Fluent Is Not Faithful: Building a Safer AI Paraphrasing Pipeline

Short summary

A developer describes a structured paraphrasing pipeline that separates rewrite constraints (mode, strength, protected keywords, output count) into validated fields rather than a single text prompt. The pipeline uses Zod schema validation, mode-specific prompt instructions, and post-generation contract checking to prevent fluent-but-unfaithful rewrites. The article acknowledges what the pipeline catches (invented facts, dropped qualifiers) and what it still cannot prove.

• •Structured rewrite payload separates mode, strength, keepWords, and outputCount as independent constraints
• •Zod validation prevents unknown modes, oversized input, and unbounded output requests
• •Post-generation contract checking verifies keyword preservation and meaning integrity
• •Pipeline catches dropped qualifiers and invented facts but cannot fully prove faithfulness

Generated with AI, which can make mistakes.