Benchmarks & research
RSI-Jev is a recursively self-improving research loop that trains and evaluates open-weight System One decision models with a Jev-compatible server.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The loop publishes failed experiments alongside releases. Its v6.1-VL models answer text and image questions without generating reasoning, stopping at earlier layers when confident enough. The pinned README separates the 7 October 4B release from the 8 October 27B release: reported public Decision Index 0.3 scores are 50.98 and 65.64 respectively; the official 27B score was pending. These are author results, not reproduced measurements. RSI-Jev is not affiliated with TypeSafe.
Choose how deeply the model processes each question. The README recommends low effort for sending messages to the right destination or finding relevant information, medium for sorting items into groups, and high for knowledge or code questions. With mixed questions, the automatic setting stops at the first stage confident enough.
The documented examples ask whether a message requests a refund and whether a customer's photo shows damage. You can supply text and images with yes/no, multiple-choice or rating questions. An existing Jev client can point to this server, but receives RSI-Jev's answers, not TypeSafe's.