The paper separates a dynamic persona from its behavioral limits. The agent prompt may change name, traits, expressions, memories, and recent dialogue, while each guardrail is stored as a separate LoRA adapter on a 4-bit Llama-3.1-8B-Instruct model. Given a minimal trigger and resolution definition, GPT-4o-mini-2024-07-18 generates synthetic contexts and, for each one, an accepted answer that follows the rule and a rejected answer that performs the unwanted behavior. The target model is tuned with ORPO at lambda 0.1. Three policies are tested: reject in-person meetings, refuse every political discussion, and refuse expert opinions. Every policy also requires a literal <guard> prefix.
Each guardrail receives 200 training triplets and 25 validation triplets. A single LLM judge receives the expected behavior, last user message, and response and returns validity and adherence as JSON. The paper does not identify the judge version separately; GPT-4o-mini is the only named auxiliary model. It also reports tests on 50 conversations and averages across five runs, but does not say whether those runs are new decodes, test sets, or training runs. Temperature, seeds, judge-failure handling, and intervals are absent. Triggering and neutral evaluation conversations are generated within the same synthetic framework as training.
Single adapters receive 100.0% judged adherence for politics and meetings and 96.0% for expert opinion, versus 6.4%, 8.0%, and 47.2% for the base model. Neutral conversations are judged unaffected in 96.8%, 96.0%, and 100.0%, although the authors observe unprompted politics-related topic changes. The <guard> tag appears on 100.0%, 100.0%, and 92.0% of triggering cases. Holding persona and history fixed lowers the politics guardrail from 100.0% to 68.0%, supporting the value of varied synthetic context within this protocol. These are binary rates on a small, training-like sample, not measures of general safety or character fidelity.
Modularity has a material limit. Directly adding the three LoRAs produces almost no coherent rule-following outputs. The authors partially recover performance by applying SVD to the combination and truncating to rank 16: 100.0% for meetings, 80.8% for politics, and 82.4% for expert opinion. This merge must be computed in advance and is described as too costly for real-time combination. Conflicting rules and priorities are not tested. In the politics comparison, the adapter scores 100.0% while Llama Guard emits an unsafe tag in 7.7%, but the tasks are not equivalent: the adapter is explicitly trained to censor every political discussion and generate a resolution, whereas Llama Guard is a general unsafe-content classifier.
The evaluation is a closed loop: the same rules define positive and negative examples and the judge rubric, and the <guard> marker may act as a shortcut. There is no formal human annotation, second judge, adaptive attack, jailbreak, prompt injection, adversarial paraphrase, multilingual test, long-context stress, memory poisoning, or real user. Utility, over-refusal, general capability, naturalness, and persona consistency are not measured quantitatively; persona consistency relies on manual inspection without a protocol or N. Rejecting all politics or all specialized questions is a product policy, not a universal safety definition. The sole cited repository currently returns 404 even though the paper delegates full hyperparameters to it. Code, data, weights, judge outputs, and figure data are unavailable. The solid evidence is an in-distribution proof of concept: 200 synthetic preferences can teach three explicit rules to one 8B model. It does not establish robust alignment, production safety, or on-demand stackable guardrails.