# Эксперимент: Orbit-Consistent Equivariant Distillation (#1118)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built a C4 orbit-consistent policy/value distillation MVP with an MLP student, explicit state/action rotations, equivariance and value-invariance penalties, numerical identity checks, and lambda sweeps. The defining geometry passed exactly, and equivariance improved from 0.01591 to 0.00635 at lambda_eq=0.3, but the effect was not monotonic, worsened at lambda_eq=1.0, and achieved only a 2.5-fold reduction rather than the predicted 5-fold reduction. Value invariance was also non-monotonic across lambda_v, so the required mechanism-level predictions were not confirmed.",
  "metrics": {
    "baseline": "lambda_eq=0: equivariance error 0.01591, transformed MSE 0.00015996, value invariance 0.00924",
    "idea": "lambda_eq=1.0: equivariance error 0.02230, transformed MSE 0.00020098, value invariance 0.01279; best observed lambda_eq=0.3 gave equivariance error 0.00635 and transformed MSE 0.00012296"
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 run_experiment.py",
  "files": [
    "run_experiment.py",
    "results.json"
  ],
  "limitations": "Only a synthetic 2D C4 toy task was tested; no image/VLA encoder, stochastic policy, RL environment, held-out orbit split, multi-seed confidence intervals, FLOP/speed comparison, or larger symmetry groups were evaluated."
}