Skip to content
AI.info

The Pulse

Licklider launches SciGround verification layer for AI-assisted research

SciGround is designed to check research designs, data, analyses and claims outside the AI model. Its free public beta is limited to study-design registration, state checks and diagnostics in Claude Web.

Licklider launches SciGround verification layer for AI-assisted research

Researchers can use a separate verification layer to check the connection between study designs, data, analyses and claims while working with an AI assistant, Licklider says. The Tokyo company launched the product, SciGround, on November 13 for research involving large language models.

Source prnewswire.com

SciGround MCP is in public beta at version 0.1.0 and connects to assistants that support the Model Context Protocol. The free public beta described on Licklider’s website starts with registering and checking a planned two-group study in Claude Web; it requires Google sign-in and is limited to design registration, state checks and diagnostics.

Checks outside the model

Researchers keep their own AI assistant as the interface and reasoning partner. The assistant calls SciGround over MCP, with the verification work taking place outside the model.

According to Licklider, calculations and checks run under recorded software versions using tools including SciPy, NumPy, R, FLINT/Arb for guaranteed-precision arithmetic, and the Z3 solver. The product is designed to check whether data match the study design, whether numbers hold within a stated error bound and whether a claim follows from the result.

The test, comparison direction and threshold are fixed before a result is produced. For unsupported designs or data, the company says SciGround pauses or refuses rather than approximating, and does not switch methods after seeing statistical significance.

Research workflows

The launch list includes Student’s and Welch’s t-tests, paired t-tests matched by ID, one-way ANOVA with pre-specified corrected comparisons, and the Mann–Whitney U test. It also covers bulk RNA-seq differential expression with false-discovery-rate control, equipment-failure missing data, three fixed batches, an original-versus-replication comparison and a two-look interim design.

Several outcomes can be evaluated with a pre-specified correction. Each workflow has defined inputs and claims, and eligibility depends on the complete design and inputs, rather than the name of a familiar test alone.

For paired comparisons, the workflow matches measurements by ID rather than row order and calculates differences from the original values before the paired test. For several outcomes, Licklider’s workflow guide says every outcome remains in the report, whether significant or not.

Records and corrections

Each result returns as a machine-readable Record containing its claim, assumptions, limits, evidence, inputs and software versions. The intended record lets others rebuild the check independently.

When data are corrected, affected results and claims are rechecked, with evidence retained for each version, the company says. The workflow guide also describes retaining the type and precision of each source when comparing a published summary with individual data from a replication.

Licklider was founded in Tokyo in 2024. Its founder and CEO is Tasuku Kobayashi.

Sources

Related stories

Explore

More articles