Stratyfy

The removal test leaves nothing. The proprietary Probabilistic Rules Engine is the entire product, applied across credit risk assessment, fraud detection and bias mitigation, and the company's whole argument is about the properties of a particular kind of model rather than about workflow around one.

Interpretability here is architectural rather than a layer added afterwards, since decisions are produced as readable probabilistic rules in the first place, which is a modelling choice and not a reporting feature.

A sixth distinct architecture in this index, and the most direct: the human does not review the model's output, the human writes part of the model. The engine pairs data driven inference with human expertise so that lenders can incorporate market conditions and emerging risk factors directly into it, and the company describes the approach as combining automated data evaluation with the wisdom of real people.

Because the model is expressed as interpretable rules rather than weights, a credit officer can read what it will do before it does it and change it, which is oversight before the decision rather than after. That also addresses drift, since a person can inject a market shift the training data has not yet seen.

Recordsure prohibits the decision, AdvisoryAI justifies the prohibition, Vouched bounds the grant, Saturn mandates review, Performativ applies the same permissions to agents as to people, and Stratyfy lets the human author the rules.

Interpretability by construction makes the model its own documentation, which is a stronger position than any post hoc explanation technique because there is no second system approximating what the first one did. Predictions of poor loan performance are stated to be explainable to customers, regulators and other stakeholders on the basis of clear rules, and a comparative study is published showing two metrics moving in opposite directions at once, more qualified applicants identified alongside a lower bad rate, which is the correct way to demonstrate a genuine improvement rather than a shift in risk appetite. The ability for lenders to write emerging risk factors into the model directly addresses the drift problem that degrades credit models between retraining cycles.

Ten million dollars was raised in a round co led by the venture arm of a large United States bank and an inclusive investing fund, with participation from a major core banking and payments provider, a community bank focused venture firm and a diversity focused fund, and the inclusive investing fund was investing for a second time. Backing from both a bank and a core provider is unusual and functions as distribution as much as capital.

A comparative study was published showing the engine identifying nearly twice as many pre qualified applicants than traditional methods used by small and midsize banks while lowering the overall bad rate. What holds this at B is that no institution is named as a customer and the study is the company's own rather than independent.

No data boundary statement was located. The architecture points toward containment without stating it, since the engine is configured per lender with that institution's own experts writing rules into the model, which implies a per customer configuration rather than a single shared model.

What is not addressed is whether performance data, learned rules or outcomes from one institution inform the engine deployed at another, and the buyer set of small and midsize banks includes many competing in the same local markets.

No data protection agreement, retention schedule, subprocessor list or deletion commitment was located. The payload is credit application and performance data on individual borrowers, processed to reach decisions that determine access to credit, and the fraud and compliance modules extend that to behavioural and screening data. Nothing published describes handling, retention or what happens to applicant data where a decision is declined.

No attestation, certification, trust centre or enumerated framework was located. Investment from a large bank's venture arm and from a core banking provider implies diligence has been conducted at an institutional standard, and none of that assurance is published, which is the obstacle a small bank's vendor management process will hit first when the product itself is otherwise designed around regulatory defensibility.

No supervisor, statute or instrument is named, though the positioning engages the regime more directly than most. Solutions are described as delivering automation without new operational or regulatory risks, and explanations are stated to serve regulators as one of three audiences alongside customers and internal stakeholders.

That is the adverse action and model explainability regime addressed by function without ever being cited, and for a product whose principal claim is regulatory defensibility, naming the rule its interpretability satisfies would be the natural next step.

Bias mitigation is a named product line here rather than a claim attached to one, sold alongside credit risk and fraud detection, and the founding argument is stated as plainly as anywhere in this index: transparency and controls are essential so that the biases of our past do not encode into our future. That is the direct counter to the institutional memory proposition other vendors sell.

The published study supports it in the right direction, showing nearly twice as many qualified applicants identified while the bad rate fell and the addressable market expanded by 70 percent, which is access widening with risk reduced rather than traded. Two investors are inclusion focused funds.

What holds this below Upstart's grade is the missing half: no outcome data by protected characteristic is published, so a company selling bias mitigation does not show the demographic evidence that would demonstrate it works.

The first A recorded on this axis, and it is earned by naming the party every other credit vendor omits. Predictions such as poor loan performance are stated to be clearly explainable to customers, regulators and other stakeholders on the basis of interpretable rules, and the customer is listed first.

Throughout this index the credit lane fails here for the same reason every time, that a declined applicant is not told why and has no route to challenge the inference, and interpretability by construction gives them exactly that: a decision expressed as rules a person can be shown, disputed against and corrected on. No commercial guarantee or indemnity accompanies it, and recourse to the subject is what this axis exists to reward.

The engine is the company's own proprietary technology, which shortens the chain at the decisive point and means no external foundation model sits inside a credit decision. Nothing else is disclosed. No data source is named for the credit, fraud or screening inputs the models consume, no bureau or alternative data provider appears, and no hosting arrangement or subprocessor list was located.

No named integration was located on either side. A major core banking and payments provider sits on the investor register, which implies a distribution relationship and possibly a technical one, and nothing published describes it. No loan origination system, core platform, bureau connection or developer documentation appears, so an institution cannot establish what deployment involves.

No hosting provider, region selection, residency commitment or private deployment option was located. Exposure is domestic and therefore simpler than for international vendors, and a bank outsourcing credit decisioning is nonetheless expected by its examiners to establish where applicant data is processed and held, and nothing published answers it.

No pricing, packaging or basis of charge was located. One cost argument is made and it is relevant to the buyer set: the platform is positioned as unlocking data driven decisioning without the need for in house data science expertise, which for a small or midsize bank is the largest cost of adopting machine learning at all. That describes what is avoided rather than what is charged.

Three functions are served by the same engine, spanning credit risk decisioning, fraud detection and compliance, which is unusual coverage for a single modelling approach and follows from the technique being general rather than task specific. The buyer focus is stated clearly as financial institutions, with small and midsize banks named as the comparison group in its own research, and the no data scientist required positioning points at institutions without internal modelling teams. The limits are geographic and scale: no international footprint is evidenced, and the largest institutions with their own model risk departments are not the target.