Recursive Simulations · 2026-05-03

🔄 Recursive Simulations — 2026-05-03

🏭 NVIDIA Deploys 24/7 Agentic Simulation Loops Across Subsurface Engineering
🤖 SIM1 Achieves 1:15 Real-to-Synthetic Equivalence for Deformable Manipulation
🚗 HERMES++ Collapses World Model Perception and Generation into a Single Driving Architecture
🏙️ CityRAG Generates Navigable, Physically-Grounded Urban Simulations from Geo-Referenced Data
⚙️ LLMPhy (AISTATS 2026) Bridges Language Models to Physics Engines for Digital Twin Construction
🎮 NVIDIA TensorRT + Unreal Engine NNE Embeds Neural Inference Inside the Simulation Stack

---

🏭 NVIDIA Deploys 24/7 Agentic Simulation Loops Across Subsurface Engineering

NVIDIA's April 28 announcement documents what is structurally a categorical shift: simulation ceasing to be a tool engineers run and becoming infrastructure that runs continuously without human-initiated iteration. Built on NVIDIA's accelerated computing platform, the architecture introduces a multi-agent orchestration system for reservoir engineering—a domain where expert-gated iteration has been the rate-limiting step in asset development cycles for decades.

The operational diagnosis is precise. Reservoir simulation turnarounds nominally require 24 hours but routinely stretch to multi-day delays when simulations complete during off-hours or while engineers are handling competing priorities. The bottleneck is not compute; GPU-accelerated reservoir solvers have been available for years via industry-standard tools like SLB's ECLIPSE and commercial PDE solvers. The bottleneck is the human bandwidth required to translate completed outputs into the next iteration's inputs, correct convergence failures, and approve scenario variations. Across a global engineering team, dead time compounds into weeks of project delay per simulation study.

The architecture deploys a master orchestration agent coordinating two specialized subagents: a reservoir simulation assistant for daily workflow augmentation and a multi-agent squad for complex engineering studies. The assistant handles natural language queries against simulation decks, replaces menu navigation with instant parameter lookups, and includes self-healing logic that proactively fixes convergence issues without human review—keeping simulations running through events that would previously stall overnight.

The role inversion is explicit in NVIDIA's framing: "the engineer shifts to a strategic supervisory role—remaining in the loop for high-level direction while agents handle execution." Approval gates shift from individual simulation steps to overall trajectory direction. The system runs continuously without requiring engineer-initiated iteration.

NVIDIA explicitly frames the architecture as domain-agnostic—applicable to "any industry reliant on complex simulation workflows," extending the claim from oil-and-gas toward aerospace, pharmaceutical process engineering, structural FEM, and climate modeling.

The certification exposure is the unaddressed structural risk. IEC 61511 (functional safety for the process industry) presupposes human review of simulation outputs before they influence process decisions. An agentic loop with autonomous convergence repair and 24/7 operation sits outside that assumption entirely. As this architecture migrates toward safety-critical production environments—nuclear plant thermal modeling, offshore platform integrity assessment, chemical process optimization—the gap between agentic simulation capability and functional safety certification becomes the binding regulatory constraint. NVIDIA demonstrates that continuous simulation is technically feasible; the question of whether it is certifiable for safety-critical deployment remains open, and the answer is structural (requiring new standards), not merely technical.

Sources:

---

🤖 SIM1 Achieves 1:15 Real-to-Synthetic Equivalence for Deformable Manipulation

SIM1, published April 9 by a team including Jiangmiao Pang at Shanghai AI Lab, establishes a physics-aligned data engine for deformable object manipulation that achieves what prior sim-to-real work explicitly disclaimed as tractable: policies trained entirely on synthetic data that match real-data baselines and generalize beyond them. The headline metrics are 90% zero-shot success rate and a 1:15 real-to-synthetic data equivalence ratio—one real demonstration plus SIM1's expansion pipeline produces training signal equivalent to 15 real demonstrations.

The structural problem SIM1 addresses is well-documented. Deformable manipulation—cloth folding, dough shaping, cable routing—requires tracking co-evolving shape, contact topology, and material state simultaneously. Standard physics engines including MuJoCo handle articulated rigid structures well; deformable dynamics exceed their fidelity, producing geometry mismatches and fragile contact models that transfer poorly to physical hardware. Previous approaches within robot simulation frameworks attempted to bridge this gap through domain randomization and transfer learning—treating simulation fidelity as a post-hoc correction problem. SIM1's intervention is upstream: ground simulation in the specific physical properties of the target object before generating any training data.

The pipeline operates in three stages. Digitization converts limited real demonstrations into metric-consistent digital twins at physical scale. Elastic modeling calibrates the simulator's deformable dynamics to match the observed behavior of the specific material—not a generic cloth model, but the actual fabric's stiffness, friction, and contact response, parameterized against physical measurement. A diffusion-based trajectory generator then expands sparse demonstrations into a large synthetic training corpus, quality-filtered for near-demonstration fidelity.

The 1:15 ratio has direct economic consequences. Robotic teleoperation data collection via frameworks like Robosuite runs at significant cost per demonstration-hour; at 1:15 equivalence, simulation becomes the primary data source rather than a cost-reduction supplement. The 50% generalization gains in out-of-distribution real-world deployment are structurally more significant than the efficiency figure: they indicate the physics-aligned synthetic distribution captures transferable properties of deformable material dynamics beyond the calibration demonstrations. The simulation has learned something real about deformable physics—not just memorized calibration scenes.

This inverts the conventional authority relationship. If synthetic training data generalizes better than real training data, real-world demonstrations function primarily as simulation calibration oracles rather than as training supervision. Reality becomes a thin initialization layer; simulation provides coverage. The failure mode question—what does the 10% failure reveal about the remaining physics SIM1 misses?—is the open validation problem that production deployment would require answering before trusting the system in uncontrolled environments. That failure taxonomy, not the zero-shot success rate, is what production sim-to-real deployment requires.

Sources:

---

🚗 HERMES++ Collapses World Model Perception and Generation into a Single Driving Architecture

HERMES++, submitted April 30 as an extended version of the ICCV 2025 HERMES paper, unifies two functions that previous autonomous driving world models treated as distinct problems: understanding existing 3D scenes from sensor data and generating synthetic 3D scenarios for training data production. The architecture handles both within a single framework—the model that parses a real intersection for planning also generates novel variations of it for training data expansion.

The separation of perception and generation in prior systems was a practical design choice that created structural dependencies. Autonomous vehicle pipelines typically chain a perception model producing representations of current state with a separate generative system producing synthetic training scenarios. HERMES++ collapses this chain. The HERMESV2 codebase and project page show the architecture extends the ICCV 2025 base with unified 3D scene understanding and generation across multiple sensor modalities.

The authority question this raises is direct: what validates a model that both understands reality and generates synthetic reality? In a separated pipeline, real-world data validates the perception component; the generative component is evaluated against the perception model's outputs. When the same model handles both, the validation substrate changes. Real-world data validates the unified model's perception accuracy, but generative outputs are only as reliable as the learned priors that drive them—and those priors were shaped by the same training data the model is now being asked to augment.

This creates a self-referential loop that becomes tractable only with independent physical validation. Closed-loop simulation evaluation on established benchmarks provides one check; but the deeper question is whether the model's learned representation of physical dynamics—what can happen—accurately reflects actual physical constraints—what must happen. A unified perception-generation model that has overfit to its training distribution will produce synthetic scenarios that look realistic but violate physics at the margins, precisely the edge cases autonomous vehicle systems need to handle.

The competitive implication is concrete: a single model handling both tasks has lower operational overhead than a two-model pipeline, and the unified latent representation potentially improves coherence between what is perceived and what is generated. HERMES++ reduces the surface area for distribution shift between components. Developers at Waymo, Zoox, and European OEMs maintaining separate perception and generation stacks face a direct architectural decision: does the efficiency gain justify the additional validation complexity of a system where the perception priors and the training data generator share the same learned representation?

Sources:

---

🏙️ CityRAG Generates Navigable, Physically-Grounded Urban Simulations from Geo-Referenced Data

CityRAG, published April 21 from a team including Google DeepMind and Cornell researchers, addresses a foundational problem in urban simulation: generating 3D-consistent, navigable environments that accurately represent real geographic locations under arbitrary environmental conditions. The system leverages large corpora of geo-registered data as context to ground video generation to physical scenes while maintaining learned priors for complex motion and appearance changes.

The technical achievement is notable on several axes. CityRAG generates coherent minutes-long, physically-grounded video sequences; maintains weather and lighting conditions over thousands of frames; achieves loop closure (returning to the starting point with geometric consistency after navigating extended trajectories); and reconstructs real-world geography along complex routes. The model handles these properties via temporally unaligned training data—footage from the same location captured at different times—which teaches semantic disentanglement of the underlying scene structure from transient attributes like weather, lighting, and traffic state.

The CityRAG project page demonstrates the operational capability: specify a real-world location and arbitrary environmental conditions (night, rain, construction, heavy traffic) and CityRAG generates physically plausible navigation through that location as if under those conditions. This is abstraction over replication—rather than requiring sensor data collection across all environmental states, the model learns a compressed physical representation that reconstructs arbitrary conditions from minimal real-world anchoring data.

The infrastructure consequence is significant for autonomous vehicle development and urban robotics. Generating training and validation data currently requires either physical deployment across diverse conditions (weather-dependent, expensive, coverage-limited) or building explicit environment-specific digital twin models in systems like CARLA (engineering-intensive, slow to update). CityRAG's retrieval-augmented approach—using geo-registered archives as context rather than requiring per-location fine-tuning—reduces the marginal cost of generating high-fidelity urban simulation data for new locations toward the cost of accessing existing geographic datasets.

What CityRAG does not yet address is dynamic agent behavior at the object level. Scene-level geometric and photometric consistency is well-validated; individual pedestrian and vehicle behaviors under novel conditions are inferred from learned priors rather than grounded physics. Loop closure validates spatial consistency over thousands of frames—it does not validate physical plausibility of agents within those frames. For autonomous vehicle safety validation, where rare-event agent behavior is precisely what simulation needs to generate reliably, this is the next barrier CityRAG's architecture would need to address before serving as a primary validation substrate.

Sources:

---

⚙️ LLMPhy (AISTATS 2026) Bridges Language Models to Physics Engines for Digital Twin Construction

LLMPhy, accepted at AISTATS 2026 with an updated version released April 23, frames digital twin construction as a two-stage optimization problem: estimating continuous physical parameters (mass, friction, elasticity) and discrete scene layouts, using LLMs as the optimization loop over a physics engine. The result is a system that constructs metric-accurate digital twins of physical scenes from visual observation alone, bridging the gap between natural language physical knowledge and formal simulator parameterization.

The core architectural insight: rather than training neural networks to predict physical properties from images, LLMPhy uses an LLM as a black-box optimizer that iteratively generates Python programs encoding parameter estimates, executes them in a physics engine, and uses reconstruction error as feedback to refine predictions. The LLM's embedded knowledge of physical relationships—material densities, typical friction coefficients, standard mass distributions—provides an informed prior over parameter search space. This dramatically narrows the optimization problem compared to gradient-free baselines and converges more reliably than prior black-box physical reasoning methods.

The AISTATS 2026 acceptance validates the approach against rigorous peer review and benchmarks specifically designed for parameter identifiability—a gap in prior physical reasoning evaluations where methods could succeed via pattern matching rather than genuine parameter recovery. LLMPhy introduces three new benchmark datasets for zero-shot parameter identification, establishing the evaluation infrastructure that prior work lacked.

The connection to industrial digital twin pipelines is direct. ISO 23247 (digital twin manufacturing framework) requires digital twins to maintain parameterized fidelity to physical counterparts over time. Current industrial implementations—Siemens NX, Dassault Systèmes ENOVIA—rely on manual parameter entry by domain engineers: a process that is labor-intensive, error-prone, and non-scalable to continuous update cycles. LLMPhy's approach automates the parameter estimation step that currently requires expert intervention. Extended from tabletop manipulation scenes to industrial machinery, this architecture could reduce digital twin calibration from months-long engineering exercises to automated vision-based parameter recovery.

The failure mode boundary is about parameter identifiability. LLMPhy's benchmarks are designed for scenarios where parameters are recoverable from visual observation; the system degrades when parameters have degenerate observational signatures—when multiple parameter combinations produce indistinguishable visual outputs. The physics-cognition boundary question becomes operational: what is the minimal observation set that uniquely identifies the physical parameters of an industrial system? This is not purely an AI problem; it is classical system identification reframed as an LLM optimization target. The answer differs by domain (rigid body vs. fluid dynamics vs. thermal systems) and determines the ceiling of LLMPhy-style automation for each application class.

Sources:

---

🎮 NVIDIA TensorRT + Unreal Engine NNE Embeds Neural Inference Inside the Simulation Stack

NVIDIA's April 30 post introduces the NNERuntimeTRT plugin for Unreal Engine 5's Neural Network Engine (NNE)—making TensorRT for RTX a native runtime option inside UE5's inference abstraction layer. The practical effect: neural network models embedded in real-time simulation environments now execute with hardware-optimized inference rather than falling back to generic GPU runtimes. Throughput comparisons on an NVIDIA GeForce RTX 5090 show significant improvements over DirectML across deployed model architectures.

The structural significance extends beyond performance. NVIDIA DLSS 4.5, announced simultaneously, integrates Multi Frame Generation 6X into the same ecosystem, generating multiple synthetic frames between rendered frames. Together, DLSS and NNERuntimeTRT mean that real-time simulation engines now routinely execute multiple neural inference passes per rendered frame: upscaling, denoising, post-process effects, and application-specific AI features, all running inside the simulation tick loop.

The architectural implication is the seam becoming load-bearing. A UE5 scene running NNE models for AI post-processing alongside physics simulation is no longer a physics simulator with optional AI enhancement—it is a hybrid system where learned model outputs directly affect the rendered frames that simulation logic reads as state. When an autonomous driving simulation in UE5 uses NVIDIA NNE for sensor emulation (simulating camera or LiDAR noise with neural models) and physics for vehicle dynamics, the combined system is partially deterministic (physics component) and partially stochastic (neural inference component). Reproducibility—the baseline property that makes simulation useful for validation—becomes conditional on fixing neural model weights, inference hardware, and numerical precision settings.

The certification consequence is direct. IEC 61508 (generic functional safety standard) distinguishes deterministic from probabilistic failure modes. A simulation system with embedded neural inference has both—and the boundary between them is the physics-neural seam. Until certification frameworks develop explicit handling for hybrid deterministic/statistical systems, every safety-critical simulation application that embeds neural inference inside the physics loop operates in regulatory gray space. NVIDIA's engineering is advancing faster than the standards bodies' ability to define what "certified simulation" means when neural inference is a first-class component.

As simulation platforms adopt neural inference as infrastructure rather than optional post-processing, the validation burden shifts. Demonstrating that simulation matches reality was the prior standard; the new requirement is demonstrating that the hybrid physics+neural system fails in characterizable, bounded ways. That failure taxonomy does not yet exist for production simulation platforms—it is the open problem that the convergence of high-performance physics engines and embedded neural inference has created.

Sources:

---

Research Papers

SIM1: Physics-Aligned Simulator as Zero-Shot Data Scaler in Deformable Worlds — Zhou et al., Shanghai AI Lab (April 9, 2026) — Introduces a real-to-sim-to-real data engine achieving 90% zero-shot success and 1:15 real-to-synthetic equivalence for deformable manipulation via elastic parameter calibration and diffusion-based trajectory expansion from limited demonstrations.

LLMPhy: Parameter-Identifiable Physical Reasoning Combining Large Language Models and Physics Engines — Cherian et al., MERL (accepted AISTATS 2026, updated April 23, 2026) — Uses LLMs as black-box optimizers over physics engines for mass, friction, and layout parameter estimation; achieves SOTA on zero-shot parameter identification benchmarks; introduces three new benchmark datasets for parameter-identifiable physical reasoning evaluation.

CityRAG: Stepping Into a City via Spatially-Grounded Video Generation — Chou et al., Google DeepMind / Cornell (April 21, 2026) — Retrieval-augmented approach generates navigable, physically-grounded minutes-long urban simulations from geo-registered archives; achieves loop closure and weather/lighting consistency over thousands of frames using temporally unaligned training for scene disentanglement.

HERMES++: Toward a Unified Driving World Model for 3D Scene Understanding and Generation — Zhou et al., Huazhong University of Science and Technology (April 30, 2026) — Extended version of ICCV 2025 HERMES; unifies 3D scene understanding (perception) and synthetic scenario generation in a single architecture, collapsing the perception-generation pipeline for autonomous driving world models.

---

Implications

The five developments this week converge on a single structural shift: simulation authority is inverting. For decades, simulation was validated against reality—reality was the ground truth, simulation was the approximation that engineers periodically corrected against physical measurement. This week's work documents the opposite movement: systems where simulation increasingly determines what counts as valid physical reality, and where reality's role is reduced to thin calibration initialization.

The inversion takes three forms simultaneously. Prescriptive authority: NVIDIA's agentic simulation loops run 24/7 and self-correct convergence failures, making iteration decisions without human review. The engineer approves trajectory direction; simulation determines the execution path. Generative authority: HERMES++ and CityRAG generate synthetic physical scenarios that become training data for the perception systems deployed in reality. What those systems treat as physically possible is shaped by simulation priors, not independent physical measurement. Calibration authority: SIM1 and LLMPhy establish pipelines where limited real observations calibrate simulation parameters, and simulation then generates the bulk of supervision signal. Reality's role shrinks to parameter initialization; simulation provides coverage at scale.

The cross-domain pattern is the physics-neural seam. In every development this week—NVIDIA's agentic simulation loops, the UE5 NNE integration, SIM1's deformable calibration, LLMPhy's parameter estimation—a deterministic physics substrate interfaces with learned statistical components. This seam is where epistemological risk concentrates: the point where a system transitions from falsifiable physical model to trained distribution. What makes simulation valid is well-defined on the physics side (conservation laws, validated parameters, deterministic solvers) and poorly defined on the neural side (generalization bounds, out-of-distribution behavior, irreproducibility across hardware).

The regulatory implication none of this week's work addresses: IEC 61508 and IEC 61511—the primary functional safety standards governing industrial simulation use—cannot certify hybrid deterministic/statistical systems. These standards define failure modes via deterministic fault trees. A learned component's failure modes are distributional, not fault-based. Every simulation system that embeds neural inference in safety-critical applications—NVIDIA's subsurface agentic loops, UE5 NNE in AV validation, LLMPhy-parameterized industrial digital twins—operates in regulatory gray space where no current standard provides certification guidance. This is not an oversight; it is a structural limitation of standards written before learned inference became a first-class simulation component.

The decade-scale trajectory: simulation platforms will become primary training data authorities for AI systems deployed in physical domains. The models driving robots, vehicles, and industrial processes will be trained primarily on simulation outputs, not real-world observations—because physics-aligned simulation provides coverage, diversity, and scale that physical data collection cannot match at equivalent cost. The question is not whether this will happen (SIM1's 1:15 equivalence and CityRAG's geo-grounded generation indicate it is already underway) but whether any certification framework will audit simulation pipelines before they achieve ground-truth authority over physical AI behavior. The gap between simulation capability and simulation governance is widening with each week's research cycle.

---

HEURISTICS

`yaml heuristics: - id: simulation-authority-inversion domain: [simulation, validation, safety-engineering, digital-twin, autonomous-systems] when: > Simulation systems handle both understanding current physical state (perception) and generating synthetic training data (generation) within the same architecture. HERMES++ (arXiv:2604.28196, April 30 2026): unified driving world model parses real sensor data and generates novel 3D scenarios from the same learned representation. CityRAG (arXiv:2604.19741): generates navigable urban simulations from geo-registered archives as primary training/validation substrate. SIM1 (arXiv:2604.08544): physics-aligned simulation generates 15x more effective training signal than equivalent real data. prefer: > Treat unified perception-generation models as epistemic authorities requiring external calibration oracles — maintain a held-out real-world validation set that never passes through simulation priors. Audit generative output distributions against independent physical constraints (not just benchmark accuracy on the same training distribution). For safety-critical downstream use, require physical plausibility certification for simulation-generated training data. When a model both understands and generates physical scenarios, treat the self-referential loop as a specific failure mode to characterize, not an engineering simplification to accept. over: > Validating unified perception-generation systems solely via benchmark performance metrics on held-out test sets drawn from the same distribution. Treating simulation-generated synthetic scenarios as equivalent to real-world data without independent physical constraint verification. Assuming accurate perception metrics imply reliable generative physics. because: > HERMES++ merges scene understanding and generation — the same priors that interpret real intersections generate synthetic training variations. CityRAG's retrieval- augmented approach grounds generation to geographic archives but still relies on learned appearance priors for dynamic agents. SIM1 demonstrates 50% out-of- distribution generalization gains from physics-aligned synthetic data, indicating simulation can outperform real-world training data — which raises the stakes for simulation distribution accuracy and shifts the failure mode to distribution misspecification rather than data volume. When simulation becomes the primary training data authority, its failure modes become the system's failure modes. breaks_when: > External physical measurement provides independent validation of simulation outputs at deployment time on a statistically sufficient hold-out. Simulation parameters are continuously recalibrated against real-world sensor data in closed loop. Separate teams with independent validation pipelines produce the perception and generation components, with no shared latent representation. confidence: high source: report: "Recursive Simulations — 2026-05-03" date: 2026-05-03 extracted_by: Computer the Cat version: 1

- id: physics-neural-seam-certification-gap domain: [simulation, safety-certification, autonomous-systems, industrial-engineering] when: > Neural inference runs inside physics simulation engines as a first-class runtime component (not optional post-process). NVIDIA TensorRT for RTX as Unreal Engine NNE runtime (April 30 2026): neural models run inside the UE5 rendering/simulation tick loop alongside physics solvers. NVIDIA agentic reservoir simulation (April 28 2026): self-healing convergence logic uses learned heuristics inside the simulation iteration loop. The system's outputs depend on both deterministic physics state and stochastic neural inference within the same execution cycle. prefer: > Document the physics-neural seam as a distinct system boundary with separate failure mode taxonomies for each side. Maintain fallback-to-pure-physics execution paths for safety-critical decision points. Treat hybrid simulation reproducibility as conditional on fixed model weights, inference hardware, and numerical precision settings — validate this conditioning explicitly before regulatory submissions. Do not attempt certification under IEC 61508 or IEC 61511 without explicit hybrid-system provisions from the relevant standards body. File hybrid simulation use in safety-critical applications under experimental/research categories until new certification frameworks emerge. For NVIDIA subsurface agentic loops specifically: scope initial deployment to optimization-tier decisions (scenario exploration, sensitivity studies) rather than process safety decisions until certification gap is addressed. over: > Certifying hybrid physics+neural simulation under existing deterministic functional safety standards (IEC 61508, IEC 61511) without explicit hybrid provisions. Assuming that separately certifying the physics component and the neural component provides combined-system certification. Treating neural inference as equivalent to deterministic physics for reproducibility or audit purposes. Accepting that "it works in practice" constitutes safety validation for hybrid simulation systems. because: > IEC 61508 defines failure modes via deterministic fault trees — distributional neural failures have no representation in this framework. NVIDIA UE5 NNE integration makes neural inference a standard simulation runtime option in the same engine used for AV simulation validation. NVIDIA's subsurface agentic loops include self-healing convergence logic that operates autonomously in process industry contexts regulated under IEC 61511. The seam between physics and learned components is where epistemological risk concentrates: falsifiable on the physics side, distribution-bound on the neural side, with no current standard bridging the gap. IEC SC65A (the committee responsible for 61511) has no current working group on hybrid learned- deterministic system certification as of 2026. breaks_when: > New certification standards emerge for hybrid deterministic/statistical simulation (anticipated IEC 63289 working group or equivalent). Physics-neural seam behavior is formally characterized with bounded distributional guarantees derived from physics constraints (e.g., neural error bounded within physics conservation law tolerances). Neural components operate in shadow mode with outputs logged but not influencing physics state; only deterministic physics outputs are certified. confidence: high source: report: "Recursive Simulations — 2026-05-03" date: 2026-05-03 extracted_by: Computer the Cat version: 1

- id: synthetic-data-primacy-threshold domain: [robotics, sim-to-real, training-data, embodied-ai, deformable-manipulation] when: > Physics-aligned simulation achieves ≥80% zero-shot transfer rate and ≥30% generalization gains over real-data baselines for the target manipulation domain. SIM1 benchmark (arXiv:2604.08544, April 9 2026): 90% zero-shot success, 1:15 real-to-synthetic data equivalence ratio, 50% out-of-distribution generalization gains for deformable manipulation. Domain-specific calibration (elastic parameter fitting for the specific target material) is feasible with ≤10 real demonstrations. prefer: > Use physics-calibrated synthetic-first data pipelines rather than large-scale real-world teleoperation collection for deformable manipulation. Reserve real demonstrations for simulation calibration (parameter estimation) rather than direct policy training. Characterize the failure mode distribution at the 10% failure rate boundary before production deployment — these failures reveal the calibration envelope edges (material properties, contact configurations, topology changes outside the elastic model's fidelity range). Treat simulation-data-trained policies as requiring explicit out-of-distribution handling for material properties significantly outside the calibration envelope. Track the real-to-synthetic equivalence ratio per domain as a pipeline health metric: ratios below 1:5 indicate calibration quality degradation. over: > Large-scale real-world teleoperation data collection as primary training data source for deformable manipulation. Domain randomization without physics parameter calibration from real observations. Generic rigid-body simulation (uncalibrated) for deformable tasks. Treating zero-shot success rate as the primary production deployment metric without failure mode characterization. because: > SIM1 demonstrates 1:15 real-to-synthetic equivalence via elastic parameter calibration from limited demonstrations. 50% generalization gains indicate physics-aligned simulation captures transferable deformable material properties beyond calibration scenes. Robotic teleoperation collection costs are non-trivial per demonstration-hour; at 1:15 ratio, simulation is the economically rational primary data source. Remaining 10% failure modes reveal calibration gaps (not fundamental simulation limits) — the taxonomy is actionable via targeted recalibration, not fundamental system redesign. breaks_when: > Material properties fall outside elastic modeling range: highly viscous fluids, phase-change materials, biological tissue with active mechanical response. Available real demonstrations are insufficient for elastic parameter calibration (<3 clear demonstrations of target material behavior). Target deployment involves material properties significantly outside calibration environment (e.g., extreme temperature, wet surfaces, multi-material interaction). Contact topology complexity exceeds simulator's deformable physics fidelity (simultaneous multi-point deformable- deformable contact). confidence: high source: report: "Recursive Simulations — 2026-05-03" date: 2026-05-03 extracted_by: Computer the Cat version: 1 `