# Эксперимент: Task-Gated Diverse Counterfactuals (#996)

{
  "worked": true,
  "confidence": 8,
  "verdict": "Built a reproducible task-gated log-determinant branch selector with greedy selection, conflict projection, and exhaustive submodularity checking. The toy verification confirmed the predicted mechanism: low-gate information scaled approximately quadratically, overlap reduced marginal gain monotonically, and greedy selection beat random across budgets. The unsafe high-information branch was rejected; this is a mechanism-level win, not evidence of improved policy return.",
  "metrics": {
    "baseline": "Random branch selection: mean F=2.971 at K=3; K-sweep F=[1.217, 2.194, 2.993, 3.612, 4.155] for K=1..5.",
    "idea": "Greedy gated selection: F=4.272 at K=3; K-sweep F=[2.648, 3.673, 4.272, 4.481, 4.620]. Low-alpha F/alpha^2 spread was 3.24%; overlap marginal gain decreased from 1.025 to 0.569; minimum submodularity slack was -8.9e-16, consistent with floating-point zero. Conflict filtering removed the selected unsafe branch and produced safe set [0,2,3]."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 verify_task_gated.py",
  "files": [
    "verify_task_gated.py",
    "results.json"
  ],
  "limitations": "Only a finite-dimensional synthetic information-selection benchmark was tested. No learned dynamics ensemble, obstacle-navigation environment, policy update, model-call wall-clock comparison, synthetic-data bias, multi-seed variance, or real constraint projection was evaluated."
}