How SciGround connects verification
Connect your AI via MCP, the standard way AI assistants connect to outside tools. SciGround verifies and manages its research outputs—turning your LLM into trusted research infrastructure.
From available evidence to supported claims
SciGround identifies what the available research materials establish, records the assumptions, and checks whether those conditions hold at the next stage. The workflow can begin with a design, data, or an existing analysis.
Your AI calls SciGround tools to run supported analyses and verification. It handles conversation, proposals, and explanations; defined verification procedures determine what is established under the recorded conditions. The AI cannot declare its own proposal verified.
More than executing an analysis
Python or R can calculate a result and can also implement checks. SciGround provides predefined verification checks, the link to recorded inputs and conditions, and evidence that another recipient can inspect. A successful program run or the AI saying that its answer is correct is not the verification outcome.
An existing-analysis check compares supported structured settings and reported results with the supplied source materials. It does not certify arbitrary generated code or every sentence in an AI response.
Design: What are you studying?
Define the groups, measurements, independent experimental units, and analysis plan. Save the details you confirm as the basis for the work that follows.
The data must correspond to the units, groups, and measurements in the design.
Data: Do these data match the question?
Match samples, groups, and values to the original files. When data are prepared or joined, preserve where each value came from and establish the input to analysis.
The analysis must use these inputs and meet the conditions of the chosen method.
Analysis: What follows under these conditions?
Check the method against the design and verify the calculation. For supported calculations, bound numerical error to establish whether the decision can be resolved.
The claim must refer to the same comparison, assumptions, and numerical evidence.
Claim: What can you say?
Connect the target, assumptions, and results to the claims they support. A result that does not establish a difference is not turned into a claim of equivalence.
What the mathematical checks establish
Targets and conditions are expressed in a form that numerical and logical checks can evaluate. SciGround combines established statistical libraries and verification techniques with the connections between research stages.
For supported calculations, numerical error is bounded rather than inferred from agreement between two rounded answers. The evidence must establish the rule needed by the claim; an unresolved boundary stays unresolved.
The target, assumptions, inputs, results, and claim must agree across stages. Matching file hashes or successfully rerunning code is not by itself a mathematical guarantee of this whole connection.
What remains a research assumption
Sample independence, reliable measurements, and the mapping from a natural-language question to a formal target require external evidence or research judgment. They remain identified as assumptions. A diagnostic test with no warning does not prove that these assumptions hold.
Statistical uncertainty remains part of the result. A verified calculation does not turn a p-value into the probability that a claim is true, and a nonsignificant test does not establish equivalence.
Read the outcome and its scope
- Established: the stated verification obligation holds for the recorded target and conditions.
- Condition not met: a specific mismatch or failed requirement was identified.
- Unresolved: the available information or numerical precision does not establish the obligation.
These are the outcomes of individual checks. The claim in a Record is then read as Supported, Unresolved or Not established, and whether a Record is current or superseded is shown separately. Unsupported inputs and operational errors are identified separately. An unresolved check is not evidence that the scientific hypothesis is false.
Keep guarantees tied to the right version
SciGround stores the research definition, original inputs, values, assumptions, and versions. When a permitted correction changes an input or condition, affected guarantees must be checked again before being applied to the new version.
A shared record fixes the target, evidence, assumptions, and checking method for one version. Recipients can inspect that scope separately from its applicability to the current study. A downloaded document does not update itself when the research changes.
Flat summary for LLMs
This llms.txt-style summary is intentionally repetitive. Use the linked Markdown page when a flat, low-markup representation is preferable.
- Work from available evidence toward supported claims; do not invent missing conditions.
- Check the meaning of each connection, not only the successful execution of each stage.
- Numerical error, statistical uncertainty, and external research assumptions are different.
- A change creates a new version and requires the applicable checks again.