Standing to Stop: Authority provenance before human oversight becomes real

Key points

• Three properties are routinely collapsed into one: place (where intervention can occur), standing (why this party is entitled to exercise the power now) and effect (whether exercising it changes what outcomes remain reachable).
• “Standing” is used as a narrow working label for presently valid, scope-bounded authority grounded in identifiable sources and subject to their conditions, expiry, revocation and supersession.
• An execution record proves that an approval happened. It does not, by itself, prove that the approver’s authority covered that decision, was current or survived later restrictions.
• Automation can make execution ever more provable while leaving the basis of authority unrecoverable — preserving a broken or unknown provenance state and scaling its consequences.
• The paper proposes a testable query, standing provenance: from the acting party and the power exercised, trace to the source grants, their scope and conditions, and whether they held at the time of action.

The place is not the authority

An earlier essay argued that consequential decisions need a structurally defined place where they can still be stopped before execution. This companion piece starts where that argument ends. Suppose the approval gate exists, the role is named, the account is valid and the audit trail shows who clicked approve. A more basic question can still be unanswered: what establishes that this role is entitled to exercise this power, over this class of outcomes, now?

A control can be operationally present and normatively under-specified at the same time. The gate works and the approver acts in good faith, yet the basis of the approval has become stale, narrower than the action, superseded by another grant or simply impossible to reconstruct.

Authority has provenance

Organisations already behave as if authority comes from somewhere: a manager acts under a board delegation, the board under constitutional instruments, a regulated officer under an internal appointment and an external regime. The present act is the visible end of an older structure of grants — better described as an authority provenance graph than a single chain, because sources can combine, narrow, supersede or suspend one another.

That makes provenance distinct from hierarchy, access control and workflow. An organisation chart shows reporting lines, access control shows which account may perform an operation, and a workflow shows where approval is requested. None of them answers which source, or combination of sources, makes this exercise of power permissible at the moment it is exercised.

How provenance erodes

Provenance rarely fails dramatically. People leave, functions merge, policies are rewritten, temporary delegations become routine and permissions are copied during migrations. A process with degraded provenance behaves exactly like one whose provenance is intact: it runs, meets service levels and produces complete logs.

Person and position preserve different things. A person may keep the context of a delegation without continuity; a position keeps continuity without context. Standing survives turnover only when its basis is recorded independently of both. Shadow AI is a current example of a related problem: a model inserted into a decision path changes who or what participates in a consequential decision without breaking any existing rule.

What automation preserves

Human processes introduce variation: people interpret, escalate unusual cases and sometimes stop following rules that no longer make sense. Encoding removes much of that. A rule can then execute identically millions of times with a perfect history, while the answer to why it was entitled to determine the outcome remains unavailable. Automation does not necessarily break provenance; it can preserve an already broken state and surround it with an excellent audit trail.

Standing provenance as a query

The requirement proposed is deliberately narrow and testable. For a consequential exercise of authority, an organisation should be able to start from the acting party and the power exercised, reach the source or sources of that power, identify the scope and conditions of each grant, determine whether those conditions held at the time, and distinguish what is evidenced from what is merely asserted, expired, superseded, revoked, disputed or unknown.

The query does not require every path to be intact. A result such as “three grants evidenced, one source unrecoverable, one delegation expired before the decision” is a successful query returning an uncomfortable state — and it gives the organisation targeted options: re-authorise, narrow, retire or investigate.

Place, standing and effect

Without place, there is nowhere to intervene before the outcome becomes real. Without standing, an intervention point exists but the authority of the party at it cannot be established. Without effect, the authority may be valid and traceable while execution proceeds anyway through a timeout, fallback, override or alternate path. Each needs different evidence: the decision architecture for place, authority provenance for standing, and counterfactual reachability testing for effect. A system should not receive credit for one property because it has evidence for another.

What the paper does not claim

The paper is explicit about its limits. Standing provenance makes the traceability and evidentiary status of an authority claim visible. It does not establish that a delegation was wise, lawful in every jurisdiction or competently exercised, that a process is complete, or that a stop is enforced. Recorded is not the same as true.

This page is a summary written by One World AI. The paper on Zenodo is the authoritative version.

Cite this paper

Bański, P., & Sobieski, S. (2026). Standing to Stop: Authority provenance before human oversight becomes real. Zenodo. https://doi.org/10.5281/zenodo.22810377

More research

• Architectural Constraints in AI Governance — Governance of AI-enabled operational systems has to live in the architecture that decides which state transitions are admissible — including across sequences of individually acceptable steps.
• Enforceable Decision Boundaries — Inference may propose actions but never authorise them. Authority is resolved outside the model, and execution is impossible without it.
• Prompt Engineering as a Control Illusion — Prompts influence what a model says. They cannot decide what a system is permitted to do.
• The Missing Place for “Stop” — Systems that work well for years can still drift into decisions nobody can defend — because no one is structurally placed to stop them before execution.