CTGT releases Concordance, an open-source tool to trace why AI agents make inconsistent decisions, revealing that pipeline inputs often cause the drift.

The transition from AI assistants to autonomous agents marks a pivotal shift for the enterprise. While a chatbot hallucinating a poem is a curiosity, an AI agent omitting a critical regulatory obligation in a banking audit is a liability. This tension between the generative nature of Large Language Models (LLMs) and the rigid requirements of corporate governance is the catalyst for a new release from CTGT.

The startup, which operates as an applied interpretability lab, has launched Concordance. This open-source tool, released under the Apache 2.0 license, is designed to help businesses investigate why an AI application produces different decisions when tasked with the same objective. By providing a testing protocol and reference implementation, CTGT aims to give enterprises a mechanism to defend the decisions their AI agents commit to in production.

The illusion of model non-determinism

Conventional wisdom in the tech industry suggests that AI inconsistency is a byproduct of the non-deterministic nature of LLMs—the fact that they construct outputs on the fly. However, CTGT's recent findings suggest that the culprit is often not the model itself, but the pipeline surrounding it. In a series of experiments, the company discovered that while the underlying model responded consistently to identical requests, the application pipeline was feeding it slightly different data each time.

To prove this, CTGT utilized a deterministic serving test. The results showed that 800 requests matched a control over the first 1,000 tokens, suggesting that the model's inherent randomness was not the primary driver of the errors. Instead, the divergence happened earlier in the process. As CEO Cyril Gorlla noted, fixing non-determinism in the model is futile if the pipeline continues to deliver varying inputs.

A banking manual and the cost of drift

The necessity for Concordance became evident during a trial involving a nine-page retail banking manual. Using the Qwen3.8-27B model with SGLang’s deterministic inference mode, CTGT processed the same document four times. The results were erratic: the runs extracted 32, 31, 33, and 33 obligations respectively. Even when the total number of obligations matched, the actual content differed across all four outputs.

The stakes of this inconsistency were high. In one specific run, the AI omitted a requirement to escalate a serious issue to the compliance chief. This omission occurred despite the model's ability to handle identical requests consistently. The failure was traced back to a prompt of 12,525 characters where only four identifier lines had changed. These minute shifts in input were enough to alter the model's interpretation of the banking policy, leading to a critical failure in compliance extraction.

Tracing the source of divergence

Concordance allows developers to locate the first point of divergence across various application stages without needing access to the model's internal weights. Through its implementation, CTGT identified three specific triggers that caused the banking manual experiment to fail:

The application kept changing what it sent: upload-specific identifiers appeared inside prompts, a 30-item routing list changed its order, and a topic registry expanded between runs, which altered subsequent requests.

By stabilizing these inputs—ensuring identifiers remained constant and list ordering was fixed—CTGT was able to achieve repeatability. In subsequent tests, five separate runs saved the exact same 34 obligations. While this stabilization fixed the repeatability of the results, the company clarified that it did not automatically establish that the resulting decisions were correct, only that they were consistent.

From basic guardrails to deterministic governance

The release of Concordance sits alongside CTGT's broader commercial mission to move enterprises from simple guardrails to comprehensive governance. Traditional guardrail solutions typically offer binary detection—flagging an output as either toxic or non-compliant. This approach often leaves users stranded without a solution and struggles with the nested, contradictory nature of enterprise policies.

CTGT's approach involves a graph-based policy engine that utilizes Neo4j to resolve collisions deterministically. This system is designed to not only detect violations but to actively remediate them. According to their technical benchmarks, this engine has achieved 89% remediation accuracy when validated against over 35,000 FINRA rules, with a P90 retrieval latency of 20ms, making it viable for synchronous production environments.

The architecture of AI interpretability

The development of these tools stems from research conducted by Cyril Gorlla following his time at UC San Diego and Stanford. By positioning the company as an interpretability lab, CTGT focuses on the intersection of AI models and the software orchestration layers that manage them. The goal is to create an audit trail for every policy decision, providing a level of transparency that is currently missing from most black-box AI deployments.

For companies moving toward AI agents that approve or record consequential business decisions, the ability to trace a decision back to a specific input change is essential. Concordance provides the tooling to ensure that when an AI agent makes a decision, the company can explain exactly why that decision was reached and ensure it happens the same way every time.

Strategic implications for US and UK enterprises

For businesses operating in the US and UK, the release of Concordance addresses a critical gap in risk management. In highly regulated sectors such as finance and insurance, the inability to produce consistent results from an AI system is not just a technical glitch but a regulatory risk. In the US, where agencies like FINRA maintain strict oversight of compliance, the ability to provide a deterministic audit trail is a prerequisite for the adoption of autonomous agents.

The UK market, which has leaned toward a pro-innovation but safety-conscious approach to AI, will find these tools useful for meeting internal governance standards. As enterprises shift from using AI for productivity (assistants) to using AI for authority (agents), the legal burden of proof shifts. If an agent fails to flag a compliance breach, the company cannot simply blame the non-deterministic nature of the LLM. The ability to stabilize pipelines and trace divergences allows firms to implement a 'human-in-the-loop' or 'policy-in-the-loop' system that satisfies both operational efficiency and legal accountability.

FAQ

---

Sources: Superpowerdaily, Venturebeat, Ctgt ·