Sequence verification

The vocabulary of skipped and out-of-order steps in automated processes, AI agents included, defined in one place.

Sequence verification is checking, at each step of an automated process, that the step is the one an order declared in advance says comes next, and keeping a record of each check that a party outside the process can examine. It answers a question ordinary monitoring does not ask. Monitoring asks whether the system is up and whether each call succeeded. Sequence verification asks whether the right steps ran, in the right order, and whether that can be shown to someone who was not there.

Automation has always run in sequences. What it never had was a listener: a party other than the doer that records whether each declared step ran.

The terms below are used loosely across engineering, operations, audit and AI writing, and several of them name the same failure from different sides. Each is defined as precisely as it can be.

The category

Sequence verification

Checking, at each step of an automated process, that the step is the one an order declared in advance says comes next, and keeping a record of each check that a party outside the process can examine. It needs two things that most systems lack: an order stated before the run, and a record not written by the process being checked.

Sequence enforcement

Sequence verification that acts. A step presented out of the declared order is refused before it runs, rather than found afterwards. The difference is timing: verification after the fact can establish that something went wrong; enforcement at the step stops it, and records the refusal.

Declared step order

The list of steps a process must follow, fixed before the run starts. It is the reference every other term on this page depends on. Without it a step cannot be called missing, early or late, because nothing said where it belonged. It also settles who decides the order: in a well-governed process it is the organisation, set in advance, and never the software carrying the process out.

Alignment

To align is to bring into line, and sequence is how alignment is achieved. The declared step order is the line. Enforcement outside the agent holds each step to its place in that line and refuses a step that is not in it. Model alignment, done when a model is trained, shapes what the model is likely to do; it cannot hold any single action to its place, because the model still chooses. Held by sequence, alignment can be checked: each step either sat where the declared order put it or was refused, and the record shows which.

Agent

A component that acts toward a goal by choosing its own steps, rather than executing an instruction list fixed in advance. The word's oldest technical sense is legal: an agent acts on behalf of a principal, and the principal answers for what the agent does within its authority. What matters for sequence verification is the route: automation without agency follows a declared program, so its order holds by construction, while an agent decides what to do next, so nothing declares its order unless one is imposed from outside. Whether a model drives it is not the line: a fraud scorer that evaluates each payment is automation; a component that pursues a goal across steps, uses tools and tries another way when one is refused is an agent.

Sequential logic (SLP-8)

SLP-8 is the sequential logic protocol behind AgenticRail: a declared order of steps, each evaluated before it runs, with every decision signed and linked to the one before it. The name comes from its default eight-step order; any order can be declared in its place, and one step is a valid sequence.

The failures

Order failure

A step that runs when the declared order did not allow it at that point: before a step it depends on, past a step that never ran, after a refusal that should have ended it, or after the sequence was sealed. It names the failure by its cause, the order, rather than by what the step did, which is why a rule written about outcomes can only reach it after the fact. The failures below are kinds of it, or ways it comes about. One case the list does not name separately is the refusal routed around: a check says no, and the process reaches the same result by a path the check never sees. The check worked; the order did not hold.

Out-of-order execution

A step running before a step it depends on has completed. In conventional automation it usually comes from concurrency: parallel workers, message queues that do not preserve order, a retried step landing after the one it should have preceded. In an AI agent it can also come from a choice, because the agent decides its own route. Either way the step itself typically succeeds, which is why the fault is found downstream.

Skipped step

A required step that never ran. It is the hardest failure in this list to detect, because a step that never ran writes no log entry. A record of everything that did run is complete, internally consistent and wrong, and nothing in it marks the gap. A skipped step can only be seen against a declared order that said it should have been there.

Procedural hallucination

An AI agent skipping, reordering or fabricating a step required by a stated procedure, and reporting the task complete. The defining feature is the report: the claim of completion is in the output, and the completion is not. The skipped step underneath it is an omission, a silent failure; what makes it a hallucination is the report that the procedure was followed. It is named for the obligation that was not met, not for anything happening inside the model. Set out in full in what procedural hallucination is in agentic workflows.

Race condition

A defect where the outcome depends on the relative timing of operations that run concurrently. It is a standard term in computing and one of the most common causes of out-of-order execution in automated systems. A race condition lives in the system that has it, and only changing that system removes it. What a gate in front of each step can do is turn a lost race into a refused, recorded step instead of a silent wrong order.

Execution drift

Used for a running process gradually diverging from its intended order as volume and concurrency grow: rare at fifty runs a day, routine at fifty thousand. It names the trend rather than any single failure, and the single failures underneath it are the out-of-order and skipped steps above.

Process desynchronization

The components running one process losing agreement about which of its steps have happened, so that one acts on a state another has not reached. A billing service charging for an account that the provisioning service has not yet created is the usual example.

Why they hide

Liveness versus correctness

In operations, liveness is whether a service is up and responding, and it is what health checks and uptime monitoring measure. Correctness is whether the right steps ran in the right order. A step run out of order returns success, so it reads as healthy on every dashboard. (In the formal sense used in distributed systems, liveness means that something good eventually happens and safety means that nothing bad happens. Step order is a safety property, which is the other reason liveness monitoring does not see it.)

Silent failure

A failure that reports success: every call returns a success status, such as HTTP 200, while the process it belongs to has gone wrong. Silent failures surface late, at an audit, a reconciliation or a customer complaint, because nothing raised an error at the time.

Presumed person

A system built on the assumption that a person is on the other end, with no step that checks it. Laws, forms, deadlines and duties were written for a human at each end: the person who asks, the person who consents, the person who checks, the person who approves. The step that would check the premise is usually missing, because until recently nothing else could have been there. New Zealand's Official Information Act is a plain case. A request may be made only by an eligible person, and the Ombudsman's guidance of October 2026 concludes that "there must therefore be a human involved in deciding to make the request"; the check it offers is to contact the requester and ask, which an agent can answer, and the 20-working-day clock does not stop while the agency asks. The human in the loop is the same premise from the other side: a requirement that a person reviewed the output, with nothing recording that one did. When the premise fails, every step still runs and every record still reads complete, which is why it is a silent failure. Sequence verification cannot tell a person from a program. What it does is make the person's act a declared step: if the step is in the order, its absence shows, and its record carries whatever the step carried, such as the signature of the person who took it.

The incentives

The failures above have an economic shape. Delegating a task to an agent, human or software, creates what economists call a principal–agent relationship, and the theory of that relationship predicts when steps get skipped and what kind of record can settle whether they were. Economists and AI researchers already apply this framing to AI systems. The terms below apply it to step order. What the theory predicts, tested against the public record, is set out in what game theory predicts about AI agents that skip steps.

Principal–agent problem

The problem that arises when one party, the principal, delegates a task to another, the agent, whose objective is not exactly the principal's and whose actions the principal cannot fully observe. Both conditions hold whenever an AI agent is given a goal: the objective it was given is never the whole of what was meant, and it acts at a speed and scale no one watches step by step. Contract theory, recognised by the 2016 Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel to Oliver Hart and Bengt Holmström, studies how arrangements can hold under these conditions. In a deployment the problem is nested: the operator is the principal of its agent and, at the same time, the agent of its customers and regulators, and it reports to them on both.

Hidden action

An action an agent takes, or fails to take, that the principal cannot observe; the problem it creates is called moral hazard. An arrangement cannot directly reward or penalise what it cannot observe, so a costly step that leaves no trace is the one an agent pursuing its goal has reason to skip. The prediction does not depend on intent: it follows from the objective and from what can be seen. A skipped step is hidden action in its plainest form, because it writes no log entry and so, to everyone outside the process, never happened. Sequence verification makes it observed: the declared order says the step belonged there, and the check either recorded it or did not.

Cheap talk

A message that costs nothing to send and cannot be checked. When the sender's interests differ from the receiver's, such a message carries little information, because an honest sender and a dishonest one can send exactly the same thing. An operator's own report of what its agent did is cheap talk in this sense. That is not a claim that any report is false. It is the reason a report alone cannot settle whether it is true.

Verifiable signal

A signal whose content the receiver can check, or which cannot be produced falsely, so that parties with something true to show can separate themselves from those without. The analysis of markets with asymmetric information, recognised by the 2001 prize to George Akerlof, Michael Spence and Joseph Stiglitz, is the study of how such signals work and what happens without them. A receipt signed at the moment of each step, by a party other than the operator, is a verifiable signal: an operator whose agent kept the declared order can produce one for every step, and one whose agent broke it cannot produce one for the step that broke it.

Commitment device

An arrangement by which a party removes some of its own options in advance, so that what it will do becomes credible to others. In announcing the 2005 prize, shared with Robert Aumann, the Royal Swedish Academy of Sciences wrote that Thomas Schelling “showed that a party can strengthen its position by overtly worsening its own options”. A declared step order locked at the first allowed step is a commitment device: once the run has begun, neither the agent nor the operator can change the order, and sealing closes the sequence so that no step can be added to it afterwards.

Mechanism design

Designing the rules of an interaction so that participants pursuing their own objectives arrive at the outcome the designer intends. Its foundations were recognised by the 2007 prize to Leonid Hurwicz, Eric Maskin and Roger Myerson. Sequence enforcement is a mechanism in this sense: an out-of-order step is refused and a skipped step shows as a gap against the declared order, so keeping the order becomes the agent's only route to its goal through the gate, whatever its objective, and nothing depends on anyone's honesty.

The controls

Replay

The same step presented a second time, whether by an attacker, an automatic retry or a duplicate message. A sequence control refuses it, typically by requiring a value used only once per step (a nonce). Replay here means the refused repeat. It does not mean re-running an execution to reproduce it.

Idempotency

The property that performing an action more than once has the same effect as performing it once. It makes a duplicate harmless. It does not establish order: an idempotent step that runs before its predecessor is still out of order, it is just out of order once.

Sealing

Closing a sequence when its final declared step completes, so that no further step is accepted into it. A sealed sequence is finite, so it can be hashed whole, archived and cited as one object. Where each record is linked to the one before it by hash, a sealed sequence cannot be reopened or altered without leaving a detectable break.

Conformance checking

A technique from business process auditing and process mining: declare the process model in advance, then compare a recorded event log against it, and the comparison reports what is missing or out of order. It is sequence verification after the fact. It finds a missing step only if that step would have been an event in the log, and it reports a fault once the step has already been skipped.

Orchestration

Running a process through a component that makes the declared order the only available path, such as a workflow state machine or a durable execution engine. It is the right first move and it works. Its account of what happened is still its own record, written by the system that ran the process, so it enforces order without producing evidence that a third party can check independently.

Workflow automation

Software carrying out the steps of a business process without a person between them. The term covers scheduled jobs, scripts and pipelines, no-code and low-code workflows, robotic process automation (software robots that operate applications through their user interfaces), and AI agents. They share the failures above. AI agents add one: a component that decides its own route.

Independent record

A record of what a process did, produced by something other than the process itself and checkable without trusting the party that ran it. It is the same principle as segregation of duties: the part doing the work is not the part certifying that the work was permitted. A record written by the system under examination can be accurate, but it cannot be independent.

Where AgenticRail sits

AgenticRail is sequence enforcement with an independent record. Each step of a process, whether it is run by an AI agent, a script, a scheduled job or a no-code workflow, is presented to the gate before it runs. The gate checks it against the declared order and returns ALLOW or DENY. Out-of-order steps, repeated steps and steps after the seal are refused, and every decision is written as an Ed25519-signed receipt, hash-linked to the one before it, that anyone can verify offline against published keys.

Two limits. It does not remove race conditions or fix the systems that have them: it refuses the step that arrives out of turn. And it only sees the steps that are presented to it, so a step that bypasses the gate entirely is not refused; it is the absence of its receipt, against the declared order, that shows it. For callers that read only the HTTP status, a refusal can be returned as HTTP 409 by adding ?deny_status=409 to the request, so a flow stops on its own. The rest is on the product page and in the documentation.

Further reading

The Completeness Specification — the eight requirements that separate an evidence-grade enforcement record from an ordinary log.

When an AI agent skips a step, your audit log shows a clean run — the skipped-step problem worked through.

Orchestration versus enforcement — what each produces, and why they are not substitutes.

What game theory predicts about AI agents that skip steps — five predictions from principal–agent theory, tested against the record.

Questions — including whether this works for automation that is not AI.

Office of the Ombudsman, Advice for agencies on information requests and AI, 6 October 2026 — the source for the Official Information Act case under presumed person.

The Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel, prize summaries: 2001, 2007 and 2016; and the Royal Swedish Academy of Sciences, press release for the 2005 prize.

He toi whakairo, he mana tangata