Claims (SN111) Puts Miners to Work Turning Scientific Papers Into Reusable Records Anyone Can Query

Scientific findings are trapped inside PDF files that AI systems and research teams cannot process at scale.

On Claims (Bittensor subnet 111), miners compete to convert those papers into structured records anyone can query.

Every extraction links back to its source, preserves the exact numbers and context, and stays open to correction over time. The output serves pharmaceutical teams, AI developers, and research institutions who frequently interact with scientific papers.

What Claims Is

Claims (SN111) transform scientific publications into a persistent evidence graph.

1. Each output is a claim-evidence record with full provenance attached: Statements pair with their exact source location, structured quantities, uncertainty intervals, and version history.

2. The system separates paper content from world truth: Claims record what a paper says without judging whether the underlying proposition holds true in reality.

3. Records stay reversible and open to correction over time: New evidence creates new versions, keeping the audit trail intact.

4. Customers span pharma, AI, review teams, and research analytics: The potential customers are every group that currently pays to reconstruct the same evidence from scratch across separate projects.

How It Runs on Bittensor

The subnet splits production work between miners and validators.

1. Miners run private extraction pipelines and submit structured records: They pick their own models, prompts, and verification routines while meeting the common output schema.

2. Validators construct tasks, adjudicate disputes, and calibrate scoring: They build the evaluation infrastructure deciding which submissions enter the final evidence layer.

3. Rewards flow only for genuinely new information reaching the final record: Multiple identities running the same method get grouped into one behavioral family for payment.

4. Independent corroboration adds bounded value on top: A second family confirming a finding earns extra, though third and later contributors receive no additional credit.

The Bronze-Silver-Gold Verification Tiers

Claims allocate verification effort across three tiers scaling from cheap and broad to expensive and selective.

1. Bronze delivers economical reference extraction on every paper: A common baseline pipeline runs across the entire corpus, giving miners a clear floor to beat.

2. Silver handles machine adjudication for disputed cases only: Stronger models get applied selectively where cheap extraction produces disagreement between miners.

3. Gold supplies sparse trusted verification through independent sources: Synthetic challenges, human review, and later-resolved cases anchor the scoring system to real ground truth.

4. Verified corrections flow backward to train cheaper layers: A Gold correction becomes training data for Silver, which produces more corrected labels, sharpening Bronze itself.

Verification runs from cheap toward expensive tiers, while learning runs the opposite direction back down over time.

Claims Versus Minos: Same Playbook, Different Truth Structure

Minos (SN107) and Claims (SN111) use Bittensor competition around scientific inference but handle ground truth very differently.

Dimension	Minos (SN107)	Claims (SN111)
Primary object	Variant-calling performance on genomic reads	Canonical claim-evidence data from scientific papers
Miner contribution	Tool configuration and custom algorithms	Structured extraction plus blinded judging
Ground truth	Hidden synthetic mutations with known labels	Partial: synthetic defects, curated subsets, and selective trusted review
Validator work	Execute configurations and score against truth	Construct baseline, adjudicate disputes, generate trusted verification
Reward logic	Winner-takes-all after score smoothing	Almost-winner-takes-all with family-level ranking
Identity defense	Concentrated prize plus blind tasks	Behavioral families, one family rank, bounded corroboration value
Compounding output	Synthetic genome database plus future consensus caller	Evidence graph, hard examples, calibrated judges, improved reference extraction

Both share the same Bittensor design DNA, running blind known-answer tests, private-method competition, and reusable data assets compounding over time.

A Permanent Evidence Layer Growing Under the Papers

Scientific literature keeps expanding while the human cost of extracting reusable facts stays fixed at expensive hours. Claims (SN111) turns that recurring cost into a permanent asset improving through miner competition, and not on any single team’s budget.

The customer base spans pharmaceutical, AI, review, and research organisations all facing the same document-to-data gap. What Claims delivers is scientific evidence extraction reshaped as infrastructure anyone can query without repeating work someone else already did.

➛ Read the Full Claims Paper Here.