Enterprise Intelligence World Model

A hospital can add beds and still see longer waits if staffing or discharge capacity becomes the bottleneck. A manufacturer can restore a machine and still miss deliveries because the work waiting behind it has changed. In both cases the outcome depends on how several parts of the organization move together, and on the decisions people make in response.

The evidence for those dynamics already exists in fragments. A shift roster, a work queue and a delivery commitment describe different parts of the same operation. Their relationships matter, as does the order in which people learned about them. We want to understand when that history can help explain how a disruption spreads, which consequences can be forecast and which decisions are worth studying.

An Enterprise Intelligence World Model (EIWM) would combine an explicit representation of an organization with a learned model of how its state changes. Craven already represents entities and relationships and links forecasts to evidence and observed outcomes. The research starts by extending that representation through time. Then we can ask whether it improves predictions and, in a separate study, whether the evidence can support estimates of what an intervention changes.

We separate three objects. The ontology defines types and relations, such as an order drawing on a stock pool. The enterprise state instantiates them at a given time, with quantities, active relationships and process stages. The dynamics model estimates a distribution over how that state may change. An explicit state lets researchers test whether particular relationships add predictive value and trace the underlying evidence when a forecast fails.

A historical state needs two clocks. One records when something happened or held true. The other records when a particular version of the evidence became available. Keeping both is called bitemporal history. Each observation also needs a stable entity identifier and a source. A replay must use the ontology and entity mappings available at the cutoff, since later merges and corrections can leak information into earlier forecasts.

Craven already connects source evidence to business entities and relationships. The research step is to make a complete decision-time view reproducible and measure how much history can be recovered. OCEL 2.0 offers an exchange format for events linked to several objects. The replay below illustrates what a forecast would have been allowed to see at two points in the same morning.

Once a replay can reconstruct the state at a decision time, the next question is whether its structure adds predictive value. A forecast based on individual order records may miss the fact that several of those orders depend on the same stock pool or production line. The proposed benchmark tests whether giving a model those relationships improves its forecasts on later data.

Let Sₜ denote the reconstructed state at forecast cutoff t, A the actions taken over a horizon h, and E external events over that horizon. The transition model estimates P(Sₜ₊ₕ | Sₜ, A, E) from observed transitions. For a forecast issued at t, the model can use commitments and external forecasts already available at t. It must integrate over unknown future actions and events or specify them as scenario inputs.

A first model can estimate discrete transitions, such as whether an order progresses from open to received within seven days. A richer model could estimate several coupled outcomes, such as the distribution of receiving workload, backlog and stock availability. These should respect domain constraints, including feasible process transitions and the accounting of items into and out of a stock pool.

We would compare a structured statistical model with process and capacity variables, a sequence model trained on event history, and a hybrid model that adds typed relationships to a learned transition model. Graph networks, as described by Battaglia and colleagues, offer one way to encode those relationships. Comparing the same architecture with and without relational inputs would help isolate what the structure adds. The simplest model that holds up on later data sets the standard for additional complexity.

For a first public benchmark, we propose predicting whether an open e-commerce order will arrive by its promised date, issuing the forecast 24 hours after purchase. Olist’s anonymized Brazilian e-commerce release contains about 100,000 orders from 2016 to 2018 with order, seller, item and delivery information. It supports a test of whether links among orders, sellers and products improve forecasting over order fields alone.

The release is retrospective. We would audit which fields can plausibly be dated to the cutoff and exclude those with ambiguous availability from the main analysis. Delivery status, reviews and updated estimates can carry information from after the prediction. Event timestamps alone cannot establish when a live system learned the fact. We would publish the eligibility rules, field availability assumptions, split dates and counts at each exclusion so the benchmark can be reproduced.

The simulated OCEL Order Management log provides a complementary test of replay and event joins across orders, items, packages and related objects. Controlled perturbations can test whether the reconstruction handles delayed records or missing links. A later operational study would require permissioned records with source availability, action logs and independently verifiable outcomes. The public benchmark and simulation serve different purposes and should be reported separately.

Train on earlier periods, select models on subsequent periods and reserve the final period for a locked comparison. In rolling-origin evaluation, use only the observations and outcome labels available at each training cutoff. Group related order rows together, and report performance separately for sellers seen during training and sellers first seen during evaluation. A single marketplace can test transfer across sellers. Transfer across organizations requires independent organizations and a separate holdout.

For binary outcomes, report Brier score, calibration by risk band, and precision and recall at a fixed review capacity. For time-to-event outcomes, account for orders whose observation window ends before resolution. Report sample sizes, outcome coverage, uncertainty intervals that respect dependencies among orders, and performance under sparse or delayed evidence. Where a comparable evidence-backed forecaster is available, run it on the same records and horizon.

When the benchmark is complete, we would publish a dated research note with the data version, sample counts, baseline scores, calibration plots and representative failures. A gain that disappears once late-arriving fields are removed, or performance that falls on new sellers, belongs in that report. Measuring operational value would require a prospective comparison with the cost of false alarms and missed events specified before the study.

Even a model that predicts delays accurately leaves a separate question open. Will a particular decision improve the outcome? Return to the factory at 09:00. The same observed state supports three different questions, each with its own evidence requirements.

Following Hernán and Robins’s target trial framework, we would define eligibility, the decision time, competing actions, the outcome and follow-up before analysis. Planners may reroute the most urgent jobs, which complicates comparisons with jobs left in place. Shared capacity adds another complication. Moving a job to Line B can affect every job in its queue, so a study could compare outcomes across eligible shifts or facilities and track delays across both lines.

The broader opportunity is to study how a change in one part of an organization propagates through shared capacity, dependencies and time delays. The table describes candidate research questions. Each domain would need its own outcome definitions, evidence cutoffs and evaluation.

The factory timeline is illustrative. The proposed public benchmark measures retrospective forecast performance for one marketplace. Its timestamps leave some questions about evidence availability unresolved, and sellers in a shared marketplace provide a narrower transfer test than independently operated businesses.

A model can score well on the next transition and still drift over longer simulated sequences. A small error in a repair-time estimate can change which orders enter a queue and compound across later shifts. We would evaluate each forecast horizon separately and test how conclusions change under different assumptions about recovery time and shared capacity.

Intervention estimates depend on how decisions were made and recorded. Rare outages, informal workarounds and shared crews can leave too few comparable cases for a credible estimate. In those cases, the research can report forecast performance or describe scenarios under stated assumptions. A causal estimate requires additional evidence.

Longer term, we would test whether recurring patterns such as demand, capacity, queues and cash conversion transfer across organizations with different ontologies. Compare a shared model adapted to each organization with models trained locally, hold out entire organizations and test under changes in policy or demand. Any joint study would require explicit agreement on data access and permitted use. A transferable representation is a hypothesis to test across settings.