Research & evidence
External claims on this website are grounded in original publications from regulators, international institutions and established industry research organisations.
Links reviewed 10 September 2026I use research to challenge and sharpen my judgement. It does not replace accountable decisions. Dates below refer to the original publication or the current official update where stated.
17 primary or institutional referencesEnterprise adoption and value
Shows that adoption is widening while enterprise value still depends on workflow redesign, leadership commitment and operational rigour.
Original sourceAI portfolio focus and value
Connects stronger AI outcomes with a focused portfolio, redesigned processes, workforce enablement and systematic measurement.
Original sourceCross-sector AI risk management
Provides cross-sector actions for governing, mapping, measuring and managing generative-AI risks throughout the lifecycle.
Original sourceEuropean AI regulation
Provides the official risk-based framework and current application timeline relevant to cross-border European operations.
Original sourceTool-using agent architecture
Documents agent modes and explicit allow, ask and deny boundaries for built-in, custom and MCP-exposed tools.
Original sourceSearch, experimentation and external evaluation
Demonstrates a research workflow that explores experiments through managed tree search, executes code and uses review feedback, while documenting important limits.
Original sourceInteractive reasoning evaluation
Introduces interactive environments that test exploration, memory, goal acquisition and planning efficiency on unfamiliar tasks.
Original sourceCurrent benchmark competition
Provides the active competition rules, data, scoring method and current leaderboard for ARC-AGI-3. Leaderboard results can change.
Original sourceIntelligence as skill-acquisition efficiency
Provides the conceptual basis for measuring generalisation, priors and the efficiency with which a system acquires new skills, beyond its score on familiar tasks.
Original sourceAbstract reasoning and efficient generalisation
Introduces a benchmark designed to remain relatively easy for humans while exposing capability and efficiency gaps in frontier reasoning systems.
Original sourceAdaptive inference-time computation
Shows that the value of additional reasoning compute depends on the problem and the strategy used, supporting selective rather than uniform allocation.
Original sourceLimits of intrinsic self-correction
Finds that asking a model to reconsider its own reasoning without external feedback can fail or degrade performance, motivating independent evaluation signals.
Original sourceFeedback, memory and agent correction
Demonstrates an agent design that converts task feedback into linguistic reflections and episodic memory to improve later attempts across several task types.
Original sourceStructured access to tools and resources
Defines capability negotiation and the resources, prompts and tools that an MCP server can expose. It is an integration layer, not a reasoning or safety mechanism.
Original sourceQuantum combinatorial optimisation
Introduces QAOA as a method for producing approximate solutions to certain combinatorial optimisation problems. This is the narrow technical basis for exploring quantum-assisted planning subproblems.
Original sourceQuantum advantage and benchmarking
Reviews the opportunity and open questions in quantum optimisation, emphasising rigorous comparison with appropriate classical methods before claiming advantage.
Original sourceQuantum constraint solving and current limits
Examines QAOA on hard constraint-satisfaction problems while documenting the substantial practical and comparative hurdles involved in outperforming classical solvers.
Original source