The whole book: test it under stress
Customers sit in five affordability bands: Secure, Ticking Over, Comfortable, Stretched, and Stressed. Raise everyone's living costs and watch customers move between bands. The money at risk updates with them, and a signal shows whether the book is within the lending policy, at its limit, or over it. The chief risk officer sees the policy and the capital it requires on one screen.
One customer: the person in the red
Behind every band is a person. Portfolio IQ opens the file: income against spending, what is left at the end of the month, and the signs of trouble a credit score does not show. Income replaced by borrowing. The share of spare spending that goes to gambling. Bank charges, debt collection activity, cash withdrawals. Beside it, ninety days of money in and out, and the affordability score next to the credit score that disagrees with it.
Where Sheaf comes in
An AI agent recommends a top-up loan. Two of the agents used two different incomes for the same customer, because one of them counted a recent loan as income. They cannot both be right. Sheaf sees the contradiction, stops the recommendation under Consumer Duty and parks it for a person, and records why before anyone acts.
Every decision checked
Each recommendation carries a signed record of whether the reasoning held together, with any conflicting claims named.
Contradictions stopped
A recommendation that rests on contradictory facts waits for a person. The agent cannot go around the stop.
Fairness across the book
Similar customers treated differently shows up across the whole book, which a check on one decision at a time cannot see.
Evidence that lasts
Stopped actions, releases, and records go into a tamper-evident record, signed and timestamped by an independent timestamp authority.
The mathematics behind it
Portfolio IQ uses sheaf cohomology across the whole loan book. Every agent's claim about a customer is one local piece, and the mathematics asks whether all of them fit into one whole. That is what lets it see a contradiction across the book that a check on one decision at a time cannot. The sheaf Laplacian gives a coherence score from 0 to 100 and a consensus answer weighted by each agent's confidence.