Groundline

Eval numbers

Each workspace below opted to publish its latest eval run: a fixed set of questions with reference answers, asked to the live assistant. Refusals and citations are checked deterministically; a judge model grades correctness with a pinned rubric. No number here is typed in by hand.

gemini-3.5-flash-lite judged by gemini-3.5-flash-lite · rubric correctness-v1 · 40 cases · 14 Oct 2026

Answer accuracy

87%

26 correct · 3 partial · 1 incorrect of 30 answerable

Cited the right source

100%

answerable questions with an expected source

Correct refusals

100%

10 unanswerable questions

Over-refusal

3%

answerable questions the assistant refused

Avg latency

6535 ms

question to full answer

Run cost

$0.0277

40 cases · list prices, free tier pays $0

How the numbers are produced

  • Answer accuracy: share of answerable questions the judge marked “correct” (every key fact present, nothing contradicting). “Partial” answers do not count.
  • Cited the right source: share of answered answerable questions whose citations include the expected document.
  • Correct refusals: share of unanswerable questions where the assistant declined instead of guessing. Over-refusal is the mirror on answerable questions.
  • The judge is a different model family from the one under test when keys for both exist, to limit self-preference. Unparseable judge output counts as incorrect, never as missing.