# Research scope and input conditions > Find the supported research structure, required inputs, and boundaries for each SciGround workflow. Status: SciGround product guide Updated: October 15, 2026 ## Match the research structure first SciGround focuses on comparison studies in the life sciences. Use the target, experimental units, pairing, outcome, and comparison plan to identify the appropriate workflow. Each workflow has its own numerical input domain and assumptions. Exact limits belong to the versioned capability returned by your connection. The entries below do not imply that every method, multiplicity adjustment, and preparation step can be freely combined. ## Compare two group means Independent groups with continuous measurements, including Welch comparisons. One continuous outcome and one specified comparison. Keep the method, group direction, assumptions, and numerical support conditions explicit. Welch uses an approximation; verifying its numerical calculation does not remove that approximation. > Outside this workflow: Paired or clustered observations treated as independent, arbitrary covariate adjustment, or switching tests until a significant result appears. ## Compare before and after Two measurements on the same individuals or samples, with complete, explicit pairs. Two time points, one continuous outcome, and one pair per independent individual or sample. Keep the difference direction and pair identities fixed. > Outside this workflow: Missing-pair imputation, three or more time points, mixed models, or an automatic causal interpretation of a before–after difference. ## Compare three or more groups Independent groups defined by one factor, with comparisons specified in advance. One continuous outcome. Distinguish the overall group test from the specified group comparisons and their multiplicity adjustment. > Outside this workflow: Multiple factors, interactions, repeated measurements, or treating an overall significant result as evidence that every pair differs. ## Compare groups using ranks A specified question about ranks or distributions in two independent groups. One ordered outcome and one comparison. State the hypothesis, tie conditions, and exact or approximate calculation used. > Outside this workflow: All nonparametric methods, paired rank tests, or an unconditional interpretation as a difference in means or medians. ## Evaluate several measurements A predefined set of outcomes in the same two groups, with multiple comparisons accounted for. The same planned sample set, multiple continuous outcomes, and one comparison per outcome. Retain all outcomes and unresolved results in the declared testing set. > Outside this workflow: Selecting only significant outcomes for correction, silent removal of failed calculations, or arbitrary combinations with the other workflows. ## Analyze bulk RNA-seq expression Raw integer count tables for two independent groups, one comparison, and one batch. Gene-level unnormalized integer counts and sample information. Keep the filtering, normalization, estimation, testing, and correction steps tied to their respective gene sets and a fixed supported pipeline. > Outside this workflow: FASTQ alignment, single-cell analysis, multiple-batch adjustment, time series, pathway analysis, causal claims, or general guarantees for every RNA-seq pipeline. ## Account for measurements lost to equipment failure Keep missing samples visible while comparing two population means under a specified equipment-failure model. Two independent groups, one continuous outcome, one Student comparison, and one documented failure mechanism independent of the results. Preserve every planned sample ID, observed or missing status, and the failure reason. The normal model, common positive variance, and observation-selection conditions must hold; calculations use observed sample counts. > Outside this workflow: General missing-data handling, imputation, mortality, detection limits, result-dependent exclusions, causal effects, or combining this workflow with batch or sequential analyses. ## Compare treatment and control across three fixed batches One comparison giving equal weight to three observed batches, with both groups represented in each batch. One continuous outcome, one value per independent experimental unit, and a fixed six-cell model. Each group-by-batch cell contains 2–200 observations. Independent normal errors, common positive variance, and positive total residual variation are required. Keep batch identities, comparison direction, units, and value provenance explicit. > Outside this workflow: General mixed models, extrapolation to future batches, automatic time-series aggregation, multiple-batch RNA-seq, or selecting a model because it gives a significant result. ## Compare an original study with a replication Compare one original study and one replication while preserving where their measurements and summaries came from. Two groups per study, one outcome, one time point, and aligned units and comparison direction. Supply mean, SD, and sample count with extraction evidence for one study, and identified individual observations for the other. The result reports each effect, their difference, source types, and unresolved assumptions. Any approximate reference interval requires a separately supported profile and retains its approximation and extraction-error limits. > Outside this workflow: Automatic replication success or failure, equivalence claims, general meta-analysis, p-value output for this workflow, or treating figure-derived estimates as original observations. ## Plan two opportunities to continue or stop A fixed two-stage comparison with the decision rules and sample plan recorded before results are used. Two independent groups, one outcome and direction, complete planned observations, and two two-sided Student tests. Fix group sample counts, stage membership, and stopping rules in advance, using an overall level of 0.05 and 0.025 at each look. The supported model assumes constant group means and a common positive variance across groups and stages. Per-look intervals are 97.5%; the result preserves the decision and execution history. > Outside this workflow: Adaptive sample sizes, three or more looks, post-hoc method changes, statistical futility rules, clinical operation, combining with missingness or batches, or guarantees of sample savings. ## Explore, then confirm Explore a set of continuous outcomes with recorded summaries, plots and comparisons, lock one candidate before any confirmation data are opened, then test it on independent samples with one two-sided Student’s t-test. Exploration results are not issued as confirmatory evidence, and recording the exploration does not adjust for selection. Confirmation needs samples independent of the exploration, not a re-split of data already seen. > Outside this workflow: selective inference on the same data, simultaneous guarantees over many candidates, sequential confirmation, and repeated confirmations without adjustment. ## Optional check: sample-removal stability This is an optional check on two-group comparisons, not a separate workflow. For supported small independent two-group comparisons, SciGround can test whether the specified mathematical decision stays unchanged for every removal in a declared family. The bounded workflow starts with Student/Welch comparisons, at most 12 samples per group, and removal of at most two samples. Results distinguish a proof of stability, a verified removal that changes the decision, and an unresolved result. No counterexample is inferred merely because a proof could not be established. The certificate identifies the mathematical target. A claim about the behavior of a particular software build needs additional evidence linking that build to the target; numerical agreement alone is not that link. > This is separate from the statistical validity of a selected analysis. It does not cover arbitrary preprocessing, RNA-seq pipelines, or general post-selection inference. ## Working with incomplete history Existing data and analyses can support limited checks even when a full chain cannot be established. The report names the checked propositions and missing evidence. Uploading an old study today does not establish that its comparisons were specified before data collection. Historical planning and selection cannot be repaired by a new timestamp. ## Flat summary for LLMs - Each workflow has its own adopted conditions; support does not imply arbitrary combinations. - The RNA-seq route is bounded to independent two-group, one-comparison, single-batch bulk RNA-seq integer counts. - Equipment-failure missingness, fixed three-batch comparisons, original-study/replication comparisons, and two planned looks have separate conditions. - Sample-removal certificates concern a specified mathematical decision and removal family, not universal statistical validity.