# Эксперимент: Support-Scenario Attention Pruning (#1279)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built a finite-support attention pruning MVP with possibilistic-collapse verification, overlap continuation repair, greedy validity-preserving deletion, and connectivity diagnostics. The support-order equivalence passed 5,000 randomized checks, and the repaired structured mask had zero continuation violations and one connected component in all 10 seeds. However, it retained only 12 edges versus 54.4 for independent thresholding and had much worse attention-output MSE (0.629 versus 0.0176), so the claimed sparse-quality benefit was not observed.",
  "metrics": {
    "baseline": "Independent epsilon=0.08 thresholding: 54.4 +/- 6.7 edges, 0.1 continuation violations, output MSE 0.0176 +/- 0.0086, 2.8 +/- 0.4 window components.",
    "idea": "Structured repair plus greedy cleaning: 12.0 +/- 0.0 edges, 0 continuation violations, output MSE 0.6288 +/- 0.0907, exactly 1.0 window component."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 support_pruning.py",
  "files": [
    "support_pruning.py",
    "results.json"
  ],
  "limitations": "This was a tiny synthetic NumPy attention experiment, not a trained Transformer or ListOps/WikiText evaluation. It measured attention-output approximation rather than perplexity, wall-clock speed, GPU memory, or real sparse-kernel acceleration; the toy restriction and continuation definitions are simplified analogues of the paper's simplicial maps."
}