STROT

STROT compiles a task description, input schema and bounded answer choices into a cached Python artifact. An LLM synthesizes configuration during compilation; the default runtime uses the 144-million-parameter Julia-1 decision model locally. Applications can classify records, score feedback or route support tickets, with low-confidence cases left for the caller to handle.

The core package supports Python 3.9 or later, while the default local-model dependencies require Python 3.11 or later. Installation uses strot-ai[local] plus the model’s bundled inference package and checkpoint. Compilation needs a supported LLM client and its credentials, or a configured local Claude Code CLI; optional hosted decision runtimes have separate provider costs. The compiler’s code validation is not a security sandbox, and its guide recommends external isolation for untrusted intent.

STROT uses the MIT license; the default Julia-1 model uses Apache-2.0. Version 0.11.1 removes automatic runtime LLM escalation and retains a local policy that abstains below its fitted threshold.