Our approach
Kenshiki does not replace the lender's decision. It asks bounded questions of approved source environments and returns a record a reviewer can inspect.
The decision, not the data
The hard part of credit is not collecting more data. It is making a decision that holds up — the same way, for the same applicant, no matter who reviews the file or when. Human judgment alone does not do that consistently, and a single score does not give a reviewer enough to decide well.
55%underwriting judgment varied by reviewer
In a study of underwriting judgment, the premiums different underwriters set independently for the same customers varied by 55% — the same person could be quoted very differently depending only on who caught the file. Inconsistency like this is a fairness and credibility problem before it is a pricing one.
Source: Kahneman, Sibony & Sunstein, Noise: A Flaw in Human Judgment (2021).
In a 2026 field experiment, loan decisions improved when the reasoning behind a recommendation was disclosed — reviewers could see when a rationale was internally inconsistent, and their overrides became more selective and more effective. The conclusion: a decision is stronger when the evidence can be opened and checked.
Most institutions already have AI governance on paper. Far fewer can prove it when it counts. In a benchmark survey of 500 financial-services leaders, only 21% were very confident they could produce complete, auditable evidence of how an AI-influenced decision was made: the exact record an examiner, auditor, or court asks for.
21%of financial-services leaders confident they could produce auditable evidence of an AI decisionSource: AAA-ICDR Institute, From Principles to Practice, 2026 (n=500).
Kenshiki is built for exactly this. We do not hand a reviewer a number to trust. We return a bounded, corroborated record they can inspect, explain, and replay — the same evidence, defined in advance, applied the same way every time. The reviewer still decides. They just decide on something that holds up to an examiner.
Consistent input. Inspectable reasoning. A decision you can explain.
Proof without possession
The lender needs a reviewable answer, not a new warehouse of raw private data. Kenshiki sends narrow questions to approved source boundaries and records only the bounded result.
Kenshiki sends the smallest possible evidence question.
The provider checks against consented data without sending a raw feed.
The answer can support, weaken, or contradict the file.
The reviewer sees what was checked and what Kenshiki did not take.
How corroboration works
Every check follows the same pattern: ask only what the credit decision needs, preserve the source boundary, and seal what came back into a reviewable record.
01
Each check asks whether the claimed person, place, obligation, and continuity hold together for the decision in front of the lender.
02
Providers compute on their side and return a bounded result. Kenshiki never ingests more detail than the question requires.
03
The lender opens a record of evidence, reason codes, gates, and the raw data Kenshiki deliberately did not take.
The coherence cascade
Each application traverses four nested tests before a coherence score is produced. A fail at any level raises a typed flag with the archived decision path; only a full pass yields the signed packet the reviewer opens.
Evidence architecture
Each axis tests one slice of whether the file and the life behind it cohere; uncertainty stays visible.
No single signal proves a life. Several bounded source patterns can show whether the story holds together, while validation measures how independent those patterns really are.
Gate
Consent-based SSA match/no-match verification before corroboration.
A match is necessary but not sufficient; it is not proof of a coherent life.
Score-back presence verification
Asks a provider-side area-and-time question without acquiring raw location data.
Carrier attestation
Checks whether line tenure and device signals cohere with the claimed identity.
Identity and consumer data graph
Checks whether address and residency signals support the claimed life pattern.
Payroll verification
Checks whether income evidence can be verified at an authoritative source.
Open banking
Checks whether account history supports the rhythm of the requested obligation.
Identity resolution
Tests whether identity elements have a seasoned, continuous history.
Behavioral biometrics
A candidate signal for human-interaction and device-consistency checks.
Kenshiki-built signal
The SDK records a behavioral-biometrics session during the application flow and returns a token.
Raw biometric data stays on device; Kenshiki receives a session token, not the typed content.
The customer API returns the already-sealed evaluation; it does not re-decide the file.
Comparing paths? See why many alternatives to credit scoring make decisions harder to explain.
Where-and-when presence, in detail
The decision may need to know whether presence fits the declared region. It does not need a street-corner trail.
Kenshiki sends the bounded question; the provider checks presence on its side. Raw location never enters our system.
Provider layer
Each provider defines what can be checked, where computation happens, and what evidence can be returned without pulling raw source data into Kenshiki.
The evidence layer
Kenshiki binds independent corroboration signals into the decision record. Named providers stay review-gated until the integration is real and public naming is approved.
Built on
Open telecom standard
Spatiotemporal presence · Carrier-attested telecom
Integrates with
Identity resolution and telecom data
Carrier-attested telecom · Identity-element trace
Integrates with
Identity, personal, and behavioral graph
Address & residency · Identity-element trace
Integrates with
Identity and consumer graph
Address & residency · Identity-element trace
Integrates with
Behavioral preference context
Address & residency
Integrates with
Identity resolution and data connectivity
Identity-element trace
We name a source only where the integration is real and the provider permits it. Where a relationship is in progress, we say so.