Time traveling companyExplore

OpenX / Technical documentation

SimW1
System card

Scope, evidence, limitations, and the conditions for responsible evaluation.

Version 0.128 September 2026Draft · Not a release assessment
Read the announcement

1. Overview and status

SimW1 is the name used for OpenX’s proposed simulation world model. This document accompanies a research-concept announcement. It records the intended decision-support scope and a proposed evaluation program; it does not document a verified trained model release.

No completed SimW1 benchmark runs, independent safety audit, or validated deployment evidence are supplied with this card. The website’s visual demonstrations are explanatory interfaces, not proof of the underlying model’s accuracy or readiness.

“Proposed” denotes a design or evaluation intention. “Not disclosed” means this document does not provide the relevant evidence; it must not be interpreted as a favorable finding. No capability or safeguard should be inferred solely from its inclusion in this document.

2. Technical disclosures

System developer
OpenX, as identified in this site’s branding. Legal entity and accountable release owner require confirmation before publication as a release document.
Model architecture and size
Not disclosed. No verified architecture, parameter count, or base-model dependency is specified.
Training and fine-tuning
Training data composition, collection periods, compute, objectives, and procedures are not disclosed.
Supported inputs and outputs
Proposed inputs include evidence, starting states, actors, constraints, and interventions. Proposed outputs include conditional trajectories and comparisons. Production interfaces, modalities, limits, and schemas are not established here.
Execution and tools
Tool access, network permissions, isolation boundaries, and execution controls require verification. No autonomous action capability is asserted.
Version and reproducibility
System-card version 0.1 is not a model version. No model checkpoint identifier, evaluation artifact hash, or reproducible run manifest is supplied.
Availability
No release date, service-level commitment, pricing, or supported deployment environment is announced by this card.

3. Intended use and exclusions

The proposed purpose is bounded decision rehearsal: exploring how alternative actions could affect a modeled environment under explicit assumptions. Candidate research tasks include inventory planning, product rollout comparisons, resource allocation, and disruption analysis.

Appropriate evaluation users would be able to inspect the assumptions, assess the domain evidence, and compare outputs with independent baselines. A scenario should be treated as conditional analysis, not as a statement that an event will occur.

Outside the supported scope of this draft

  • Autonomous medical, legal, financial, employment, credit, or other high-stakes determinations.
  • Safety-critical control of infrastructure, vehicles, or industrial equipment.
  • Inferring sensitive traits or making consequential judgments about identifiable individuals from synthetic personas.
  • Covert behavioral manipulation, surveillance, or targeting of vulnerable groups.
  • Claims of guaranteed outcomes, universal causal understanding, or validated prediction across industries.

These exclusions define the intended scope of the proposal. They are not a claim that technical enforcement has been implemented or independently tested.

4. Data governance and privacy

Before an evaluation uses external or customer data, it should document source, license, permitted purpose, consent where applicable, collection dates, and relevant exclusions. Access to a dataset is not equivalent to permission to train on it, redistribute it, or use it to profile individuals.

Training, development, and held-out evaluation data should be separated at the appropriate unit. Participants, organizations, time periods, and related scenarios can leak information across an apparently random split. Public benchmarks also need contamination checks before they support broad conclusions.

Data retention, deletion, encryption, residency, and access-control practices are not established in this card. Sensitive or confidential information should not be submitted on the assumption that this draft specifies those protections. A production data-handling policy and verified controls are prerequisites to such use.

Synthetic outputs may still reveal source information or reproduce bias. Calling a population synthetic is not a privacy guarantee. Any behavioral evaluation should examine reidentification risk, small-group disclosure, and representativeness.

5. Evaluation status

The program contains twelve tracks. None has a measured SimW1 result in this document. External benchmark names indicate proposed evaluation resources, not successful participation, endorsement, or compatibility already demonstrated.

Evaluation mock-up

Five examples of what a benchmark report could show.

Every value below is fictional interface data. These examples do not report a real SimW1 run, predict a current election, or establish comparative model performance.

United States · Illustrative backtest

Presidential election

Predict the winning candidate in each state from a frozen pre-election information cutoff.

Primary metric
State winner accuracy
Proposed split
Historical election cycles held out by year
SimW1 conceptGPT-6 AstraClaude Opus 5.5Specialist baseline
0%25%50%75%100%88%SimW184%Astra81%Opus 5.586%SpecialistSTATE WINNER ACCURACY · FICTIONAL VALUES
The concept model leads this fictional comparison, but eight incorrect states could still change the electoral outcome.

Before execution, each track needs a fixed task, dataset version, leakage assessment, comparator configuration, scoring method, and treatment of invalid outputs. Sample sizes should follow a precision or power rationale appropriate to the task, not be chosen after observing favorable results.

Reports should include uncertainty intervals, costs, latency, failure counts, and domain-level breakdowns. Averages must not erase poor performance on consequential subgroups or difficult scenarios. Comparison with a commercial model requires matched conditions and actual evaluation runs.

No relative performance claim against GPT‑6 Astra or another named system is supported here. Results from unrelated leaderboards are not interchangeable with a controlled SimW1 comparison.

6. Risk assessment

Model error and false causality

A plausible mechanism may be wrong or incomplete. Hidden confounders can change the apparent effect of an action, and assumptions appropriate to one organization may fail elsewhere. Counterfactual outputs should identify the assumptions required for their interpretation.

Overconfidence and automation bias

Detailed trajectories, animated agents, and precise percentages can imply more certainty than the evidence supports. Users may favor a visually compelling scenario over a better-grounded but less vivid baseline. Interface testing should examine how people interpret uncertainty, not only whether a disclaimer is present.

Bias and representation

Behavioral data may exclude people or contexts relevant to a decision. Aggregate fit can coexist with systematic errors for smaller groups. Simulated preferences must not stand in for consultation, consent, or a person’s own account of their needs.

Distribution shift and feedback

An intervention can alter the environment that produced the original data. People and institutions may respond to predictions themselves. A model that performs well retrospectively can fail after deployment, especially when its recommendations become part of the system being modeled.

Security and misuse

If a future implementation uses retrieval, uploaded documents, or executable tools, it will need testing for prompt injection, data exfiltration, unauthorized actions, and poisoned inputs. Documents and tool results should be treated as untrusted evidence, not instructions granting additional authority.

Objective misspecification

A plan can optimize a stated measure while harming an omitted outcome. Evaluations should examine binding constraints and distributional effects, and make objective choices explicit. A high simulated reward is not itself evidence that an action is acceptable.

7. Proposed safeguards

The following controls are requirements for investigation, not verified features of a released system:

  • Separate observations, assumptions, interventions, and generated outcomes in the interface and run record.
  • Record evidence provenance, model configuration, seeds where applicable, and the conditions needed to reproduce a run.
  • Show uncertainty, sensitivity, and failure cases alongside favored scenarios.
  • Check explicit constraints independently of generated explanations.
  • Restrict tool permissions and require appropriate human authorization before external actions.
  • Identify unsupported domains and abstain or escalate when evidence is insufficient.
  • Apply purpose limitation, data minimization, and access controls before handling sensitive inputs.

Safeguard effectiveness requires its own evidence. A proposed abstention mechanism, for example, should be evaluated for both missed dangerous cases and unnecessary refusals. Human review must be designed around qualified, informed reviewers rather than treated as a universal remedy.

8. Proposed release criteria

A release assessment should be scoped to specific tasks and users. It should not inherit permission for every application mentioned in the announcement. Before expanding beyond concept demonstrations, the responsible team should:

  1. Identify the accountable owner, actual model configuration, dependencies, and supported scope.
  2. Complete data rights and privacy reviews for the relevant sources.
  3. Freeze task-specific acceptance criteria before evaluating the held-out test set.
  4. Run applicable benchmarks against credible baselines and publish limitations and negative findings.
  5. Test security boundaries, misuse scenarios, and the effectiveness of proposed safeguards.
  6. Establish incident handling, monitoring, and rollback procedures.
  7. Update this card with evidence, residual risks, and an explicit deployment decision.

No numerical acceptance thresholds are specified here because neither a validated operating context nor a risk tolerance has been established. Those thresholds require domain-specific justification. This card must not be used as a substitute for that work.

9. Monitoring and revision

A deployed version would need monitoring of forecast drift, calibration, constraint failures, user-reported errors, and shifts in the population or operating environment. New data sources, model changes, or changes in permitted use should trigger a review of whether prior evaluation evidence still applies.

Material failures should lead to a documented response: constrain the scope, pause the affected function where appropriate, investigate the cause, and evaluate the correction before reinstatement. The current draft does not establish an operational incident-response service or response-time commitment.

Questions about this proposal can be directed through the OpenX contact page. Do not include sensitive data in an initial inquiry. A dedicated security-reporting channel and accountable response owner remain to be specified.

10. Revision history

v0.1 · 28 September 2026. Initial research-concept card. Added intended scope, disclosure gaps, twelve planned evaluation tracks, risk categories, proposed safeguards, and release criteria. No empirical results or independent audit findings accompany this version.

Future versions should identify changes in evidence and scope, not merely changes in presentation. A dated record of limitations is part of responsible documentation.

Return to Introducing SimW1