Search the Atlas

Search risks, controls, and glossary terms

MediumOutput DriftDemonstrated

Output drift & performance degradation

Model Accuracy & Reliability

Description

Model or pipeline changes (provider updates, prompt edits, data shifts) silently degrade output quality over time.

Example scenario

A silent provider model update degrades the summariser, and errors accumulate before anyone notices.

Real-world evidenceDemonstrated

The 'Lost in the Middle' paper (Liu et al., 2023) empirically demonstrated that transformer models systematically underweight information positioned in the middle of long contexts, confirming silent information loss as a real model behaviour. This has been reproduced across multiple models and context lengths, but specific production incidents where context truncation caused documented decision-making harm in a financial or legal workflow have not been publicly confirmed.

Primary mitigations

  • Continuous evaluation harness
  • regression test suites
  • canary prompts
  • change gating
  • provider-version pinning.

Detection signals

Eval-score trend monitoring; canary regression alerts; drift dashboards.

Mitigating controls

9
Non-agentic controls

Related risks in Model Accuracy & Reliability