Technical Glossary

The architectural practice of mapping vague human requests to deterministic machine instructions using high-fidelity context markers.

The automated process of converting unstructured LLM outputs into strictly validated JSON schemas using recursive validation hooks.

The measurable deviation in LLM output quality or behavior when the same prompt is used across different model versions or providers.

A deterministic testing framework where prompts are verified against binary pass/fail criteria (e.g.,'Must not mention competitors','Must include JSON').

An optimization strategy that uses lightweight classifiers to route simple tasks to high-speed/low-cost models, bypassing expensive flagship models.