1. Scope and audience
The Lupus Atlas is educational research infrastructure for people living with lupus, caregivers, advocates, clinicians, and researchers. It does not provide diagnosis or individualized treatment advice. Material clinical claims require direct sources; only records reviewed by an appropriately qualified human may carry an expert-reviewed label. The current launch corpus is citation-audited, not clinician-reviewed.
2. Source hierarchy
Source choice follows the question. Guidelines and regulator documents anchor current care and labels. Systematic reviews and randomized trials anchor efficacy questions. Well-designed cohorts can address prognosis and uncommon harms. Human mechanistic studies, animal models, and in-vitro work inform biology without being presented as clinical proof.
3. Two-axis evidence grading
Every material claim is graded on both study design and directness. This prevents a sophisticated laboratory study from appearing equivalent to a measured patient benefit.
| Public label | Typical support | What it does not mean |
|---|---|---|
| Established care | Current high-quality guideline and/or regulatory support in a defined context | Best or safe for every individual |
| Supported clinical evidence | Replicated or strong patient-outcome evidence | Universal response or no uncertainty |
| Observational | Association in human populations | Causation or individual prediction |
| Human mechanism | Human biological evidence | Demonstrated clinical benefit |
| Preclinical | Animal, in-vitro, or model-system evidence | Safety or efficacy in people |
| Site hypothesis | Explicit synthesis proposed by this project | A finding, recommendation, or protocol |
4. The plain-language layer
The cause candidates and fix attempts pages restate the same audited evidence without jargon. They are a reading path, not a second corpus, and they operate under one rule: the simplification may not outrun the evidence.
- Every cause and fix resolves to the same source ledger as the technical records. The build fails if it does not.
- Every cause candidate must state the case against it, and every fix attempt must state what remains unknown. These are enforced by the validator, not by editorial habit.
- Candidates that have been weakened or ruled out stay published. Removing them would leave the map looking more settled than it is.
- Confidence rungs and outcome verdicts are labels in words, never colour alone.
- Where the strongest evidence for something is a conference presentation, it is described but not cited, and the page says so in place of a source list.
Plain language is not the same as reassurance. Where the honest answer is that nobody knows, that is the answer these pages give.
5. Agent-assisted workflow
Specialized agents may search defined evidence streams, extract structured study fields, search for contradictory results, propose connections, audit claim entailment, and translate reviewed material into plain language. Agent output is a draft or review signal—not publication authority.
- Scout: locate relevant new and foundational sources.
- Extract: capture population, design, comparison, endpoints, results, harms, follow-up, and limitations.
- Challenge: search for null results, failed trials, bias, alternative explanations, corrections, and retractions.
- Connect: propose graph edges with explicit inference labels.
- Audit: verify each claim against the cited source.
- Human review: approve, revise, reject, or request specialist review.
6. Connection rules
A graph edge is a claim. It must state direction, population or system, evidence design, directness, uncertainty, supporting sources, conflicting sources, and review date. Inferred edges are visually and structurally distinct from directly studied relationships.
7. Updates and corrections
Automated monitors can create diffs, never public conclusions. Material changes receive editorial review, a dated verification record, and—when needed—a visible correction entry. Silent edits are reserved for spelling, formatting, or broken links that do not alter meaning.
8. Review and conflicts
Reviewer identity, relevant expertise, date, and declared conflicts should accompany reviewed records. Kidney, pregnancy, pediatric, neuropsychiatric, and other specialist content should be routed to appropriately qualified reviewers.
9. Reproducible data exports
Versioned JSON and CSV snapshots expose claims, sources, trials, therapies, outcomes, hypotheses, cause candidates, fix attempts, and typed graph edges. Each file carries an as-of date, schema version, review state, and reuse caveat. Start with the checksum manifest or download the evidence ledger.