The Right to a Reproducible Past: Measuring the Model-Lifecycle Contestability Gap in Foundation-Model Services

Foundation-model services change continuously, yet contestability debates still focus primarily on individual decisions rather than on the temporal conditions under which those decisions can later be reconstructed, challenged, and audited. This article introduces the Model-Lifecycle Contestability Gap: the distance between a provider’s power to update, replace, or retire a model and an affected person’s practical ability to identify which model acted, preserve relevant evidence, validate a replacement, and reproduce a contested output after change. Using public primary sources available on 12 July 2026, the study audits six enterprise-facing services—OpenAI API, Anthropic API, Google Vertex AI, Microsoft Foundry, Amazon Bedrock, and Databricks Foundation Model APIs—and three consumer interfaces—ChatGPT, Claude.ai, and the Gemini app. Ten lifecycle dimensions are coded on a 0–2 scale and aggregated into an equal-weight Model-Lifecycle Contestability Index. Enterprise services average 12.83/20, compared with 4.67/20 for consumer interfaces. Across both groups, evidence preservation is more common than executable historical access, and business migration support is substantially stronger than appeal continuity. The article argues that logs and changelogs cannot by themselves preserve meaningful contestability. It develops a proportionate Right to a Reproducible Past and proposes six Minimum AI Continuity Duties for providers and deployers. This is the accepted manuscript of an article accepted for publication in the Journal of Ethics and Emerging Technologies. The article is scheduled for publication in the first issue of 2027.