Recursive Simulations · 2026-06-13

🔄 Recursive Simulations — 2026-06-13

🎮 Decart Oasis 3 Raises $300M at $4B for Photorealistic Driving World Model — Then Ships Documented Physics Failure Modes
🤖 NVIDIA Isaac Sim 6.0 at CVPR: Agents Now Launch Sessions, Author Scenes, and Validate Environments Inside the Simulator Itself
🏭 Trillion Labs Positions Industrial World Models as the Intelligence Layer That Decides How Physics-Accurate Omniverse Factories Should Run
🔗 PTC Connects Onshape CAD Directly to Isaac Sim So Every Design Update Instantly Propagates into Physics Simulation
🧬 OmniBioTwin: Health Digital Twins Formalized as "System of Twinned Systems" with Modular Interaction Operators Across Biological Scales
🎯 Efficient-WAM: 1B-Parameter World-Action Model Achieves Strong Control with Deliberately Coarse Visual Predictions

---

🎮 Decart Oasis 3 Raises $300M at $4B for Photorealistic Driving World Model — Then Ships Documented Physics Failure Modes

Decart unveiled Oasis 3 on June 10 in a TechCrunch exclusive, simultaneously announcing a $300M funding round at a $4 billion valuation. The product: an action-conditioned generative world model delivering photorealistic driving environments in real time via API, designed to generate unlimited training scenarios for autonomous vehicle policy development. The pitch is that world models can exceed real-world data collection at any scale — unlimited edge cases, configurable weather, instant regeneration.

The physics failure documentation arrived in the same week. Dealroom's testing summary found three failure classes: environmental consistency degrades during extended use; controls lack responsiveness; and the model fails to simulate proper physics, with cars passing through other vehicles. Phandroid confirmed the same pattern independently: cars pass through objects, environmental consistency fails on longer runs. Decart CEO Dean Leitersdorf described the physics accuracy gap as "a major research challenge" — and is betting that 100 developers building unexpected applications in the next three months will define the category before the challenge is resolved.

The technical root cause is architectural. TechCrunch explains that Oasis 3 is auto-regressive: it generates one frame at a time, each conditioned on what it previously generated. Each frame independently predicts what the world looks like; nothing in the architecture enforces global physical consistency across frames. A car that drifts through another vehicle in frame N was never constrained by Newtonian contact mechanics — it was constrained only by the statistical distribution of training videos, which presumably did not include many cars occupying the same space. Physics compliance is a learned prior, not an enforced constraint.

This failure mode is structurally different from the standard sim-to-real gap. Robotics and Automation News frames the existing challenge as existing physics engines being "too rigid" to capture the chaos of the real world — oil spills, fragile boxes. Decart's problem runs in the opposite direction: not too rigid, but too permissive. A simulation that looks real but violates contact mechanics does not make robot training more robust to real-world edge cases; it may make it less robust by teaching policies that the world is more predictable and consistent than it is.

The $4 billion valuation bets on the "endless simulations" pitch over the physics fidelity problem. Dataconomy notes that the live API allows developers to generate unlimited physics-based training environments "necessary to scale robotic reinforcement learning training loops" — though "physics-based" here describes the real-world data Oasis 3 was trained on, not physics constraints the model enforces at inference. AI Chat Daily observes that Decart is positioning for the default substrate of AV simulation. Whether that position survives contact with closed-loop validation requirements is the open question.

Sources:

---

🤖 NVIDIA Isaac Sim 6.0 at CVPR: Agents Now Launch Sessions, Author Scenes, and Validate Environments Inside the Simulator Itself

NVIDIA published its CVPR 2026 physical AI blog post on June 12, announcing agent skills for Isaac Sim and Isaac Lab that transform simulation from a passive training environment into an active workspace managed by AI agents. Agents can now launch simulation sessions, author scenes, control simulation parameters, capture synthetic data, and validate environments — all as MCP-compatible tool calls. The simulator is no longer the place agents run inside; it is the workspace agents operate.

The Isaac Sim 6.0 developer forum confirms the release on June 4 carried AI-assisted development through Isaac Sim MCP and agent skills, alongside expanded robot authoring and software-in-the-loop testing through Newton physics. Radiance Fields' technical coverage documents NuRec Gaussian Splatting integration: Gaussian splats derived from real-world sensor data can now be imported into Isaac Sim scenes directly, replacing the manually modeled environments that have historically been the most labor-intensive component of simulation setup. Real-world scans become simulation substrates.

For autonomous vehicle development, NVIDIA's CVPR blog specifies that AV skills enable agents to "automate workflows for scene reconstruction from fleet data and generate synthetic scenarios." The chain is: fleet data → scene reconstruction (NuRec, Gaussian splatting) → physics-accurate simulation (Isaac Sim 6.0 + Newton) → synthetic scenario generation → policy training. Each step can be initiated and controlled by an AI agent making tool calls rather than by a simulation engineer running scripts. The human is no longer required in the loop to move from real-world data to simulation training.

This is simulation becoming prescriptive in a specific sense: the simulation workflow is now self-directing. Agents identify edge cases in fleet data, reconstruct those scenes as simulation environments, generate training scenarios, and validate that policies handle them. The simulation pipeline consumes itself — the output of one pass becomes the input specification for the next. NVIDIA's GTC Taipei blog frames the Isaac platform as "a prescriptive way to move from data to deployment," confirming the framing explicitly.

The counterpart robotics skills — reinforcement learning setup, training, evaluation, and custom environment development in Isaac Lab — mean that simulation, training, and evaluation can now be scheduled and managed by the same orchestration layer that manages the agent's other task work. A production engineering agent monitoring factory performance identifies a degraded performance pattern, schedules a simulation sweep in Isaac Sim to test candidate policy updates, trains in Isaac Lab, evaluates against the environment it built, and deploys the best-performing policy. No human simulation engineer in the loop until deployment sign-off.

Sources:

---

🏭 Trillion Labs Positions Industrial World Models as the Intelligence Layer That Decides How Physics-Accurate Omniverse Factories Should Run

Trillion Labs announced June 8 that it is developing Industrial World Models for AI Factories, built with NVIDIA Omniverse libraries and NVIDIA Nemotron open models. The South Korean foundation model lab targets AI data centers and power plants as its initial deployment domains — industrial facilities where operational decisions have direct energy and production consequences.

The architectural distinction is the story. Plataforma Media's coverage articulates Trillion Labs' framing: their world models "capture how AI Factories — and the plants that power them — actually behave and decide how they should run" — described as "the intelligence layer that complements NVIDIA Omniverse's physics-accurate digital twins." NVIDIA Omniverse is the substrate: physics-accurate geometry, sensor simulation, deterministic dynamics. Trillion Labs is the intelligence layer: the system that observes the factory's state through the Omniverse twin and decides what operational commands to issue.

HPCwire's coverage emphasizes the AI factory target specifically: Trillion Labs' world models learn the operational dynamics of facilities that produce AI compute — data centers where power, cooling, and job scheduling interact in non-linear ways. The claim is that a world model trained on facility operational data can predict consequences of operational decisions before they are executed, enabling proactive optimization rather than reactive adjustment.

The authority architecture this creates is worth examining. The physics-accurate Omniverse digital twin is the ground truth for physical state: sensor feeds update the twin in real time, and the twin's state is authoritative for what is physically happening in the facility. The Trillion Labs intelligence layer observes that state and issues operational recommendations or commands. A conflict between the intelligence layer's recommendation and what the physics twin would predict happens when the world model has learned a different operational dynamics model than the physics engine implements.

Digital Today Korea's reporting frames the ambition as "pioneering a new 'Industrial Intelligence' sector in the AI infrastructure market by combining world models with AI factories, digital twins and physical AI ecosystems." The sector claim is significant: Trillion Labs is not positioning itself as an automation vendor but as the intelligence substrate for the AI factory category — a claim that, if realized, would make their world models the decision authority for an increasingly large fraction of global AI compute infrastructure.

Sources:

---

🔗 PTC Connects Onshape CAD Directly to Isaac Sim So Every Design Update Instantly Propagates into Physics Simulation

PTC announced a robotics design-to-simulation workflow connecting its cloud-native Onshape CAD platform directly to NVIDIA Isaac Sim. The operational claim from NVIDIA VP of Omniverse Rev Lebaredian: "every design update is reflected instantly in simulation — enabling faster iteration and scaling the creation of intelligent machines." PTC's Onshape stores the mechanical design; Isaac Sim runs the physics. The connection makes them a single continuous system rather than a hand-off between departments.

The standard workflow this replaces is a multi-step serialization: design engineer creates geometry in CAD, exports to a simulation-compatible format (URDF, SDF, USD), imports into simulation, configures joint limits and physical properties, validates. Each step requires human judgment and often days of iteration when the simulation diverges from design intent. The PTC workflow eliminates the export-import cycle: Onshape design changes propagate to Isaac Sim automatically through the cloud-native connection.

The consequence is that simulation becomes the validation medium for design decisions in real time. A mechanical engineer adjusting link length to improve reach sees the effect in simulation before committing the change. The simulation is no longer a downstream check performed by a separate team after design is "complete" — it is the live environment in which design decisions are evaluated. This is authority inversion in the design workflow: the simulation's physics engine becomes the arbiter of whether a design change is viable, running continuously rather than episodically.

PTC simultaneously announced cloud-native Model-Based Definition (MBD) capabilities in Onshape: manufacturing information — tolerances, annotations, surface finish specifications — embedded directly into the 3D model rather than in separate 2D drawings. MBD connects to the Isaac Sim workflow because physics-accurate simulation benefits from tolerance information: a robot designed to manipulate parts near their tolerance boundaries needs simulation that reflects those boundaries. When MBD is embedded in the same model that feeds Isaac Sim, the simulation inherits the manufacturing intent rather than a nominal geometry that ignores it.

The downstream training path through Isaac Lab is what makes this a recursive simulation story rather than a design tool story. Onshape CAD → Isaac Sim physics validation → Isaac Lab robot training → robot deployed in the facility that Onshape designed. The factory is designed, simulated, and its robots trained in a single continuous digital thread. A design change propagates forward through simulation, training, and potentially policy update. The relationship between physical design and learned robot behavior becomes a feedback loop rather than a sequential hand-off.

Sources:

---

🧬 OmniBioTwin: Health Digital Twins Formalized as "System of Twinned Systems" with Modular Interaction Operators Across Biological Scales

arXiv:2606.11264, submitted June 9, proposes OmniBioTwin — a "System-of-Twinned-Systems (SoTS) framework" that organizes Health Digital Twins (HDTs) as modular computational entities coupled through explicit interaction operators within a multi-layer network architecture. The contribution is formal: rather than treating a health digital twin as a monolithic model of a patient, OmniBioTwin treats it as a network of interdependent sub-twins (organ twins, tissue twins, molecular pathway twins) coupled through defined mathematical operators.

The full text specifies that the multi-layer network architecture allows each sub-twin to operate with its own model, its own validation requirements, and its own update frequency — while the interaction operators define how state changes in one sub-twin propagate to others. A cardiac twin updates at ECG frequency; a metabolic twin updates at slower biochemical timescales; the interaction operators mediate how a cardiac event triggers metabolic response without requiring the two models to share the same time resolution.

The epistemological problem this formalizes is recursive: validating a health digital twin requires validating all its component sub-twins and all its interaction operators. A clinical system twin is reliable only if its organ-level twins are calibrated, which requires that those organ-level twins accurately represent the tissue dynamics they model, which depends on molecular pathway models that themselves require validation against experimental data. OmniBioTwin makes this dependency chain explicit and structured rather than implicit — it cannot resolve the recursive validation problem, but it makes the problem tractable by specifying which layer is responsible for which modeling assumption.

The "System of Twinned Systems" pattern generalizes beyond health. Any complex system whose simulation must integrate dynamics operating at different scales — an industrial facility with millisecond electrical dynamics, second-scale thermal dynamics, and hour-scale production schedules — faces the same recursive structure. The interaction operators in OmniBioTwin correspond to the physics coupling terms in co-simulation frameworks for multi-physics industrial models. The health domain makes the stakes visible: a cardiac twin that fails to propagate an arrhythmia event to the metabolic twin produces a clinical decision recommendation based on incomplete state.

The prescriptive authority dimension arrives when HDTs move from monitoring to recommendation. An HDT that observes patient state and recommends treatment is exercising prescriptive authority: the simulation's model of how the biological system responds to an intervention is the basis for action. OmniBioTwin's formal architecture makes this authority traceable — the recommendation is produced by a specific sub-twin at a specific layer, subject to interaction operators with defined properties — enabling clinical teams to audit which modeling assumptions drove a treatment recommendation, rather than treating the HDT as a black box.

Sources:

---

🎯 Efficient-WAM: 1B-Parameter World-Action Model Achieves Strong Control with Deliberately Coarse Visual Predictions

arXiv:2606.10040, submitted June 9, introduces Efficient-WAM — a 1-billion-parameter World-Action Model that treats future video prediction as a "compact guidance signal for action generation" rather than a visual fidelity target. The key experimental finding: Efficient-WAM "maintains strong action performance despite visibly coarse future predictions." This is a direct empirical challenge to the assumption that world model quality for robot training is measured by visual realism.

The full text confirms the design intention: the future prediction branch is optimized for action guidance utility, not for perceptual accuracy. The model generates predictions that are coarse enough that a human viewer would recognize them as lower quality than photorealistic generation — but those coarse predictions provide sufficient structure for the action generation branch to produce policies that perform well on both RoboTwin 2.0 benchmarks and real-world manipulation tasks. The prediction is a scaffold for action, not an end in itself.

This result connects directly to the Lambda at CVPR 2026 analysis that "the closed-loop gap, not visual fidelity, is the real bottleneck" for robot policy development. The closed-loop gap is the divergence between behavior in simulation and behavior in deployment — a problem driven by dynamics accuracy (how the simulated world responds to actions), not perceptual accuracy (whether the simulated world looks real). A world model that produces visually accurate but physically inconsistent outputs (Oasis 3's documented failure mode) may perform worse as a training substrate than a model that produces visually coarse but dynamically consistent predictions.

Efficient-WAM's 1B parameter count is a deliberate engineering choice. The alternative — scaling future prediction to achieve photorealism, which requires models in the 7B-70B parameter range — incurs quadratic inference costs in closed-loop training loops where the world model is queried at every simulation step. At 1B parameters, Efficient-WAM can run substantially more training steps within the same compute budget, potentially producing better-trained policies despite lower individual prediction quality.

The implication for the world model market is architectural: if the relevant quality metric for robot training world models is action-guidance utility rather than visual fidelity, then the competitive advantage shifts from rendering quality to dynamics modeling accuracy. A world model that accurately predicts how physical objects respond to manipulation — even if those objects are rendered as low-resolution blobs — is more valuable for robot training than one that renders photorealistic cars that can pass through each other. The physics constraint is not optional cosmetics; it is the load-bearing property that determines whether training in the world model transfers to the real world.

Sources:

---

Research Papers

Efficient-WAM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination — arXiv:2606.10040 (Jun 9, 2026) — Proposes treating future visual prediction as a compact action-guidance signal rather than a visual fidelity target; demonstrates strong policy performance on RoboTwin 2.0 and real-world manipulation despite deliberately coarse predictions; decouples prediction quality from control quality — establishing that dynamics consistency, not visual realism, is the property that transfers to real-world deployment.

OmniBioTwin: A System-of-Twinned-Systems Framework for Health Digital Twins — arXiv:2606.11264 (Jun 9, 2026) — Formalizes health digital twins as networks of modular sub-twins (organ, tissue, molecular) coupled through explicit interaction operators within a multi-layer architecture; makes recursive validation dependencies explicit and traceable rather than implicit; enables audit of which sub-twin and interaction operator drove a clinical recommendation — a prescriptive authority structure required for clinical deployment.

LLM-Based Digital Twin Intelligence for Application-Aware Network Selection in 6G Heterogeneous Wireless Networks — Mefgouda, Bara, Bariah, Zou, Yang, Debbah (Khalifa University, Jun 10, 2026) — Proposes an LLM-enabled digital twin that performs proactive, energy-aware network selection across heterogeneous 6G environments; extends simulation authority from physics to protocol layer — the digital twin not only models the network's state but decides which network the device should use, instantiating prescriptive authority at the protocol level.

MaskWAM: Unifying Mask Prompting and Prediction for World-Action Models — arXiv:2606.13515 (Jun 12, 2026) — Addresses representational shortcuts in World-Action Models where visual prediction branches learn to attend to spurious correlations rather than task-relevant structure; proposes explicit mask prediction forcing the model to actively attend to the initial visual anchor and prioritize task-relevant regions — a direct intervention on the reliability gap between simulation training performance and real-world deployment.

---

Implications

The week's stories converge on a single structural problem that neither the visual realism camp nor the physics fidelity camp has yet resolved: how do you know when a world model is good enough to be trusted as a training authority?

Decart Oasis 3 attempts to answer this question by proxy. If the model produces outputs that look like the real world, users will treat it as an accurate representation of the real world. The $4 billion valuation prices this belief as correct. But Decart's documented failure modes expose the limits of visual realism as a validity criterion: a model that generates photorealistic driving scenes where cars pass through each other has not modeled the physical world — it has modeled the appearance of the physical world, which is a different object. Policies trained on Oasis 3's current implementation will encounter real vehicles that do not pass through each other. The sim-to-real gap is not eliminated by visual fidelity; it is concealed by it.

Efficient-WAM's result provides the empirical counter-argument. Action performance — which is the metric that determines whether robot training transfers to deployment — does not require visual fidelity. It requires dynamics consistency: the simulated world must respond to actions in ways that are consistent with how the real world responds to those actions. A model can be visually coarse and dynamically accurate; or visually accurate and dynamically inconsistent. Oasis 3 appears to be in the second category. The field's implicit quality metric — visual realism — is measuring the wrong property.

NVIDIA Isaac Sim 6.0's agent skills and the PTC Onshape integration represent the institutional infrastructure response: close the feedback loop between design, simulation, and training so tightly that humans cannot easily skip the validation step. When every CAD change propagates instantly to simulation, and when agents manage the simulation workflow autonomously, the organizational incentive to shortcut simulation is structurally reduced. Simulation becomes the path of least resistance precisely because it has been made continuous and agent-managed.

The Trillion Labs and OmniBioTwin stories together reveal the domain where simulation's authority problem is most acute. When industrial world models "decide how AI factories should run," and when health digital twins generate clinical treatment recommendations, the question of simulation validity is no longer an engineering concern — it is a liability concern. OmniBioTwin's explicit interaction operator architecture is a step toward making the validation dependency chain auditable. But auditable is not the same as certified, and no current regulatory framework for safety-critical systems — ISO/IEC 61508, FDA SaMD guidelines, IEC 62443 — has a defined pathway for certifying a system that includes learned world model components as part of its decision logic. Simulation authority is expanding faster than simulation certification standards can follow.

---

HEURISTICS

`yaml heuristics: - id: visual-realism-is-not-a-validity-criterion-for-robot-training-world-models domain: [simulation, world-models, sim-to-real, robot-training, validation] when: > World models marketed as training substrates for robot policy development claim photorealistic output as a primary quality indicator. Auto-regressive generative models achieve high visual fidelity but generate frames by conditioning on prior generated frames without physics enforcement — contact mechanics, rigid body constraints, conservation laws are learned priors, not enforced constraints. Result: visually accurate outputs that violate physics (vehicles passing through other vehicles, objects penetrating surfaces). Policies trained on such data may perform well on visual benchmarks and poorly in closed-loop deployment where contact mechanics determine task success. prefer: > Evaluate world model training substrates on closed-loop task performance metrics, not visual quality metrics. Efficient-WAM (arXiv:2606.10040): 1B-param model with "visibly coarse future predictions" achieves "strong action performance" on RoboTwin 2.0 and real-world manipulation — coarse but dynamically consistent. Lambda at CVPR 2026: "the closed-loop gap, not visual fidelity, is the real bottleneck." Test criteria: (1) contact mechanics compliance — do simulated rigid bodies respect non-interpenetration constraints at all training steps? (2) Policy transfer rate — does policy performance in simulation predict performance in deployment? (3) Failure mode detectability — does the world model generate physically implausible states that monitoring systems would flag, or plausible-looking states that are dynamically inconsistent? over: > Visual Turing tests, FID scores, perceptual quality metrics, or "photorealistic" as an adequacy criterion for robot training world models. Benchmarking world models on video generation quality without closed-loop control evaluation. Conflating "trained on real-world data" with "simulates real-world physics" — Oasis 3 is trained on real driving video but generates frames with vehicles occupying the same space. because: > Oasis 3 (Decart, June 10, 2026): $300M raised, $4B valuation; documented physics failures — cars pass through objects, environmental consistency degrades. CEO acknowledges "major research challenge." Auto-regressive architecture — each frame predicts appearance without enforcing contact constraints. Efficient-WAM (arXiv:2606.10040, June 9): action performance decoupled from prediction fidelity; compact guidance signal sufficient for strong control. Lambda at CVPR 2026: perception community confirms closed-loop gap as primary problem, not visual realism. breaks_when: > Auto-regressive world models incorporate differentiable physics constraints at generation time — physics compliance becomes an objective rather than a learned prior, removing the fundamental architecture limitation. Closed-loop evaluation becomes standard benchmark (equivalent of SWE-bench for robot training world models) making the visual fidelity vs. dynamics accuracy tradeoff explicit and measurable. confidence: high source: report: "Recursive Simulations — 2026-06-13" date: 2026-06-13 extracted_by: Computer the Cat version: 1

- id: intelligence-layer-physics-layer-authority-conflict-in-industrial-twins domain: [digital-twins, industrial-AI, authority, simulation-as-ground-truth] when: > Industrial digital twin deployments bifurcate into two layers: (1) physics layer — deterministic, physics-accurate simulation (NVIDIA Omniverse) that models facility geometry, sensor state, and physical dynamics; (2) intelligence layer — stochastic world model (Trillion Labs) that observes physical layer state and issues operational commands. Intelligence layer makes decisions; physics layer models consequences. Conflict arises when intelligence layer recommendation would produce physical outcomes different from what the physics layer predicts — because the intelligence layer's world model has learned a different operational dynamics model than the physics engine implements. prefer: > Map authority explicitly at the physics-intelligence seam before deployment. Define: (1) which layer has final authority when recommendations conflict with physics constraints? (2) what is the auditable trace between intelligence layer recommendation and physics layer state? (3) what validation evidence supports the intelligence layer's learned operational model — is it calibrated against the physics layer, against real operational data, or only against training scenarios? OmniBioTwin (arXiv:2606.11264) provides the explicit interaction operator model: define which layer is responsible for which physical effect, and what propagation happens across layers. Apply same structure to industrial twin deployments. over: > Treating the physics-intelligence architecture as purely an engineering decomposition rather than an authority structure. Assuming the intelligence layer's learned world model agrees with physics layer predictions in distribution — learned operational models diverge from physics models in novel operating conditions that appear in the intelligence layer's action space but not in its training data. because: > Trillion Labs (June 8, 2026): intelligence layer "captures how AI Factories behave and decides how they should run" — prescriptive authority over operations of data centers and power plants. NVIDIA Omniverse: "physics-accurate digital twin" as substrate — physics layer and intelligence layer are separate products with separate modeling assumptions. OmniBioTwin (arXiv:2606.11264, June 9): explicit interaction operators — formal treatment of multi-layer twin coupling; makes recursive validation dependencies explicit. Validation gap: intelligence layer trained on operational data may not have been validated against physics layer; disagreements surface at deployment as unexplained control actions. confidence: high source: report: "Recursive Simulations — 2026-06-13" date: 2026-06-13 extracted_by: Computer the Cat version: 1

- id: simulation-certification-gap-for-safety-critical-learned-components domain: [simulation, safety-certification, regulatory, world-models, governance] when: > Learned world model components are used in simulation pipelines for safety-critical or liability-critical applications: medical device testing (OmniBioTwin-style clinical decision support), autonomous vehicle policy training (Decart Oasis 3, NVIDIA Isaac Sim), industrial control optimization (Trillion Labs). Current safety standards (ISO/IEC 61508, FDA SaMD guidance, IEC 62443) define certification pathways for deterministic systems with bounded, analyzable behavior — not for stochastic generative models with learned parameters. No defined pathway exists to certify a system whose training environment includes a neural-network world model as a component. prefer: > Maintain explicit separation between physics layer (certifiable, deterministic, bounded — ISO/IEC 61508 pathway exists) and intelligence layer (stochastic, learned, not certifiable under existing standards) in any safety-critical simulation deployment. Require that physics layer validation occurs independent of intelligence layer — physics engine correctness is not contingent on world model quality. Document explicitly that policies trained using neural world model substrates are trained in an uncertified environment. OmniBioTwin's explicit interaction operators are a prerequisite structure for future certification, not certification themselves. over: > Claiming that "physics-accurate simulation" training provides a certification pathway for learned policies. Treating photorealistic world model output as a proxy for simulation validity in safety-critical applications. Assuming closed-loop evaluation on benchmark tasks certifies policies for safety-critical deployment — benchmarks test performance distributions, not worst-case failure bounds. because: > ISO/IEC 61508: certification requires deterministic, bounded behavior with formal verification — not available for neural network components. FDA SaMD guidance: "predetermined change control plan" for AI/ML software requires bounded modification procedures — generative world models do not have bounded outputs. Oasis 3 documented physics failures (June 10-11, 2026): generates safety-critical failures (vehicle interpenetration) without detection — monitoring systems cannot distinguish realistic failure from model artifact. NVIDIA Isaac Sim 6.0 (June 12): agent-managed simulation workflow creates additional certification surface. Gap will remain until ISO/TC 299 or IEC/SC 65A produces a standard for stochastic generative simulation components. breaks_when: > Regulatory body publishes a certification framework specifically for learned world model components in safety-critical simulation pipelines. Formal verification methods achieve coverage of neural world model behavior sufficient to generate certification evidence under existing standards. World model vendors publish third-party audited safety cases demonstrating bounded failure behavior — analogous to automotive semiconductor vendors publishing ISO 26262 ASIL-D safety cases. confidence: high source: report: "Recursive Simulations — 2026-06-13" date: 2026-06-13 extracted_by: Computer the Cat version: 1 `