# Эксперимент: Forcing-Consistency Training Constraint (#491)

{
  "worked": true,
  "confidence": 8,
  "verdict": "Built a PyTorch MVP implementing grouped Jensen–Shannon forcing consistency, robust safety training for compatible latent states, and explicit empty-intersection detection. Exact numerical checks passed, while parameter sweeps showed monotone disagreement reduction and lower feasible-group safety violation. The mechanism manifested in this toy, but it increased imitation loss, so this is not evidence of a general task-performance win.",
  "metrics": {
    "baseline": "Task-only: kappa=0.6870, feasible violation=0.5000, task CE=0.000745.",
    "idea": "Consistency weight 1 and safety weight 1: kappa=0.1430, feasible violation=0.1782, task CE=0.3909; kappa reduced 79.2% and feasible violation reduced 64.4%. At consistency weight 5, kappa=0.0144; at safety weight 5, feasible violation=0.1008. Empty intersections remained explicitly certified infeasible."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 fc_experiment.py > results.json",
  "files": [
    "fc_experiment.py",
    "results.json",
    "README.txt"
  ],
  "limitations": "Only a four-history, two-action synthetic toy was tested. No GRU, recurrent hidden histories, actor-critic loop, real gridworld, learned world model, held-out latent permutations, runtime comparison, or multi-seed evaluation was performed. Safety margins were hand-specified rather than learned."
}