Version 2026.7.26
A clearer LLM Arena
A redesigned, evidence-first Arena makes model rosters, run status, provisional results, verified trades, and comparison history easier to read. Arena remains SIM research and does not place live orders.
Safer review and recovery
Execution stays bound to the terms and account context you reviewed. Corrected newsletter trade alerts now show a durable warning in proposal views without changing proposal state.
Smoother model setup
ChatGPT subscription sign-in now opens reliably from the desktop app. Curated model eligibility checks also keep unsupported providers out of the Arena roster.
A cleaner desktop package
Release gates now reject developer-only SDKs and deferred server components before signing. This is a macOS-first release; Windows availability is handled separately.
Also fixed
- Run now keeps the selected comparison roster, survives restarts, and reports model readiness and billing state accurately.
- Completed same-symbol SIM scale-in history is archived only after filled orders balance and no open position remains.
- Corrected same-ID newsletter alerts are deduplicated while preserving an owner-scoped warning in Proposal Card and Chat.
- LLM Arena publication fails closed on stale or incomplete evidence instead of presenting an unsupported performance claim.