Shieldstral is a 3 billion parameter open-source safety classifier that treats content moderation as a question-answering task. Instead of baking fixed harm categories into model weights, you supply moderation policies as plain-language queries at inference time—"Does this promote violence?" or "Is this safe for minors?"—and it returns a calibrated probability score. The claimed advantage: it works on text and images without retraining, runs on a single 16GB GPU, and reportedly matches guardrail models seven times its size on standard benchmarks. The source benchmarks four dimensions (text safety, refusal detection, policy adaptability, multimodal) but doesn't show the actual numbers—just relative performance claims. For practitioners: this trades the retraining burden of traditional guardrails for prompt-level policy specification, which could reduce friction if your safety definitions shift. The real test would be how well it generalizes to *your* specific policy boundaries and edge cases, which the source doesn't address.