LLM Guidance Routing and Evaluation

This project assembles guidance for language models. Its earlier working name was reasoning-layer; LLM Guidance Routing and Evaluation more accurately describes what it does. It routes prompt material rather than implementing a model’s reasoning.

Guidance assembly

I developed a deterministic Python router for portable guidance modules, respecting dependency and context-budget constraints and producing provider-specific bundles. The implementation was AI-assisted under my direction.

Testing the premise

Controlled comparisons used frozen criteria, held-out cases, and explicit call budgets. In one retained negative result, structured revision did not outperform the baseline. That result belongs in the project record: more elaborate prompting is only useful when the comparison supports it.

This remains experimental work. It is not a claim of improved model reasoning or general performance gains.