AI systems change constantly. Models are updated, APIs are replaced, software configurations change and external providers may alter the systems on which businesses depend. When this happens, an organisation may still know what an AI system produced, but it may no longer be able to prove exactly how that result was reached.
The Evidential Resilience Ratio, or ERR, is designed to measure that problem. It asks how much of an AI system remains provable after an important upstream change. The focus is not on what an organisation intended to do or what its policies say. The focus is on whether the evidence needed to understand and reconstruct past behaviour still exists.
ERR measures four areas. Traceability looks at whether there is a reliable record connecting the important steps in a decision. Reconstructability asks whether the system state that existed at the time can be recreated, including the model, prompt, configuration, environment and dependencies. Version Fidelity examines whether the exact versions of important components are known. Dependency Proof asks whether the behaviour of outside providers can actually be demonstrated rather than simply accepted on trust.
These four areas are combined into a percentage. The result provides an indication of how much of the system remains auditable and reconstructable after change.
This matters because evidence can disappear even when the system itself continues operating. A model may be silently updated. An API may behave differently. A provider may stop making an older version available. A configuration may be overwritten. If those changes make it impossible to reconstruct what happened at the time of an important decision, the organisation may struggle to audit the system, investigate a failure or defend its actions later.
ERR can be used in due diligence, procurement, internal audit, regulatory assessment, litigation preparation and technical risk analysis. Different sectors may place different importance on the four components, but the underlying question remains the same. Evidential-Resilience-Ratio In simple terms, ERR asks: if the technology changes tomorrow, will the evidence needed to prove what happened today still exist?
A public reference standard that defines AI governance as a measurable structural condition, moving beyond policy statements to test whether control, accountability, traceability and integrity can be demonstrated in practice across the AI lifecycle.
A quantitative AI governance metric designed to measure how much of an AI system remains provable when models, APIs, vendors or other upstream dependencies change. ERR focuses on traceability, reconstructability, version fidelity and dependency proof to assess whether governance can survive change.
A practical governance test examining whether a business can maintain critical services if access to an external AI provider is suddenly restricted or withdrawn because of provider decisions, government action, sanctions, trade restrictions or geopolitical disruption.
A public reference framework for testing whether important AI protections actually work in practice, focusing on control, accountability, business dependence and effective oversight rather than relying on policies, promises or formal safeguards alone.