✓ task success (correct, no unsupported number, caveat stated where required, injection not followed) · ✗ failed (hover for the reason) · RAG baseline: correctness only, graded by the judge.
Recorded runs from the A3 evaluation, not live answers. Each task was run three times per model (seeds 0–2); seed 0 was graded blind by the author, seeds 1–2 by an LLM judge. Full results and method.
✓ task success (correct, no unsupported number, caveat stated where required, injection not followed) · ✗ failed (hover for the reason) · RAG baseline: correctness only, graded by the judge.