Most content-moderation models are trained against a fixed list of harm categories, which means adapting one to a new product surface usually means fine-tuning it again, and a message that is fine in a cybersecurity research tool might get flagged as dangerous in a mental-health app running the same model. Shieldstral flips that setup around: instead of baking categories into the weights, it takes the policy as a plain-language yes/no question at the moment of inference, alongside the content to be judged (text, an image, or a prompt-response pair), and returns a calibrated safety score from a single forward pass. Built on top of Mistral's small Ministral-3B model with a vision encoder attached, and released with open weights under Apache 2.0, it is small enough to run on one consumer-class GPU with 16GB of memory, which matters for teams that want moderation running inside their own infrastructure rather than sent to a third-party API, a common requirement for anything touching health data, minors, or financial records. On benchmark text-safety scores it reportedly matches models many times its size, and its multimodal scores lead the open baselines Mistral tested against, though it is noticeably weaker on lower-resource languages and on adversarial or heavily obfuscated inputs. For developers, the practical upshot is that guardrails become a per-tenant, per-surface configuration problem instead of a retraining problem: a single deployed checkpoint can enforce different rules for different customers or products just by changing the question it is asked, and because the output is a continuous score rather than a hard yes/no label, teams can route borderline cases to human review instead of auto-blocking everything near the threshold. That is a meaningfully different operating model from most safety classifiers on the market today, which usually ship as either a fixed taxonomy or an expensive reasoning-based judge.