Supply Chain Digital Twin MCP Server

> View on ApifyForge | Use on Apify Store

---

Quick Start

Add to your MCP client (Claude Desktop, Cursor, Windsurf):

{
  "mcpServers": {
    "supply-chain-digital-twin-mcp": {
      "url": "https://ryanclinton--supply-chain-digital-twin-mcp.apify.actor/mcp"
    }
  }
}

---

Supply chain digital twin simulation via 8 quantitative algorithms — connect any AI agent to live corporate, trade, sanctions, hazard, and financial data across 17 public data sources. This MCP server turns natural-language supply chain questions into structured risk intelligence: cascade failure probabilities, optimal transport routes, adversarial interdiction games, supplier survival curves, Leontief multipliers, and multi-horizon demand forecasts.

Built for supply chain risk analysts, operations researchers, and AI agents that need real-world data fused with rigorous algorithms — not static databases. Each tool runs up to 17 actors in parallel, assembles a live supply chain network graph, then applies a specific algorithm from network science, operations research, or econometrics.

⬇️ What data can you access?

| Data Point | Source | Example use |
|---|---|---|
| 📋 Global corporate registrations and ownership | OpenCorporates | Tier-1/2/3 supplier identification |
| 🏢 UK company filings and financial data | UK Companies House | UK entity verification and financials |
| 🔗 Legal Entity Identifiers and parent-child chains | GLEIF LEI | Corporate ownership mapping |
| 🌐 International bilateral trade flows | UN COMTRADE | Inter-sector IO coefficient estimation |
| 📊 Economic development indicators | World Bank | Country-level macro risk factors |
| 🏛️ Government contract awards and opportunities | SAM.gov | Defense and federal supply chain exposure |
| 🚫 OFAC sanctions and blocked persons list | OFAC | Sanctioned entity flagging |
| 🌍 Global sanctions, PEP, and watchlist data | OpenSanctions | Multi-jurisdiction sanctions screening |
| 🌋 Earthquake activity and seismic magnitudes | USGS Earthquake | Natural hazard proximity scoring |
| 🌩️ Severe weather events and climate data | NOAA | Hazard exposure for geographic nodes |
| 🆘 Global disaster alerts and severity scores | GDACS | Multi-hazard disruption triggers |
| 🔴 US federal emergency declarations | FEMA | Domestic disaster exposure |
| 📍 Geocoding and lat/lon resolution | Nominatim | Haversine distance-based hazard mapping |
| 💱 Foreign exchange rates and currency data | Exchange Rate Tracker | FX risk in cross-border trade flows |
| 🔐 Internet host certificates and vulnerabilities | Censys | Supplier cyber exposure scoring |
| 🌫️ Air quality index by country and region | OpenAQ | Environmental operational risk |
| 🖥️ Website uptime and availability monitoring | Website Monitor | Supplier operational continuity signals |

Why use Supply Chain Digital Twin MCP?

Manual supply chain risk analysis means pulling data from a dozen government portals, building spreadsheets, and applying heuristics. A thorough analysis of a single supplier portfolio can take a team of analysts a full week — and still miss cascade dynamics, game-theoretic attack surfaces, and multi-horizon demand signals.

This MCP server replaces that workflow. Paste a query into any MCP-compatible AI agent. The server calls up to 17 live data sources in parallel, builds a weighted network graph with Leontief technical coefficients, and runs the requested algorithm. You get structured JSON results in seconds — not weeks.

- Scheduling — run on any cadence via the Apify platform to keep supply chain models current
- API access — trigger tool calls from Python, JavaScript, or any HTTP-based MCP client
- 17 parallel data sources — corporate, trade, sanctions, hazard, cyber, environmental, and financial data assembled in one call
- Monitoring — receive Slack or email alerts when runs fail or return anomalous results
- Integrations — connect to Claude Desktop, Cursor, Windsurf, LangChain, or any MCP-compatible client

Features

- Buldyrev-Parshani-Stanley mutual percolation — models two interdependent supply networks; iteratively removes non-giant-component nodes and propagates failures until a stable state, detecting first-order phase transitions at the critical threshold p_c
- Cross-entropy importance sampling — optimizes the Monte Carlo proposal distribution q(theta) to focus simulation effort on catastrophic rare events, delivering variance reduction ratios orders of magnitude better than naive Monte Carlo
- Sinkhorn-Knopp entropy-regularized optimal transport — solves min_{P} <C,P> - epsilonH(P) subject to marginal constraints via iterative scaling u = mu/(Kv), v = nu/(K^Tu), computing the Wasserstein-2 distance as total logistics cost
- Tri-level Stackelberg attacker-defender game — models adversarial interdiction as max_x min_y max_z f(x,y,z); solved via Benders decomposition with optimality cuts; reports Nash equilibrium resource allocation and convergence gap
- Competing risks Cox proportional hazard model — five cause-specific hazards (financial, natural disaster, sanctions/regulatory, cyber, quality/compliance) with cumulative incidence F_k(t) = integral h_k(u)S(u)du and Harrell's concordance index
- Algebraic connectivity maximization — computes lambda_2 (Fiedler value) of the graph Laplacian via power iteration with deflation; greedy edge additions scored by gain_ij = (v_i - v_j)^2 using the Fiedler vector
- Leontief input-output via Neumann series — inverts (I-A) as L = sum_{k=0}^{K} A^k up to 15 terms; computes forward linkage (column sums), backward linkage (row sums), and Ghosh supply-side multiplier; identifies keystone industries with above-average linkage in both directions
- Fuzzy regression discontinuity design — estimates tau_FRD = (E[Y|X>=c] - E[Y|X<c]) / (E[D|X>=c] - E[D|X<c]) with Imbens-Kalyanaraman (2012) optimal bandwidth h_IK = C_k (sigma^2 / (f(c)m''(c)^2))^{1/5} n^{-1/5} at geographic and temporal thresholds
- Rao-Blackwellized particle filter — samples nonlinear regime component via particles, marginalizes the linear state-space (trend, seasonal, cycle) analytically via Kalman filter per particle; systematic resampling when effective sample size ESS < N/2
- Stochastic DP via Bellman equation — solves V(s) = min_a { c(s,a) + gammaE[V(s')] } for optimal base-stock inventory policy with uncertainty quantification
- Haversine distance-based hazard exposure — assigns hazard exposure to supplier nodes within 500 km of a hazard event (1 - dist/500), falling back to seeded probabilistic assignment (15% chance, 0.3 weight) when coordinates are unavailable
- Leontief matrix normalization — enforces sub-stochastic rows (spectral radius < 1) by scaling rows with sum >= 0.9 by factor 0.85/rowSum, guaranteeing Neumann series convergence
- Seeded deterministic pseudo-randomness — all stochastic components use a string-hash-seeded linear congruential generator for reproducible results across identical inputs
- 17 actors called in parallel — all data fetches use Promise.all for maximum throughput; each actor call has a 180-second timeout with graceful empty-array fallback on failure

Use cases for supply chain digital twin simulation

Supply chain stress testing and resilience scoring

Risk managers at manufacturing firms and diversified industrials use simulate_disruption_cascade to stress-test their supplier networks before quarterly board reviews. Input a company name or sector description. The tool assembles live corporate, sanctions, hazard, and trade data, constructs a two-layer interdependent network, and runs up to 5,000 Monte Carlo simulations with cross-entropy importance sampling to estimate the probability of catastrophic cascades. Output includes the percolation threshold, first-order transition flag, and a ranked list of critical suppliers.

Logistics network cost optimization

Operations teams at 3PLs and large manufacturers use optimize_logistics_transport to identify the lowest-cost routing and sourcing configuration given current supply and demand distributions. The Sinkhorn algorithm solves the entropy-regularized optimal transport problem, returning Wasserstein-2 distance as the optimal cost, dual shadow prices for each capacity-constrained node, and a full set of optimal routes with utilization rates.

Critical infrastructure and defense supply chain protection

Government contractors and defense procurement teams use model_adversarial_interdiction to identify which supply chain links an adversary would most rationally target and where defensive investment yields the highest return. The tri-level Stackelberg game produces a Nash equilibrium that specifies which edges are interdiction targets, how much protection each deserves, and the Benders convergence gap as a quality indicator.

Supplier portfolio risk management

Procurement teams and supply chain finance professionals use estimate_supplier_survival to assess the 12-month failure probability of every supplier in a portfolio, decomposed into five competing failure causes. Output includes per-supplier survival curves, cumulative incidence by failure mode, expected number of failures over the next year, and Harrell's concordance index as a model fit metric.

Economic impact and sectoral contagion analysis

Policy researchers and corporate economists use compute_input_output_impact to estimate how a production shock in one sector or country propagates through the full input-output network. The Leontief multiplier tells you total output impact per unit of direct demand change. Keystone industry identification flags the sectors whose disruption would most damage the entire network — essential for both business continuity planning and macroeconomic policy analysis.

Demand forecasting and inventory optimization

Supply chain planners use forecast_multi_scale_demand to generate uncertainty-quantified demand forecasts across multiple horizons (weekly, monthly, quarterly) and compute the optimal base-stock level for each. The Rao-Blackwellized particle filter decomposes demand into trend, seasonal, and regime components, and the stochastic DP solver finds the inventory policy that minimizes expected holding plus shortage costs.

How to connect this MCP server

This server runs in Apify Standby mode — it stays alive and accepts MCP connections over HTTP. The endpoint is always available.

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "supply-chain-digital-twin": {
      "url": "https://supply-chain-digital-twin-mcp.apify.actor/mcp"
    }
  }
}

Cursor

Add to your Cursor MCP settings:

{
  "mcpServers": {
    "supply-chain-digital-twin": {
      "url": "https://supply-chain-digital-twin-mcp.apify.actor/mcp"
    }
  }
}

Windsurf / other MCP clients

Any client that supports the MCP streamable HTTP transport can connect to:

https://supply-chain-digital-twin-mcp.apify.actor/mcp

No authentication headers required. The Apify token is managed server-side.

⬆️ MCP tools reference

| Tool | Algorithm | Data sources called | Output highlights |
|------|-----------|---------------------|-------------------|
| simulate_disruption_cascade | Buldyrev mutual percolation + cross-entropy IS | 17 actors | Cascade steps, percolation threshold, first-order transition flag, recovery days, variance reduction ratio |
| optimize_logistics_transport | Sinkhorn-Knopp entropy-regularized OT | 8 actors | Optimal routes, Wasserstein-2 distance, shadow prices, Sinkhorn convergence |
| model_adversarial_interdiction | Tri-level Stackelberg + Benders decomposition | 8 actors | Game value, Nash equilibrium, interdiction targets, defender reinforcements, Benders gap |
| estimate_supplier_survival | Competing risks Cox PH model | 11 actors | Per-supplier survival probability, 5-cause cumulative incidence, expected failures 12m, concordance index |
| reinforce_network_resilience | Algebraic connectivity (Fiedler value) | 7 actors | Lambda_2, redundancy score, Fiedler vector, optimal edge additions with lambda_2 gain |
| compute_input_output_impact | Leontief I-O via Neumann series | 6 actors | Leontief inverse, forward/backward linkages, Ghosh multiplier, keystone industries, spectral radius |
| identify_causal_disruption_paths | Fuzzy RDD + IK bandwidth | 11 actors | Causal paths, RDD estimates, IK bandwidth, total causal effect, confounders |
| forecast_multi_scale_demand | RBPF + stochastic DP Bellman | 7 actors | Multi-horizon forecasts with 95% CI, optimal base-stock, service level, effective particles |

Tool parameters

| Parameter | Type | Required | Default | Applies to |
|-----------|------|----------|---------|------------|
| query | string | Yes | — | All 8 tools |
| simulations | number | No | 1000 | simulate_disruption_cascade only (max 5000) |

The query parameter accepts free-text descriptions of supply chain scenarios, company names, industries, commodities, or geographic regions. The server extracts intent and passes it to each underlying data actor as a search query.

Output examples

simulate_disruption_cascade

{
  "nodesAffected": 23,
  "totalImpact": 4.87,
  "percolationThreshold": 0.312,
  "firstOrderTransition": true,
  "criticalSuppliers": [
    "Pinnacle Semiconductor GmbH",
    "Meridian Electronics Co Ltd",
    "Vantage Logistics Holdings"
  ],
  "recoveryTimeDays": 47,
  "ceProposalDistortion": 0.0234,
  "monteCarloRuns": 1000,
  "varianceReduction": 18.4,
  "cascadeSteps": [
    {
      "node": "Pinnacle Semiconductor GmbH",
      "tier": 1,
      "failureProbability": 0.78,
      "cascadeDelay": 3,
      "economicImpact": 1.24,
      "networkA": true
    },
    {
      "node": "Meridian Electronics Co Ltd",
      "tier": 2,
      "failureProbability": 0.61,
      "cascadeDelay": 8,
      "economicImpact": 0.89,
      "networkA": false
    }
  ],
  "networkSize": { "nodes": 94, "edges": 187 }
}

estimate_supplier_survival

{
  "portfolioSurvival": 0.71,
  "expectedFailures12m": 4.2,
  "concordanceIndex": 0.724,
  "competingRisks": [
    { "cause": "financial",          "cumulativeIncidence": 0.142 },
    { "cause": "natural_disaster",   "cumulativeIncidence": 0.081 },
    { "cause": "sanctions",          "cumulativeIncidence": 0.034 },
    { "cause": "cyber",              "cumulativeIncidence": 0.058 },
    { "cause": "quality_compliance", "cumulativeIncidence": 0.091 }
  ],
  "suppliers": [
    {
      "supplier": "Acme Fabrications Ltd",
      "survivalProbability": 0.84,
      "hazardRate": 0.024,
      "medianLifetimeYears": 8.3,
      "riskFactors": [
        { "factor": "sanctioned",       "hazardRatio": 3.41, "coefficient": 1.227 },
        { "factor": "hazardExposure",   "hazardRatio": 2.18, "coefficient": 0.780 },
        { "factor": "cyberExposure",    "hazardRatio": 1.64, "coefficient": 0.495 }
      ],
      "causeSpecificHazards": [
        { "cause": "financial",        "hazard": 0.018, "cumulativeIncidence": 0.098 },
        { "cause": "natural_disaster", "hazard": 0.009, "cumulativeIncidence": 0.049 }
      ]
    }
  ],
  "networkSize": { "nodes": 71, "edges": 132 }
}

optimize_logistics_transport

{
  "totalCost": 2847300,
  "totalTimeDays": 18.4,
  "wasserstein2Distance": 0.3812,
  "sinkhornIterations": 47,
  "sinkhornConvergence": 0.0000083,
  "dualVariables": [
    { "constraint": "Shanghai-Port",   "shadowPrice": 142.50 },
    { "constraint": "Rotterdam-Hub",   "shadowPrice": 98.20 },
    { "constraint": "Chicago-DC",      "shadowPrice": 67.80 }
  ],
  "optimalRoutes": [
    {
      "from": "Shenzhen Manufacturing Zone",
      "to": "Los Angeles Distribution Center",
      "mode": "sea",
      "cost": 1240000,
      "timeDays": 21,
      "capacity": 5000,
      "utilization": 0.87,
      "sinkhornWeight": 0.412
    }
  ],
  "networkSize": { "nodes": 58, "edges": 103 }
}

❓ How much does it cost to run supply chain digital twin simulations?

This MCP uses pay-per-event pricing — you pay per tool call. Compute costs are included. The pricing per tool reflects the number of underlying actor calls and algorithm complexity.

| Tool | Price per call | Typical monthly budget (20 calls) |
|------|---------------|-----------------------------------|
| simulate_disruption_cascade | $0.040 | $0.80 |
| optimize_logistics_transport | $0.040 | $0.80 |
| model_adversarial_interdiction | $0.045 | $0.90 |
| estimate_supplier_survival | $0.035 | $0.70 |
| reinforce_network_resilience | $0.040 | $0.80 |
| compute_input_output_impact | $0.035 | $0.70 |
| identify_causal_disruption_paths | $0.035 | $0.70 |
| forecast_multi_scale_demand | $0.030 | $0.60 |

For AI agent workflows running dozens of tool calls per session, a typical session cost is under $1.00. You can set a maximum spending limit per run in your Apify account settings to prevent unexpected charges — the server checks this limit before executing and returns a budget-exceeded message rather than proceeding.

Compare this to enterprise supply chain risk platforms (Resilinc, Riskmethods, Everstream) at $50,000–$200,000 per year for similar analytical capabilities. This MCP delivers quantitative simulation on live data for a few cents per analysis.

How Supply Chain Digital Twin MCP works

Phase 1: Parallel data assembly

Each tool call dispatches between 6 and 17 actor calls via Promise.all, all running in parallel with a 180-second timeout. Corporate data comes from OpenCorporates, UK Companies House, and GLEIF LEI. Trade flow data comes from UN COMTRADE and World Bank. Sanctions screening queries OFAC and OpenSanctions. Natural hazard data comes from USGS Earthquake, NOAA, GDACS, and FEMA. Location resolution uses Nominatim. Cyber exposure uses Censys. Environmental risk uses OpenAQ. Operational monitoring uses Website Monitor. Government contract exposure uses SAM.gov. Any actor that times out or returns an error is treated as an empty array — the algorithm proceeds with available data.

Phase 2: Network graph construction

buildSupplyChainNetwork assembles all actor results into a SupplyChainNetwork object. Corporate entities are assigned to tiers 1, 2, or 3 based on their position in the corporate data array. Hazard nodes are linked to supplier nodes using haversine distance — exposure weight = 1 - dist/500 for nodes within 500 km; probabilistic assignment (15% chance, weight 0.3) otherwise. Air quality risk is mapped by country code. Cyber exposure is matched by name prefix. Supply edges between tiers carry random IO coefficients (0.05–0.40) seeded deterministically from a string hash of the node pair. The Leontief technical coefficient matrix A is built from these IO coefficients, then row-normalized to enforce spectral radius < 1 (Neumann series convergence condition).

Phase 3: Algorithm execution

Each of the 8 algorithms operates on the assembled SupplyChainNetwork:

- Percolation partitions nodes into two networks by node index, builds a dependency map, and iterates removal of non-giant-component nodes across both layers. Cross-entropy importance sampling perturbs edge retention probabilities by ceProposalDistortion to concentrate simulation mass on rare cascade events.
- Optimal transport constructs a cost matrix from haversine distances between supplier and distributor nodes, then runs Sinkhorn iterations until convergence (tolerance 1e-7, max 1000 iterations).
- Interdiction enumerates supply edges as potential attack targets, scores each by criticality * weight, and solves the bi-level assignment via greedy Benders cuts. The game value is the expected flow reduction at Nash equilibrium.
- Survival analysis computes cause-specific Cox hazard rates from node attributes (sanctioned status, hazard exposure, cyber exposure, air quality risk, contract dependency), then integrates cumulative incidence functions numerically.
- Algebraic connectivity builds the graph Laplacian L = D - A, then runs 50 iterations of power iteration with deflation against the constant eigenvector to estimate lambda_2 and the Fiedler vector.
- Input-output computes the Leontief inverse as a Neumann series truncated at 15 terms, checks spectral radius via 30-step power iteration, and identifies keystone sectors as those with above-average forward and backward linkages.
- Causal paths applies fuzzy RDD at geographic thresholds (500 km bands) and temporal thresholds, computing IK optimal bandwidth and local linear regression estimates on either side of each discontinuity.
- Demand forecasting runs a particle filter with 50 particles, Kalman filter per particle for linear state components, systematic resampling when ESS < 25, then backward-passes Bellman recursion to find the optimal base-stock level.

Phase 4: Structured JSON output

All results are serialized to JSON and returned as MCP tool call content. Floating point values are rounded to 2–4 decimal places to keep responses readable in agent contexts. Large matrices (Leontief inverse) are truncated to 6×6 for display.

Tips for best results

1. Be specific in your query. "Semiconductor supply chain for automotive electronics, Taiwan and Japan suppliers" produces a richer network than "electronics supply chain." Specificity drives more relevant actor results, which produce more accurate network edges.

2. Use simulate_disruption_cascade for unknown unknowns. If you are not sure which algorithm to start with, cascade simulation is the broadest diagnostic. It surfaces critical suppliers, percolation thresholds, and whether the network sits near a first-order transition — information that shapes which other tools to run next.

3. Follow cascade simulation with reinforce_network_resilience. Once you know which suppliers are critical (from cascade output), use resilience reinforcement to identify which new supplier relationships (edge additions) would most increase algebraic connectivity at lowest cost.

4. Combine estimate_supplier_survival with model_adversarial_interdiction. Survival analysis identifies suppliers most likely to fail naturally. Interdiction modeling identifies suppliers most likely to be deliberately targeted. Together they give a complete vulnerability picture.

5. Use compute_input_output_impact for board-level reporting. Leontief multipliers translate technical supply chain risk into financial language — total output change per unit of disruption. Keystone industry identification is directly presentable to non-technical stakeholders.

6. Set a spending limit per session. For exploratory agent workflows, set a per-run budget in your Apify account. The server checks the limit before each tool call and stops gracefully rather than continuing to charge.

7. Chain forecast_multi_scale_demand with inventory decisions. The optimal base-stock output is directly usable in ERP or planning systems. Export the forecasts array as CSV via the Apify dataset API for import into Excel or Google Sheets.

Combine with other Apify actors

| Actor | How to combine |
|-------|---------------|
| Company Deep Research | Run deep research on the critical suppliers identified by simulate_disruption_cascade to get full financial, legal, and operational profiles before escalating to procurement leadership |
| Website Contact Scraper | Extract contact details from the websites of high-risk suppliers flagged by estimate_supplier_survival to initiate proactive supplier diversification outreach |
| Trustpilot Review Analyzer | Cross-reference supplier quality compliance risk scores with customer reviews and complaint patterns for triangulated supplier health assessment |
| B2B Lead Qualifier | Score potential replacement suppliers identified through trade data before adding them to a shortlist for RFQ |
| Website Tech Stack Detector | Enrich cyber exposure data from Censys with technology stack analysis of supplier web properties to produce a more accurate cyberExposure score |
| WHOIS Domain Lookup | Verify domain registration authenticity for suppliers flagged as high-risk in the interdiction model |
| Website Change Monitor | Monitor supplier websites for changes (office closures, product discontinuation notices, restructuring announcements) that may precede the failure events predicted by the survival model |

Limitations

- No proprietary ERP or inventory data — the network graph is built entirely from public sources. Internal BOM data, actual inventory levels, and proprietary lead times are not incorporated. For enterprise use, the scoring algorithms can be adapted with private data inputs via the Apify API.
- Leontief model assumes linear production technology — fixed technical coefficients do not capture substitution effects, economies of scale, or non-linear production functions. Disruptions that cause fundamental structural shifts will be underestimated.
- Cox proportional hazards assumes time-invariant hazard ratios — if the relationship between risk factors and failure probability changes over time (e.g., sanctions risk growing due to geopolitical escalation), the competing risks model will not capture this.
- Cross-entropy importance sampling requires valid proposal distribution — for very small networks (fewer than 10 supplier nodes), the CE proposal may not have enough structure to improve on naive Monte Carlo.
- Benders decomposition convergence is not guaranteed for all game structures — very large or degenerate interdiction games may produce high Benders gap values, indicating the reported game value is approximate.
- Haversine-based hazard mapping requires geocoded nodes — when supplier coordinates are unavailable (lat/lon = 0), the system falls back to probabilistic hazard assignment, which is less precise.
- Actor data freshness depends on upstream sources — government databases (SAM.gov, OFAC, USGS) are updated on varying schedules. The model reflects data available at the time of the call, not real-time streaming feeds.
- Network construction is capped at 20 sectors for the Leontief matrix, and haversine edge creation is bounded to supplier pairs within a sliding window of 6 — large queries may produce sparser networks than smaller, focused queries.

Integrations

- Claude Desktop — connect via claude_desktop_config.json; ask Claude to "run a disruption cascade simulation for the automotive semiconductor supply chain"
- Cursor — add to MCP settings and use in agent mode for code-adjacent supply chain analysis
- Apify API — trigger tool calls programmatically from Python, JavaScript, or any HTTP client via the standby endpoint
- LangChain / LlamaIndex — use as a tool in LangChain agent pipelines for automated supply chain monitoring workflows
- Webhooks — trigger downstream actions (Slack alerts, CRM updates) when a run completes or a cascade simulation crosses a risk threshold
- Make — build no-code supply chain risk workflows: trigger on schedule, run simulation, push results to Google Sheets or HubSpot
- Zapier — connect supply chain simulation results to email, Slack, Airtable, or any Zapier-supported destination

Troubleshooting

Receiving Spending limit reached in tool output — the per-run budget set in your Apify account has been exhausted. Go to the actor's run settings and increase the maximum spend, or run the tool independently with a higher limit.

Network returns very few nodes (fewer than 10) — the query may be too generic or too narrow for the underlying actors to return sufficient corporate and trade records. Try adding geographic or industry context: "semiconductor suppliers in Taiwan and South Korea for consumer electronics" rather than "electronics suppliers."

Sinkhorn algorithm not converging (high sinkhornConvergence value) — this occurs when supply and demand distributions are very imbalanced. The tool still returns a result using the best available approximation after 1,000 iterations. Consider narrowing the logistics query to a more balanced geographic scope.

Cascade simulation returns firstOrderTransition: false for every run — the network may be too sparse or the query may not be returning enough hazard or sanctions data to produce interdependent network layers. Add explicit geographic context to your query to improve hazard data coverage.

Actor timeout errors in logs — individual underlying actors have a 180-second timeout. Slow responses from government APIs (SAM.gov, GLEIF) can occasionally cause timeouts. The server degrades gracefully — timed-out actors return empty arrays and the algorithm runs with available data.

Results differ slightly between identical queries run at different times — the network graph uses deterministic seeding for all stochastic components, but actual live data from the 17 upstream actors changes over time. Differences in results reflect real changes in the underlying data (new corporate registrations, updated sanctions, new hazard events), not randomness in the algorithm.

Responsible use

- All 17 data sources accessed by this MCP are publicly available government databases, international trade datasets, and open corporate registries.
- Sanctions and watchlist data (OFAC, OpenSanctions) should be used only for legitimate compliance and due diligence purposes.
- Supply chain intelligence derived from this tool should not be used to facilitate unlawful discrimination, market manipulation, or targeted harassment of individuals or companies.
- Comply with GDPR, applicable trade compliance laws, and data protection regul