In this piece. The most common question we get from risk and credit leads is not whether the model works. It is what the signal actually is, and how they would explain that to a risk committee. This piece answers it directly: what the model reads, what happens between capture and score, whether the model that scores your borrowers is trained on your own data, why excluding protected attributes is only half a fairness answer, and how you can check any of it yourself. A long FAQ at the end covers the rest.
There is a question we brace for at every Vision Score session, and it almost never comes.
Nobody asks whether it works.
What risk and credit leads ask instead, in one form or another, is this: what is it actually reading, and how would I explain that to the people who have to sign off on it?
That is the harder question, and it is the better one. A lift number is worth very little if the person accountable for the decision cannot describe the model to a risk committee, to an internal auditor, or to a customer who has just been declined.
It is also a fair question, because almost nobody has looked at this signal before. When we poll risk teams on how they currently use the biometric capture their onboarding already collects, the answer is overwhelmingly the same: identity verification and liveness only. Capture, confirm, move on.
And when we ask how they would feel about using it as a credit input, the dominant answer is not rejection. It is a conditional. Open, if proven and compliant. That was the majority position again at our most recent session, The Full View: Vision Score in the Field.
Fair enough. So here is the proving part.
A loan application lands at 11pm. On paper, it checks out. Steady income, a clean record, every document in order. The model approves. Six months later the account is in collections, and nothing in the file saw it coming.
Every risk team has some version of this story, and its mirror image: the applicant who looked thin or unscoreable, got declined, and would have paid every installment on time. Both have the same problem wearing different clothes. The information you underwrite describes what the borrower did before they applied. It says very little about the application in front of you.
That gap is where Vision Score works.
Vision Score runs in five stages, and it is worth walking through them, because most of the difficult questions are really questions about one specific stage.
Here is the part most lenders have not clocked. That signal is already in your workflow.
Vision Score runs in five stages, and it is worth walking through them, because most of the difficult questions are really questions about one specific stage.
One, biometric signal capture. The signal is collected inside the onboarding workflow you already run, as part of standard identity verification. Nothing new is asked of the borrower.
Two, liveness verification. The capture is checked for deepfakes, masks, and spoofing before anything else happens.
Three, risk parameter extraction. Computer vision models read the risk-relevant properties of the capture. What gets read is deliberately narrow: non-sensitive properties such as image integrity, consistency, and resistance to spoofing, alongside onboarding cues like liveness, age estimation, and environmental context. These parameters are validated against real loan performance data.
Four, score calculation. This is the stage most people are actually asking about when they ask where the data comes from, and the answer is more specific than most expect. The score is trained and calibrated on your institution's data. The generic model is validated on more than a million real-world loan outcomes, but the model that scores your applicants is fitted to your borrowers, your products, and your definition of a bad account.
That matters for the question we get asked most often about provenance, which is whether the model is trained on Philippine data. The honest answer has two parts. Trusting Social operates across Vietnam, Indonesia, India, Singapore, and the Philippines, and the underlying research draws on that scale. But the model that ends up scoring your book is calibrated on your book, in your market. Locally, Trusting Social Philippines has been operating since 2019 with more than 50 institutional partners and over 110 million scored customers in the country, which is where the local validation depth comes from.
Five, decision integration. The score, 0 to 100, is returned into your lending workflow through a single API. Your decisioning logic decides what to do with it.
If you want the fuller treatment of why this signal exists in your workflow at all and why nobody has been reading it, we wrote that up separately in Your Model Reads Their Past. Here Is What It Looks Like to Read Their Present.
Here is the thing people believe before we say a word, and it is worth naming plainly because it is the real source of the hesitation.
Most people assume a model like this is judging appearance.
It is not. Vision Score is not based on physical appearance, and it is not trained to link any physical characteristic to creditworthiness. Protected attributes including race, ethnicity, skin color, and nationality are excluded from the model by design.
That sentence, on its own, is where a lot of vendor conversations stop. It should not, and if you are evaluating any model of this kind you should not let it.
Excluding an attribute from a model's inputs does not by itself guarantee a fair outcome. A model can still learn a proxy for something you removed. That is a well-understood problem in credit modelling generally, and it applies to a new signal exactly as it applies to an old one. Anyone who answers a fairness question purely by listing what they left out is answering half of it.
The other half is evidence. Vision Score is tested against the EEOC Four-Fifths Rule across race, gender, age, and skin tone, re-tested on every model update, and it has been through 30 independent adverse-impact checks, passing all 30 with zero violations.
The model's iterative development process. To understand the current version of the model, it is helpful to look at how we built it, as the development sequence is just as critical as the final outcome.
The model's development followed a rigorous, multi-stage architecture. We began by constructing an initial version and verifying its predictive power. Subsequently, we performed a deep-dive assessment for bias correlation, identifying any hidden links to protected attributes and stripping them from the logic. These steps were not a one-off; they were an iterative cycle we performed repeatedly to refine the current scoring engine.
Build, validate, assess, and exclude—this loop defines the model's provenance. The version scoring your borrowers today is the product of continuous refinement, not a static algorithm with a simple list of omitted fields. We only shipped the iteration that maintained its integrity and fairness through every subsequent test.
This approach reflects the transparency we would demand ourselves. It moves beyond the standard claim of "excluding the obvious" and replaces it with a proactive search for proxy bias against a rigorous external benchmark, backed by the resulting evidence.
The most reasonable response to everything above is that we would say all of it either way.
Which is why the part risk teams tend to appreciate most is not a claim at all. Because the signal lives in the capture rather than the moment, Vision Score can be read from onboarding images you already hold. That includes customers who applied years ago, whose loans have since run their course.
So a risk team can score borrowers whose repayment behaviour is already sitting in their own data, and see whether the score ranked them the way it says it will. Your customers, your outcomes, checked by you, before anything touches a live decision. If the ranking does not hold on your book, that is a real answer too.
A new credit signal earns its place by being explainable, not by being impressive. The risk officers we talk to have already worked that out, which is why they ask what the model reads long before they ask what it delivers.
The reason it matters beyond any single model is that a very large share of Filipino adults still have little or no credit history, which makes them invisible to a system built to read a record. Reaching them means using signals that have never been used this way before, and signals like that only get adopted if the people accountable for the decision can explain them to the people they answer to.
So we would rather over-explain than oversell.
Want the rest of the answers? Watch the on-demand session, The Full View: The Biometric Signal Your Model Is Missing, or read the deep dive on how the signal works. If your question is not covered in either, send it to us and we will answer it properly, the same way we have tried to here.