Iterative prompt refinement does not reliably improve code quality. Across the evaluated models, performance does not consistently improve with additional prompt iterations. Instead, different models exhibit different trajectories:

• Claude 4 Opus shows a steady decline in correctness over iterations.

• GPT-4.1 demonstrates diminishing returns after the first iteration, with later iterations sometimes degrading results.

• Gemini 2.5 Pro follows a dip–recovery pattern, where performance initially worsens before partially improving.

Correctness is the dominant factor determining overall code quality. Maintainability and security scores remain consistently high (near ceiling levels) across tasks and models. Because these dimensions show little variation, the overall quality score is largely driven by correctness.

Cross-dimensional relationships between evaluation metrics are weak. Changes in one quality dimension rarely correlate with changes in others. Iteration-to-iteration correlations between correctness, maintainability, and security are small or statistically insignificant, suggesting that improvements or declines in one metric do not reliably predict changes in another.

Most actionable improvement occurs in early iterations. The most meaningful changes in correctness typically occur between iteration 0 → 1, and occasionally 1 → 2. Beyond this point, additional iterations tend to yield minimal or negative returns.

Task type influences performance for some models but not all.

• GPT-4.1 shows a moderate task-category effect, performing better on data structures and algorithms tasks compared to OOP tasks.

• Claude 4 Opus shows a weaker trend of task dependence.

• Gemini 2.5 Pro shows no statistically reliable task-category effect.

Structural quality metrics exhibit ceiling effects. Maintainability and security scores remain consistently high regardless of task category or iteration, limiting their usefulness as discriminative evaluation metrics under the current setup.

Model-specific refinement strategies are necessary. The results suggest that prompt iteration strategies should be model-aware and bounded:

• Keep Claude’s iteration loops short due to declining correctness.

• Limit GPT-4.1 to one or two targeted refinements.

• Apply incremental monitoring and rollback strategies for Gemini.