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": "scientific-method",
  "primary_value_surface": "localization",
  "primary_metric": "fault_localization_accuracy",
  "secondary_metrics": ["hypothesis_quality", "discriminating_observation_quality", "falsification_rigor", "localization_precision"],
  "distractors": {
    "off_target_set": "Problems with single obvious cause; situations where direct inspection suffices; debugging where no observation is possible; meta-questions about the scientific method",
    "description": "Items where hypothesis-differential debugging adds no value over direct inspection or where the method cannot be applied"
  },
  "calibration_range": [0.40, 0.70],
  "calibration_exception": null,
  "min_sample_primary": 100,
  "min_sample_replication": 100,
  "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",
    "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 hypothesis_quality < 0.60",
    "ceiling_needs_harder_data": "placebo_baseline > 0.70"
  },
  "eval_family": "Debugging/fault-localization"
}