# Эксперимент: Noncrossing Brace Attention (#1186)

{
  "worked": false,
  "confidence": 8,
  "verdict": "Built a PyTorch MVP implementing ordered, nonoverlapping interval routing for two lightweight experts, plus a numerical support/count and softmax-convexity check. The brace mask matched the expected 210 ordered interval pairs for L=6 and had zero routing collisions, but it did not improve the toy task: accuracy was lower than the independent router and runtime was about 3.6x higher. Therefore the structural effect was observed, but the claimed accuracy/stability win was not demonstrated.",
  "metrics": {
    "baseline": "Independent routing: test loss 0.6933, accuracy 0.5469, collision rate 0.3086, 15.84 s, 276 parameters",
    "idea": "Noncrossing brace routing: test loss 0.6925, accuracy 0.5156, collision rate 0.0000, 57.68 s, 276 parameters; math check found 210/210 valid ordered pairs and softmax sum 0.9999999"
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 brace_mvp.py",
  "files": [
    "brace_mvp.py"
  ],
  "limitations": "Only a small synthetic classification task was tested; no transformer language modeling, real attention workload, FLOP-normalized comparison, gradient-variance measurement, multi-seed confidence interval, GPU execution, or larger-scale evaluation was performed."
}