CTGT releases open source tool Concordance to trace AI decision inconsistencies

Source published 23h 52m agoTerminal published 23h 50m ago

Summary

CTGT has launched an open-source tool called Concordance, designed to help businesses understand and trace the reasons behind inconsistent decisions made by AI agents. This development addresses the challenges posed by the non-deterministic nature of generative AI, which can yield different outputs from the same inputs, complicating policy enforcement as companies shift from AI assistants to autonomous agents responsible for critical decisions. Concordance enables enterprises to analyze discrepancies in AI outputs by capturing and comparing exact model requests and responses, reinforcing the importance of consistent decision-making in compliance and operational scenarios. Released under the Apache 2.0 license, this tool can be freely modified and integrated into existing AI systems, enhancing debugging capabilities without the need for overhaul.

Analysis

Jev: Jev is the decision-model implementation of TypeSafe’s System One architecture that returns constrained choices, scores, and probabilities instead of open-ended text. It is designed to keep workflow control and business rules in deterministic code while using AI only for narrower judgments on unstructured data. The approach is presented as complementary to post-hoc testing tools for reducing certain classes of AI reliability issues. CTGT: CTGT is an applied interpretability lab focused on understanding AI model and software behavior in regulated industries including banking and insurance. The company maintains a separate commercial policy enforcement engine while releasing tools to diagnose inconsistencies in AI applications. CTGT developed and open-sourced Concordance specifically to help enterprises locate sources of decision variation in workflows that process documents or approve actions. Qwen3.8-27B: Qwen3.8-27B is the large language model used in CTGT's experiments on obligation extraction from banking documents. The model was run under SGLang’s deterministic inference mode to isolate sources of output variation. Tests showed it produced identical responses to repeated identical requests while application-level changes produced differing results. Cyril Gorlla: Cyril Gorlla is the founder and CEO of CTGT. He has guided the company's research into AI interpretability following his academic work at UC San Diego and Stanford. Gorlla has positioned Concordance as a free addition to existing tracing platforms that enables repeatable testing of AI decision pipelines. AI Non-Determinism: Generative AI systems are inherently non-deterministic, producing varying outputs even for identical or similar inputs, which complicates consistent policy enforcement in enterprise settings. Open Source Debugging: New open-source releases under permissive licenses like Apache 2.0 allow enterprises to integrate targeted comparison procedures into existing AI tracing and evaluation platforms without replacing current infrastructure. Enterprise AI Deployment: Companies are moving from AI assistants to autonomous agents that make consequential decisions such as approvals, flags, and compliance obligations, increasing the need for tools that verify output repeatability.

Summary

Analysis

Categories