Is every long event liquid?

No. A long completed task may contain smaller periods that do not themselves constitute that task.

Can a liquid predicate imply truth outside its known interval?

No. The basic property supports contained subintervals, not unrestricted extension.

Does a liquid property hold between two positive sensor readings?

Not automatically. The inference requires a justified assertion about the whole interval, not only its sampled endpoints.

AI Glossary / Graphs and symbolic AI / Intermediate

Liquid event

Definition

A liquid event is an event type or temporal property that, when true throughout an interval, is also true throughout its subintervals. The term comes from temporal knowledge representation and event semantics. It describes how an event predicate behaves over time, not whether the event involves a physical liquid.

What it means

Some descriptions refer to an ongoing condition, such as a lamp being on throughout a period. Restrict attention to a smaller period inside it, and the description still applies. Other descriptions require a complete change or accomplishment, such as turning the lamp on. A smaller interval within that change does not necessarily contain the whole event.
Liquidity therefore belongs to the chosen description and formal model. The same real-world episode can support a liquid state-like description and a nonliquid completed-event description. It is important to specify what the predicate means rather than classify an informal verb without context.

How it works

An interval-based model records that a predicate holds over a period. For a liquid predicate, the inference to contained subintervals is licensed by its semantics. Some formulations additionally require that truth on overlapping intervals extends to their combined span, provided the same predicate and participants are involved.
This creates a representation issue. With continuous time, one interval contains infinitely many subintervals. A system should not necessarily enumerate each one. Compact interval representations can capture families of valid intervals. Compound event descriptions, especially those requiring a sequence or transition, need not retain liquidity even when their component predicates are liquid.

Three practical examples

A video-reasoning model can describe an object as supported throughout an interval and infer support during shorter contained intervals.
A monitoring model can represent a device being active during a continuous period, subject to the accuracy and granularity of its evidence.
A language-understanding system can distinguish an ongoing process description from a completed accomplishment requiring a beginning or endpoint.

Think of it this way

Think of a uniformly colored strip of paper: any smaller piece cut from inside it retains the same color. In contrast, a strip showing an entire before-and-after sequence may lose that sequence when cut. The analogy is limited because temporal predicates depend on modeled semantics and observations, not literal material that can be divided.

Worked example

Consider an invented idealized record in which lamp L is on throughout the closed interval [2, 8], measured in minutes. Treat On(L) as a liquid property and assume the model's interval assertion is correct.
It follows that On(L) holds throughout [3, 5], [2, 4], and [6, 8], because each is contained in [2, 8]. It does not follow that the lamp is on throughout [1, 4], since that interval extends before the known period. Nor does the assertion alone tell us what happened after minute 8.
For a finite counting exercise, restrict endpoints to the seven integers 2, 3, 4, 5, 6, 7, and 8, and count only intervals with distinct start and end values. Choosing any two endpoints determines one such subinterval. The count is 7 times 6 divided by 2, or 21. If point intervals are also allowed, seven additional intervals raise the count to 28. With arbitrary real endpoints, there are infinitely many possible subintervals instead.
Now consider the different description: L changes from off to on during [1, 2]. That does not imply that L changes from off to on during every subinterval of [1, 2]. A subinterval containing only the already-on portion does not contain the transition.
Finally, if a model licenses the overlapping-span property and separately states On(L) throughout [2, 6] and [5, 9], it can combine them into [2, 9]. The overlap matters; knowing [2, 4] and [6, 9] does not justify filling the unobserved gap.

Illustrative case study

In a fictional video-annotation project, Elm Events labels an entire clip as picking up a box. A developer treats that label as liquid and copies it to every short subclip. Some subclips show only the stationary box before contact, while others show it already held.
The team separates ongoing predicates such as supported or held from the compound pickup event. This invented case illustrates why subinterval inference must follow the event definition. It does not claim that a particular vision system can reliably infer all such predicates from images.

Common misconceptions and mistakes

Liquid does not mean fluid motion, a water-related event, or simply something that lasts a long time. Duration alone does not establish the subinterval property.
Do not apply liquidity to a completed event merely because its components persist over time. Nor does truth at two sampled instants establish truth throughout the interval between them. That requires additional evidence or modeling assumptions. Endpoint conventions and temporal granularity also matter: a formal continuous-time description may be an idealization of noisy measurements rather than a direct guarantee about arbitrarily small physical intervals.

Business use

This concept helps teams designing temporal annotations, event-recognition rules, and process-monitoring representations. Define which labels mean holds throughout and which mean an entire transition or accomplishment occurs. Those meanings lead to different valid inferences.
Avoid generating redundant subinterval records when a compact representation can preserve the same information. When combining observations, retain uncertainty and gaps rather than treating all nearby intervals as continuous. Evaluate annotation rules against short clips or time windows that cut across event boundaries. That often exposes semantic mistakes that remain hidden when only complete episodes are reviewed.

Check your understanding

Why are there 21 nonzero-length integer-endpoint subintervals in the worked interval [2, 8]?

Answer: There are seven permitted endpoints. Each choice of two different endpoints gives one start-before-end interval, so the count is 7 times 6 divided by 2. Point intervals were explicitly excluded.

Related terms

Fluent
Interval algebra
Compound Events
Event logic
temporal interval
process
event semantics
subinterval
spanning interval
compound event
temporal granularity

Learn next

Compare being open, opening, and completing an opening-and-closing cycle. For each description, ask whether it must remain true on every subinterval. Then study compact interval representations and how they preserve the difference between continuous conditions and ordered compound events.

Compound Events

Sources

Jeffrey Mark Siskind, Grounding the Lexical Semantics of Verbs in Visual Perception using Force Dynamics and Event Logic, original research paper: https://arxiv.org/pdf/1106.0256

Previous: Lipschitz continuity Next: Liquid State Machine Network