Machine-Readable Policy and Accountable Decision Systems
Organisations want AI to help staff apply their own rules, from bursary eligibility to claims and permits. That only works if the rules exist in a form systems can apply consistently and people can defend. This eight-part series is for architects, engineers, policy owners and risk teams. It separates prose policy from structured rules, decision tables and code, shows what makes a rule authoritative and how to apply the version in force on the date that matters, handles conflicting rules and genuine ambiguity, defines the case file every recommendation needs, designs appeals as reassessment and systemic correction, keeps the model out of the role of the law, tests rules, models and whole decisions separately, and puts every change that can alter an outcome through one gate. It is the technical companion to Rewiring Government for the AI-Native Era, but the mechanisms apply equally to banks, insurers and other regulated decisions. It follows Amatola Bursary Agency, a fictional provincial bursary agency; its policy, rules, cases and figures are illustrative. Legal references are general information, not legal advice. Research claims were checked on 3 November 2026.
8 articles
- Part 1A policy document is not an executable ruleOrganisations want AI to help apply their policies, but a policy PDF is written for people to interpret, not for systems to execute. Part 1 of Machine-Readable Policy and Accountable Decision Systems separates prose, structured policy, decision tables and code, and shows where interpretation must be owned.9 min read
- Part 2The rule must be authoritative before the agent can helpA rule in a codebase is not authoritative because it runs. Part 2 of Machine-Readable Policy and Accountable Decision Systems sets out the record that gives a rule authority: source instrument, approval, effective dates, scope and supersession, and why systems must apply the version in force on the date that matters.7 min read
- Part 3When rules conflictReal policy libraries contradict themselves, and an AI assistant will resolve those contradictions silently unless the system stops it. Part 3 of Machine-Readable Policy and Accountable Decision Systems covers precedence, exceptions, genuine ambiguity and the mandatory escalation path.6 min read
- Part 4Every automated recommendation needs a case fileLogs scattered across five systems are not an explanation. Part 4 of Machine-Readable Policy and Accountable Decision Systems defines the case file: the inputs, evidence, rule versions, retrieved sources, model and tool versions, recommendation and official decision that let any outcome be reconstructed and defended.7 min read
- Part 5An appeal is a reassessment, not a second opinionAn appeal path only works if the system can reassess a case under the rule that applied at the time, correct facts everywhere they were copied, preserve the original evidence and fix every other case a faulty rule touched. Part 5 of Machine-Readable Policy and Accountable Decision Systems covers the mechanics.7 min read
- Part 6The model should not be the lawA language model is useful for reading documents and drafting explanations, and unsuitable as the place where rules live. Part 6 of Machine-Readable Policy and Accountable Decision Systems separates probabilistic interpretation from deterministic rules and from human authority, and shows how the three hand over.6 min read
- Part 7Test the rule, test the model, test the whole decisionAn AI-assisted decision can fail in the rule, in the model's reading of the evidence, in the joins between them or in the outcomes it produces for real people. Part 7 of Machine-Readable Policy and Accountable Decision Systems sets out a layered testing approach in which each layer catches what the others miss.6 min read
- Part 8Continuous learning must not silently change the rulesA new model version, a reworded prompt or a refreshed retrieval index can change decisions as surely as a policy amendment, without anyone approving it. Part 8 of Machine-Readable Policy and Accountable Decision Systems puts every kind of change through one gate: shadow run, difference report and owner approval.6 min read
Working through any of this in your own organisation?
These articles come out of delivery work rather than theory. If the problems in this series look familiar, a short scoping conversation is usually the quickest way to work out what is worth doing first.
Talk to CloudNala