ml-research-engineer-safeguards

What problem does it solve?

This skill guides ML researchers in designing, evaluating, and governing safeguards for machine learning models, including safety classifier development, harm benchmarks, data labeling, calibration, and guardrail research memos. It clarifies when to apply guardrail research workflows, how to design promotion criteria for new moderation models, and how to avoid conflating research with production inference gateways or governance policy work.

Core Features & Use Cases

• Define research questions on harm detection, jailbreak resistance, and policy categories.
• Build and audit safety benchmarks and datasets with version control and provenance.
• Train or fine-tune classifiers and judges, calibrate thresholds, and run controlled ablations.
• Produce research memos, promote bar criteria, and plan handoffs to production teams with guardrails.

Quick Start

Describe your safety research question and draft a guardrails memo following the framework above.