A Multi-Level Validation and Traceability Framework for AI-Generated Telescope Scheduling Decisions

Research Problem

As time-domain astronomy grows, the need for rapid, automated telescope scheduling has increased. While AI models show promise in managing complex, multi-constraint observational tasks, they are prone to "hallucinations," inconsistent data references, and non-executable commands. These reliability issues have historically prevented the adoption of AI in high-stakes astronomical operations where precision is critical.

The Proposed Framework

This paper introduces a validation and traceability framework designed to sit between an AI model and the telescope execution system. Instead of attempting to fix the AI's internal generation process, the framework treats the AI output as a series of structured claims. It forces the AI to output its reasoning in the form of Atomic Reasoning Units (ARUs)—the smallest indivisible causal steps—which are then organized into a Directed Acyclic Graph (DAG).

This structure allows the system to perform:

• Data Reference Validation: Checking that target priorities and time windows are consistent with actual database records.
• Constraint Verification: Ensuring that proposed observations are physically possible given current instrument status and environmental conditions.
• Logical Consistency Checks: Verifying that each step in the reasoning chain is supported by valid evidence.

By decomposing complex AI decisions into verifiable units, the framework enables a "fail-safe" mechanism. If a reasoning step fails validation, the system can block the erroneous command before it reaches the telescope, preventing the waste of precious observational time. Furthermore, because the reasoning is stored as a DAG, the system provides a clear audit trail, allowing astronomers to understand exactly why a specific scheduling decision was made. This approach bridges the gap between the flexibility of modern AI and the rigorous reliability requirements of professional astronomical observatories.