Skip to content

AI uses for records

Business world model training data: what real records add to a world model

By SourceX Editorial · Updated

9 sections

Short answer

Real business records give a world model long, timestamped sequences of an operation's states, the decisions people took and the consequences that followed under real constraints. To use them, map each source to state, action or consequence, then check time and units, provenance, impossible states, off-system interventions, gaps and outcome evidence. Records show what followed the actions taken, not what other actions would have caused.

Video and sensors

Business world model

This page covers business world model training data: what real records add to a world model.Illustrative

Key takeaways

  • Business records supply real transitions: states, logged decisions and observed consequences under real capacity and policy limits.
  • Map every source to the part of a transition it can honestly support; many systems describe state well but prove consequences poorly.
  • Check time, units, provenance, state constraints, interventions, gaps, rights and outcome evidence before training.
  • Actions in records follow a policy, so record the decision rule by date and report which transitions are rarely observed.
  • Treat episodes that end without a closing event as truncated, not complete.

What can real business records contribute to a world model?

Real business records contribute what simulators and web video rarely have: long, timestamped sequences of an actual operation's states, the decisions people made in those states, and the consequences that followed, all under real constraints such as capacity, staffing, contracts and equipment wear. A warehouse system logs where stock was, which move was ordered and when it arrived; a field service system logs the job, the dispatch and whether the first visit fixed it.

What records do not contribute on their own is a clean causal story. Each system logs only what it was built to log, on its own clock, and the actions in it were chosen by people following a policy, not at random. The useful contribution is therefore a set of linked observations that a world model can learn transitions from, once each record has been mapped to a state, an action or a consequence and checked for the errors that would teach the model something false.

This page gives that mapping and the checks, so a buyer can tell which records are worth requesting and what the model can and cannot learn from them.

How do states, actions and consequences show up in operational systems?

In reinforcement learning terms, a world model learns how the state of an environment changes after an action. Ha and Schmidhuber's World Models paper describes a model that learns a compressed spatial and temporal representation of an environment, and the DreamerV3 paper describes an agent that learns a model of the environment and improves its behavior by imagining future scenarios. Both treat the environment as observations, actions and what follows.

Standard interfaces make the pieces explicit. Gymnasium's step function returns the next observation, a reward, a terminated flag for reaching a terminal state of the task, a truncated flag for an ending outside the task such as a time limit, and an info dictionary. RLDS stores episodes as steps holding an observation, the action taken in it and the reward, with flags for first, last and terminal steps.

Operational records arrive in a different shape. Process mining tools such as PM4Py read event logs as a case ID, an activity name and a timestamp, the convention inherited from the XES standard. The OCEL 2.0 object-centric event log specification goes further: events relate to many objects, objects relate to each other, and object attributes change over time. That is much closer to a world state than a flat event log, because an order, a truck, a technician and a part can all be tracked through the same event.

Mapping records to states, actions and consequences

Map each record source to the part of the transition it can honestly supply. Some records describe state, some record a decision, and only some prove what happened afterward. The table is the core asset; the system names are common examples, and the same logic applies to similar products.

Mapping records to states, actions and consequences
Record sourceState it describesAction it recordsConsequence evidenceMain caveat
Warehouse management scansLocation and quantity of each item, slot occupancyMove, pick and putaway tasks issuedConfirmation scan at the destinationScans batch-upload late; unscanned moves leave no trace
ERP orders and inventory, for example NetSuiteStock on hand, open orders, committed quantitiesPurchase orders, transfers, allocationsReceipts, shipments, backordersPosting date can differ from the physical event
Field service work orders, for example ServiceTitanOpen jobs, technician skills and locations, equipment on siteDispatch and scheduling decisionsJob outcome, callbacks, return visitsJobs may be closed at shift end rather than at completion
Maintenance management recordsAsset condition, last service, open defectsPreventive and corrective work ordersFailure codes, downtime, repeat failuresFailure codes chosen from a dropdown under time pressure
Machine historians and PLC logsSensor readings, setpoints, interlock statesSetpoint changes and operator commandsDownstream readings, alarms, scrap countsSampling rates, units and clock drift vary by tag
Support tickets, for example ZendeskCustomer issue, priority, queue loadRouting, replies, refunds, escalationsResolution, reopen, satisfaction responseOutcomes after the ticket closes are unseen
Issue trackers, for example JiraBacklog, assignee load, issue statusStatus transitions and assignmentsResolution, reopen, linked incidentsStatus changes may be cleanup, not real work
Fleet telematicsVehicle position, speed, engine stateRoute and dispatch changesArrival time, fuel use, faultsGaps where devices lose signal or are unplugged

Which evidence checks must each mapped record pass?

Each mapped record should pass eight checks before it is used to train or evaluate a world model. They come straight from the ways operational data misleads: wrong time, wrong units, a device nobody trusted, a state that cannot exist, an action taken off the system, a consequence that was never observed.

Run them on a sample first. If many records fail the same check, the fix is usually upstream, in how the source system is configured, rather than in cleaning.

Which evidence checks must each mapped record pass?
CheckQuestionSign of failureWhat to request
Time and unitsDo all sources share a clock, time zone and units?Consequences that precede actions; values off by a constant factorClock source, time zone and unit dictionary per field
Sensor and system provenanceWhich device or user produced each value?Values with no device or user ID; firmware changes mid-periodDevice registry, user roles, change log
State constraintsDoes every logged state respect capacity and physical rules?Two items in one slot; a line running while stoppedCapacity tables, interlock rules, travel times
InterventionsWere actions taken outside the system?State changes with no logged actionRadio, phone or paper override logs
Occlusions and gapsAre there periods with no observation?Long silences inside active workUptime logs, known outage dates
Capacity and policyWhat rules limited which actions were possible?Actions never seen in some statesDispatch rules, staffing levels, policy changes by date
RightsMay these records be licensed for model training?Customer or vendor terms restrict reuseSource ownership and contract review
Consequence evidenceIs the outcome observed, inferred or assumed?Outcomes filled by default valuesThe field that proves the outcome and how it is set

Could source error, missing observations or impossible states explain the pattern?

Often they do, and a world model trained on unchecked records will faithfully learn the logging system rather than the operation. A pattern such as orders shipping before they are picked usually means two clocks disagree; a sudden fall in machine output at midnight can be a counter rollover; a technician who appears at two sites at once is a record that was edited later.

Time is the most common trap. The MCAP file format used in robotics separates the time a message was recorded from the time it was published, and business systems have the same split under other names: event time, entry time and posting time. Decide which one is the state time for each source and keep the others as evidence.

Missing observations are harder, because a gap leaves nothing to inspect. Treat an episode that ends without its expected closing event as truncated rather than complete, in the way Gymnasium and RLDS distinguish truncation from a true terminal state, and do not let a model learn that work orders simply stop.

Why records rarely show what would have happened otherwise

Business records are logged under a policy. Dispatchers send experienced technicians to hard jobs, planners expedite late orders, and agents escalate angry customers. The record shows what followed the action that was taken, not what would have followed another one, so a model can learn that experienced technicians have more callbacks when the real cause is that they get the hardest jobs.

Three habits keep this honest. Record the decision rule and its changes by date, so the model and its users know which actions were possible in which states. Look for natural variation, such as periods when a rule changed, a shortage forced different choices or two sites used different policies, and label those periods. Report which transitions are well covered and which were almost never observed, because predictions for rare actions rest on very little evidence.

None of this makes records less valuable; it decides how they are used. Records are strongest for learning how an operation usually evolves and for evaluating predictions against what actually happened, and weakest for predicting the effect of actions the business seldom took.

Worked example: a field service dispatch model built from work orders

Treat the following case as illustrative only: the contractor, the records and the percentages are hypothetical. A buyer wants a world model that predicts whether a dispatched job will be fixed on the first visit, given the job, the technician and the parts on the truck. A heating and cooling contractor offers two years of work orders, truck inventory records and vehicle telematics.

Mapping turns each work order into an episode: the state is the open job, equipment type, technician certifications and truck stock at dispatch time; the action is the dispatch; the consequence is the job outcome and any return visit within the following weeks. Checks find that about 12 percent of jobs are closed more than a day after the telematics show the van leaving the site, so the team uses telematics departure as the completion time. About 4 percent of episodes have truck stock that would exceed van capacity, traced to a parts transfer that was logged twice, and are excluded with a note.

The exception: during one winter month the contractor dispatched by phone from a shared spreadsheet while the scheduling system was being migrated. Work orders exist, but the dispatch action and its time are missing. Those episodes are kept as truncated observations for state statistics and left out of transition training, and the gap is disclosed in the dataset documentation.

Which warning signs mean the records need narrowing or an expert?

  • Stop when the consequence field is set by default or by a later clean-up, because the model would learn the clean-up rule.
  • Narrow the scope to one site, one equipment family or one period when policies, layouts or systems changed often.
  • Bring in an operating domain expert to write the state constraints and to explain off-system interventions before training.
  • Bring in a world model researcher when the target needs continuous control signals that business records only summarize.
  • Bring in rights reviewers when records include customer, patient or employee details, or when vendor terms govern system exports.

Writing a request for business-record world model data

A useful request names the transition the model must learn and asks for the fields that make each part of it checkable: state fields with timestamps and units, the action log with the decision rule in force, the consequence field and how it is set, and the device and system context behind each value. Ask for two or more linked sources for the consequence wherever possible.

World-model teams can put this kind of request to SourceX, which keeps nothing in stock and looks for U.S. businesses whose records match; the request form records scope, volume, format, timing and licensing needs. Each request is assessed on its own merits, with a usual starting point of at least two years of consistent records and tens of thousands of them in a single system.

Frequently asked questions

What is business world model training data?

It is operational data arranged as transitions: the state of an operation, an action taken in that state and the consequence that followed, linked by time and by the objects involved. It usually comes from several systems, such as an ERP, a warehouse or field service system and telematics, joined on shared identifiers.

Can event logs train a world model?

They can supply the action and timing backbone, since process mining logs already record a case, an activity and a timestamp. They usually need state fields from other systems and a clear outcome field before they describe full transitions.

How is this different from world model video data?

Video world models learn how scenes change visually after actions. Business-record world models learn how an operation's measurable state changes after decisions, such as stock levels, job status or machine readings, and can be combined with video when both share a clock.

Do records show cause and effect?

Only partly. They show what followed the actions that were taken under the business's own rules, so effects of rarely taken actions are poorly supported. Periods when rules changed or constraints forced different choices are the best evidence for causal questions.

How much history is needed?

Enough to cover the transitions the model must predict, including seasonal and rare states. As a general guide, buyers tend to look for two or more years of consistent history in one system, but coverage of the right states matters more than volume.

Sources

  • The World Models paper describes a model that learns a compressed spatial and temporal representation of an environment. Source
  • DreamerV3 learns a model of the environment and improves its behavior by imagining future scenarios. Source
  • Gymnasium's step returns an observation, reward, terminated and truncated flags and an info dictionary, with truncation covering endings outside the task such as time limits. Source
  • RLDS episodes are steps holding observation, action and reward with first, last and terminal flags; an episode ending without a terminal step is truncated. Source
  • PM4Py reads event logs as case ID, activity and timestamp columns, following XES naming conventions. Source
  • OCEL 2.0 relates events to multiple objects, records object-to-object relationships and tracks object attribute changes over time. Source
  • MCAP separates the time a message was recorded (log time) from the time it was published. 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