A member academy of the International Academies of Social Sciencesldass.org
LDASSLondon

LDASS research record

January–March 2023 to September 2026

The digital record of this branch, listed from the most recent entry backwards.

September 20262 researches

  1. L090LDASSSeptember 2026

    The Secure Knowledge State: London, European Research Networks and the Algorithmic Governance of Strategic Exposure in an Era of US–China Technology Competition

    The strategic contest surrounding China is increasingly extending from physical supply chains into the architecture of knowledge itself. Universities, laboratories, artificial-intelligence companies, venture-capital networks and multinational research collaborations simultaneously generate economic value and create pathways through which advanced technologies can circulate internationally. By late September 2026, this tension has become particularly consequential for London: the United Kingdom seeks to preserve an internationally connected scientific ecosystem while developing stronger mechanisms for identifying foreign-state risks, protecting dual-use research and maintaining interoperability with European and American technology-security frameworks. This article investigates whether a new “secure knowledge state” is emerging in Europe: a governance model in which research openness is increasingly mediated by computational assessments of institutional relationships, technological sensitivity and downstream application risk. The question is especially timely because the European Commission updated its dual-use control list on 14 September 2026 to incorporate controls covering additional technologies. In Britain, contemporary research-security analysis similarly emphasizes the difficulty of protecting strategically sensitive science without undermining the international collaboration on which British research strength depends. A September 2026 Chatham House study argues that the UK’s position also depends increasingly upon deeper integration with European science, technology and defence-industrial networks. The article consequently treats London not simply as a geographic research centre but as an information-processing junction connecting British universities, European institutions, American technology-security priorities, international finance and China-related expertise. It examines how policymakers could distinguish ordinary academic collaboration from relationships warranting additional scrutiny without allowing automated systems to substitute nationality, ethnicity or political assumptions for evidence.

    Method notes are released to subscribers.

  2. L089LDASSSeptember 2026

    From De-Risking to Dependency Intelligence: Mapping Europe’s China Exposure Across Critical Minerals, Advanced Manufacturing and Strategic Technology Networks

    By September 2026, European economic security has moved beyond the relatively simple language of “de-risking” that characterized the first half of the decade. The central strategic problem is increasingly one of dependency visibility: European governments can identify politically sensitive suppliers and headline trade concentrations, but they possess a less complete picture of the second-, third-, and fourth-order dependencies connecting Chinese industrial capacity to European manufacturing, energy transition infrastructure, artificial intelligence, defence production and advanced research. This article develops a machine-learning-assisted framework for reconstructing those hidden dependency networks and evaluates their significance for the United Kingdom, European Union and wider transatlantic economic-security architecture. The study situates the September 2026 European debate within three simultaneous developments: increasingly interventionist European industrial policy; continuing attempts to diversify critical-material and technology supply chains; and the emergence of economic security as an increasingly important component of Western strategic competition with China. The EU continues officially to characterize its China strategy as de-risking rather than decoupling, while identifying Chinese export controls, industrial subsidies and supply-chain dependencies as important economic-security concerns. September developments sharpen this problem. European strategic critical-mineral projects have reported financing and implementation difficulties even as policymakers seek to reduce Chinese dependencies, while the proposed Industrial Accelerator Act would use procurement and local-content mechanisms to strengthen European production in strategically important sectors. Rather than measuring dependency through bilateral trade statistics alone, the article conceptualizes Europe as a multilayer industrial graph connecting mines, refiners, component manufacturers, logistics nodes, patents, corporate ownership, research partnerships and final manufacturers. It asks which apparently European supply chains remain indirectly dependent upon Chinese processing, intermediate components, intellectual property or production capacity after first-tier suppliers are removed from analysis.

    Method notes are released to subscribers.

August 20262 researches

  1. L088LDASSAugust 2026

    The Machine-Mediated Transatlantic Order: Measuring Strategic Convergence After Agentic AI

    The machine-mediated transatlantic order is contested in Europe less through raw capability than through the institutions, rules and published positions that different capitals bring to it. The study measures strategic convergence after agentic AI using six classes of evidence — matched US, UK and EU public strategy corpora, agent-generated analyses, source-ablation runs, citation graphs, frame labels and human comparative assessments — assembled for 2026 and linked throughout to collection date and source. Analysis uses counterfactual institution-specialist agent ensembles with source swaps, narrative-distance metrics, citation-graph analysis, disagreement preservation and expert comparative adjudication; holdouts follow the geographic, temporal or institutional structure of the evidence so that performance is not inflated by leakage. Legal obligations, political framing and material conditions are coded independently, since they frequently move on different timetables. Disagreement between specifications is preserved instead of resolved, as it marks the boundary of what the evidence can currently support. The output is an August 2026 reference point for the machine-mediated transatlantic order, defined tightly enough that subsequent periods can be compared against it.

    Method notes are released to subscribers.

  2. L087LDASSAugust 2026

    Agentic Compliance After the AI Act: Auditing Autonomous Systems Against Applicable European Obligations

    By August 2026, agentic compliance after the AI Act had become a recurring point of divergence between London, Brussels and Washington, each working from different legal instruments and strategic assumptions. This research audits autonomous systems against applicable European obligations across a purpose-built evidence base covering 2026: applicable AI governance texts, bounded agent workflows, action logs, risk categories, obligation mappings, evidence citations and expert audit labels. The modelling strategy is compliance-auditor agents with retrieval-grounded legal memory, action-trace inspection, obligation-to-event mapping, adversarial workflow tests and expert legal-policy review, checked against simpler baselines to confirm that the added complexity earns its place. Representation is held as a separate analytical layer: US, UK and EU positions are compared as distinct institutional voices rather than averaged into a single Western view. Residuals are treated as evidence: the cases the model misses tend to mark infrastructure openings, regulatory changes or reporting shifts that the covariates do not contain. Taken together, the study provides a bounded August 2026 account of agentic compliance after the AI Act and an evidence architecture that later batches can reuse.

    Method notes are released to subscribers.

July 20262 researches

  1. L086LDASSJuly 2026

    World Models for Strategic Infrastructure Resilience: Testing European Counterfactuals With Explicit Uncertainty

    World models for strategic infrastructure resilience is documented in Europe mainly through texts — legislation, guidance, parliamentary record and official commentary — which makes provenance as important as content. Working from 2020–2026, the study tests European counterfactuals with explicit uncertainty through an evidence base combining energy, transport, digital and logistics networks, disruptions, weather, demand, interventions, geospatial layers and uncertainty annotations. Estimation relies on uncertainty-aware graph world model with multimodal network states, intervention rollouts, conformal scenario intervals, historical backtesting and causal-assumption logging, with results reported alongside their uncertainty rather than as point claims. What an institution says is recorded separately from what can be measured, so that shifts in framing can be dated against shifts in observable conditions. Systematic error is reported rather than smoothed away, because clustered residuals usually point to an omitted institutional or infrastructural factor. The result is a reproducible July 2026 baseline for world models for strategic infrastructure resilience that later work can extend without re-specifying the evidence base.

    Method notes are released to subscribers.

  2. L085LDASSJuly 2026

    Policy Agents With Institutional Identity: Preserving London, Brussels and Washington Context in Shared Workflows

    Debate over policy agents with institutional identity in the United Kingdom and the European Union moves faster than the evidence base behind it, which is why a dated, source-controlled record is useful. The article examines preserving London, Brussels and Washington context in shared workflows, drawing on matched UK, EU and US public policy corpora, institution-specific retrieval stores, collaborative tasks, message logs, citations and expert frame annotations for 2026 and recording provenance and missingness for each variable. The estimator is identity-conditioned multi-agent collaboration with isolated retrieval indexes, structured message passing, source-boundary enforcement, context-fidelity scoring and expert adjudication, with validation structured around the way the data are actually generated rather than random splitting. Legal obligations, political framing and material conditions are coded independently, since they frequently move on different timetables. Where the model fails, the failure is documented and located, since patterned error is often the most informative output of the exercise. Predictive performance is not treated as causal proof; the contribution is a dated, documented July 2026 benchmark for policy agents with institutional identity.

    Method notes are released to subscribers.

June 20262 researches

  1. L084LDASSJune 2026

    Transatlantic Strategy After Agentic AI: Auditing Machine-Mediated Convergence and Divergence

    Transatlantic strategy after agentic AI is contested in Europe less through raw capability than through the institutions, rules and published positions that different capitals bring to it. The study audits machine-mediated convergence and divergence using six classes of evidence — matched US, UK and EU public strategy corpora, agent analyses, source-ablation runs, citations, frame annotations and human comparative assessments — assembled for 2026 and linked throughout to collection date and source. Analysis uses institution-specialist research agents with isolated retrieval, counterfactual source swaps, narrative-distance metrics, citation-fidelity checks and expert frame adjudication; holdouts follow the geographic, temporal or institutional structure of the evidence so that performance is not inflated by leakage. What an institution says is recorded separately from what can be measured, so that shifts in framing can be dated against shifts in observable conditions. Disagreement between specifications is preserved instead of resolved, as it marks the boundary of what the evidence can currently support. The output is a June 2026 reference point for transatlantic strategy after agentic AI, defined tightly enough that subsequent periods can be compared against it.

    Method notes are released to subscribers.

  2. L083LDASSJune 2026

    Negotiating With Evidence: Human-Agent Teams for European Strategic Planning

    By June 2026, negotiating with evidence had become a recurring point of divergence between London, Brussels and Washington, each working from different legal instruments and strategic assumptions. This research examines human-agent teams for European strategic planning across a purpose-built evidence base covering 2026: bounded strategic-planning cases, source packets, human/agent actions, evidence graphs, confidence reports, decisions and evaluator rubrics. The modelling strategy is human-agent team experiments with adaptive task allocation, shared provenance graphs, calibrated confidence, disagreement protocols and blinded quality evaluation, checked against simpler baselines to confirm that the added complexity earns its place. Legal obligations, political framing and material conditions are coded independently, since they frequently move on different timetables. Residuals are treated as evidence: the cases the model misses tend to mark infrastructure openings, regulatory changes or reporting shifts that the covariates do not contain. Taken together, the study provides a bounded June 2026 account of negotiating with evidence and an evidence architecture that later batches can reuse.

    Method notes are released to subscribers.

May 20262 researches

  1. L082LDASSMay 2026

    Causal World Models for Infrastructure Decisions: Stress-Testing Counterfactual Claims in Europe

    Causal world models for infrastructure decisions is documented in Europe mainly through texts — legislation, guidance, parliamentary record and official commentary — which makes provenance as important as content. Working from 2020–2026, the study tests stress-testing counterfactual claims through an evidence base combining energy/transport networks, disruptions, weather, demand, interventions, policy changes and carefully documented causal assumptions. Estimation relies on structural causal graph constraints combined with learned graph dynamics, intervention simulations, sensitivity analysis and historical backtesting, with results reported alongside their uncertainty rather than as point claims. What an institution says is recorded separately from what can be measured, so that shifts in framing can be dated against shifts in observable conditions. Systematic error is reported rather than smoothed away, because clustered residuals usually point to an omitted institutional or infrastructural factor. The result is a reproducible May 2026 baseline for causal world models for infrastructure decisions that later work can extend without re-specifying the evidence base.

    Method notes are released to subscribers.

  2. L081LDASSMay 2026

    Continual-Learning Policy Agents: Updating European Knowledge Without Catastrophic Forgetting

    Debate over continual-learning policy agents in the United Kingdom and the European Union moves faster than the evidence base behind it, which is why a dated, source-controlled record is useful. The article models updating European knowledge without catastrophic forgetting, drawing on time-versioned EU/UK policy corpora, sequential update batches, superseded rules, evaluation questions, model checkpoints and forgetting scores for 2024–2026 and recording provenance and missingness for each variable. The estimator is parameter-efficient continual learning with retrieval memory, rehearsal buffers, temporal benchmarks, forgetting metrics and rollback tests, with validation structured around the way the data are actually generated rather than random splitting. Legal obligations, political framing and material conditions are coded independently, since they frequently move on different timetables. Where the model fails, the failure is documented and located, since patterned error is often the most informative output of the exercise. Predictive performance is not treated as causal proof; the contribution is a dated, documented May 2026 benchmark for continual-learning policy agents.

    Method notes are released to subscribers.

April 20262 researches

  1. L080LDASSApril 2026

    Machine-Readable Accountability for AI Agents: Logging Decisions Across European Public Services

    Machine-readable accountability for AI agents is contested in Europe less through raw capability than through the institutions, rules and published positions that different capitals bring to it. The study examines logging decisions across European public services using six classes of evidence — synthetic administrative workflows, tool calls, evidence references, decision events, permission states, replay traces and auditor labels — assembled for 2026 and linked throughout to collection date and source. Analysis uses event-sourced agent logging with structured rationale fields, provenance graphs, anomaly detection, replay testing and human audit scoring; holdouts follow the geographic, temporal or institutional structure of the evidence so that performance is not inflated by leakage. What an institution says is recorded separately from what can be measured, so that shifts in framing can be dated against shifts in observable conditions. Disagreement between specifications is preserved instead of resolved, as it marks the boundary of what the evidence can currently support. The output is an April 2026 reference point for machine-readable accountability for AI agents, defined tightly enough that subsequent periods can be compared against it.

    Method notes are released to subscribers.

  2. L079LDASSApril 2026

    Governed Agent Swarms for European Crisis Analysis: Coordination Without Unbounded Autonomy

    By April 2026, governed agent swarms for European crisis analysis had become a recurring point of divergence between London, Brussels and Washington, each working from different legal instruments and strategic assumptions. This research examines coordination without unbounded autonomy across a purpose-built evidence base covering 2026: fictionalised crisis scenarios, public-policy-derived rules, evidence packets, agent roles, permissions, interaction logs and expert assessments. The modelling strategy is governed multi-agent orchestration with role permissions, shared evidence blackboards, verifier gates, budget constraints and reproducibility testing, checked against simpler baselines to confirm that the added complexity earns its place. Legal obligations, political framing and material conditions are coded independently, since they frequently move on different timetables. Residuals are treated as evidence: the cases the model misses tend to mark infrastructure openings, regulatory changes or reporting shifts that the covariates do not contain. Taken together, the study provides a bounded April 2026 account of governed agent swarms for European crisis analysis and an evidence architecture that later batches can reuse.

    Method notes are released to subscribers.