Kenshiki Labs builds source-side evidence systems for consequential decisions — the kind a person will be asked, later, to defend. We started in credit because that is where the problem is most exposed. It is not where it ends.
Why we built this
AI rarely fails the way software fails
There is no crash, no red error state, no stack trace for the person relying on the answer. AI fails by sounding authoritative before anyone has established whether the authority is real. The public record bears this out: the MIT AI Risk Repository catalogs documented risks and real-world incidents across false information, fraud, and other harms.
That produced a conviction the rest of this company serves: anyone who has to rely on a consequential AI answer is owed the ability to inspect why it was allowed to leave. That is a duty, not a feature.
Most systems that judge people judge a photograph of a person, not the person. The error is cheap to make and concentrates on the people who can least absorb it — the applicant frozen at the worst eight months of a decade, the work history with a bounded gap, the citizen reduced to an irregularity. We decided that building a machine to judge people without the obligation to show why was unacceptable to us. So we moved authority out of the model entirely.
Where it comes from
Inspection, or construction
The conviction is not new to us. It comes from the quality tradition the founders trained in — the Deming and Toyota lineage that draws a hard line between inspecting a finished part and constructing quality into the process. Inspection is seductive because it is cheap and legible: measure the part at the end, accept or reject. But inspection cannot tell you why a part is out of spec — so it cannot tell a part that is bad from one that merely looks bad at the moment of measurement.
Most automated judgment is inspection applied to people. It takes one reading and mistakes the reading for the person. We build the other way: authority is constructed from evidence and policy before an answer is allowed to leave, and the final check merely confirms what the construction already built in.
“Build to be examined. State uncertainty honestly. Prove before claiming.”
What we believe
Convictions, and what they cost us
A value is only credible when it costs something. Ours do.
Authority lives outside the model — and we would rather emit less than pretend otherwise. A model can propose language. It cannot establish its own authority, invent its own provenance, or decide its own answer is safe to rely on. The cost: when the evidence is not there, we return nothing. We accept fewer answers as the price of defensible ones.
We govern the generation layer; we do not replace it. The model stays. We govern what it can rely on and what it is allowed to emit. The cost: we forgo the larger, simpler market of being one more model endpoint.
Every consequential answer is reduced to claims that can be checked. Not tone, not topic — claims, against evidence, before emission. The cost: this is slower and more expensive to build than emitting fluent prose unchecked.
We corroborate the shape of a decision without surveilling its contents. The cure for too little, frozen data is not a camera that never stops filming. Continuity can be established without recording the merchant, the purchase, or the neighborhood. The cost: we deliberately leave predictive lift on the table because to descend below that line would rebuild old exclusions with better tools.
Who it is for
The desk where a fluent mistake outlives the moment
We build for the desk where someone will be asked, later, to show why a decision was allowed — where a fluent mistake can move money, shape care, expose intelligence, or create legal and regulatory risk. That desk exists in more than one industry.
Regulated lenders and financial institutions that must defend an approval or decline under audit, litigation, or examination
Government and public-sector systems that face oversight and disclosure pressure
Healthcare and life-sciences teams that need evidence-backed recommendations
Defense and intelligence workflows where sourcing must survive review
Credit is our wedge, not our company. It is the most exposed, most regulated, and most fixable instance of the problem — which makes it the right place to start, and a poor place to stop.
How to evaluate us
Judge the record, not the prose
Our standard is not a sentiment. It is a specification: before the system acts, it must show what evidence was in scope, what claims were made, what held up, what failed, and why an answer was allowed to leave at all.
Kenshiki was founded by operators whose disciplines are not a list of accomplishments — they are where the conviction came from. Statistical process control, Good Laboratory Practice, continental-scale data infrastructure, and regulated transaction systems all share one rule: you answer for the output under review later.
Stephen learned the difference between inspecting quality and constructing it at the source: he began his career at Japan Management Association Consultants in the late 1980s, on statistical process control and quality function deployment in the lineage of Deming and Juran. The same discipline runs through the rest of his work — a Bell Labs researcher and a member of Microsoft’s Windows Base Team; the architect of the connected-vehicle platform behind Toyota’s Entune and inventor of its peer communications framework (US Patent App. 2011/0126014); and, as a chief data scientist, the builder of a continuous pattern-of-life pipeline resolving home, work, and visit locations for 120 million mobile subscribers. That last system is also why he is unwilling to build one without limits.
He holds an M.S. in computer science and mathematics from the Tokyo Institute of Technology, with a thesis in the cryptanalysis of block ciphers, and is bilingual in English and Japanese.
Larry owns the core of the platform, Kura and Kadai, and brings nearly four decades of building systems that have to answer for their output later. He began at Battelle delivering regulated systems for the US Air Force and Army under Good Laboratory Practice certification — work where provenance was not optional.
He spent seventeen years at a regulated consumer-data and verification business serving the mortgage and credit industry, rising to VP of Software Engineering, where he architected the API behind its primary transaction system, scaling it from 100,000 transactions per hour to roughly seven million at ~100ms response times, led its migration to REST, implemented OAuth 2.0, encrypted payloads, and SCIM 2.0, and served as the engineering organization’s security champion. He studied computer and information sciences at The Ohio State University.
The founders carry the build. Our advisors carry the outside read — a pressure-test on buyer readiness and the discipline of shipping, and the capital-markets and corporate-development path of scaling a regulated enterprise.
Eric brings the operator’s discipline of turning a sharp question into shipped results. He spent his formative years at McKinsey and BCG before moving from advising to building — as Chief Revenue Officer at Flashfood, he helped scale the company from 4 to more than 2,000 retail stores between 2019 and 2022. Through Tribe Ventures he now helps leaders build and ship new ventures from zero to one in the AI era, with a particular focus on cutting through vendor over-claiming to the judgment that only comes from having operated. For Kenshiki, that means a relentless pressure-test of the real question behind every decision — frame, validate, build, ship — and an outside read on buyer readiness in a market where credibility, not noise, wins.
Andy brings two complementary careers to the table. He spent a quarter-century as an investment banker in New York — at JPMorgan, UBS, and DLJ — advising on complex M&A and capital-markets transactions across technology, media, and telecom. He then moved from advising to operating, serving as Chief Commercial Officer of OUTFRONT Media (NYSE: OUT), where he led digital transformation, strategic partnerships, the technology platform, and corporate development for a publicly traded company. For Kenshiki, Andy’s perspective spans the two things a governance-infrastructure company has to get right at once — the capital-markets and corporate-development path to scale, and the operating discipline of running a large, regulated enterprise.
Why the names
A language where the distinction has a word
The names are Japanese. We know they cost us instant comprehension and easy search visibility. We keep them because a system that enforces the boundary between verified evidence and fluent assertion should be named in a language where that distinction already has a word.
Kenshiki (検識) — verification through inspection: examining something to establish its true nature. Not trust, not belief — inspection.
Kura (蔵) — a storehouse: the governed evidence store that keeps evidence intact and retrievable under policy.
Kadai (課題) — the formal problem that demands a rigorous, structured response: the governed inference engine that returns bounded answers from governed evidence.
