assess-sensitivity

assess-sensitivity

Purpose

Perturb a specified input, assumption, model choice, analysis choice, or weight and measure the resulting output sensitivity.

Input contract

required: [baseline_input, perturbation_axes, output, comparison_metric]
optional: [assumptions, perturbation_design, uncertainty_model, baseline_ranking]
constraints: [each perturbation is attributable to one declared axis; the output comparison metric is fixed before evaluation; preserve the caller's scale and direction]

Procedure

1. Define the baseline input, output, perturbation axes, and comparison metric.
2. Generate the caller-specified perturbation scenarios, including weight or leave-one-out variants where applicable.
3. Recompute the output for each scenario and retain the scenario-level evidence.
4. Compare scenarios, identify unstable regions and dominant drivers, and classify convergence or instability.
5. Return a sensitivity report with rankings, effect magnitudes, and rationale.

Output contract

produces: [sensitivity_report, scenario_results, instability_regions, dominant_drivers, stability_verdict]
delta_fields: [findings, evidence_updates, decisions, uncertainties]

Quality gates

• Preserve source weight perturbation of +/-20% per dimension, weight-vector sum 1.0 with +/-0.001 tolerance, at least 4 scenarios, and stability labels stable/sensitive/unstable with $\tau$ bands >=0.8, 0.5-<0.8, and <0.5.
• Rank comparison reports at least one Kendall $\tau$ or Spearman $\rho$ and all alternatives differing by >=2 positions.
• Conclusion sensitivity rates every assumption, identifies critical assumptions, reports interaction effects, and gives an overall robustness rating.
• Fragility index retains 0.0 as robust and 1.0 as extremely fragile; do not reinterpret it as a probability.

Parameterization

The caller must provide the baseline object/output, perturbable axes, scenario generator or bounds, comparison metric, stability labels, and any fixed statistical bands. For ranking calls provide the normalized weight vector and scoring matrix; for conclusion calls provide assumptions and challenged variants; for meta-analysis calls provide included studies, outliers, and subgroup variables.

Failure and counterexamples

Reject when the baseline is undefined, scenarios change more than one undeclared axis, the comparison metric is missing, or a stability verdict is given without scenario results. Do not infer causal dominance from sensitivity alone.

Provenance map

• resolved: hypothesis-formation/weight-perturbation
• resolved: convergence/rank-comparison
• resolved: convergence/method-sensitivity-report
• resolved: deep-insight/conclusion-sensitivity-measurement
• resolved: stress-test/fragility-measurement
• resolved: convergence/conclusion-sensitivity
• resolved: knowledge-acquisition/sensitivity-analysis-design
• intermediate: Pass3/assess-rank-robustness
• intermediate: Pass3/measure-sensitivity
• intermediate: Pass3/design-meta-sensitivity
• resolved: deep-insight/convergence-assessment

Verbatim source criteria excerpts

• weight-perturbation line 22: The elements of the input weight vector must sum to 1.0 (+/-0.001 tolerance allowed)
• weight-perturbation line 23: The number of rows in the scoring matrix (number of gaps) must be >= 2
• weight-perturbation line 24: At least 4 perturbation scenarios must be generated (+/-20% per dimension)
• weight-perturbation line 35: stable (all scenarios $\tau$ >= 0.8) / sensitive (any scenario 0.5 <= $\tau$ < 0.8) / unstable (any scenario $\tau$ < 0.5)
• rank-comparison line 27: Must report at least one rank correlation metric (Kendall tau or Spearman rho), and must list all alternatives with ranking differences >= 2 positions.
• conclusion-sensitivity line 24: Must consider interaction effects between assumptions.
• fragility-measurement line 33: fragility_index: Overall fragility (0.0 = robust, 1.0 = extremely fragile)

Preserved source criteria ledger