Skip to content

28 eval-informed mental models and critical-thinking skills for Claude Code, GitHub Copilot, Codex, Cursor, and other Agent Skills-compatible tools

1.4 KB
Raw Blame
{
  "skill": "bayesian",
  "primary_value_surface": "calibrated classification",
  "primary_metric": "posterior_calibration_error",
  "secondary_metrics": ["prior_elicitation_quality", "likelihood_assessment_accuracy", "update_correctness", "base_rate_adherence"],
  "distractors": {
    "off_target_set": "Questions with definitive answers; situations with no meaningful prior; meta-questions about Bayesian reasoning",
    "description": "Items where probability estimation adds no value over direct knowledge or where priors cannot be meaningfully established"
  },
  "calibration_range": [0.40, 0.70],
  "calibration_exception": null,
  "min_sample_primary": 60,
  "min_sample_replication": 60,
  "planned_comparison": "placebo",
  "decision_gate": {
    "elevate": "primary_delta_pp >= 5 AND paired_test_p < 0.05 AND replication_delta_pp >= 5 AND replication_paired_test_p < 0.05 AND replication_direction_matches_primary AND posterior_calibration_error < 0.10",
    "directional_not_replicated": "primary_delta_pp >= 5 AND paired_test_p < 0.05 AND (replication_delta_pp < 5 OR replication_paired_test_p >= 0.05 OR replication_direction_opposite)",
    "no_lift": "primary_delta_pp < 5 OR paired_test_p >= 0.05",
    "quarantine_redirect": "primary_delta_pp < 0 OR base_rate_adherence < 0.50",
    "ceiling_needs_harder_data": "placebo_baseline > 0.70"
  },
  "eval_family": "Quantitative/uncertainty"
}