discover-improve

Improve a blueprint's discovery score by applying deterministic Phase A fixes to clean prose, followed by LLM-driven semantic Phase B refinements. The skill orchestrates a halt-loop that re-scores the blueprint after each iteration until the target verdict is reached.

A halt-loop orchestration for iterative blueprint improvement is an autonomous cycle that enforces safety checks and progress validation. It repeatedly applies fixes and re-scores blueprints using the discover-confidence scorer until the desired quality verdict is achieved.

Fix hedge words and loopholes in blueprints automatically using the scripts/apply_fixes.py Phase A deterministic fixes. This process cleans prose and marks missing or cited material before initiating LLM-driven semantic refinements for structural gaps.

You can use LLM-driven semantic refinements to address structural gaps, citation density, and plan alignment in blueprints. Phase B prompts produce final adjustments after deterministic fixes clean the prose, ensuring comprehensive quality assurance.

Deterministic fixes for blueprint quality assurance are limited to cleaning prose, removing hedge words, and marking missing material. They cannot resolve structural gaps or citation density issues, which require subsequent LLM-driven semantic refinements to achieve the target score.