CASE STUDY

Case Study: from divergent outputs → trustworthy selection → guided refinement

A product-style walkthrough that shows the workflow and the decision points—not a research summary.

Context

Code assistants often force a binary accept/reject decision with limited transparency. Different models produce divergent solutions and verification shifts to humans.

Job-to-be-done
  • • Pick the best output quickly
  • • Understand trade-offs beyond correctness
  • • Improve weak dimensions without guessy prompting
Scenario

Task: calculate gender distribution by country, create a bar chart, and save as distribution.png.

Step 1 — Compare functional structure across models

The comparison view summarizes each model solution into functional blocks, making divergence visible without line-by-line scanning.

Compare view screenshot
Step 2 — Diagnose quality trade-offs with evidence

Users inspect diagnostics to understand why a solution is risky: execution signals, efficiency indicators, and maintainability cues.

Individual view screenshot
Step 3 — Refine with structure and diffs

Guided refinement helps translate feedback into concrete edits. Users review diffs and keep control over accepting or reverting changes.

Home view screenshot
Outcome

CGM Comparator turns verification into a repeatable product loop: compare structure → diagnose trade-offs → refine with guidance → track outcomes over time.

Why it matters
  • • Faster comprehension
  • • Higher confidence
  • • Lower iteration overhead