Critic Self-Improvement Loop

The Core Loop

Meridian generates a candidate response, then immediately passes it through a critic pass. The critic identifies flaws — factual gaps, tone mismatches, structural weaknesses — and feeds them back as a refinement prompt. This loops until a quality threshold is met or a budget is exhausted.

Critic Dimensions

• Factual accuracy— cross-checks claims against retrieved context
• Completeness— ensures all sub-questions are addressed
• Tone alignment— verifies voice matches the user persona
• Concision— flags redundancy and filler

Budget & Guardrails

Each loop iteration consumes tokens. Meridian enforces a hard cap (default 3 refinement passes) and a cumulative token ceiling. Early-exit triggers when the critic score exceeds 0.92, preventing diminishing returns.

Why It Works

LLMs are better at evaluating than generating. By decoupling creation from judgment, Meridian catches errors the generator would otherwise miss. The critic prompt is itself a structured rubric, eliminating ambiguity in self-assessment.