From calculationto a supported claim.
Run an eligible analysis on approved inputs, check covered numerical and logical conditions, and return the claim the evidence supports. Assumptions, unresolved checks and unsupported requests stay visible.
Three questions behind the answer.
| Question | What is checked | What comes back |
|---|---|---|
| Is this the intended analysis? | Design, comparison and approved source inputs refer to the same research target. | The identified method, inputs and version. |
| Are the covered numbers established? | Numerical error bounds and decision conditions are checked. | An established check, a failed condition or an unresolved boundary. |
| Does the claim follow? | Direction, uncertainty and stated assumptions support the interpretation. | The supported claim and its recorded scope. |
Numerical error bounds, statistical uncertainty and external research assumptions are different. A verified calculation does not establish biological truth.
A useful answer includes the next step.
- 01
Established
Use the result within its stated target, conditions and checked scope.
- 02
Condition not met
Inspect the specific mismatch and any permitted correction.
- 03
Unresolved
Keep the uncertainty visible; review what information or precision is missing.
Unsupported requests and operational errors are reported separately. These are explanatory outcomes, not a live verification response.
Calculation, claim and stability checks.
Trusted libraries. Independent checks.
Standard methods run on established scientific libraries such as SciPy and NumPy, with versions and settings recorded for every run. An independent checker then re-verifies the covered results.
Numbers with error bounds.
Key numbers are computed with guaranteed error bounds using high-precision tools such as FLINT/Arb, a library for arithmetic with guaranteed error bounds. If a p-value is too close to its threshold to call, SciGround says so instead of rounding it into a verdict.
Claims that fit the evidence.
Each result comes with the claim it supports. SciGround rejects the classic leaps: reading non-significance as no effect, a mean difference as causation, a reversed group direction, or a quietly changed target.
Design to claim, one chain.
Design, data, analysis, and claim are checked as one connected chain with formal logical checks. A correct calculation on the wrong samples, or for a different comparison than the one planned, does not pass.
When it stops, it shows the way forward.
Every diagnosis comes in four parts: what was confirmed, what could not be and why, what you can still use and save, and the next safe step—or why it is right to stop. Where a fix is possible, SciGround suggests it with reasons, runs it within the permitted steps, rechecks, and returns a new version with a diff.
Would the conclusion survive losing a sample?
For small two-group t-test comparisons (Student or Welch), SciGround can prove that removing up to two samples does not change the decision, show a concrete removal that does change it, or explain why neither can be established.
The result and the conditions it carries
- The result with method, library version, and settings
- The supported claim and the assumptions it depends on
- A clear check outcome: established, condition not met, or unresolved—with unsupported requests and operational errors reported separately
- Optional sample-removal stability evidence for small two-group t-test comparisons