Gauss

Gauss is a scientific decision-support workstation. It helps a team compare candidates using evidence, objectives, assumptions, and uncertainty, but it is not a molecule editor, evidence retrieval system, simulation engine, or general-purpose conversational tool. Numerical conclusions come from the deterministic, versioned decision engine.

Write down the decision question and the intended use of the comparison. Specify what a candidate means, what outcome the team wants to improve, which constraints are mandatory, and which trade-offs are acceptable. If the question cannot be stated clearly, a score will make ambiguity look precise without resolving it.

Load the candidate records and measurements for the analysis. Inspect candidate identity, units, direction, missing values, uncertainty, and source provenance before running a method. Use a separate field for every measurement that represents a distinct scientific outcome; do not merge endpoints simply because they share a numerical range.

Check whether a larger or smaller value is preferable for each objective. Confirm unit conversions and any normalization method explicitly. Keep the input evidence table versioned so that a later analysis can explain whether its ranking changed because of new data, changed assumptions, or a different method.

Use the decision canvas to define each objective and why it belongs in the comparison. Record objective direction, units, weights, and any required thresholds. Distinguish a hard constraint, which eliminates an infeasible option, from a preference, which contributes to a trade-off. Document the stakeholder or scientific reason for a weight rather than treating it as an inherent property of the data.

If multiple teams have different priorities, save separate scenarios rather than negotiating a single hidden compromise. Compare how the rankings respond to plausible weight changes and preserve the scenario inputs with the result.

Run the deterministic engine only after checking that the candidates, measures, units, and criteria form a coherent comparison. Review the output table alongside the objective definitions and missing-evidence indicators. Open the method details to inspect the version, inputs, and provenance that generated the ranking.

A ranking is conditional on the data and assumptions supplied. It is not a validated safety claim, clinical recommendation, or proof that a candidate is superior in the real world. Use it to organize reasoning and identify where additional evidence could change the decision.

The Pareto view helps show candidates that are not dominated across all selected objectives. It makes competing objectives visible instead of compressing every trade-off into one score. Inspect the frontier with the underlying measurements and units, especially when a display rescaling could make a small difference look large.

Sensitivity views help determine whether a ranking remains stable when assumptions or weights change. Test values that stakeholders consider plausible, and record the rationale for the tested range. If a minor change reverses the preferred candidate, report the decision as sensitive and identify which new evidence or agreement would resolve it.

Gauss keeps method version, input version, seed where relevant, units, weights, and provenance with the analysis. Export the comparison together with those inputs and assumptions. A summary without the underlying context may be convenient for a presentation but is not sufficient for reproduction.

The bundled investigation is illustrative. Gauss supports research decision support, not clinical recommendation, validated safety, or formal GxP/21 CFR Part 11 claims. Optional language-model features may interpret selected context, but they cannot manufacture scores, sources, or scientific provenance.

Start with an investigation question that describes the choice and the authority that will make it. Define candidates as records, then create objective dimensions for the measurements that inform the choice. For each objective, record its measurement unit, whether larger or smaller values are preferred, any normalization settings, and the weight that expresses its relative contribution.

Separate hard feasibility constraints from preferences. A constraint determines whether an option meets a declared boundary; an objective expresses a trade-off within the feasible set. Inspect missing data and unknown values before computing. Gauss will not fill gaps with invented measurements, and a default candidate score should not be interpreted until the decision model's scope is understood.

Add candidate measurements with their units, source object versions, evidence class, context, and any shared-source grouping. A source may support multiple candidate rows, but those rows do not become independent evidence merely because they are stored separately. Inspect conflict records and source dependence before interpreting aggregate support.

When the same observation is reported through multiple publications or imports, preserve the common origin where known. Gauss distinguishes independent study groups from repeated claims linked to shared sources. If a source is contradictory, keep the conflict visible and inspect the underlying values and context instead of averaging away the disagreement.

The decision engine includes weighted and constraint-first approaches, evidence-weighted ranking, probabilistic analysis, Pareto comparison, robustness, sensitivity, dependency, and confidence views. Each method answers a different question. Weighted ranking compresses a configured trade-off into a score; constraint-first methods apply declared feasibility before ranking; evidence weighting accounts for source metadata; Pareto views expose nondominated trade-offs without collapsing them to one number.

Open the method detail before a run and review its version, seed, objective configuration, input manifest, and assumptions. The same inputs and deterministic method should reproduce a result, but reproducibility is not proof that the chosen decision model is scientifically appropriate. Compare methods as alternative lenses, explain why one is relevant, and keep rejected candidates and reason codes in the record.

Decision Studio shows the leading candidate, score and margin when available, evidence state, stability analysis, next-best-action status, score composition, objective weights, and candidate matrix. Inspect the retained candidate set and filters before interpreting the ranking. An excluded or screened-out option should remain distinguishable from an option that was evaluated and ranked last.

Open score composition to see which objectives contributed and how their weights affected the result. Review conflict counts, shared-source evidence, units, and evidence status. The application can label a record illustrative; those records are useful for exploring the interface but must not be presented as validated evidence or as a real operational recommendation.

The Pareto view displays candidates that are not dominated across the configured objectives. A Pareto front is a set of trade-offs, not a recommendation to choose one point. Compare candidate positions and objective directions, and confirm that constraint handling matches the decision model before discussing a preferred option.

Sensitivity analysis varies objective weights or parameters and shows how the ranking changes. Inspect the configured ranges, sampled points, warnings, and any single-objective edge cases. A stable point ranking under one scan does not prove robustness to every unmodeled assumption. Use the result to ask which judgment or measurement matters most, then record the answer as a review note or a new scenario branch.

Robustness analyses can evaluate candidate behavior across uncertainty settings and run as version-pinned background jobs. Track their job progress in the bottom console and open result-specific detail panels when complete. Confidence, uncertainty, and stability depend on the recorded distributional or scenario assumptions; never interpret a numeric range without reading those assumptions.

Formal expected value of perfect information (EVPI) and expected value of sample information (EVSI) analyses estimate how additional information might affect a configured decision model. Their usefulness depends on well-specified outcomes, uncertainty, costs, and sampling assumptions. Inspect the exact input manifest, method version, and assumptions. If the action inputs are incomplete, the application can report “not assessed” rather than fabricate a next experiment.

A scenario branch lets you explore changed objective weights, assumptions, evidence selections, or candidate subsets without modifying the base case. The scenario records its parent and context so a reviewer can identify the difference. Use the comparison view to review changed inputs and results before carrying a branch into a decision discussion.

Named context groups and version-pinned bookmarks can save an exact candidate, evidence, and filter selection. Rename or reapply a group when you want to return to a known analysis context. Archive controls should preserve history; changing a context group does not change the underlying scientific data or prove that the selected records are appropriate.

The decision graph displays persisted typed edges among scientific objects and is lazy-loaded so the main workspace remains responsive. The timeline and activity views help follow edits, scenario creation, analyses, reviews, and decisions. Inspect the object records behind a node or edge; graph presence indicates a recorded relation and not causation.

The scientific console accepts domain commands and does not provide shell access. The interface can call analysis, context, and project actions from menus, the command palette, or the console. Inspect a command's affected objects before accepting a mutation; numerical ranking remains the deterministic engine's responsibility, not an AI-generated substitute.

AI tools, when configured, receive a minimized context manifest built from selected project objects and attachments. Inspect which candidates, evidence, and source excerpts are visible to the analyst. Unselected records should not be assumed to be included. The AI can interpret context and invoke approved tools, but it cannot author numeric scores or fabricate provenance.

A consequential change remains pending until a human approves it. Check citations against their original source and treat AI assertions as interpretation rather than validated evidence. Attachments can be uploaded and anchored for review; confirm the file, page, and text location before relying on an excerpt. Sensitive context should be limited to the minimum authorized scope.

Keep ranking, recommendation, decision, override, and approval as distinct states. A reviewer can comment, challenge an input, or record an override with a reason. Independent review and approval should identify the exact analysis version and its context, not a mutable project label. The record retains human authority rather than automatically promoting the top-ranked row.

When evidence or weights change, generate a new analysis and review it against the prior result. Preserve both the original and changed states so a team can see what changed and why. Any operational policy—such as who may approve or which evidence state blocks an action—must be reviewed against the application's configured controls and organizational process.

Snapshots and exports retain method metadata, input versions, seed, units, and provenance needed to examine a decision later. Preserve the hash-linked audit events and scenario parentage when handing off. Use a report to explain the decision question, objectives, method, rejected candidates, uncertainty, conflicts, and human review—not just the top score.

The bundled GAUSS investigation is illustrative. Its values and outcomes demonstrate the workflow and should not be described as real scientific evidence. GAUSS is decision support rather than a molecule editor, simulation engine, or literature retrieval product; it does not provide clinical recommendations, validated safety claims, formal GxP controls, or 21 CFR Part 11 compliance.