Editorial policy

1. Who runs this atlas

The atlas is an independent civic-technology project, built and maintained by an independent operator. The operator is not a clinician, not a researcher, and not a journalist. The atlas is software that makes ME/CFS research easier to find; it is not a medical service.

The atlas does not represent any research group, patient organisation, clinic, or public body. Correspondence: /contact.

2. What is machine-drafted

The per-study text on this site — plain-language summary, advanced summary, why it matters, what it does not prove, observed findings, inferred conclusions, remaining questions, methodological strengths and limitations — is generated by an LLM (currently Claude Haiku) from the paper’s title and abstract. It is not written by a clinician. It is not peer-reviewed. It is a structured re-reading of the abstract, produced by a program with a fixed prompt.

Every study page carries a permanent disclosure panel stating this fact. Treat each page as a reading aid, not as medical advice.

3. What is hand-corrected (and what is not)

The atlas does not routinely review study content by hand. No study on this site has been reviewed by a clinician, a domain expert, or a peer reviewer. The atlas makes no such claim.

A small number of studies — 93 of 6,145 public entries at last count — have had specific text re-generated under a tighter generator prompt, or had a specific problematic sentence corrected. Those changes are written to the study’s revision history as a technical change log. They are not expert editorial review. A machine re-summarisation is still machine output; a hand correction fixes one sentence, not the rest of the page.

An earlier version of the atlas carried areviewedStatuslabel with three tiers (machine / editor / expert). That field was bulk-applied to most of the public corpus before any genuine editorial review existed. It has been retired from every public-facing surface — pills, sort, crawl priority, metrics — and no replacement review label is shipped.

4. What is provisional

Everything on this site should be treated as provisional until proven otherwise. In particular:

• A single study is not evidence of a treatment or a mechanism.
• A finding in a long-COVID or comparator cohort does not automatically apply to ME/CFS.
• A preprint has not been peer-reviewed.
• A case report describes one patient.

Where a study page fails to make these caveats explicit, that is a bug. Please report it — see §7.

5. AI disclosure

The atlas uses large language models in the following places:

• Summarisation. Abstract → 9 structured fields (Claude Haiku).
• Classification. Abstract → provenance enums (Claude Haiku).

The atlas does NOT use LLMs to:

• invent findings that are not in the source paper;
• re-score or re-grade papers by opinion;
• generate clinical advice;
• answer patient questions.

The paradigm labels on the homepage paradigm charts (biomedical / neutral / psychosomatic) are output of the Classification step above — assigned per study from title and abstract by the same LLM, not by human review. They describe how a paper frames itself, not what kind of evidence it provides: a biomedically-relevant cohort or review can still land in “neutral” if the abstract takes no explicit paradigm stance. Misclassifications happen; if a label is visibly wrong, please report it (see §7).

6. What this atlas is not

• Not a clinical resource. Nothing here is medical advice.
• Not a systematic review. It does not weigh studies against each other or produce recommendations.
• Not a guideline. It does not tell clinicians what to do.
• Not a patient-support service. It cannot answer personal medical questions.
• Not a replacement for reading the primary study.

7. Reporting a mistake

If you see a study page that makes a claim the study does not support, or that misrepresents ME/CFS, please tell us via the contribute page. Confirmed problems lead to the study being removed from public listings, the summary being re-drafted, or the generator prompt being patched so the same error cannot happen again.

8. Active trust mechanisms

These are in effect as of this version of the policy:

• Generator v1 — tightened summariser prompt that prefers association language, hedges mechanism claims, and always bridges non-ME/CFS populations back to ME/CFS with an explicit qualifier.
• Scanner v1.4 — deterministic linter across 8 wording-risk patterns (causal leakage, efficacy leakage, population mismatch, preprint-settled, etc.), version-controlled.
• Lint gate in enforce mode — any draft with a CRITICAL wording flag is routed to a HOLD queue and does not auto-publish.
• Worst-first backfill — the historical studies with the most severe wording risk are re-summarised under v1 in rounds. Each round is backed up and reversible.
• Data-invariant checker — integrity checks that catch stale cached counters, orphaned studies after taxonomy changes, and drift between stored and derived counts.

9. How to read a single study page

Every study page surfaces: evidence level (E0–E3), research approach (review / trial / cohort / mechanism / etc.), PEM status in the case definition, publication type, and the explicit “What it does not prove” caveats. Together these framing signals tell the reader what kind of evidence they are looking at, before they read the summary text.

See also /process for the exact pipeline every study goes through.

10. Corpus at a glance

• 6,145 public study pages in the atlas today.
• 15 research topics.
• 93 studies with at least one machine-regenerated or hand-corrected field recorded in revision history (not expert review).
• Every public study carries a machine-draft disclosure panel directly beneath its title.

11. Governance

This policy is edited directly by the operator. Any substantive change is recorded in git history. If the atlas ever takes funding from a party with a stake in ME/CFS research or treatment, that relationship will be disclosed here and on every affected page. As of 2026-04-18 there is no external funding and no commercial relationship.