What Is Human in the Loop (HITL)?

🧩 Where the checkpoint sits, and what each placement buys

Three placements are common and they are not interchangeable. Approval before the effect: nothing proceeds until a person releases it. A withdrawal window: the action proceeds and can be recalled for a defined period. Sampling: a portion of the flow is reviewed and the rest proceeds untouched, which measures a process rather than gating it.

Each buys something different and each has a characteristic failure. A blocking approval buys prevention and costs latency; its failure is the reviewer who approves everything, because approving is the fast path through a queue. A withdrawal window buys throughput and loses the cases that turn out to be irreversible sooner than the window closes. Sampling buys visibility and gates nothing at all — which is exactly what it is for, and exactly what it is mistaken for.

The phrase has a second established sense, and the two are worth keeping apart because a specification naming the wrong one describes a different system. Everything above is an operational control: a person standing in a live process, deciding one case at a time, present at the moment the case arrives. In machine learning the same phrase names a data-collection arrangement — people supplying labels, corrections, or rankings of candidate outputs, which are then used to train. Reinforcement learning from human feedback is the familiar instance: rankings are collected from people, a reward model is fitted to them, and training optimises against that model. Nobody is present when the trained system afterwards runs. The judgement was gathered once and generalised, so it applies to every future case at once, which is the exact inverse of a checkpoint's defining property of applying to precisely one.

Only the first two can stop anything, and only the first can stop everything. A design that calls itself human-in-the-loop while using the last row is describing an audit, which is a useful thing to have and a different thing to claim.

🗃️ How the Registry classifies an approval-gated task

Routing a refund decision to a reviewer before the payment is released is filed as a transaction in this registry's classification. The class covers purchases, bookings, payments and orders, and the presence of an approval step does not move a task out of the class the task belongs to.

The Registry's canonical brief for this entry is filed as: Route every outbound refund decision to a human reviewer for approval before the payment is released. Submitted for assessment it is classified as Transaction, and its wording is hashed once — to 3c7154ac4e7abcb3…, the first sixteen of sixty-four hexadecimal characters — with the wording itself never stored. The hash is what the derivation reads. That class's own page is /tasks/transaction.

Classification is one of three inputs. The other two are the configuration submitted with the task, and the permanent chart derived from that configuration — fixed by the model name, the training cutoff and the temperature alone, and never reading the task at all. The same ascendant, ruling planet and harmony therefore appear on every assessment a given configuration receives, whatever it was asked to do. The derivation is published in full at /method.

❓ Questions about human in the loop

What is human in the loop?

A human-in-the-loop design places a person at a defined point in an automated process with the authority to approve, correct or reject what reaches them before it takes effect. The checkpoint sits upstream of the consequence rather than reviewing it afterwards.

What is the difference between human in the loop and human on the loop?

On-the-loop generally describes supervision of a running process — a person monitoring it, able to intervene — while in-the-loop describes a person the process passes through. The terms are used loosely enough that a specification should state where the checkpoint sits rather than rely on the label.

Does a human checkpoint slow the process down?

A blocking approval does, by design: nothing proceeds until it is released, and that latency is the price of the prevention. A withdrawal window or an exception route trades some of that prevention back for throughput. There is no placement that gates everything and costs nothing.

When should a decision require a person?

Where the effect is irreversible, where it reaches someone outside the operator, or where a rule requires a person to make it. The first two are design judgements about blast radius. The third belongs to whoever holds the obligation and is not settled by a page like this one.

What does human in the loop mean in machine learning?

There the phrase names a training-time arrangement rather than a runtime gate: people supply labels, corrections or rankings of candidate outputs, and those judgements are used to train. Reinforcement learning from human feedback is the common instance. The difference that matters is when the person acts — once, before deployment, with the result generalised across all later cases, rather than during a single live decision.