Contract-Governed Multi-Agent Graph Orchestration for Long-Horizon Autonomous Research Pipelines

This submission focuses on the backend and architectural theory of the Quarks framework, specifically how complex research processes can be represented as graph-structured, contract-driven multi-agent workflows. The main research object is not only implementation performance but the formalization of orchestration: how agent roles, interfaces, validators, and phase semantics can be defined so that long-horizon tasks remain stable, interpretable, and controllable. The proposed approach separates responsibilities across three primary agents: (1) an orchestrator that manages global graph state and sequencing policies, (2) an executor that performs phase-local actions under explicit input/output contracts, and (3) an AI-scientist agent that provides domain-aware adaptation, remediation, and meta-level intervention. A higher-level meta-orchestration layer governs rerouting, backtracking, and exploration strategies (including Monte Carlo-style branching) when validation fails or uncertainty increases. The theoretical contribution targets contract semantics, validator composition, and graph transition rules for dynamic vs. static prompt/instruction templates. Particular attention is given to bounded interfaces, failure recovery, and policy constraints required for multi-day or multi-week autonomous runs. The memory model includes persistent filesystem-backed context to preserve cross-phase continuity while maintaining traceability and reproducibility of decisions.