Catalyst Releases Compiled Code Variable Verification Tool for AI Agents

Catalyst, a tool that enables AI agents to identify and verify truly important variables in complex compiled systems, has been released. This tool converts Rust code into LLVM IR and then directly analyzes compiler output to calculate the impact of input variables on the final result as derivatives.

Key Features and Performance

• Ultra-fast analysis: Completes all derivative calculations and independent verification of 105 points within 1.37 seconds for a 72-hour simulation model with 52 input variables.
• Precision guarantee: Compared with the Complex-step reference, the maximum relative error is approximately 2.1e-15, verifying numerical accuracy.
• Reusable artifacts: Generates verified results as artifacts, with subsequent re-run times being very fast at approximately 12ms. It also includes integrity checks that automatically reject tampered artifacts.

Case Study: Environment-Specific Sensitivity Analysis

In a data center cooling/power plant demo, it was proven that even with the same source code, the dominant variables change completely depending on the operating environment.

• Heatwave scenario (38°C): Fan control (fan_target_c) ranks first in cost impact, and 17 out of 52 variables determine 90% of the cost.
• Mild weather (18°C): UPS and PDU efficiency become the main drivers, and the influence of fan_target_c converges exactly to 0.

Practical Value

Agents can make optimization decisions based on measured data instead of guessing code. In the demo, this approach reduced modeled operating costs by 18.1%. It is also applicable to various engineering workflows, such as independent export to Go/R code and production error reproduction reduction.