Another $40 million into AI evaluation. The headlines are predictable: a16z leads the A round for Vals AI, an evaluation tool startup. The narrative is seductive — reliable AI assessment is the missing piece for enterprise adoption. But if you’ve spent years dissecting smart contracts and ZK proofs, you see the pattern differently. Evaluation tools without cryptographic verification are just another layer of trust. And trust, in the world of code, is a bug waiting to surface.
Context: The Evaluation Gold Rush
The AI industry is shifting from training to deployment. In 2024-2025, the bottleneck moved from model performance to model governance. Enterprises need to know if a model is safe, accurate, and compliant before they let it touch customer data. Enter the evaluation tool layer: platforms that benchmark, test, and monitor AI outputs. The market is crowded — LangSmith, Galileo, Patronus AI, Arthur AI, Confident AI. Each pitches a slightly different flavor of the same promise: "We measure what matters."
Vals AI’s $40 million raise from a16z is a signal that capital believes this layer is infrastructure, not just a feature. But infrastructure for what? The current paradigm relies on centralized evaluation servers, proprietary datasets, and opaque scoring algorithms. The evaluator becomes the new oracle. And oracles, as we learned in DeFi, are single points of failure.
Core: The Verification Gap
Based on my audit experience with ZK circuits and smart contracts, I see a fundamental gap in the AI evaluation stack: verifiability. Current tools measure model outputs, but the measurement itself is not independently verifiable. If Vals AI runs a benchmark on your GPT-4o agent, you get a score. But can you prove that score was generated correctly? Can you replay the exact evaluation inputs? Can you verify that the evaluation model wasn’t biased or compromised?
In 2026, I built a prototype using a ZK-Circuit to verify that an AI model’s output was generated without tampering. The goal was to create a "verifiable inference" proof — a cryptographic attestation that the input data and model weights were authentic, and that the inference ran on a specific compute environment. The same principle applies to evaluation. Instead of trusting Vals AI’s backend, we could generate a ZK-proof that the evaluation was performed correctly, using a known dataset and a deterministic scoring function.
The math is clear: evaluation without verification is just opinion. Math doesn’t negotiate.
Vals AI’s technical positioning is in the evaluation tool layer, not the model layer. The article doesn’t disclose their methodology, but the industry standard is "LLM-as-Judge" — using a high-quality model (e.g., GPT-4o) to evaluate another model’s outputs. This creates a recursive trust problem: who evaluates the evaluator? If the judge model is hosted on a central server, the evaluation results are only as trustworthy as that server’s operator. A single bug in the evaluation pipeline can corrupt months of testing.
From the five sparse data points extracted from the original article, we know that Vals AI’s core narrative is "reliable AI evaluation tools." But reliability in the traditional sense means statistical accuracy. What we need is cryptographic reliability — the ability to independently verify the evaluation process. This is where blockchain-based verification can add value. By committing evaluation results on-chain, we create an immutable audit trail. By using ZK proofs, we can preserve privacy (the evaluation inputs and outputs can be proved without revealing the raw data).
Contrarian: The Blind Spots of Evaluation-as-a-Service
The contrarian angle is that AI evaluation tools, despite their promise, introduce new attack surfaces. First, there’s the risk of "evaluation hijacking" — an attacker could manipulate the evaluation inputs to produce a favorable score. Second, there’s the problem of benchmark overfitting: models are optimized to score well on specific tests, masking real-world weaknesses. Third, and most relevant to my background, is the conflict of interest. Vals AI is funded by a16z, which also invests in numerous AI model companies. Can an evaluation tool be truly independent when its investors have stakes in the models being evaluated? This is the same dynamic we saw in the crypto audit space — firms paid by the projects they audit. The result is often a rubber stamp, not a rigorous assessment.
In my 2021 post-mortem of the LUNA crash, I traced the exact integer overflow in the redemption oracle. The code was open, but the evaluation was flawed. The same thing can happen in AI evaluation: the metrics can be gamed, the test sets can be leaked, the evaluation pipeline can be biased. Code is law, but bugs are reality.
Moreover, the evaluation tool industry itself suffers from a "liquidity fragmentation" problem — not of capital, but of attention. Enterprises are faced with a dozen evaluation platforms, each with its own metrics, datasets, and scoring methods. The result is not scaling, but slicing — the same small user base spreads across multiple tools, diluting the ability to establish a universal standard. This mirrors the Layer2 fragmentation in blockchain: dozens of rollups, but the same users.
Takeaway: The Verifiable Future
I have a strong conviction that the next evolution of AI evaluation will be on-chain. Not because blockchain is a panacea, but because evaluation needs a decentralized verification layer. The winner in this space will not be the tool that produces the prettiest dashboard, but the one that provides a cryptographic proof of evaluation correctness. Privacy is a feature, not a bug — ZK proofs let enterprises share evaluation results without revealing proprietary data.
Vals AI’s $40 million raise is a bet on the importance of evaluation, but the real opportunity lies in making evaluation trustless. The question is: will Vals AI pivot to cryptographic verification, or will a new entrant eat their lunch by building on a ZK foundation? Based on my experience auditing institutional custody solutions, I’ve seen how quickly the disconnect between marketing claims and actual security can be exposed. The AI evaluation industry is at a similar inflection point.
The next time you read a benchmark score, ask yourself: can I verify it? If the answer is no, the score is just a number. Math doesn’t negotiate. And neither should your evaluation.
