Lecture 2.7

The States of Authority

A decision structure does not only return 'authorised'. It returns one of a small set of states, and the discipline requires each to be explicit: deterministic authorisation, conditional authorisation, a referral to human authority, and an unresolved state where authority cannot be established. This lecture treats those states as first-class outputs and argues that a machine must never convert the absence of an answer into a fabricated one.

Supports Law 7 — The Boundary of Authority Must Be Enforced, Not Assumed. →

Lesson 13 / 40Start the course to keep your place and get the rest by email.

Learning objectives

After this lesson, the reader should understand:

01Enumerate the outcome states an authority artifact can return and what each asserts.
02Distinguish a denial (authority says no) from an escalation (authority is undetermined).
03Design a rule so that 'human judgement required' is an authored outcome, not a fallback.

Concept framework

Four states

01Deterministic — conditions resolve; the artifact authorises or denies
02Conditional — authorised unless a further condition holds; otherwise escalate
03Discretionary — the artifact routes to a named human authority by design
04Unresolved — authority cannot be established; the action stops and returns to the institution

Case study

The missing 'no'

An artifact evaluates a hardship waiver. The facts needed for one condition are not in the record, and the rule has no branch for that. What should it return?

Not a denial, and certainly not a grant. A denial asserts the institution considered the case and refused it; the institution did no such thing. The correct output is an unresolved state — authority could not be established on the evidence available — which halts the automated path and hands the case back with the evaluated context attached. Institutions resist this because an escalated count looks like a system that does not work. It is the opposite: an artifact that always returns authorised or denied has been built to hide the cases it cannot actually decide. 'The machine is not authorised to decide this' is itself a statement of computable authority, and a valuable one. A reference runtime shows how the four states are carried in practice. Its literal verdicts are three — AUTHORIZED, DENIED, ESCALATED — and the conditional state is not a fourth verdict but a disclosure attached to the others: the transcript reports how many higher-priority rules were left unresolved for want of facts, so a caller can tell whether the verdict was sealed with every relevant condition actually decided. The unresolved state itself splits by cause^[1]. A rule with no branch for the facts present escalates and returns the case. A request whose issuer cannot be recognised at all fails before any rule is consulted, as a distinct error rather than a verdict — because re-presenting better evidence of authority can fix the second and never the first. The strongest objection is practical. An honest unresolved state moves work back to people, and the people it moves work to are already the bottleneck automation was meant to relieve. Worse, it can be gamed: an applicant who expects a stricter outcome from the machine than from a caseworker has every reason to withhold the fact that would decide the case. Both halves are real. The answer to the first is that an escalation should carry the decisive question, not just the case: a runtime that knows which missing fact could still change the verdict can ask for that fact and nothing else, and stop asking once the verdict is stable — the reference runtime's intake planner works this way. The answer to the second is that withholding a fact should not be cheaper than supplying it; an escalation for missing evidence is a request for evidence, not a route to a more lenient reader. There is a sharper critique of the referral itself. Research on mandated human oversight of government algorithms finds that reviewers frequently defer to the machine's framing^[2], and that oversight requirements can legitimise a system without making its decisions better. An unresolved state routed to a person who rubber-stamps it has only relocated the fabrication. The discretionary and unresolved states therefore need what the deterministic ones have: a recorded decision, a named decider, and a trace that shows what the human actually changed. Finally, the unresolved state defers; it does not abolish. Legal systems generally refuse a court the option of declining to decide — the doctrine of non liquet is mostly a doctrine of its prohibition^[3]. An institution may tell its machine not to decide. It may not tell itself that.

Discussion questions

Is an escalation a failure of the artifact, or a designed feature of it?
Should the four states be visible to the decision subject, or only to the institution?
What proportion of escalations is healthy, and what would too few tell you?
Who is accountable for a case that sits unresolved because no one owns the referral path?
How would you know whether the humans receiving escalations are deciding them or merely approving the machine's framing?

Discussion (0)

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

Start the course to join the discussion.

Exercise

Assign an outcome state to every path in one rule.

01List every combination of conditions the rule can encounter, including missing facts.
02Assign each to one of the four states.
03For every discretionary path, name the human authority it routes to.
04For every unresolved path, state what context travels with the referral.
05Flag any path currently resolved by a silent default.

Footnotes

^1 Computable Authority, §10 — 'Ambiguity and Discretion'.
^2 Ben Green, "The Flaws of Policies Requiring Human Oversight of Government Algorithms," Computer Law & Security Review 45 (2022): 105681.
^3Julius Stone, "Non Liquet and the Function of Law in the International Sphere," British Year Book of International Law 35 (1959) — whether a tribunal may decline to decide.