The Coravian Institute for Machine Intelligence has published its first safety-audit framework, setting out how AI systems used in public-sector decisions will be tested for accuracy, bias and failure behaviour — voluntarily for a year, and as a binding requirement from late 2027.
The framework requires documented evaluations before deployment, incident reporting within 72 hours, and — its most-discussed provision — a “contestability” standard: any citizen affected by an automated decision must be able to trigger human review.
“We are not auditing mathematics. We are auditing consequences,” said institute director Dr Sena Vall, positioning the rules deliberately between the Accord’s lighter guidance and the prescriptive licensing regimes emerging elsewhere.
Public bodies use an estimated 300 automated decision systems, from benefit triage to traffic enforcement; a survey by the institute found fewer than half had any documented evaluation. The framework’s first-year voluntary phase comes with free audit tooling and, Vall noted dryly, “the certainty that year two exists”.
Industry response split on familiar lines: public-sector suppliers called the timeline “workable”, while startup groups warned the contestability standard could freeze smaller vendors out of procurement. The institute publishes its first audit league table next summer — a document likely to be read as closely in ministries as in boardrooms.