A test bench for better data stories

We now run the story skill blind on real questions, have a model from another vendor critique the results, and turn the lessons into skill edits that we re-test on new questions, with the owner's blind preference as the measure that counts. The story lab page explains what we optimise, how the loop works and what the first rounds have shown, including the first open-data story run (WWII) and a France A/B now awaiting the owner's blind judgement.