From GitHub · · 1m
GitHub - strot-ai/Project-STROT: STROT (arXiv:2505.01636): compile-time LLM code synthesis with deterministic, self-routing NLU runtimes
Compile once with LLM, decide forever with type-safe JSON outputs.
In brief
Compile once with LLM, decide forever with type-safe JSON outputs. STROT compiles LLM intent into type-safe JSON decisions at compile time, reducing routing errors by 18 percentage points compared to naive methods. Originally reported by GitHub.
STROT Overview
STROT compiles intent once with an LLM to create type-safe JSON decisions for routing records.
“Compile your intent once with an LLM. Then decide on every record forever, classification, scoring, routing, as type-safe json with probabilities and confidence, on your own hardware, at zero marginal cost”
How It Works
STROT uses compile-time schema synthesis to generate a frozen decision model that runs on every record without lldm calls.
“COMPILE (once per intent + schema) intent + sample ─▶ synthesize ─▶ golden gate ─▶ teacher labels ─▶ calibrate”
Key Features
STROT provides type-safe JSON outputs with confidence scores, zero marginal cost, and no structural hallucinations.
“0% structural hallucination, output is selection + computation over your registry; nothing is generated at runtime”
Performance Metrics
STROT achieves 92.5% routing accuracy versus 74.8% for naive methods, with 86.7% semantic text accuracy.
“Grounded routing accuracy (398 records) 92.5% compiled vs 74.8% naive, the compiler is worth +18pp on the same model”
Installation
Install STROT via pip with options for local or hosted decision models and nlu check points.
“pip install strot-ai # core: pure stdlib”
Why STROT
STROT uses compile-time schema synthesis to eliminate runtime hallucinations and improve decision accuracy.
“Golden-record promotion, artifacts must pass your known answers before they serve”
The headline
STROT compiles LLM intent into type-safe JSON decisions at compile time, reducing routing errors by 18 percentage points compared to naive methods.
