Lecture 5.6
If authority is executable, it must be callable. This lecture covers the interface through which systems request an authority evaluation and receive a bound, traceable outcome.
Supports Law 4 — Authority Must Precede Automation. →Learning objectives
After this lesson, the reader should understand:
Concept framework
Evaluation interface
Case study
Six systems apply the same policy. Should each implement it?
No — six implementations are six interpretations, and they will diverge on the first amendment. A single evaluation service means one artifact, one version in force, and one record format across every calling system[1]. The organisational difficulty is not technical: it is deciding which body owns the artifact when six teams previously owned six copies. The NOMOS reference implementation exposes this as one call. A published SDK primitive — can() — resolves a request three ways from a single return shape: against a public authority (free, or paid with a key), against an artifact the caller owns, or against an artifact presented with its chain of certificates and a pinned root. The response carries the verdict, the reasoning trace, any obligations, the artifact's own sealed version, and whether it has been revoked. A request that cannot pay raises a specific payment-required error rather than being silently charged; a chain that will not resolve raises a specific not-recognised error rather than returning a fabricated verdict. The keyless path behind it is the same one a regulator or a counterparty can call with no prior relationship to the operator. The strongest objection is that one evaluation service is one point of failure, and one point of capture. If it is down, six systems stop; if its operator changes a rule or a price, six systems change with it. Both are real, and each needs a specific answer rather than reassurance. For availability, every caller must decide in advance what it does when the service cannot be reached, and for consequential actions that answer should be to fail closed — hold the action — rather than proceed unchecked. The reference implementation's gateway for agent tool calls does exactly that, blocking every call while its policy service is unreachable[2]. That is safe and also costly, which is why the artifact being self-contained and deterministic matters: in principle it can be evaluated anywhere, so the service is a convenience, not a necessity. Whether a given client actually evaluates locally is a property of that client and should be checked, not assumed. For capture, the protection is the same property: an artifact anyone can verify and re-execute without the operator gives callers a way to leave. The subtler limit is that one service does not by itself mean one interpretation. Each of the six systems must translate its own records into the facts the artifact reads — its 'customer' into the artifact's 'applicant', its risk score into the artifact's risk band. Those six adapters are where interpretation re-enters, quietly, in six different places. The shared service removes six copies of the rule; it does not remove six mappings of the data, and a divergence in a mapping produces exactly the inconsistency the service was meant to end. The adapters need the same discipline as the artifact: named owners, declared mappings, and tests. Organisations tend to reproduce their own communication structure in their systems[3], and an evaluation service with six unowned adapters will reproduce six teams.
Discussion questions
Discussion (0)
No responses yet. Be the first to answer one of the questions above.
Start the course to join the discussion.
Exercise
Specify one governance endpoint.