Growing the pilot to 93 institutions across 27 countries exposed a problem the sample size itself created: the live relationship graph was built when this pilot tracked about fifteen institutions, and a flat ring of 93 names at a fixed radius had degraded into overlapping text β readable in principle, not in practice. Two changes ship today to fix that, plus a new tool built on top of the same real dataset.
The graph now groups by category instead of scattering alphabetically
Institutions are grouped into arcs by category β Bank, Insurer, Asset Manager, Financial Services, Payment Network β each with its own gold label and a visible gap from its neighbors, so the shape of the ring itself now shows relative category size at a glance. The ring's radius scales with the sample size instead of staying fixed, and past roughly twenty institutions, individual name labels stay hidden until you hover a dot rather than all rendering at once. Fewer things on screen by default, the same full detail one hover away.
A comparison panel that grows with the data instead of just the circle
Below the graph, three new panels break the current (filtered) selection down by model family, by category, and by country β real counts, not estimates, recalculating live as you use the existing filters. At the full 93-institution sample right now: Microsoft (Azure OpenAI) leads with 47 of 120 relationships, banks make up 52 of the 93 institutions, and the US remains the single largest country at 28 β with Japan, the UK, and Australia rounding out the next tier.
New: a portfolio simulator built on the real data, not a hypothetical
The chance-baseline calculator on the methodology page has always worked with abstract numbers β pick N models and K options, see the theoretical odds. The new portfolio simulator, live now on the data page, does the same calculation on institutions you actually pick. Search any of the 93 tracked institutions by name, or one-click add an entire category, and it computes β from the real, sourced relationships already in the dataset β how many distinct model families are present, how concentrated the group is on its single most common one, and how that compares to what pure chance would predict using the same (1/k)^(n-1) formula already used on this site.
A concrete run: adding every tracked insurer (19 institutions) shows 57.9% of them sharing Microsoft (Azure OpenAI) as a confirmed model family, against a pure-chance baseline of under 0.01% β a +57.9 point gap. That's not a new statistical model invented for this feature; it's the exact same observed-vs-chance logic behind the correlation test, applied to real vendor-relationship data instead of simulated model outputs.
What this doesn't claim
The simulator counts documented vendor relationships, the same Confirmed/Probable tiers as everywhere else on this site β it says nothing about whether those institutions' AI systems would actually fail in a correlated way, only that they share the same underlying model family on paper. See the methodology for what this index does and doesn't measure, a distinction that matters more, not less, now that the numbers are easier to generate on demand.