Consistency Lab
Does few-shot prompting make the classifier more consistent? Run the same hard articles under each prompt configuration and compare.
Consistency needs repeats — 1 run measures nothing.
Comparison
| Configuration | Category guide | Few-shot | Agreement | Accuracy |
|---|
Per-article results
Each cell shows the label the configuration settled on and how often it agreed with itself across runs. correct · wrong · a striped cell means the configuration disagreed with itself between runs.