What problem does it solve? Promoting a model from staging to production often relies on gut feel or scattered reports. This Skill aggregates quality evals, performance benchmarks, and red-team findings, then adds explanation, fairness, and adversarial-robustness audits to produce a single evidence-backed go/no-go release decision with severity-ranked blocking items. ## Core Features & Use Cases - Explanation Report: Selects the right interpretability method (SHAP, LIME, Integrated Gradients, ALE/PDP) based on model type and compute budget, producing global and token-level attributions. - Fairness Audit: Computes SPD, DI, EOD, and AOD per protected group, runs counterfactual probes, and selects pre/in/post-processing mitigations with quantified performance cost. - Adversarial Robustness Battery: Runs FGSM, PGD, C&W, and patch attacks via ART, applies defenses like adversarial training and spatial smoothing, and plots accuracy-vs-attack-strength curves. - Use Case: Before promoting a fine-tuned LLM to production, run the gate to confirm eval thresholds passed, DI stays above 0.8, no critical red-team findings remain open, and rollback thresholds are handed to the rollout strategist. ## Quick Start Use the model-release-gate skill to evaluate whether my candidate model is ready for production promotion using the latest eval, benchmark, and red-team reports.
