator recovery",
    "mechanism_checks": "Softmax entropy matched analytic values across tau=1,0.5,0.25,0.1,0.05; shared-latent residual 3.13e-32; measured noise MSE for sigma=0.01,0.03,0.1 was 0.00010046,0.00090141,0.0100255 versus predictions 0.0001,0.0009,0.01"
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py",
    "results.json",
    "REPORT.md"
  ],
  "limitations": "Only one seed, one small synthetic kinetics setup, and one 500-epoch optimization configuration were tested. The exact shared-latent and noise checks validate algebraic consequences rather than proving that gradient training discovers the intended denominator; no hard pruning/fine-tuning or mutation search was evaluated."
}