Strot：编译时一次推理，本地永久运行决策

开源 Python 工具库 Strot 通过将大语言模型推理从运行时转移至编译阶段，使生产环境能够在本地硬件上以零成本运行分类、评分、路由等决策任务1。该工具在编译时调用一次 LLM（如 Claude Opus 4），生成冻结的自包含决策制品，避免了幻觉问题并确保了输出的有界性和可靠置信度1。

生产环境中的本地决策模型采用 144M 参数的 Julia-1，单条记录的推理耗时约 60ms1。系统生成的输出为类型安全的 JSON 格式，包含概率和置信度信息，不含任何生成内容1。Strot 支持分类、评分、路由和类型化问答等多种决策任务1，并内置温度缩放和分割共形预测机制，确保错误率不超过指定阈值 α1。该项目采用 MIT 许可，默认决策模型为 Apache-2.0 许可的 SupersonicLabs/Julia-11，其实现基于 arXiv:2505.01636 的研究思想1。

Strot is an open-source Python toolkit that shifts large language model inference from runtime to compile time, enabling production systems to execute classification, scoring, routing, and other decision tasks on local hardware with zero computational cost 1. The tool uses an LLM such as Claude Opus 4 at compile time to generate a frozen, self-contained decision artifact that eliminates hallucinations, guarantees bounded outputs, and provides reliable confidence scores 1.

In production, the system deploys a lightweight decision model with 144 million parameters—Julia-1 from SupersonicLabs—that processes individual records in approximately 60 milliseconds 1. Outputs are delivered as type-safe JSON containing probabilities and confidence levels, with no generated text 1. The toolkit supports classification, scoring, routing, and typed question-answering workflows, and incorporates temperature scaling and split conformal prediction to ensure error rates remain at or below a specified threshold 1. Strot is distributed under the MIT license, with the default decision model available under Apache 2.0 1, and implements research concepts from arXiv:2505.01636 1.