- 0:00–0:06
Start from a real question • Support ticket • Test set • Case A
Case A sits in a fixed badge that never moves through the film. A blank ticket card sits in the Support ticket source. A cyan route draws to an empty slot in the Test set frame, the ticket travels along it into the slot, and the slot is outlined: Case A has joined the test set. No answer is shown.
- 0:06–0:12
Agree what it should do • Reference answer agreed • Case A • Real questions • Required passage • Should answer • Should decline
The answer-or-decline figure shows Real questions: one leads to a Required passage and the tag Should answer, the other to an empty dashed document and the tag Should decline. A small cyan marker rides the answer lane from the first question into the document, where the required passage is outlined, then stops beside Should answer, which is outlined. The decline lane stays visible as the alternative. Reference answer agreed appears: an expectation written before any run, not a result.
- 0:12–0:18
Label it twice, separately • Reviewer 1 • Reviewer 2 • Labelling guide • Case A
Reviewer 1 and Reviewer 2 slide in from opposite sides, and each draws a line to its own empty label slot. A dashed route then runs from the gap between the two slots back to the Labelling guide. Both slots stay empty; no label value appears.
- 0:18–0:24
Check retrieval and the claim apart • Two separate checks • Case A • Retrieval • Generation • Passage found? • Claim supported?
The figure shows Retrieval feeding Generation, each with its own open ring: Passage found? and Claim supported?. A cyan trace runs from Retrieval to the Passage found? ring, which is outlined; a second trace then runs from Generation to the Claim supported? ring. Neither ring fills and no score appears. Two separate checks appears.
- 0:24–0:30
Read it in its slice • Arabic numeric questions • Count beside each rate • Case A
A slice column headed Arabic numeric questions slides in under the Case A badge beside a second, unlabelled column, and its outline turns cyan. Each column holds an empty rate slot with an empty count slot beside it. Count beside each rate appears and both count slots are outlined; no numbers, rates or bars appear. The final state holds to the end.
An illustrative Arabic expense-policy question enters the test set, gets an agreed expectation, two independent labels and separate retrieval and claim checks, then waits in its slice with every result still pending.
Read the article: How to Evaluate LLM Hallucinations: Test Sets, Labels and Metrics