MARVIS-Agent puts credit-risk modeling behind human gates and deterministic tools

MARVIS-Agent is a local-first, governed platform for credit-risk work, positioned by its README as something other than a chatbot wrapped around scripts. A user describes a business outcome in natural language, the agent asks for missing files and definitions, builds a reviewable plan, pauses at responsibility gates for human confirmation where required, and then runs deterministic tools. The output is intended to be datasets, evidence, models, strategy code, and reports that stay attached to the task for review and reuse. The intended audience is risk teams and business users who need to respond when a segment deteriorates or new fraud signals appear, without rebuilding the same analysis each time. The README frames the problem around traceability: metric definitions, field meanings, historical performance, and decision records should persist beyond a single conversation, while people retain supervision over consequential choices. Existing workflows connect data preparation, feature analysis, model comparison, validation, strategy assessment, backtesting, and reporting in one task workspace. How V3 structures the work V3, documented in a linked V3.0.0 release note, centers on a task workspace that groups work by what needs attention, what is running, and what was done recently. The interface supports English and Chinese and light and dark themes, with task names and original values preserved. Diagnostic cards explain when work stops and offer supported recovery options, including action history, budgets, and undo where available. Dependency installation remains an explicit user action rather than an automatic step. For modeling handoffs, the README calls out sample maturity checks, score direction