LLM Orchestration Speeds Energy Crisis Model Integration

Energy-crisis analysis has a coordination problem before it has a modeling problem. Oil, gas, shipping, water, fertilizer, helium, and macroeconomic models often already exist, but they sit in different institutions, languages, file formats, and assumptions. This paper argues that the binding constraint in a fast-moving crisis is the cost of connecting those models quickly enough for policy use.

The authors introduce an LLM-orchestrated framework for the hypothetical 2026 closure of the Strait of Hormuz. The important design choice is narrow: the language model plans, translates, routes, and summarizes, but it is not allowed to create quantitative outputs. Reported values must come from model runs, with provenance and analyst approval attached to each stage.

Core Contribution

The paper’s contribution is an integration layer rather than a new economic model. The framework coordinates 16 heterogeneous models spanning oil, natural gas, LNG, shipping, water, helium, fertilizer, agriculture, electricity, and macroeconomic equilibrium. It builds internally consistent scenarios, converts shared assumptions into model-specific inputs, executes models in dependency order, and composes a policy-facing synthesis from source-attested outputs.

That is a practical claim: a crisis desk could read a portfolio of models instead of whichever one is easiest to run. The paper’s strongest argument is not that the resulting forecasts are correct, but that the orchestration makes disagreement and coverage visible. In Appendix B, the coverage map shows that model participation varies by scenario: Scenario C uses all 16 models, while Scenarios D and E use only 5. That variation is treated as part of the evidence, not hidden behind a single aggregate answer.

Technical Approach

The system works as a controlled pipeline. Scenario definitions, a model registry, forwarding rules, and consistency tolerances constrain what the orchestrator can do. The LLM maps narrative assumptions into model inputs, schedules dependent runs, flags values that violate consistency checks, and writes briefs whose numbers remain traceable to underlying model outputs. Analyst approval is part of the process; the architecture is designed to block silent numerical fabrication rather than trust the model to be careful.

The model suite includes a mix of real configured models, minimum viable integrations, and stubs. The appendix lists examples such as the Global Gas Model, LNG Spreadsheet Tool, World Helium Model, SimRLFab, MAgPIE, APSIM, AISdb, airstream, NEMS, MAM, MPSGE.jl, OpenCGE, pycge, OSeMOSYS, MESSAGEix, and TEMOA. This heterogeneity is the point. The orchestration layer has to bridge optimization models, spreadsheets, agent-based models, CGE models, time-series tools, spatial databases, and Python, GAMS, Julia, Excel, Fortran, R, and other execution environments.

Figure 5 shows the regional macroeconomic asymmetry that the system surfaces: higher oil prices can produce a net GDP gain for the MENA producer aggregate even as the same shock raises instability and import-dependent losses elsewhere.

Results and Analysis

The main experiment applies the framework to five Strait of Hormuz closure scenarios. Across the matrix, the authors report that the five all-scenario anchor models run in every scenario, while the full model count ranges from 5 to 16 depending on relevance and available assumptions. That is useful evidence for graceful degradation: the pipeline can still produce a bounded reading when only part of the model suite applies.

The paper also tests whether the analysis can be refreshed as information changes. It performs eight weekly reruns from 15 February to 5 April 2026, using news and U.S. Energy Information Administration indicators to update the information set while holding scenario definitions, the model registry, forwarding rules, and tolerances fixed. The weekly refreshes track changing headline quantities such as oil-price impulse, effective shipping-fleet-capacity loss, GDP impact, CPI inflation, and article coverage. The authors interpret week-to-week shifts as controlled remeasurement when the information set changes, with regressions flagged when identical inputs produce different behavior.

A second test swaps the orchestrating LLM backend. Of eight candidate backends, four produced complete analyses and four returned no usable brief. Among the four that ran, Sonnet 4.6 averaged 26.3 minutes per weekly rerun, while Llama-4-Scout, gpt-oss-120b, and gpt-oss-20b took about 7.0, 8.5, and 7.9 minutes. The trade-off is coverage versus latency: Sonnet 4.6 reached a clean synthesis every week but completed fewer model runs on average, while the faster open backends completed roughly twice as many runs with a lower weekly success rate.

The evidence is a serious systems demonstration, but it is not a validation of crisis forecasts. The scenarios are constructed, some listed components are stubs or MVP integrations, and long-run geopolitical and financial spillovers are outside the matrix. The paper is best read as evidence that LLMs can reduce the manual integration burden for existing scientific models under strict provenance controls, not as proof that the resulting policy judgments are correct.