SIRF tackles the primary deployment requirement of industrial content risk control: achieving ultra‑low latency (second‑level or lower) while maintaining very high precision. Conventional metrics focus on average accuracy, but real‑world scenarios care about the proportion of risk that can be auto‑handled under a strict precision threshold.
The model converts the platform’s complex policies into training signals using EntiGraph, MAGA rewriting, and account‑level chain‑of‑thought (CoT), then applies continued pretraining (CPT) on top of an existing model so that the policies are directly embedded in the weights. At inference time only a verdict is emitted, eliminating extra policy reasoning and thus achieving ultra‑low latency.
In a controlled same‑source comparison, Qwen3‑8B‑SFT and SIRF‑8B‑SFT share identical policy injection and verdict‑only output; the only difference is the policy‑grounded CPT in SIRF. Results show SIRF‑8B‑SFT reaches 71.3% Black Recall@P95, a +15.1‑point gain over the baseline, using only ~70 M CPT tokens without harming general ability, and it matches or exceeds much larger systems among log‑prob‑available models.
Deployed as a tree‑model adjudication layer, SIRF recovers roughly 20% of mis‑penalized samples. In a freezing transfer scenario it reduces relative mis‑penalization by about 70% at low cost.
Review: By internalizing policies into model weights, SIRF delivers high‑precision, low‑latency automated risk handling, offering a scalable and cost‑effective solution for industrial‑grade content safety.