Note Try our deterministic citation checker in your browser. Code: https://github.com/tonydzi/verbatim-citation-gate . A matching quote is not proof that the claim is true.

Note Eight authored research-reading questions, reference answers and rubrics from seven version-pinned primary papers. All evidence excerpts checked against fetched abstracts. No model score or benchmark pass claimed.

Note 91 author-written action tasks, expected answers and measured per-case outputs of a pinned deterministic baseline. 100% must-ask routing and 50% over-refusal. Code, conditions and reproduction included; no SWE/GAIA score claimed.

Note Reference benchmark. This collection item reports no agent run or score. Our separately measured 91-task baseline: https://huggingface.co/datasets/TonyDzi/agent-action-policy-eval

Note Reference benchmark. This collection item reports no agent run or score. Our separately measured 91-task baseline: https://huggingface.co/datasets/TonyDzi/agent-action-policy-eval

Note Agent leaderboard we track

pyannote/speaker-diarization-community-1

Automatic Speech Recognition • Updated Sep 29, 2025 • 5.11M • 3.93k

Note Diarization in our voice pipeline

openai/whisper-large-v3

Automatic Speech Recognition • 2B • Updated Aug 12, 2024 • 3.54M • • 6.77k

Note Speech-to-text in our voice pipeline

Note ReAct: background on interleaving reasoning and actions. Primary source for our research-reading tasks. This collection entry reports no reproduced model run or agent score.

Note BrowserGym ecosystem: primary paper about comparable web-agent environments and AgentLab. Source for our research-reading tasks; no local installation or benchmark result is implied.