team1

team2

Live at the Louvre

There are 300 million scientific papers. Zero are structured as machine-readable data.

Frontier AI models hallucinate references more than half the time. Retrieval tools surface what authors said — not whether it's true, contradicted, or failed to replicate. R&D teams in pharma, biotech, and defence spend months and millions reconciling papers that shouldn't have passed peer review.

Claims is a Bittensor subnet that turns the scientific literature into a structured claim-evidence graph: every claim made, the evidence that supports or contradicts it, and a calibrated confidence score. The team behind it built 

 — an AI research assistant achieving 0% hallucinated references on a Nature-published benchmark. A first paid pilot with a drug-discovery startup is already underway.

Other tools retrieve papers. Claims tells you which findings are credible, where the field disagrees, where the hidden links and opportunities lie, and what that means for the question you're actually trying to answer.

$3 trillion is spent on R&D globally every year. The intelligence layer that should sit on top of that spend doesn't exist. Decisions are made on incomplete, sometimes fraudulent data - and the tools built to help haven't solved it, they've just made the surface area of the problem larger.

The dataset Claims produces can't be built inside a single company. Coverage needs to be vast, extraction quality extremely high, scoring needs to be adversarially robust, and quality needs to compound continuously. That's not a product problem. It's an incentive problem - which is exactly what Bittensor was built to solve.

Every surfaced disagreement sharpens the signal. Every new miner expands the corpus. Every linked ontology sharpens understanding. The resulting dataset is a critically missing part of the AI infrastructure. And Bittensor is the only place that allows building an optimal incentive system to create it.

 Miners extract source-grounded claims from the scientific literature across four dependency-gated channels: source coverage, atomic extraction, claim-graph construction, and evidence appraisal. Foundational work is cheap to verify - miners who fabricate or skip it earn nothing on higher-value tasks.

Score - validators verify against gold data. 

Validators grade submissions against a reference model and expert-curated gold data, both invisible to miners. Adversarial probes catch dropped negations, sign flips, and false concept merges - the errors that summaries hide.

Graph - contradictions are mapped, not buried.

 Extracted claims are linked into a graph: support, contradiction, and replication status across papers. A confidence score reflects what the field has verified, not just what one paper claimed.

Compound - the moat deepens automatically.

 A public scorecard tracks each validator's accuracy over time. Stakes follow the validators doing the best work. Every new claim enriches the training signal — and the signal grows every day.

The scientific record belongs to everyone. Right now, almost no one can use it efficiently.

The replication crisis, the hallucination problem, the months wasted chasing contradictory findings - these aren't failures of individual researchers. They're what happens when 300 million papers exist with no structure between them.

Claims builds that structure on a foundation that can't be gamed, captured, or shut down. Decentralised competition produces calibrated confidence. The network gets smarter as data grows - and the data grows every day.

Claims

sdfsdf

Compute became high-capex infrastructure before anyone could verify delivered work.

Delivered compute has no portable, independent grade - yet.

The industry is building compute capacity faster than the work it delivers can be independently checked. A run finishes - ml_training, inference, an HPC simulation - and there's no standard answer to a simple question: what actually arrived, and was it any good? The missing layer isn't a hardware inventory; plenty of those exist. It's an independent grade for the 

Shannon, built by Symmetric Research (SymRes), is a Bittensor subnet built around that grade. SymRes measures a real run and publishes a re-derivable receipt - the Effective_SCU - and the subnet turns auditing and forecasting that receipt into a competitive market.

Shannon gave information a portable unit in 1948 - the bit - and once information had a unit, prediction became an engineering target you could push down year after year. Compute hasn't had that move yet. Capacity is measured in GPUs and watts, but delivered work still has no unit that travels across providers. SymRes is built to give it one.

MLCommons already gives the industry a shared benchmark language through its MLPerf suites. SymRes adds the layer underneath: a runtime receipt for what a specific run actually delivered, carrying its workload context, a delivery ratio, and a telemetry hash anyone can recompute. The unit is rail-agnostic by construction - the same Effective_SCU receipt travels across decentralised networks, centralised cloud providers, and managed clusters alike.

For the Bittensor reader: this is a measurement surface that sits 

 existing subnets rather than competing with them. Nodexo proves hardware identity, ComputeHorde reports capacity, Chutes serves inference at scale - Shannon could add a shared, comparable grade for the quality of work delivered across any of them.

 The SymRes engine takes a real run's read-only telemetry plus open benchmark anchors and produces an Effective_SCU receipt - workload family, delivered quantity, and a re-derivable telemetry hash. SymRes publishes this; it never scores.

 Before a run, miners commit a P50/P90 band for the delivered Effective_SCU - a forward view of what compute will actually arrive. These are graded on calibration.

 After the run, miners re-derive the receipt - arithmetic, telemetry hash, boundaries - returning either "clean" or a specific discrepancy.

 Validators score forecasts and audits against the published receipt, and validator weights reward useful verification. Honest auditors earn weight; rubber-stampers and noise earn nothing.

Independent measurement is only credible from outside the loop it grades.

If SymRes earned emissions for the scores it produced, the score would become the target and the measurement would rot - Goodhart's law. So SymRes stays external: no token, no emissions for the measurement layer. A referee cannot also be a player.

Delivered compute needs a unit, and Bittensor's incentive layer can make verifying that unit a competitive market. 

SymRes

Proof of Pitch

 in exchange for 97,455 Alpha of their subnet.Pledgers can choose between 

In addition to the discounts listed below, all pledgers will receive yield (APY) on their pledged Alpha, accrued directly from the owner key validator, meaning yield is earned on top of the initial discount.

This raise sets the base price of 1 Alpha at 

Four teams entered. Two finalists remain. 
One subnet is up for grabs.

Following live pitches on stage, the judges have selected two teams to advance to the final round.

Pledge TAO to the team you believe should launch next on Bittensor. Every pledge moves the funding bar in real time as the finalists compete head-to-head to raise 1,000 TAO.

The winner will be announced live on stage at Proof of Pitch.

If neither team reaches the target, the team with the highest total when the crowdfund closes wins.

Not at the Louvre? Watch the livestream here:

Bitstarter helps ambitious teams launch on Bittensor by combining capital, incubation, partnerships, and community support.

Most startups raise capital behind closed doors. We wanted to try something different: letting founders pitch live and giving the ecosystem a direct role in deciding what launches next.

Learn more about bitstarter.ai and the teams we’ve launched:

Four pitches. Two finalists. One subnet.

Agenda