| --- |
| pretty_name: HUD UI Validation Pilot |
| license: other |
| task_categories: |
| - image-classification |
| tags: |
| - gameplay |
| - hud |
| - ui |
| - validation |
| --- |
| |
| # HUD/UI annotation validation pilot |
|
|
| This public artifact compares 10 gameplay screenshots across four columns: |
|
|
| 1. raw frame; |
| 2. GPT-5.6 Sol X-High final reference; |
| 3. GPT-5.6 Terra High refined annotation; |
| 4. confidence-routed final annotation. |
|
|
|  |
|
|
| Files: |
|
|
| - [`analysis.md`](analysis.md): aggregate metrics and per-game error table; |
| - [`per_sample_metrics.csv`](per_sample_metrics.csv): machine-readable sample metrics; |
| - [`overlay_comparison_contact_sheet.jpg`](overlay_comparison_contact_sheet.jpg): full comparison sheet. |
|
|
| The 0--100 confidence value is a conservative pipeline routing score, not a |
| calibrated probability of correctness. On this 10-sample pilot, pure Terra High |
| element F1 is 0.7755 and mean slot Jaccard is 0.6875. The sample is too small for |
| a final model-selection conclusion. |
|
|
