How the library is built

New records are collected around the clock from scholarly indexes, publishers, and preprint servers: PubMed, Europe PMC, OpenAlex, Crossref, Semantic Scholar, Elsevier's Scopus, Springer Nature, the Directory of Open Access Journals, bioRxiv, medRxiv and arXiv, and PsyArXiv. Clinical trials come from ClinicalTrials.gov and are refreshed daily.

Every record is checked against the library before it is added, matched by DOI and then by title, so the same paper arriving from several sources is stored once. Duplicates that slip through, such as a preprint and its published version under different titles, are merged when they are found.

The library covers psychedelics and related compounds, altered states, meditation and other contemplative practice, dreaming, and the science and scholarship of consciousness. A keyword filter written in plain code settles most keep-or-drop decisions on its own.

Records the filter cannot settle go to a language model, which answers against one written rubric: is the paper's primary subject one of those areas? A passing mention is not enough, and a compound used only as an anesthetic or a cough medicine does not count. The model returns a verdict with a one-line reason, and the reason is stored. Records it rejects are not added.

When a paper already in the library is later judged off-topic, it is hidden, not deleted, so the verdict can be reviewed and reversed. A model error never hides a paper.

Titles, abstracts, authors, journals, publication dates, and DOIs are shown as the source published them, with only markup and formatting cleaned up. Every article links to the original publication. Citation counts come from OpenAlex.

Editorials, letters, and book chapters often have no abstract. For those, the abstract field holds the opening sentences of the work's own text. A paper with nothing to read at all is not shown.

Language models write two things you can read on this site, and one you cannot.

Study facts

When a paper is added, a language model reads its title and abstract and extracts its design, sample size, population, interventions, dose, duration, outcome measures, and key points. These fill the study-at-a-glance card on the article page, which sits beside the abstract so you can check one against the other. The model is told to report only what the abstract states, to keep the direction and size of the results, to state null and negative findings plainly, and never to invent a number.

State of the evidence

Each topic page, and each page pairing a compound with a condition or practice, opens with a synthesis written by a language model from a set of studies chosen by code: the 15 most-cited and the 10 most recent studies whose title or keywords name the subject. With fewer than 8 such studies, no synthesis is written.

The model sees only those abstracts. It is told not to use outside knowledge and not to invent numbers, and a check afterwards drops any citation to a study it was not given. The strength of the evidence is rated by code from the studies' designs, number, sample sizes, and agreement, not by the model. Retracted papers are never used. A synthesis is regenerated once it is more than 30 days old, and each one shows the date it was written.

No person reviews a synthesis before or after it is published. Read it as a starting point and follow its citations to the studies themselves. If one misstates a study, tell us and we will correct or regenerate it.

The same engine writes the monthly recaps and answers questions in the synthesis tool, under the same rules.

Behind the scenes

The same pass that extracts study facts also writes a short summary of each paper. It is never displayed; it only helps search rank results. Numerical embeddings of each paper power related reading, the map of the library, and the choice of studies for the synthesis tool.

We check the papers we index against Crossref's records of retractions, expressions of concern, and corrections, and count the comments each one has on PubPeer. Affected articles are marked on their own pages and listed together on the retractions page. Retracted papers stay readable, because hiding them would hide the record, but they are never used in a synthesis.

If a study fact, a synthesis, or a paper's details look wrong, email admin@consciousnesslibrary.org with a link to the page, or use the contact page. Corrections from readers are the fastest way this library gets better.

Why and how we use AI at all, and what we are working toward, is set out in our AI policy. For the technical detail, see how an article flows through the pipeline.