New Tool Cracks AI Inconsistency

Generative AI offers significant business advantages but struggles with output consistency due to its non-deterministic nature. Stealthy startup CTGT has launched Concordance, an open-source system under the Apache 2.0 license, to help businesses ensure AI outputs align with their goals. This tool enables companies to investigate the root causes of varying AI decisions when faced with identical inputs.

Concordance addresses a critical need for enterprises moving AI from simple assistants to agents making significant decisions. The system was inspired by CTGT's own compliance application, where processing the same banking manual resulted in different obligations each time. Concordance allows users to pinpoint where discrepancies emerge within the AI application and its serving software.

CTGT's initial experiment revealed that even with a deterministic inference mode, application-level factors like varying identifiers, disordered lists, and expanded registries could cause inconsistent outputs. By stabilizing these elements, CTGT achieved consistent results. Concordance captures and compares exact model requests and responses, distinguishing between model instability and application code altering requests.

While generative AI models can produce identical textual responses under deterministic configurations, this does not guarantee correctness. Concordance helps identify where differences begin, but ensuring decisions remain faithful to policies is a subsequent challenge. The tool complements other monitoring systems like LangSmith and Braintrust, focusing on repeatable testing and aligning stages to explain divergences.