An Interview With Claims (SN111) on Scaling Scientific Data, Miner Competition, and Commercial Revenue

Editor’s Note: This interview was originally conducted by Bitstarter with Philipp Koellinger, co-founder and CEO of DeSci Labs. The TAO Daily is republishing the conversation, with the original interview credited to Bitstarter.

Philipp Koellinger, co-founder and CEO of DeSci Labs, discusses how Claims (SN111) uses Bittensor to turn scientific papers into usable data, its new mining competition, enterprise pricing, and plans to generate revenue in 2027.

In 2024, researchers published a study in Quantitative Science Studies examining the growing volume of scientific publications. They found that the number of articles indexed in Scopus and Web of Science had increased by approximately 47% between 2016 and 2022, significantly increasing the amount of research available for scientists to examine.

The challenge extends beyond keeping up with new publications. Scientific findings are largely communicated through lengthy papers, where important results, supporting evidence, and limitations are embedded in paragraphs, tables, and figures. Researchers and AI systems must identify that information before they can use it to answer questions or connect findings across studies.

This is the problem Claims (SN111) is trying to solve.

The Bittensor subnet turns scientific literature into structured records that connect individual research claims to the evidence supporting them. Miners compete to improve how accurately that information is extracted, while validators evaluate the results.

In an interview with Bitstarter, Philipp Koellinger, professor of economics at VU Amsterdam and co-founder and CEO of DeSci Labs, discussed how miner competition is improving the system, why Claims is changing its incentive mechanism, and how it plans to commercialize the resulting data.

The discussion comes ahead of a new mining competition scheduled for November 9, with the team targeting its first paid enterprise contracts in 2027.

When miners outperform the team’s own system

The clearest examples of Claims’ progress came from a paper published in Nature Medicine in February 2026.

The study examined the effects of dimethyltryptamine (DMT) on depression in a clinical trial involving 34 adults. It reported that a single 10-minute intravenous dose, administered with psychological support, produced a greater reduction in depressive symptoms than placebo after two weeks.

The paper contained approximately 9,500 words. Claims processed it into 14 structured research claims, each connected to supporting evidence from the original publication.

Koellinger explained that the team’s own extraction system initially identified only five claims. Ten competing miners found nine additional claims, supporting evidence the system had missed, and an important qualification: a second DMT dose administered two weeks later produced no statistically significant additional reduction in depression scores.

The example shows why extracting individual claims matters. A summary might capture the headline finding while overlooking qualifications that could change how researchers interpret the results.

“That’s why this runs on Bittensor: We’re rewarding miners to be better than us and everyone else they’re competing with, and it works!” Koellinger said.

He also expressed surprise at how quickly miners were improving the system.

In the current competition, miners run their own processing systems and submit the results. The Claims team evaluates those outputs without necessarily knowing how each miner produced them.

Koellinger described the experience of designing incentives and watching independent miners respond as particularly exciting.

“Not a lot of people get a chance to experience the power of incentives so directly and immediately,” he said.

Extracting evidence is not the same as verifying science

There is an important distinction between accurately extracting what a scientific paper reports and establishing whether its findings are scientifically reliable.

Claims currently focuses on the first task.

“At the moment, we’re just extracting accurately what the original paper reported,” Koellinger explained.

That means a record can faithfully represent a study’s findings without independently proving those findings are correct.

The team plans to introduce assessments of evidence strength, including whether researchers have made their data and code available for independent examination.

Koellinger also sees opportunities to use statistical information such as sample sizes, p-values, effect sizes, and research methods to estimate how likely particular findings are to be replicated.

For now, Claims relies on judges that independently evaluate miners’ submissions without knowing which miner produced each result.

The team has made these judges more critical over time. Koellinger said rising rates of unanimous acceptance, alongside manual spot-checks, suggest that submitted records are improving.

However, agreement among judges does not automatically establish accuracy. Claims plans to develop a further evaluation process using synthetic data with known answers, miner consensus results, and manual reviews to improve its judges.

Koellinger expects to build this reinforcement learning process before Christmas 2026.

A new competition designed to scale production by 10–100x

Claims is preparing a major change to how miners compete.

Under its initial system, every queried miner processed the same papers, and validators evaluated every submission. This helped the team test its mechanism but created substantial duplication of computational work.

The new competition, scheduled to launch on November 9, will require miners to privately submit complete data-processing pipelines.

Claims will evaluate the submissions and run the winning pipeline in production.

This changes the competition from repeatedly paying miners to process the same documents to rewarding the development of better systems for processing them.

The team has also introduced rules to discourage miners from flooding the competition with identical submissions.

Equivalent or nearly equivalent pipelines will be grouped, with one randomly selected from each group for evaluation. Miners can still submit multiple genuinely different approaches.

A challenger must also outperform the current winning pipeline by more than 2.13 standard errors to replace it. Otherwise, the incumbent remains in production and continues earning the reward.

The intention is to reward meaningful improvements while avoiding constant changes driven by small or uncertain performance differences.

Koellinger expects the redesign to increase Claims’ production capacity by 10–100x. The team is also becoming more selective about how much computing power it spends on each task.

“At the beginning, it was OK to throw a lot of tokens at everything,” he said.

Now, simpler cases can be processed using cheaper methods, reserving more expensive AI models for difficult tasks.

The goal is to increase output without compromising the quality requirements already established by the subnet.

How Zils (SN127) is helping Claims reduce costs

Part of Claims’ efficiency improvements comes from a collaboration with Zils (SN127), another Bittensor subnet developing specialized decision models.

Zils has trained a model for Claims that evaluates simpler cases at the beginning of its verification process.

According to Koellinger, the model handles these routine decisions at a small fraction of the cost of Claims’ existing judges, without reducing evaluation quality in the team’s tests.

More difficult cases continue through the original evaluation process.

As Claims produces more data, the Zils model can be retrained to improve its performance and take on a larger share of the workload.

Koellinger said this integration is one of the main reasons Claims can remove restrictions on how many miners it evaluates per round and support its planned production expansion.

The collaboration also illustrates how specialized Bittensor networks can contribute to one another’s operations.

“It just makes a lot of sense to use a smaller, cheaper, specialized model for many routine tasks and keep the expensive frontier LLMs for special cases,” he said.

Selling scientific data to AI developers and biotech companies

Claims plans to generate revenue through two commercial channels.

The first is metered access through an API and MCP server. Developers would be able to connect AI agents to Claims’ structured scientific records, build applications, or use the data to train models.

The second is enterprise contracts, particularly for pharmaceutical and biotechnology companies that need specialized scientific datasets.

Koellinger explained that data sovereignty is especially important for these customers. Companies working on confidential drug development projects may not want their research activities exposed through external AI services.

Claims intends to process the scientific literature relevant to each customer’s work and deliver a tailored dataset.

During setup, the team would also organize the information using specialized terminology and classification systems relevant to the customer’s research area, creating a knowledge graph that connects related scientific concepts, claims, and evidence.

New publications would then be incorporated through ongoing updates.

Koellinger compared this approach to the Palantir Foundry business model applied to science.

He outlined the following proposed enterprise pricing:

1) Initial setup: $10,000–$500,000, depending on customer requirements.

2) Monthly subscription: $1,000–$10,000 for ongoing data updates.

These are proposed prices, not revenue already being generated.

The team plans to develop a customer demonstration in the second half of November 2026, then return to companies it has previously contacted.

Its next commercial steps include hiring an experienced salesperson and conducting pilot projects with two or three companies in early 2027.

The pilots would be supported by letters of intent specifying the results required before customers commit to purchasing the product.

Claims is targeting its first purchase contracts and meaningful commercial revenue from the second quarter of 2027.

What the Bittensor community should watch

For Koellinger, the most useful measures of progress over the next 90 days are straightforward: the number of scientific papers processed, the quality of the resulting records, and how well AI agents perform when using Claims data compared with other sources.

These measures would help establish whether the subnet is producing information that is both reliable and useful.

“Quickly growing data at high quality and convincing improvements in agentic accuracy mean we’ve hit a gold mine,” he said.

The planned competition upgrade is particularly important because it tests whether Claims can improve production efficiency while maintaining its evaluation standards.

Commercial adoption will provide another test in 2027, when the team expects to move from demonstrations and pilot projects toward paid contracts.

Building beyond Bittensor’s existing AI infrastructure

Toward the end of the interview, Koellinger described Claims as an example of the kind of company he hopes Bittensor will attract in the future.

Much of the network’s development has centered on infrastructure such as computing resources, model training, and inference.

He believes the next opportunity is for founders to use those building blocks to create commercial data products and applications.

With infrastructure already available across the ecosystem and new economic mechanisms such as Gamma being developed, Koellinger sees an opportunity for more entrepreneurs from outside Bittensor to build businesses on the network.

“Picks and shovels will remain important and keep making money, but the really exciting and important stuff is what gets built with them,” he said.

Claims is attempting to demonstrate that possibility through scientific data.

Its miners have already shown that competition can uncover research information missed by the team’s own extraction system. The next challenge is producing that information at a much larger scale, maintaining its quality, and turning it into something researchers, developers, and biotechnology companies will pay to use.

The November 9 competition will test the production side of that ambition. The planned customer pilots in early 2027 will test the commercial side.

➛ More on Claims (SN111):

Inside Claims (SN111)’s Mission to Structure Scientific Knowledge