Skip to main content

A work of speculative fiction. The Department of Strategic Continuity is not a real government agency, and this is not a government website. About this project

Department of
Strategic Continuity

Research program · FSEP

Frontier System Evaluation Program

Evaluates the capabilities, failure modes and infrastructure dependencies of frontier computational systems, with a focus on where advanced AI has become embedded in the control of energy, communications, finance and logistics.
Lead office
Office of Advanced Artificial Intelligence
Established
March 2026
Partner offices
Office of Strategic Risk Assessment

Program overview

The Frontier System Evaluation Program (FSEP) studies the most capable computational systems in use and the ways in which critical infrastructure has come to depend on them. It is led by the Office of Advanced Artificial Intelligence, with the Office of Strategic Risk Assessment as partner.

The Program’s premise is that the risks posed by advanced AI to long-term continuity arise less from any single system than from dependence: control functions in grids, networks, markets and supply chains increasingly rely on automated systems whose behavior under unusual conditions is poorly characterized, and whose failures may be correlated because they share architectures, training data, vendors or compute providers. Evaluating a system in isolation does not reveal these risks. FSEP evaluates systems in the context of the infrastructure that relies on them.

The Program’s methods are set out in DSC-TR-2026-007, Evaluation Framework for Frontier Computational Systems in Critical Infrastructure. Evaluation campaigns run on ALPHA-1 at Facility Alpha in isolated environments. Findings feed directly into the Computational Systems and AI module of the Integrated Civilizational Systems Model, where they inform the coupling parameters that the Civilizational Systems Outlook identified as a principal driver of systemic stress.

FSEP is an evaluation and research program. It does not certify or approve products and has no regulatory function. Findings concerning specific deployed systems are shared with the responsible operators before any public summary is released, and public reports describe classes of vulnerability rather than exploitable detail.

Research objectives

  1. Maintain and extend the evaluation framework in DSC-TR-2026-007, including test protocols for robustness, distribution shift and behavior under degraded inputs.
  2. Map the dependence of critical infrastructure control functions on frontier computational systems, and identify concentrations of shared architecture, vendor and compute supply.
  3. Characterize correlated failure modes across systems that share components or training lineage.
  4. Provide evaluated parameters for the Computational Systems and AI module of ICSM.
  5. Develop measurement methods whose results remain comparable as systems change from one generation to the next.

Current research areas

Select a research area to read a summary of current work.

Capability and failure-mode evaluation

Structured testing of frontier systems against task suites representative of infrastructure control, with emphasis on how performance degrades under stress, ambiguous inputs and adversarial conditions rather than on peak capability.

Infrastructure dependency mapping

Identification of where automated decision systems sit in the control paths of energy, communications, finance and logistics, and of what fallback capacity exists if they fail or must be withdrawn.

Correlated failure analysis

Study of how shared model families, common training data and concentrated compute supply can cause independent-seeming systems to fail together, converting local faults into system-wide events.

Evaluation measurement science

Development of evaluation methods with known reliability: test-retest consistency, contamination controls, and procedures for comparing results across system generations without the benchmarks themselves becoming training targets.

Human oversight and fallback

Assessment of whether human operators retain the skills, staffing and authority to resume manual control of systems that have been automated, and how quickly that capacity erodes after automation is introduced.

Developmental milestones

Status as of October 2026.

  1. July 2026 Complete

    Evaluation framework published

    DSC-TR-2026-007 issued on 14 July 2026, establishing the Program’s evaluation protocols and dependency taxonomy.

  2. August 2026 – 2027 In Progress

    First evaluation campaign on ALPHA-1

    Initial evaluation campaign in isolated environments at Facility Alpha, focusing on grid-operations and network-management task suites.

  3. 2027 Planned

    Infrastructure dependency atlas

    A sector-by-sector account of automated control dependencies and fallback capacity, prepared with OSRA for use in ICSM.

Research publications

Related facilities

CRF-A

Continuity Research Facility Alpha

Neural computation and high-performance computing

The Department's primary computing site. Facility Alpha hosts the ALPHA-1 system, runs production ensembles of the Integrated Civilizational Systems Model, and carries the Neural Continuity Research Program's emulation benchmark suite. Initial operating capability was reached on 18 August 2026.

Southern NevadaOperational