Why Policies Fail Machines

After this lesson, the reader should understand:

• 01Why traditional policy documents depend on human interpretation.
• 02Why AI systems create a new requirement for machine-readable authority.
• 03The difference between information and executable authority.
• 04The foundations of computable authority.

The EU AI Act

If an AI system must determine whether an organization complies with the EU AI Act, what representation of the regulation does the machine require?

The published Act is a 140-page instrument written for lawyers and regulators. A machine cannot act on it: the risk classifications are defined in prose, obligations are scattered across articles and annexes, and the conditions that trigger them are expressed as legal tests rather than evaluable predicates^[1]. What the machine requires is a representation in which each classification is a named definition, each obligation is a rule bound to that definition, and each trigger is a condition that can be evaluated against recorded facts^[2] — issued under a version, so a determination made today can be replayed unchanged in five years. This is not hypothetical. Regulation (EU) 2024/1689 has been compiled into exactly that shape^[3] — 851 rules, each traced to a specific Article, Recital, or Annex, sealed under version 1.0.0 — and published where anyone can query it. A keyless request carrying only 'risk_classification: high_risk' returns a determinate verdict of ESCALATED, names the rule it turned on (rule-002--d46fe9, sourced to Recital 51), and reports the 123 higher-priority rules still unresolved for want of facts. That transcript is public and re-fetchable at nomosprotocol.com/queries/qry_a5e0603057d9b0c4373f — a third party checks it without the operator's help. See also /examples. A precision about that transcript: the request in fact carried three facts — risk_classification, together with ai_system_processes_personal_data and gdpr_obligations_applicable — but the matched rule turned on the first alone, which is why one fact was enough to produce the verdict. The strongest objection is that the premise has dated. Large language models now read 140-page instruments fluently; why compile what a model can simply read? Because reading is not the property at issue. A model that reads the Act produces an interpretation each time it is asked, and empirical work on legal question-answering finds that interpretation unstable and, on harder questions, frequently wrong^[4]. More fundamentally, even a model that answered correctly every time would leave no version to cite, no rule to name, and no transcript a third party could replay. The objection is right that reading has become cheap; it misses that what institutions need from the representation is not comprehension but attributable, repeatable evaluation. The compiled form has limits of its own, and they should be stated. The 851 rules are one publisher's reading of the Regulation — the transcript says so in its own standing field — and a faithful-looking compilation can still be wrong in ways that then repeat consistently. Compilation does not settle what the Act means; it fixes one reading, names its author, and exposes it to correction.

• The transcript above escalates rather than authorising outright, with 123 higher-priority rules still unresolved for lack of facts. Is escalate-on-incomplete-information the right default, or should some class of those 123 be treated as non-blocking?
• rule-002--d46fe9 turned on a single fact and is sourced to a Recital rather than an Article. Should a recital-derived rule carry the same operative weight as one derived from an Article?
• The transcript's own standing field reads 'reference — the publisher's interpretation, not the issuing authority's own answer.' What would it take for the EU itself to publish the authoritative version of this artifact?
• If a language model could answer every question about the Act correctly, what would the institution still lack?

No responses yet. Be the first to answer one of the questions above.

Analyse an existing policy from your own organisation.

1. 01Identify the rules — the operative must, may, and must-not statements.
2. 02Identify the definitions the rules depend on.
3. 03Identify the exceptions and the conditions that activate them.
4. 04Identify the decision points where an outcome is produced.
5. 05Identify the unresolved ambiguity, and note who would resolve it.