Independent AI research, benchmarks, and practical insights for capital markets professionals.
Oct 18, 2026
•
3 min read
Before you trust an AI score, ask what produced it.
Oct 11, 2026
2 min read
The Dutch markets regulator named reconciliation and settlement breaks. Its answer to "is the AI right?" was "when properly validated." By whom?
Oct 4, 2026
The Governor named testing as the starting point for AI governance. Nobody named the tester. Here’s what checking our own tests turned up.
Sep 27, 2026
Whether a reasoning model actually reasons can depend on which server answers, and nothing in the response tells you.
Sep 20, 2026
One category confusion accounts for a model's entire deficit, and on a third of those cases it answers differently each time you ask.
Sep 13, 2026
Four 2026 papers independently found that a single benchmark score hides most of what a model can, and can't, actually do.
