Hypotheses Scaled. Verification Didn't. SciGround Debuts as a Mathematical Verification Layer for LLM-Driven Research

News provided by

Licklider, Inc.

Nov 13, 2026, 08:43 ET

SciGround keeps the researcher's AI assistant as the reasoning partner and moves verification outside the model, checking design, data, analysis and claim as one chain. SciGround MCP (Public Beta) connects any MCP-capable assistant

TOKYO, Nov. 13, 2026 /PRNewswire/ -- Licklider, Inc. today launched SciGround, a mathematical verification layer for research that uses large language models (LLMs). Researchers keep their own AI assistant as the reasoning partner and interface. SciGround, outside the model, checks whether design, data, analysis and claim agree. SciGround MCP (Public Beta, v0.1.0) connects it to any assistant that supports the Model Context Protocol.

Continue Reading
SciGround, launched today by Licklider, Inc., is a mathematical verification layer for LLM-driven research. It checks design, data, analysis and claim as one chain, outside the model. Image: Licklider, Inc.
SciGround, launched today by Licklider, Inc., is a mathematical verification layer for LLM-driven research. It checks design, data, analysis and claim as one chain, outside the model. Image: Licklider, Inc.

AI now produces hypotheses, analysis code and manuscript text at scale. The ability to check them has not kept pace. A July 2026 report from the White House Office of Science and Technology Policy notes that "the cost of generation has decreased exponentially, the cost of verification has not." A September editorial in Nature Computational Science observes that LLMs make reproducibility "arguably more challenging."

How SciGround works

  • Verification outside the model. The assistant calls SciGround over MCP. Calculations and checks run in established software, including SciPy, NumPy, R, FLINT/Arb (guaranteed-precision arithmetic) and the Z3 solver, under recorded versions.
  • No result before the design is fixed. The test, comparison direction and threshold are fixed in advance. If the design or data do not fit a supported workflow, SciGround refuses or pauses instead of approximating, and it never switches methods after seeing significance.
  • Design to claim, as one chain. It checks that the data match the design, that numbers hold within a stated error bound, and that the claim follows from the result.
  • Evidence others can check. Each result returns as a machine-readable Record with its claim, assumptions, limits, evidence, inputs and software versions, so others can rebuild the check independently. When data are corrected, affected results and claims are rechecked, with evidence kept per version.

Supported workflows at launch:

  • Student's and Welch's t-tests
  • Paired t-tests matched by ID
  • One-way ANOVA with pre-specified, corrected comparisons
  • The Mann–Whitney U test
  • Several outcomes with a pre-specified correction
  • Bulk RNA-seq differential expression with FDR control
  • Equipment-failure missing data
  • Three fixed batches
  • Original-versus-replication comparison
  • A two-look interim design

Each workflow has defined inputs and claims: https://licklider.ai/sciground/analyses/

"An LLM is an excellent partner to think with," said Tasuku Kobayashi, founder and CEO of Licklider. "But the proof, not the model, should speak for a conclusion. SciGround makes AI-assisted research something others can check."

About Licklider

Licklider, Inc. (Tokyo, founded 2024) builds SciGround, mathematical verification for AI-assisted research. https://licklider.ai

Media Contact
Tasuku Kobayashi
Founder & CEO, Licklider, Inc.
Phone: +81-80-7482-7670
Email: [email protected]
Website: https://licklider.ai

SOURCE Licklider, Inc.