The Station: An Open-World Environment for AI-Driven Discovery

Update to Station v2.0.0 For mathematical results published since v1.5.0, see news/2026-06-08-math-update.md, news/2026-06-08-kissing.ipynb, news/2026-06-08-epoch-book.ipynb, and news/2026-06-14-kissing.ipynb. Major release highlights since v1.5.0: Added multistart for both station initialization and the stagnation protocol. Added independent auditing to Research Center coder workflows, allowing cheaper coder models without giving up independent scientific checking. Added the Question Room for agents to propose, discuss, vote on, and solve research questions, with supervisors able to retire and restore questions. Added Station Tools for managing, monitoring, initializing, restoring, and operating multiple stations. Added automatic context compaction and parallel Research Center submissions. Added the Research Center Seed Bank and remote shared storage as optional features, both disabled by default. Added a dashboard Archive Surveyor so users can survey Station literature directly from the frontend. Expanded the stagnation protocol with lane-specific prompts, a strategy lane, randomized lane assignment, and continuous supervisor assignment. Removed sequential tick execution and Theory Room, making concurrent orchestration the only supported path. Model and policy updates: Added GPT-5.6 Sol, Luna, and Terra and Claude Opus 5 model presets. Updated the default roster to two Gemini 3.1 Pro agents, two GPT-5.6 Sol agents, and two Claude Opus 5 agents. Set GPT-5.6 Luna as the default Research coder and GPT-5.6 Sol as the default for the reviewer, Surveyor, meta-reflection, External Counter, and multistart administrator. Allowed Claude Opus 5 agents to become supervisors. Increased immature-agent isolation to 40 ticks, restored the Research Center concurrent submission limit to 2, and made ten CPUs per official Research attempt the default. Research task and example updates: Reorganized public Research tasks into AlphaEvolve, Epoch, and miscellaneous groups and reorganized station templates into named, selectable bundles. Expanded the public mathematical task catalog with the v2 AlphaEvolve suite, the Jacobian benchmark, the book-Ramsey task, new Epoch tasks, and updated evaluators and baselines. Infrastructure and UI improvements: Added shared worker management, coder replay, bounded retries, session resume, and infrastructure recovery for coders, auditors, Archive Surveyors, and multistart administrators. Added CPU and GPU allocation coordination for multi-station Research workloads. Separated top-score tracking from meaningful-breakthrough tracking and improved dashboard statistics and indexes. Overhauled the External Counter for current model-provider APIs and more reliable retry handling. It remains disabled by default. Bug fixes and robustness: Improved Research Center scheduling, official attempts, audit repair, restart recovery, storage handling, notifications, and resource utilization. Improved action parsing, OpenAI streaming and reasoning behavior, Claude refusal and truncated-response detection, and general provider compatibility. Expanded test coverage across multistart, Question Room, Station Tools, context compaction, Research auditing and Seed Bank behavior, concurrent orchestration, Archive Surveyor, External Counter, dashboard behavior, and Research recovery.