Large language models can dissect quantum mechanics, generate production-ready code, and distill research papers into crisp summaries. They can also hallucinate anywhere from 22% to 94% of the time on complex knowledge tasks depending on the model and domain, at least according to Stanford's AI Index 2026. Between those two capabilities lies what may be the most expensive unsolved problem in artificial intelligence.
A wave of startups and research labs now think they've identified a fix—not in scaling models larger or crafting better prompts, but in overhauling the provenance of training data itself. The core pitch is deceptively simple: verify every training sample before it ever touches a GPU.
Whether that approach can escape academic benchmarks and scale across the sprawling heterogeneity of real-world deployments is another question entirely.
When Confidence Meets Unreliability
The data tells two stories, and neither is particularly reassuring. Stanford's AI Index team measured hallucination rates across leading models using the HHEM benchmark for summarization tasks and found rates between 1.8% and 5.4%—manageable, perhaps, for many production use cases. Switch to the AA-Omniscience knowledge QA benchmark, which spans multiple domains, and those rates detonate: anywhere from 22% to 94%, depending on the model and subject matter.
That's not statistical noise. A March 2026 preprint posted to ResearchGate found that methodological choices—how you score abstentions, refusals, ambiguous responses—can swing measured hallucination rates by a factor of 3.5, though the findings remain subject to peer review. The benchmarks themselves remain inconsistent. But the underlying pattern holds: models are fluent. They're confident. And they're frequently wrong.
Vectara, which maintains one of the industry's most-cited hallucination leaderboards, tightened its HHEM benchmark criteria in November 2025. The result was predictably grim: higher measured rates, renewed alarm. Research examining Google's AI Overviews feature, published last spring, found troubling variance in claim fidelity and source quality, further exposing the fragility of current guardrails.
The conventional mitigations—retrieval-augmented generation, grounding with search, citation layers—help. They don't solve it. Multiple papers published in 2026 documented incremental improvements: MEGA-RAG showed over 40% hallucination reduction in public-health QA. ISFJ-RAG hit 10.39% on RAGEval, down 2.5 percentage points. Hyper-RAG used hypergraph-driven retrieval to squeeze out further gains. Even with these techniques, the floor remains uncomfortably high for compliance-sensitive or safety-critical applications.
A May 2026 Nature paper titled "Evaluating LLMs for accuracy incentivizes hallucinations" proposed that the problem might be structural. The optimization pressure toward accuracy, even with ideal data, may create inherent incentives for models to hallucinate rather than abstain—a hypothesis that requires further validation. If that thesis holds, the frontier labs have a training-data problem that model architecture alone won't fix.
Correctness by Construction
Enter PerfectBit, a San Francisco startup that emerged from Y Combinator's Summer 2026 batch with a provocative tagline: "Correct by construction AI training data."
The company—founded by Seiji Yamamoto, formerly of Meta's Superintelligence Labs, and Peter Vajda, the ex-Meta director who led the Emu and Movie Gen teams—is building something that resembles a data foundry more than a labeling service. Every sample PerfectBit generates is verified against what they call an oracle: a physics simulator, a scientific database, a formal proof system. Only then does it ship.
The concept borrows from formal methods in software engineering, where "correct by construction" means embedding mathematical guarantees during development rather than debugging after deployment.
"Models pre-trained on noisy web scrapes produce confident, fluent statements that are wrong," the company's launch post declares. "Every sample is generated against an oracle ... and verified before it ships."
The pitch targets LLMs broadly, but zeroes in on specific verticals: robotics, where physics simulators can ground spatial reasoning; and AI-for-science, where ground-truth databases and computational models already exist. PerfectBit's messaging emphasizes "verifier-grounded training data" and "train on reality, not noise." Human annotation, they argue, "does not scale to superintelligence."
They're not alone in this thinking. SRI International's NuSCI group published work in 2022 on "dehallucinating LLMs using formal methods guided iterative prompting" for autonomy-relevant tasks—self-monitoring loops paired with formal checking. A May 2026 SSRN working paper proposed a formal mathematical lens for hallucinations and demonstrated precision improvements via "anchoring" in arithmetic settings. The conceptual groundwork has been laid. The question now is commercialization, and whether the idea can survive contact with messy, production-scale reality.
RLVR: Rule-Based Rewards and the Verifier Revolution
The technical foundation for verifier-grounded data coalesced around a method called Reinforcement Learning from Verified Rewards—RLVR. The term originated with Allen AI's Tulu-3 model in November 2024, but the approach gained serious momentum in 2025 after DeepSeek-R1 documented dramatic reasoning improvements by training with rule-based, verifiable rewards on math and code tasks.
DeepSeek's Nature paper, published in 2025, showed that verifiable feedback loops—where you can deterministically check whether an answer is correct—consistently outperform learned preference rewards. The team also documented reward hacking: when models trained on learned preferences optimize for signals that don't align with ground truth, they drift. Rule-based rewards tied to formal verifiers don't have that problem. Or at least, they have different problems.
A 2026 monograph published at rlvrbook.com lays out the taxonomy: outcome rewards (did you get the right answer?), process rewards (did you follow valid reasoning steps?), and the verifiers that serve as oracles for both. Where deterministic or ground-truth verifiers exist—math proofs, code execution, structured QA with known answers—RLVR shows consistent gains over approaches that lack verifiability.
Recent papers from late 2025 and 2026 have pushed the frontier further: soft self-verified RL rewards, verifier-noise modeling, masked-and-reordered self-supervision for RLVR, reference-based verifiable rewards for open-ended generation. The academic activity suggests the paradigm has legs beyond narrow benchmarks, though whether it generalizes remains an open question.
PerfectBit's positioning nods to this lineage. "Reinforcement Learning from Verified Rewards (RLVR) was an early step," their YC profile notes. "We go further." The implication: RLVR addresses post-training; verifier-grounded pretraining data addresses the problem upstream, before bad patterns calcify.
Scarcity Economics and Market Timing
Timing matters. Market conditions favor verified approaches right now, and the data scarcity problem is real.
Fortune Business Insights valued the global AI training dataset market at $3.59 billion in 2025, projected to hit $4.44 billion in 2026 and $23.18 billion by 2034—a projected compound annual growth rate of 22.9%. A separate report from ResearchAndMarkets pegged the broader data collection and labeling market at $4.94 billion in 2025, climbing to $6.12 billion in 2026 and $22.71 billion by 2032.
Those growth curves intersect with a looming scarcity problem. Epoch AI's updated analyses, widely cited through 2024–2026, project high-quality text exhaustion somewhere in the 2026–2028 window. The original internet corpus that powered the first wave of LLM scaling is running dry, pushing labs toward licensing deals—Reddit and OpenAI in May 2024, News Corp and OpenAI that same month, Shutterstock expanding its AI partnerships in March 2026—and toward synthetic data generation.
But synthetic data introduces its own risks. Generate training samples from models trained on web scrapes, and you risk compounding existing biases and errors. A February 2026 report from Info-Tech Research Group highlighted persistent gaps in data quality and governance that undermine enterprise AI readiness. Grand View Research's March 2026 market analysis on data quality tools noted North America's dominance—over 40% revenue share in 2025—driven partly by regulatory and compliance pressures.
The scarcity of reliably verifiable, high-signal data is creating an opening for new entrants. If you can't scrape more internet and you can't trust purely synthetic generation, the logical next move is programmatic generation with rigorous controls. That's the niche PerfectBit and similar plays are targeting.
An Ecosystem Coalescing Around Quality
PerfectBit may be the most explicitly "formal methods" play in the market, but the broader data ecosystem is consolidating around themes of quality and verification. In January 2026, Handshake acquired Cleanlab, a startup that raised a $25 million Series A in October 2023 to build "confident learning" tools for detecting label issues in training data. The acquisition signals that strategic value is being placed on upstream data quality—and that customers are willing to pay for verifiable cleanliness.
Established players are evolving their offerings accordingly. Snorkel AI has long emphasized programmatic labeling and data-centric workflows. Labelbox acquired Upcraft in January 2026 to scale "the human expertise powering frontier AI," positioning itself as a "data factory" for top labs. Scale AI expanded its evaluation and alignment platform through 2025 and 2026. Prolific, a participant marketplace, published a case study in April 2026 on a 76,000-participant AI persuasion study, showcasing its capacity for large-scale behavioral data collection.
On the evaluation side, companies like Vectara (HHEM leaderboard, RAG services, HCMBench toolkit), Patronus AI (automated hallucination evaluators including multimodal capabilities), and Giskard (LLM red-teaming and RAG probes, with documentation updated as recently as May and June 2026) have built businesses around measuring and mitigating hallucinations. Their existence reflects enterprise demand for guardrails—but also a recognition that fixing hallucinations requires control over training data, not just post-hoc filtering.
Infrastructure enablers are emerging, too. NVIDIA's Apollo open models, announced in November 2025, provide scientific simulation capabilities that anchor physics-grounded AI. SRI's formal-methods flows for autonomous systems offer reference architectures. These aren't consumer products. They're the plumbing that makes verifier-grounded data generation feasible at scale, assuming the whole thing works.
Regulatory Pressure and Legal Uncertainty
Regulatory forces are accelerating the shift toward traceable, verifiable data—whether the industry is ready or not.
The EU's AI Act, with general-purpose AI obligations entering application in August 2025 and enforcement ramping through 2026, requires foundation model providers to publish training-data summaries, maintain copyright policies, and provide technical documentation. The European Commission released templates and guidelines in 2025. Compliance is no longer optional for anyone deploying models in the EU market.
In the U.S., Executive Order 14110 (issued in October 2023) and subsequent NIST AI Risk Management Framework resources—including the Generative AI Profile published in July 2024 and dual-use foundation model guidance updated in January 2025—emphasize supply-chain risk and evaluation rigor. While less prescriptive than the EU's approach, the framework pushes organizations toward auditable data provenance.
Ongoing litigation amplifies the pressure. The New York Times' cases against OpenAI, Microsoft, and Perplexity were still active as of mid-2026, with NYT executives expressing confidence in an Axios interview. Getty Images' case against Stability AI saw an appeal granted in December 2025 in the UK, with parallel proceedings continuing into 2026 in the U.S. These cases will shape training-data IP norms for years, one way or another.
"By-construction correctness" plus provenance could reduce both legal and compliance risk. If every sample comes from a licensed database or a physics simulator you control, you sidestep many of the copyright and fair-use questions currently being litigated. For regulated industries—healthcare, finance, defense—that certainty has economic value, assuming it can be delivered at scale.
The Limits of Verification
The enthusiasm around verified data has limits. Those limits matter.
Verifier-grounded approaches shine where oracles exist: math (proof checkers), code (compilers and test suites), physics (simulators), structured scientific knowledge (curated databases). DeepSeek-R1's success on mathematical reasoning and the RLVR literature's focus on math and coding benchmarks aren't coincidental. Those domains have clean ground truth. They have right answers you can check deterministically.
Open-ended generation, creative tasks, subjective judgment, nuanced social reasoning—these present harder problems. You can't run a physics sim to verify whether a product description is compelling or whether a chatbot response exhibits appropriate empathy. The May 2026 Nature paper proposed that even with perfect data, optimization dynamics can incentivize hallucination over abstention in tasks that lack formal verification. If that's true, verified data solves a subset of the problem, not the whole thing.
RAG architectures and citation-heavy systems offer partial solutions for knowledge-retrieval use cases, but they don't eliminate hallucinations. They relocate them. Research examining Google's AI Overviews, published last spring, documented variance in how well retrieved sources actually support generated claims. Retrieval quality and attribution fidelity remain active research problems, not solved ones.
PerfectBit, for its part, hasn't published case studies, datasets, or quantified impact metrics. The company's YC launch post states they're "talking to frontier labs" for pilot engagements, but no customers are disclosed. Pricing, dataset catalogs, the technical details of their verifier stacks—all remain undocumented. The website emphasizes philosophy and contact forms over demos.
That opacity is typical for an early-stage startup courting enterprise pilots. But it also reflects the immaturity of the verified-data category. The question isn't whether verifier-grounded data reduces hallucinations in specific domains—DeepSeek and the RLVR literature have already shown that it does. The question is whether it scales across the messy, heterogeneous reality of production LLM applications, and whether customers will pay a premium for data that's verifiably correct in some contexts but unavailable or inapplicable in others.
Where Reality Has Rules
The AI training dataset market is growing at a projected annual rate of nearly 23% because demand for models is outrunning supply of quality data. Compute spent on inference is now roughly two-thirds of total AI compute, according to Deloitte's 2026 predictions, which means production deployments are scaling faster than training runs. That deployment velocity intensifies the cost of hallucinations. Every wrong answer in a customer-facing chatbot or compliance application carries reputational and legal risk.
If PerfectBit and its conceptual peers can deliver on the promise of correct-by-construction data for high-stakes domains, they'll tap into a market desperate for reliability. The technical foundation exists—the RLVR literature demonstrates that much. The regulatory and litigation landscape favors provenance. The data scarcity problem creates urgency.
What remains to be seen is whether formal verification can escape the lab and scale to the breadth of tasks modern LLMs are asked to handle. For now, it works best where reality has rules—and perhaps that's exactly where it's needed most. The rest of the problem, the messy, subjective, open-ended parts that resist deterministic checking, will require different solutions. Or maybe just different expectations.
The hallucination crisis isn't going away. But at least someone's trying to build guardrails before the models confidently explain their way into the next expensive disaster.
