CEQUS Systemic Risk Control Plane: agent dependency mapping for underwriting and risk transfer

What this page is, and is not

A response to the report, not an endorsement by it

We added the graph because of the report. This is a C=US implementation proposal. It is not an endorsement by the authors, not a finding that the report prescribes this design, and not a guarantee of insurance. Neither Swiss Re nor LSE reviewed, endorsed or took part in this work, and nothing here represents their views. The report is summarized below in our own words; read the original for its exact findings. Source: LSE & Swiss Re Institute Systemic Risk Report, September 2026.

The insurance and reinsurance layering further down is the student's proposal. It builds on the report's recommendations but is not taken from it. The graph's numbers are descriptive exposure measures from synthetic data, not loss probabilities or prices.

What the report found

More connections, more concentration, and AI as a new connector

• Risks are more connected than before. Comparing the 10-K risk disclosures of 91 Fortune-100 companies in 2019 and 2026, the authors found about 24% more links between risks. AI and new-technology risk is now reported well beyond the tech sector, including by food producers.
• The network decides how far a shock spreads. Small shocks can have outsized effects when critical nodes are concentrated or transmission channels reinforce each other. Dense but diversified networks absorb shocks instead.
• Digital infrastructure is concentrated. Three firms controlled about 70% of global cloud infrastructure in 2024. For re/insurers, that is accumulation risk: one outage can trigger claims at the same time across unrelated policyholders and lines of business.
• AI can deepen concentration. AI's high fixed costs favour a few providers; reliance on similar models reduces the diversity of behaviour that normally stabilizes markets; and automated decisions can make responses faster and more synchronized.

From the report's priorities to working code

What the dependency graph does about each one

A synthetic example

One provider outage, 20 agents

In the demonstration, 20 synthetic maple-grading agents use two model providers, two runtimes and one identity directory. Failing provider A gives:

No agent names provider A directly, yet 60% depend on it through their model. Concentration that only shows up one level down is the kind of hidden link the report warns about. The directory is a single point of failure for 95% of agents, which says as much about the design of C=US as about anyone else's.

The student's proposal

Insurance and reinsurance by depth of network risk

C=US's directory already has layers. The root lists states, states endorse organizations, and each farmer controls their own agents. Risk can be layered the same way, by how far a failure travels through the graph:

The monetization question is whether signed, checkable dependency evidence is worth money. Three services could be billed separately:

1. A subscription for operators and policyholders: continuous monitoring of their own dependencies and controls.
2. Underwriting evidence for carriers and reinsurers: permissioned exposure analysis across many insureds, to see accumulation.
3. Scenario and risk-engineering analysis: showing how a fallback or diversification reduces modeled losses.

These are hypotheses to test with insurers, not evidence of demand or of premium credits.

Next milestone (planned, not built)

From a dependency map to underwriting evidence

The graph above is built and tested. The next step would turn it into evidence an underwriter could use:

• Exposure units: link agents to the critical operations, insured organizations and coverage lines they affect, without treating an agent count as money at risk.
• Evidence for each control: mTLS identity, scoped and expiring grants, revocation freshness, failover and recovery tests. Each item would carry a "not assessed" state, so a certificate or a failover claim alone can't lower a price.
• A redacted package for insurers: concentration, scenarios and residual exposure, with drill-down under access control.
• Kept separate: "modeled loss", "possibly within a policy trigger" and "confirmed covered by the insurer". Policy wording would be supplied by insurers, never written into the code.
• A change feed: new shared dependencies, provider switches or expired evidence, to support renewals.

Credible pricing would still need what the graph cannot supply: loss history, frequency, severity and correlation calibration, policy terms, exposure values and each insurer's risk appetite. The system could support actuarial and underwriting decisions; it cannot replace them.

Limits

What this does not show

• All agents, providers and outages here are synthetic, and the keys are software development keys. No production directory or insurer is involved.
• Exposure counts describe reachability. They are not calibrated failure probabilities, loss estimates or systemic-risk scores.
• Dependencies are what each operator declares, attributed by signature; a signature shows who said it, not that it is true.
• The audit trail is tamper-evident only against a checkpoint kept elsewhere. It is not immutable or legally conclusive.

Sources and attribution

What this page draws on