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

1.3 KB
Raw Blame
{
  "skill": "model-selection",
  "primary_value_surface": "routing/discoverability",
  "primary_metric": "routing_accuracy",
  "secondary_metrics": ["top_k_routing_accuracy", "model_fit_justification_quality", "fallback_appropriateness"],
  "distractors": {
    "off_target_set": "Problems with obvious single-model solutions; problems where model selection adds no value over direct invocation; meta-questions about model selection itself",
    "description": "Items where the model choice is obvious or where selection overhead is unwarranted"
  },
  "calibration_range": [0.40, 0.70],
  "calibration_exception": null,
  "min_sample_primary": 80,
  "min_sample_replication": 80,
  "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 model_fit_justification_quality < 0.60",
    "ceiling_needs_harder_data": "placebo_baseline > 0.70"
  },
  "eval_family": "Routing/abstention"
}