Open alpha one person vibe coding

Scry · Benchmark evidence

DeepSearchQA: the whole answer, every transcript

Google's DeepSearchQA is 900 research questions graded exact-set: find every member of the answer and add nothing. Google's Kaggle leaderboard reports its Gemini Deep Research Agent at 66.1% fully correct. An agent doing its research through Scry scores 71.8% — same judge model, same published grading prompt. This page is the complete evidence: every question, every grade, and the full tool-call transcript behind each answer.

Setup

One agent — GPT-6 Astra — with Scry's tools: programmatic search, page fetching, and SQL over the corpus. At most 20 tool calls per question. The configuration was frozen before the run and every question was answered under it by the single solver; grading is the official method — gemini-2.5-flash with Google's published grading prompt. The instruction policy was developed on 100 previously-burned questions and frozen before touching the rest; the frozen config, agent instructions, and raw grades are all linked below.

summary frozen config agent instructions raw grades (900) index

Loading 900 questions…

Dataset: google/deepsearchqa. There is no official submission process; every published number, including Google's, is a self-run harness. Grading here is the published prompt verbatim on gemini-2.5-flash. Full disclosures ship in the summary.