Rayan Krishnan and the problem with a perfect score

A free ice cream cone is a surprisingly practical instrument of climate policy. In 2017, while other people argued about international agreements, Rayan Krishnan was working on a transportation app for students at Tesla STEM High School in Redmond, Washington. Walk, cycle, catch the bus or share a ride, and you could earn points toward rewards. Coffee cards and cones from Molly Moon’s were among the proposed inducements. Saving the planet had acquired a loyalty program.

Krishnan was a sophomore and the technology director of Schools Under 2C, a student organization trying to reduce emissions. The students had a pledge, a launch kit and a carbon calculator. His app was in beta testing. The ambition was large; the mechanism was pleasingly ordinary. Make a different choice on the way to school. Record it. Give people a reason to repeat it.

That small project is an illuminating place to begin with the co-founder and CEO of Vals AI. His present business evaluates artificial intelligence on work people actually need done. The distance between a school commute and a software agent is considerable. Still, both projects require somebody to translate a grand objective into actions that can be observed and counted.

A city small enough to understand

Krishnan grew up in Washington, spending time outdoors in its parks, and came from Sammamish. His early projects gave that attachment a practical outlet. Alongside fellow students, he developed Operation Sustain, an educational simulation for children in third through fifth grade. Players built a city, chose energy sources, developed land and grew food. Their decisions affected pollution, public finances and citizens’ happiness.

The attraction of a simulation is that consequences arrive while the player is still paying attention. A real city can take years to discover the bill for a bad decision. A game has better manners. It shows you what happened, gives you another turn and lets you test a different idea. Operation Sustain made a sprawling subject small enough for a child to experiment with.

The team also did the less glamorous work that gets software into classrooms. They made teacher handbooks and curriculum supplements: four one-hour sessions, with activities, quizzes, worksheets and discussion questions. They worked with Washington Green Schools on academic standards. The project received a President’s Environmental Youth Award, and Krishnan later took its ideas to TEDxYouth@Seattle in October 2019. The classroom mattered as much as the code.

Two students, two models, no universal winner

At Stanford, Krishnan studied computer science with minors in mathematics and philosophy. His experience included Microsoft, Palantir and the Stanford Artificial Intelligence Laboratory. His university activities also included KZSU and Rewired, the student publication where he was listed as an editor. It is a background with room for programming, arguments and the occasional sentence intended for human beings.

An early collaboration with Langston Nashold, his eventual Vals co-founder, is particularly revealing. In a 2020 project, the pair tried to forecast computer processor usage in cloud infrastructure. They compared a neural-network approach, LSTM, with a statistical method, SARIMA, using Azure data. They looked at different prediction windows rather than declaring one technique the winner of everything.

The result refused to become a tidy advertisement for novelty. SARIMA performed better over the longer forecast window; LSTM did better over the shorter one and was more robust. A model’s usefulness depended on the question and the conditions. Years before Vals became a business, the two future founders had produced a compact demonstration of why a ranking needs context.

The exam leaves the classroom

Vals became publicly visible in a 2024 account of its attempt to independently test AI services. Krishnan and Nashold were working on a problem that would become more awkward as models improved: impressive results on familiar tests do not settle whether a system can finish unfamiliar work. Buyers have workflows, deadlines and bills. A leaderboard has a column of numbers. Connecting the two requires a great deal of engineering.

The Finance Agent Benchmark offers a concrete example. Krishnan co-authored the study with Antoine Bigeard, Nashold and Shirley Wu. The team assembled 537 expert-written questions across nine categories, with input from professionals in banking, hedge funds and private equity. Agents received tools including search and access to SEC filings. The test asked them to investigate financial questions using the documents where answers actually live.

In that study, the best-performing model reached 46.8 percent accuracy, at an average cost of $3.79 per question. Those are historical findings from that evaluation, rather than a verdict on today’s models. Their value lies in the specificity: a defined set of tasks, a measured result and an associated expense. An executive deciding what to deploy can ask considerably better questions with those three things on the table.

A working app is a demanding critic

For Krishnan, coding raised another version of the problem. A person who cannot program might still have a perfectly clear idea for an application. Could an AI build it from a written request? Vibe Code Bench, which he co-authored, tested the complete journey from specification to deployed web application. A functioning product offers fewer opportunities for bluffing than a handsome explanation of how one might be built.

The published benchmark contained 100 application specifications and 964 browser workflows, broken into 10,131 individual checks. Across the 16 models studied, the highest test accuracy was 61.8 percent. An automated browser agent interacted with deployed applications, checking whether the requested features worked. The paper also examined agreement between evaluators and human judgments. Even the examiner needed examining.

That last detail is easy to overlook. Building a test creates another system that can misunderstand instructions or make inconsistent judgments. A result becomes useful when the evaluator’s behavior is investigated too. For a founder selling measurement, the awkward questions reach all the way back to the measuring instrument. Publishing a score is the visible part; deciding what deserves a point is the real occupation.

The answer has to survive checking

Krishnan’s research collaborations extend beyond software. FormalProofBench asks models to produce mathematical proofs accepted by the Lean 4 proof checker. The problems cover advanced undergraduate and graduate mathematics, including analysis, algebra, probability and logic. A persuasive paragraph is insufficient. The proof must satisfy a system built to check its steps.

In the paper’s evaluation, the strongest foundation model solved 33.5 percent of the problems. The study also examined tool use, failure modes, cost and latency. Those measurements pull apart abilities that a single score can compress together: reaching an answer, reaching it efficiently and producing something that another system can verify. Mathematical correctness is an exacting customer.

A more recent collaboration takes that concern into legal research. Legal Research Bench, submitted on September 30, 2026, contains 413 expert-written American legal questions. Responses must satisfy every required criterion and cite authorities that verify. The strongest model in the reported experiment was fully correct on 42.9 percent of questions. Extra turns and higher inference costs did not predict better accuracy across the models.

For Krishnan’s project, this is consequential. A long, fluent answer can feel like hard work accomplished. Legal research adds an unforgiving condition: the supporting authority must exist, apply and remain valid. The benchmark makes completeness part of the grade. Looking busy has never been quite the same thing as finishing the assignment, although offices have historically been generous about confusing them.

The scorekeeper gets a bigger room

On August 13, 2026, Krishnan announced a $40 million Series A for Vals at a $400 million valuation. Andreessen Horowitz led the round, joined by returning investors 8VC, Pear VC and Bloomberg Beta, and new investors HRT Ventures and Next Ladder Ventures. The announcement also introduced the general availability of Vals Smith, a tool for creating custom coding benchmarks from GitHub repositories.

The investment made a particular kind of business visible: one that benefits from asking whether the celebrated new thing works. That carries obligations of its own. An evaluator needs access, technical competence and the ability to deliver an inconvenient conclusion. Customers must understand enough about the test to decide what its result means for them. Independence is a practice that has to survive commercial relationships.

By September, Krishnan was discussing these questions with Ben Horowitz, Jennifer Li and Erik Torenberg on the a16z podcast. The conversation covered benchmarks that age, evaluations that stretch over longer periods and the costs of using agents. The ambitions had expanded from comparing models to giving companies and policymakers a shared way to discuss what those models can do.

Who checks the checking?

At Vals’ first Public Sector Summit, held at STATION DC on September 23, 2026, Krishnan joined congressional staffers Will Burns and Dylan Irlbeck for a discussion about government’s role in AI evaluation. The questions included what evidence independent evaluators should obtain, how freely they should communicate findings and how their own work should be scrutinized.

In his opening remarks, Krishnan argued that credibility depends on the work’s quality and on being clear about a model’s limits. It is a useful standard for a founder whose business depends on other people accepting a judgment. The willingness to describe a limit is part of the product.

“Our credibility has to come from the quality of the work”Rayan Krishnan · September 2026

There is a family resemblance between this work and the projects he built as a teenager, though it would be too neat to pretend a school game contained a company’s entire future. Operation Sustain let children try choices and inspect consequences. Vals asks models to attempt work and makes their performance inspectable. In both cases, the interesting moment comes after an appealing idea meets a stubborn result.

The free cone has disappeared from the incentive scheme. The questions have become more expensive. Krishnan is still building ways to see what happens when someone, or something, actually tries.