CTGT traces its AI pipeline’s inconsistent decisions to changing inputs

CTGT says it made its AI model repeatable, but its application still produced four different answers from the same banking manual. On October 21, the company released Concordance, an open-source tool for tracing such discrepancies. Its experiment points to a failure outside the model: software that quietly changed the requests it sent.

Concordance compares exact requests, responses and results across application stages to locate the first divergence, without requiring access to model weights. CTGT’s research release includes the testing protocol, test software, experimental data and a toy pipeline under the Apache 2.0 license.

The model repeated; the requests did not

CTGT tested Qwen3.8-27B using SGLang’s deterministic inference mode, intended to make identical requests produce identical text. In a separate serving test, all 800 requests matched a control over the first 1,000 tokens. The application experiment used one synthetic banking manual in CTGT’s own system, not a customer deployment.

Four runs saved 32, 31, 33 and 33 obligations, with a different set each time. Across those runs, 74 model requests recurred exactly, and each received the same answer. Every changed answer followed a changed request.

One run dropped an obligation to escalate an unresolved conflict to the Head of Compliance. Its request differed from another run only in four identifier lines within a 12,525-character prompt. The model then treated five conflict-resolution rules as background. Exactly how those identifiers changed its interpretation remains untested.

Three changes around the model

• Identifiers: Upload-specific asset IDs appeared in 53 of 459 model calls. CTGT replaced them in prompts with short, stable labels.
• Ordering: A 30-item routing list changed order between runs. The application now sorts it.
• Shared state: A topic registry grew between runs. Each test now starts in a fresh organizational environment.

After those changes, five runs saved the same 34 obligations. Comparing stored results required normalizing upload-specific identifiers; intermediate parsing and interpretation outputs matched without that adjustment. The experiment covers one document and one pipeline, not the prevalence of these failures across enterprise AI.

A reliable replay can preserve a mistake

Before the fixes, CTGT replayed the incorrect-classification request eight times and got the same wrong answer every time. Its consistency check missed the error because the misread rules agreed with one another. Stabilizing inputs can therefore preserve an error as reliably as a correct decision.

Gorlla told VentureBeat that teams should preserve complete requests and responses for consequential decisions, retest when models, prompts or documents change, and maintain trusted examples with known answers. Those examples provide a check on correctness beyond whether a decision repeats.