Benchmarking System One Models in Online Moderation

Abstract: Online moderation systems must apply changing platform policies, community rules, and prior decisions while producing decisions that can be audited and routed to human review. We evaluate whether System One Models, which accept natural-language context but return typed choices, probabilities, or scores, can support this setting. Across five moderation benchmarks, Jev is competitive with specialized reference systems, matching or exceeding them in several policy-grounded and harmful-content settings. We then use controlled information conditions to separate written rules, retrieved precedents, and restrictions on the candidate answer space. Jev generally benefits from retrieved precedents, improving exact policy selection and harmful-content discrimination when the answer space is held fixed. Laya is less consistent: retrieval often shifts its positive prediction rate or no-violation rate without improving discrimination. Jev confidence can support selective review in several settings, although it is not consistently calibrated; Laya confidence is less useful for ranking errors. These results show a promising future for SOM-powered content moderation.