# Эксперимент: Channel-aware attention-head pruning (#916)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built a runnable channel-aware attention similarity MVP with Aitchison content/sink diagnostics and a toy pruning-selection test. The mechanism manifested: content distance was invariant to sink mass, equal-content full-distance scaling matched 2m/(m+1) to numerical precision, and channel thresholds rejected a raw-cosine false redundancy while accepting the true duplicate. However, the authoritative full-distance decomposition in the idea is mathematically incorrect as written: its median relative error was 0.59–0.96, whereas the corrected decomposition including the content geometric-mean term had less than 3e-15 error, so worked=false under the required math criterion.",
  "metrics": {
    "baseline": "Raw cosine selected the intended duplicate first but also ranked a same-sink/different-content pair next (cosine distance 0.000877; content distance 5.2082).",
    "idea": "Channel-aware thresholds accepted only the true duplicate pair [0,3] (content distance 0.00004, sink distance 0), rejecting the false same-sink pair; content sink-invariance error was <=1.8e-15 and corrected decomposition error was <=2.5e-15."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py",
    "results.json",
    "run_output.txt"
  ],
  "limitations": "Only a synthetic attention-row experiment was tested; no Transformer surgery, ablation importance estimation, recovery fine-tuning, perplexity, latency, FLOP, or memory measurements were run. The stated formula should be corrected before applying this method to real head pruning."
}