Scene QA beta
Ask the story, get the receipts
Retrieval finds the scenes. The model writes a short answer and numbers every passage it used, so each claim links back to the line it came from. If the story does not cover it, the answer says so instead of inventing one.
Try
Answer prompt edit it — it changes wording, never retrieval
Saved in this browser only.
Evaluation harness labelled questions, real pipeline
Runs every labelled question through this Worker's retrieval and answer path, then scores the rank of the expected scene, whether the answer cited it, and whether it refused the questions the story cannot answer. Uses the prompt above, so you can measure your edits.