Systems and records
World model operational constraints: encoding capacity and safety limits
By SourceX Editorial · Updated
9 sections
Short answer
World model tasks should represent capacity and safety constraints as explicit, versioned rules tied to measured state variables: each with a unit, a limit, the document it comes from, its effective dates and a declared effect when crossed, whether the episode ends, is cut short or accrues cost. Operating logs rarely show limits being crossed, so a model cannot infer them unaided.
Video and sensors
World model operational
Key takeaways
- Store each operating limit as a typed object with a measured variable, limit, unit, source document and effective dates.
- Operating logs rarely contain breaches, so a world model cannot learn a limit it never sees crossed without explicit encoding.
- Declare in advance whether each violation terminates the episode, adds cost against a budget or truncates it as out of scope.
- Rule out unit errors, stale ratings, unobserved steps and impossible states before scoring any breach.
- Flag constructed boundary cases and keep them out of test sets meant to measure real behavior.
How should world model tasks represent capacity and safety constraints?
World model tasks should represent capacity and safety constraints as explicit, versioned rules attached to measured state variables, with a stated unit, a stated limit, the document the limit comes from, and a defined consequence when the limit is crossed. A constraint that lives only in the pixels of a video or in an operator's habits is invisible to training and to evaluation.
A world model predicts how a scene or system evolves given actions. NVIDIA describes its Cosmos-Predict2 models, for example, as world foundation models that specialize in future state prediction, including video-to-world generation. Prediction alone does not tell an agent which futures are allowed. The operating limits of a warehouse, a plant floor or a fleet decide that, and they rarely appear in the raw record as anything a model could read.
This guide gives a constraint encoded scenario pack: a field-level template for turning operational logs into scenarios where every capacity and safety limit is a typed object, every violation has a declared effect on the episode, and every number can be traced to a source. It is written for teams that train or evaluate world models and agents on real operating data, and it covers the checks that separate a true violation from a logging artifact.
Why a learned model will not infer a limit its training data never crosses
A model learns the limits it observes, and well-run operations almost never show the limit being crossed. Operators stop before a rack is overloaded, dispatchers do not send a truck past its rated capacity, and the logs record the decisions that respected the rule. The boundary itself leaves no trace, so a model trained only on those logs has no evidence of where the cliff is.
This matters more for world models than for ordinary predictors because agents trained inside a learned model will search for its weak spots. In their World Models article, Ha and Schmidhuber report that their agent found an adversarial policy that kept simulated monsters from ever firing during some rollouts. They explain that the learned model only approximates the real environment and will occasionally generate trajectories that break the real environment's laws, so the controller can exploit flaws that do not exist in reality, usually in states far from the training distribution.
Operational limits are exactly that kind of state. A policy that learns to stack one more pallet or route one more job through a saturated dock looks good inside a model that never saw what happens next. The fix is not more of the same footage. It is to state the limit explicitly, to include scenarios near and across the boundary with labelled provenance, and to score agents on whether they respect it.
Is it a capacity limit, a safety rule or a site policy?
Classify each constraint before encoding it, because the class decides where the limit comes from, how it can change and what a violation should do. The five classes below cover common operating limits, and a single scenario often carries more than one.
U.S. workplace rules show why the source matters. OSHA's general requirement for walking-working surfaces says the employer must ensure each surface can support its maximum intended load. The powered industrial truck standard says only loads within the truck's rated capacity may be handled, and that modifications affecting capacity need the manufacturer's prior written approval, with capacity plates changed to match. Those are limits with a documentary source, an owner and an effective date, which is how a scenario pack should record them.
| Constraint class | Example in operating records | Where the limit comes from | How it changes |
|---|---|---|---|
| Rated physical capacity | Forklift rated load at a stated load center; rack beam rating | Manufacturer plate or specification sheet | Only with a documented modification or replacement |
| Regulatory safety rule | Maximum intended load on a mezzanine floor; stable loads only | Statute, regulation or code adopted by the site | When the rule or the site's compliance plan changes |
| Site operating policy | No pedestrian in an aisle while a truck is lifting; speed limit in a zone | Written standard operating procedure | When management revises and re-issues the SOP |
| Throughput capacity | Dock doors, staging lanes, technicians per shift, queue length | Physical layout plus staffing records | Daily or weekly with schedules and layout changes |
| Contractual or service limit | Delivery window, temperature range for a shipment | Customer agreement or service level | Per customer and per contract version |
The constraint encoded scenario pack, field by field
The pack is one record per scenario, built from linked operating records, with the constraint objects stored separately so the same limit can be reused and versioned. Use the field groups below as the schema. A field marked required that cannot be filled means the scenario is not ready for training or evaluation, not that the field can be guessed.
At a glance
The constraint encoded scenario pack, field by field
| Field group | What it holds | Validation rule | Typical failure it catches |
|---|---|---|---|
| Scenario identity | Scenario id, pack version, site or layout id, time window | Ids stable across pack versions; window has start and end timestamps | Two scenarios silently merged after a layout change |
| Time base and units | Clock source, time zone, sampling rate, unit system for every numeric field | Every numeric field names its unit; one clock per scenario or a stated offset | Pounds read as kilograms; sensor clock drift against the WMS |
| State variables | Measured quantities the constraints refer to: load weight, load length, lift height, zone occupancy, queue length | Each variable links to the sensor or record that measured it | A limit defined on a variable nobody measures |
| Constraint objects | Expression, limit value, unit, constraint class, source document, effective dates, owner | Limit and unit match the cited document; effective dates cover the scenario window | A stale capacity plate after a truck attachment was added |
| Cost and violation signal | Per-step cost, cumulative budget, or terminal flag for each constraint | Every constraint declares exactly one violation handling rule | Violations folded into reward and lost |
| Actions and interventions | Operator or agent actions, supervisor overrides, emergency stops | Each action has an actor type and timestamp | An override treated as the policy's own choice |
| Observation coverage | Camera and sensor coverage per step, occlusion flags, gaps | Gaps are labelled, never filled with interpolated values | A violation scored on a frame no sensor observed |
| Consequence evidence | Incident report, damage record, near-miss note, or none observed | Consequences link to a record id; none observed is a value, not a blank | Inferring harm that the record does not show |
| Provenance of the scenario | Recorded, edited from a recording, or constructed; generator and editor ids | Constructed scenarios are flagged and kept out of real-only test sets | Synthetic boundary cases scored as observed history |
| Rights and privacy | Source owner, license reference, people visible, de-identification status | Every scenario names its source license; faces and badges handled before use | Footage of employees used without a recorded basis |
Should a violation end the episode, cut it short or add a cost?
Decide the effect of each violation before any agent runs, because the same breach can be a terminal state, an out-of-scope stop or a cost to budget, and mixing them corrupts learning. Gymnasium, the reference reinforcement learning interface, separates these cases: terminated means the agent reached a terminal state defined by the task, which can be positive or negative, while truncated means a condition outside the task's Markov decision process ended the episode, typically a time limit or an agent going out of bounds.
Safe reinforcement learning tools add a third channel. OpenAI's Safety Gym reports constraint costs in the step information, separate from reward, and the Safety-Gymnasium project returns cost as its own value alongside reward, terminated and truncated. The underlying framework, described in the Constrained Policy Optimization paper by Achiam and colleagues, is a constrained Markov decision process: an ordinary decision process augmented with auxiliary cost functions and a limit on each.
In practice, a pack can use a simple rule. A physically impossible or catastrophic state, such as a tip-over, ends the episode as terminated with its consequence recorded. A breach of a policy or budget, such as entering an occupied aisle, adds cost and lets the episode continue so the agent can learn recovery. A state the scenario does not model, such as a truck leaving the mapped area, is truncated and excluded from scoring.
- Terminated: the breach is a terminal outcome the task defines, and the consequence record is attached.
- Cost: the breach counts against a per-episode budget declared in the constraint object.
- Truncated: the episode left the scenario's scope; no score, and the reason is logged.
- Never: a violation that changes nothing in the episode record, which is the silent failure the pack exists to prevent.
Could a flagged breach be a logging error rather than real behavior?
Often it could, so rule each cause out before treating a breach as real behavior. Operating data contains three common impostors: values that are wrong at the source, events that happened but were not observed, and combinations of state that cannot physically exist. A scenario that trips a constraint because of any of these teaches the model a rule that is not true.
At a glance
Could a flagged breach be a logging error rather than real behavior
Run the checks below on every scenario flagged as a violation, and record the outcome in the scenario so reviewers can see why it was kept, corrected or dropped.
- Unit and scale check: confirm the measured value and the limit share a unit, and that no field switched units partway through the window.
- Constraint version check: confirm the limit in force on that date; a truck modified with an attachment may have a different rated capacity than the plate in the original records.
- Coverage check: confirm a sensor or record observed the state at the step where the breach is scored; otherwise label it unobserved.
- Impossible state check: flag negative stock, two loads in one slot, or a truck in two zones at once as data errors, not violations.
- Clock alignment check: confirm events from different systems are ordered correctly; a scan logged before the lift that caused it signals clock skew.
- Override check: confirm whether a supervisor or emergency stop intervened, so the breach is attributed to the right actor.
- Consequence check: if a consequence is claimed, link the incident or damage record; if none exists, record none observed rather than inventing harm.
Worked example (illustrative): a forklift put-away with a load-moment limit
This worked example is illustrative. It shows how one recorded put-away becomes several scenarios, using OSHA's own load-moment explanation as the constraint source. OSHA's non-mandatory stability appendix explains that trucks of 30,000 pounds capacity or less are normally rated at a 24-inch load center, and gives the example of a 3,000-pound truck whose maximum load moment is 72,000 inch-pounds, so a 60-inch load with a 30-inch load center may weigh at most 2,400 pounds.
The constraint object therefore limits load weight times load center, not weight alone. A recorded put-away of a short pallet within capacity becomes the base scenario. Variants then test whether a model represents the limit correctly, and each variant carries its provenance so nobody mistakes a constructed case for history.
| Variant | What changes (illustrative) | Provenance flag | Expected handling |
|---|---|---|---|
| Base | Recorded short pallet, weight and load center inside the load-moment limit | Recorded | No cost; outcome compared with the logged result |
| Near boundary | Same pallet with load center moved so the load moment sits just under the limit | Edited from recording | No cost; tests whether the model predicts reduced stability margin |
| Long load | 60-inch load whose weight exceeds the limit for its 30-inch load center | Constructed | Terminated as a rated capacity breach; never placed in a real-only test set |
| Stale plate | Attachment added; the record still shows the original rating | Recorded with correction | Scenario held until the modified rating is documented |
| Occluded lift | Camera gap during the lift; weight known from the scale | Recorded | Breach not scored on unobserved steps; gap labelled |
| Aisle occupied | Pedestrian badge read in the aisle during lift | Recorded | Cost added under the site policy; episode continues |
Where the pack stops being enough
The pack encodes limits; it does not certify that a trained policy is safe. Use it to train and test whether a model represents constraints, then stop and bring in a qualified safety engineer or the equipment manufacturer whenever a decision would affect real equipment or people.
One exception deserves its own rule. Some limits are dynamic: OSHA's stability appendix notes that a load at maximum should be carried as low as possible, with slow, even acceleration and cautious fork tilt, because motion changes the truck's stability. A static load-moment check cannot capture that. For dynamic limits, narrow the pack to scenarios with the sensor data needed to compute the margin over time, or mark the constraint as partially encoded so evaluation results are not over-read.
Narrow the scope as well when the source document is missing, when two documents disagree about a limit, or when the scenario relies on footage of identifiable employees without a recorded basis for use. Each of those is a reason to hold the scenario, not to estimate.
Requesting operating records for constraint encoded scenarios
Teams that need real operating records for packs like this should ask for the linked set, not footage alone: the event logs, equipment ratings and their change history, site procedures with effective dates, and incident or near-miss records, with sensor coverage described. Without those, the constraint objects cannot be filled from evidence.
Operating records like these are found for each request among U.S. businesses rather than drawn from a SourceX inventory, and SourceX handles rights review and data preparation with the companies that own them. A request can state the operating domain, the constraint classes needed, the volume, format, timing and licensing terms, and whether scenario construction will happen on the buyer's side.
Frequently asked questions
What is a constraint encoded scenario?
It is a training or evaluation episode in which each operating limit is stored as an explicit object with a measured variable, a limit, a unit, a source and a violation rule. The model and the evaluator read the same definition, so a breach is detectable rather than implied by footage.
Should safety limits be part of the reward function?
Folding them into reward hides how often they were crossed and lets a large reward outweigh a breach. Safe reinforcement learning tools such as Safety Gym and Safety-Gymnasium report cost separately from reward, which keeps violations countable and lets a team set a budget for each constraint.
Can synthetic boundary cases replace real violation records?
They can fill gaps, because real operations rarely record breaches, but they should be flagged as constructed and kept out of any test set meant to measure real-world behavior. Real near-miss and incident records remain the evidence for what actually happens after a limit is crossed.
How do you handle a limit that changed during the recording period?
Store each limit with effective dates and attach the version in force at each step. Equipment modifications, re-rated racks and revised procedures all change limits, and a scenario scored against the wrong version teaches a false rule.
What operating records are most useful for world model constraints?
Event logs from systems such as a warehouse management system, equipment specifications with change history, written procedures, sensor coverage maps and incident or near-miss reports. Footage helps, but without these linked records the limits cannot be stated or checked.
Sources
- NVIDIA describes Cosmos-Predict2 as part of its Cosmos world foundation models, specializing in future state prediction through world modeling, including video-to-world generation. Source
- Ha and Schmidhuber report an agent that found an adversarial policy inside a learned world model, and explain that an approximate model can generate trajectories that break the real environment's laws and be exploited in states away from the training distribution. Source
- Gymnasium defines terminated as reaching a terminal state of the task, positive or negative, and truncated as a condition outside the MDP, typically a time limit or going out of bounds. Source
- OpenAI's Safety Gym reports constraint costs in the step info dictionary separately from reward and supports environments with constraints on entering hazard areas. Source
- Safety-Gymnasium's step method returns next observation, reward, cost, terminated, truncated and info, with cost as an extra field. Source
- The Constrained Policy Optimization paper frames constrained reinforcement learning as a constrained MDP: an MDP augmented with auxiliary cost functions and limits on each. Source
- OSHA 1910.22(b) requires the employer to ensure each walking-working surface can support the maximum intended load for that surface. Source
- OSHA 1910.178 allows only loads within the truck's rated capacity and requires the manufacturer's prior written approval for modifications that affect capacity, with capacity plates changed accordingly. Source
- OSHA's non-mandatory stability appendix gives the load-moment example of a 3,000-pound truck at a 24-inch load center (72,000 inch-pounds), limiting a 60-inch load to 2,400 pounds, and advises carrying maximum loads low with slow, even acceleration. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.