Evaluation

VisionStudio uses LLM-as-a-Judge to evaluate spec quality against profile-specific rubrics.

Evaluation Process

1. Open a spec in the editor
2. Ask the LLM panel to evaluate
3. Review scores and findings
4. Iterate on the spec
5. Re-evaluate until passing

Scoring

Specs are scored on a 0-10 scale:

Score	Decision	Meaning
8.0+	Pass	Ready for next step
6.0-7.9	Conditional	Needs minor improvements
< 6.0	Fail	Needs significant work

Rubric Categories

Each spec type has specific evaluation criteria:

MRD Rubric

• Problem Definition (25%)
• Target Users (20%)
• Business Goals (20%)
• Market Context (15%)
• Document Quality (20%)

PRD Rubric

• User Stories (25%)
• Requirements Clarity (25%)
• Acceptance Criteria (20%)
• Traceability (15%)
• Document Quality (15%)

TRD Rubric

• Architecture Clarity (25%)
• API Design (20%)
• Data Model (20%)
• Security (15%)
• Document Quality (20%)

Findings

Evaluation results include findings with severity levels:

Severity	Action
Critical	Must fix before proceeding
High	Should fix
Medium	Consider fixing
Low/Info	Optional improvements