# Эксперимент: Sink-content Aitchison distillation (#915)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built a reproducible numerical verifier and sink-heavy toy distillation comparison in experiment.py. The exact pairwise Aitchison definition exhibits the predicted quadratic sink and clr-content scaling, and the entropy identity is exact, but the authoritative decomposition is wrong by a factor of 2 for the sink term (observed 0.8 vs claimed 0.4); therefore the required mechanism as stated does not fully validate. In the toy optimization the two-channel loss improved final KL (0.005845 vs 0.014767) and eliminated sink error, but had substantially larger peak gradients (8.37 vs 0.485), so this is only a secondary promising signal, not enough to override the failed claim.",
  "metrics": {
    "baseline": "KL: final KL=0.0147666, sink absolute error=0.0221955, content Aitchison squared=1.00569, max gradient=0.485",
    "idea": "Sink-content Aitchison: final KL=0.0058452, sink absolute error=0.0, content Aitchison squared=0.988942, max gradient=8.369"
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py",
    "results.json"
  ],
  "limitations": "No full Transformer, language-model, GLUE, or equal-FLOP experiment was run. The toy uses a shared low-capacity student and does not test masking, multiple layers/heads, loss-weight tuning, or GPU execution."
}