Skip to content

Systems and records

World model causality evidence: did the event before a failure cause it?

By SourceX Editorial · Updated

10 sections

Short answer

World model data should not treat an event that happened before a failure as its cause. Grade each link instead: verified timing on a checked clock, a mechanism recorded by someone who inspected the equipment, a pattern that repeats and scales across assets, and finally an intervention that changed the suspected cause. Label links by the level they reach, and rule out clock, sensor and recording errors first.

Video and sensors

World model causality

This page covers world model causality evidence: did the event before a failure cause it.Illustrative

Key takeaways

  • A link that rests only on ordering in a log is a sequence and should be labeled as one, not as a cause.
  • Grade event-to-failure links on a seven-level hierarchy from contradicted to designed intervention, and record why each link stopped where it did.
  • Check timestamps before mechanisms: source time versus server or ingest time, clock offsets, after-the-fact entries and historian compression.
  • Operator actions taken because of a symptom look like causes; tag actions with the reason recorded at the time.
  • Hill's nine viewpoints and refutation tests are review prompts that map onto the levels, not a score.
  • Control logic makes some sequences causal by design, but the label should say design, because logic changes when it is edited.

How should world model data separate event sequence from causation?

Give every event-to-failure link in the dataset an evidence level instead of a yes or no. A link that rests only on one event being logged before another is a sequence, and should be labeled that way. It earns a causal label only as stronger evidence accumulates: verified timing, a physical mechanism recorded by someone who inspected the equipment, a pattern that repeats and scales across assets, and finally an intervention that changed the suspected cause and observed the result.

The hierarchy below makes that grading explicit for plant, warehouse and fleet records. It sits beside the timing and source checks that come first, because a link whose order is wrong never reaches the first rung, and it travels with the data so a model team can choose which levels to train or evaluate on.

This is written for AI teams scoping or buying operational data for world models, and for the operations engineers who would label it. It does not replace a root cause investigation; it records how far the evidence in the files actually goes.

Why a world model learns order unless the data records interventions

A world model trained on logs learns what tends to follow what, and that is not the same as learning what would happen if someone acted differently. Ha and Schmidhuber's World Models paper describes a generative model that learns a compressed spatial and temporal representation of an environment, which an agent's controller then uses. When the environment is a simulator, the agent's actions are real interventions. When the environment is a historian export and a maintenance system, most of the actions were chosen by people reacting to what they saw, and the data only shows the order things were recorded in.

That gap matters most for the uses buyers care about: predicting the consequence of an action, ranking interventions, or explaining a failure. A model that has absorbed a spurious link, such as a cleaning cycle that is always scheduled after a quality alarm, will confidently predict alarms from cleaning. Labeling each link's evidence level lets the model team weight, filter or hold out links by strength instead of discovering the problem in deployment.

The causal evidence hierarchy for industrial datasets

Assign the highest level whose conditions are all met, and record which condition failed when a link stops short. Levels are cumulative: a link at level 4 has also passed levels 1 to 3. The source column names typical places the evidence lives; the hierarchy works the same whether the records come from a historian, a CMMS or a telematics feed.

Level 3 leans on the way maintenance records already describe failures. ISO 14224, the reliability and maintenance data standard for the petroleum, petrochemical and natural gas industries, treats failure data such as cause and consequence as core content and names failure events and mechanisms among its uses. Coded findings in that style are far stronger evidence than a free-text note saying a part was replaced.

The causal evidence hierarchy for industrial datasets
LevelWhat the records must showWhat a model may learn from itTypical evidence source
0: ContradictedThe suspected cause is logged after the failure once clocks are checked, or the states involved are physically impossibleNothing causal; keep only as a data quality exampleClock audit, state constraint check
1: Recorded sequenceEvent A is logged before failure B in at least one system; time base and units not yet verifiedCo-occurrence and ordering onlyAlarm journal, CMMS work order times
2: Verified sequenceA precedes B on a checked time base, units and scaling are confirmed, and the window between them has no observation gapReliable ordering and delay distributionsHistorian tags with source timestamps, synchronized logs
3: Mechanism recordedA person who inspected the equipment recorded a failure mechanism that connects A to B, and the intermediate states are visibleA plausible pathway worth testingTechnician findings, teardown report, failure codes
4: Repeated and gradedThe pattern recurs across episodes or assets, larger or longer A goes with earlier or more frequent B, and B is rarer when A is absent in comparable conditionsA robust association that survives comparisonMulti-asset history, comparable fleet or line
5: Natural interventionA was changed for a reason unrelated to B's warning signs, such as a schedule or a supplier switch, and B changed afterward relative to comparable assetsAn estimate of the effect of changing A, with stated assumptionsChange logs, management of change records, schedules
6: Designed interventionA was deliberately varied in a test or trial with B measured, and the result was documentedA causal effect for the tested conditionsBench tests, plant trials, OEM test reports

Is the order even real? Clocks, units and timestamp provenance

Before any link is graded, confirm that the order in the file is the order in the plant. Industrial data often carries more than one timestamp per value, and they mean different things. In OPC UA, the sourceTimestamp is applied by the originating data source and is not changed by servers that relay the value, while the serverTimestamp records when a server received the value, so each server in a chain can apply its own. A historian that stored only server time can place a downstream reading ahead of the upstream event that caused it.

Robot and vehicle logs have the same issue. The MCAP specification distinguishes a message's log_time, when it was recorded, from its publish_time, when it was published, and requires publish_time to fall back to log time when no separate value is available. Know which one your export uses.

Where devices on a network are meant to share a clock, ask how. IEEE 1588-2019 defines the Precision Time Protocol for synchronizing clocks in networked measurement and control systems; whether a site actually runs it, and on which devices, is a question for the controls engineer, not an assumption.

  • Record the time zone and daylight saving handling for each source, and the clock each device follows.
  • Flag records entered after the fact: work orders closed at shift end, operator logs written from memory, batch uploads.
  • Check units and scaling on every tag that feeds a link, including changes after instrument replacement or recalibration.
  • Check the historian's compression or deadband settings, which can hide short excursions between stored points.
  • Estimate the residual clock offset between sources and refuse to order events closer together than that offset.

Data faults fake causal links often enough that each explanation below needs its check before anyone argues about mechanisms. The table pairs the common data-side explanations with what they look like and what they do to a link's level.

Before blaming the event: data faults that fake a causal link
Alternative explanationWhat it looks likeCheck to runEffect on the link
Clock offset between systemsConsistent lead or lag between two sources across many eventsCompare a shared event, such as a trip, seen by both systemsHold at level 1 until corrected
Ingest time used as event timeBursts of identical timestamps, values arriving in batchesInspect timestamp fields and gateway bufferingHold at level 1; reorder with source time if it exists
Sensor dropout or stale valueFlat line, repeated last value, bad or uncertain quality codesRead quality flags and look for gaps in the windowCannot pass level 2 across the gap
Event never loggedFailure with no preceding record of any kind, or a manual action with no entryCompare with shift logs, camera clips or interviewsMark as unobserved, not as no cause
Unit, scaling or tag remapStep change in a value on a date with no process changeMatch against instrument and configuration change recordsLevel 0 for links built on the wrong scale
Impossible stateValve open and closed together, speed above design limit, negative levelApply physical state constraints from the equipment ownerLevel 0 for the affected window

Actions taken because of a symptom, shared causes and saturated capacity

Even with clean timing, three patterns produce strong sequences that are not causes. The first is reverse direction: operators act because they see an early symptom, so their action appears just before the failure it was trying to prevent. A world model that learns from that record will treat the response as a cause. Tag operator and maintenance actions with the reason recorded at the time, and never grade an action higher than level 2 when the reason was a symptom of the failure.

The second is a shared cause. A heat wave, a new raw material lot or a change in product mix can raise the chance of both the event and the failure. Record known shared drivers alongside each link and compare within strata, for example same product and same ambient band, before claiming level 4.

The third is capacity. When a line, queue or cooling system runs at its limit, small events have outsized consequences and the same event at lower load does nothing. Record load or utilization at the time of the event; a link observed only at saturation should say so, because the model will otherwise apply it at every load.

Using Hill's viewpoints and refutation tests as review prompts

Two established toolkits help reviewers argue about level 3 and above without pretending to precision they do not have. In 1965, Austin Bradford Hill set out nine viewpoints for judging whether an observed association is causal: strength, consistency, specificity, temporality, biological gradient, plausibility, coherence, experiment and analogy. Several map directly onto the hierarchy: temporality is level 2, gradient and consistency are level 4, experiment is levels 5 and 6. Use them as prompts in the review, not as a score to add up.

When a team does estimate an effect from observational records, it should also try to break the estimate. The DoWhy library organizes effect estimation into four steps, model, identify, estimate and refute, and its refutation API is designed to test the causal assumptions behind an estimate. A link that fails a refutation check, such as adding a random common cause or substituting a placebo treatment, should drop a level, and the dataset card should say which check it failed.

Illustrative case: a compressor trip after a pump changeover

All values in this example are illustrative. A food plant's refrigeration compressor tripped on high discharge temperature on 3 occasions in one year, each within about an hour of a changeover between two cooling water pumps. The maintenance system shows the changeover first, so the raw link starts at level 1.

The clock check finds that changeover times in the maintenance system were typed in at shift end. The team replaces them with the pump run-status tags from the historian, which carry source timestamps. On that time base the changeover still comes first, and discharge temperature climbs steadily in between with no data gaps. The link reaches level 2.

The technician's findings for each trip record reduced cooling water flow after the changeover because one pump's discharge valve had been left partly closed. That is a recorded mechanism with visible intermediate states, so level 3. Looking across all 14 changeovers that year, only those onto one particular pump preceded a trip, and the trips clustered on the hottest days. The team records level 4 for changeovers onto that pump in hot weather, not for changeovers in general.

No one changed the valve procedure or ran a test during the period, so the link stops at level 4 with the reason stated. A fourth trip that year, with no changeover nearby, came on a day when the discharge temperature sensor reported a stale value for 20 minutes; it cannot be placed on the hierarchy at all and is labeled unobserved.

When a sequence is causal by design, and where the hierarchy stops

The exception is engineered cause and effect. When a control system's documented logic says that a given high reading trips a given unit, and the configuration in force at the time is known, a trip that follows that reading is causal by design, and it can be labeled that way at once. The label should say design rather than evidence, because what the model learns is the plant's logic, which changes when the logic is edited, not the underlying physics.

The hierarchy also has limits. It grades links that someone proposed; it does not find causes that no one thought to test. It depends on the honesty of maintenance notes, which vary by crew. And a high level for one asset, product or season does not transfer to another without new evidence. Stop and bring in a process or reliability engineer when the links that matter most to the model sit at level 3 or below, when state constraints are unknown, or when the records disagree about what happened.

Asking for records that can carry a causal label

Causal labels are only as good as the records behind them, so a data request for world model work should ask for those records explicitly: source timestamps and quality codes, configuration and change logs, maintenance findings, operator action reasons and any test or trial reports, with the rights to use them for training and evaluation.

World model teams can hand those requirements to SourceX, which looks for matching records at U.S. businesses when asked rather than keeping stock. Buyers set out scope, volume, format, timing and licensing needs on the request form, and SourceX then works through qualification, rights review, contracting, preparation and delivery with whichever companies hold the records.

Frequently asked questions

What is causality evidence in world model training data?

It is the record of how strongly the data supports the claim that one event produced another, rather than just preceding it. In industrial data that ranges from bare ordering in a log, through verified timing and recorded mechanisms, to documented interventions and tests.

Can a world model learn causes from observational plant data alone?

It can learn reliable sequences and associations, and with strong assumptions some effect estimates. Without records of interventions it cannot tell an action taken in response to a symptom from an action that caused the failure, which is why each link's evidence level should travel with the data.

Why do timestamps matter so much for causal labels?

Because a cause has to come first, and industrial systems often record several times for one value, such as when the device measured it and when a server received it. Using the wrong one, or ignoring clock offsets, can reverse the apparent order of two events seconds or minutes apart.

What should a buyer ask a supplier for to support causal labels?

Source timestamps with quality codes, configuration and change records, maintenance findings with failure codes, the reasons recorded for operator actions, and any test or trial reports. The license should cover using them for training and evaluation.

Is a root cause analysis report enough to label a link as causal?

It is strong evidence of a recorded mechanism, which places a link at level 3 of the hierarchy. It does not by itself show the pattern repeats or that changing the cause changes the outcome, so higher levels need comparison across events or an intervention.

Sources

  • Ha and Schmidhuber's World Models paper describes a generative model that learns a compressed spatial and temporal representation of an environment, used by an agent's controller. Source
  • OPC UA defines sourceTimestamp as applied by the data source and passed unchanged by relaying servers, and serverTimestamp as the time a server received the value. Source
  • The MCAP specification defines log_time as when a message was recorded and publish_time as when it was published, set to the log time if not available. Source
  • IEEE 1588-2019 defines a precision clock synchronization protocol for networked measurement and control systems. Source
  • ISO 14224:2016 standardizes collection and exchange of reliability and maintenance data, including failure data such as cause and consequence, with failure events and mechanisms among its uses. Source
  • A 2015 review in Emerging Themes in Epidemiology lists the nine viewpoints Hill published in 1965 for judging whether associations are causal. Source
  • DoWhy organizes effect estimation into model, identify, estimate and refute, and its refutation API tests causal assumptions behind an estimate. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies