Your reference library of psi research. Open for inspection, from a single data point to the big picture.
The experimental literature of parapsychology is real, large, and effectively unreachable — scattered across journals and books no citation index covers well. The ESP-Nexus database holds those documents, nearly all of them in full text, so you can inspect a single result in a single study, or stand back and question an entire line of research at once — ganzfeld, presentiment, remote viewing — with a statement of how much of that literature is actually held — measured, never assumed, and where it has not been measured yet, the answer says so. That second view has not existed in this field before: the ability to look for signal in the noise.
Follow one claim to its page
Open a single study, a scientist’s record, or one citation, and keep going until you are looking at the sentence in the source document the claim came from.
Dean Radin — 119 documents held, 116 in full text, 1983–2026Open his record → Open this record or browse all scientists →Ask the whole library
Put a question to a whole line of research at once. Every figure in the answer traces to a study record, and the answer says how much of the literature it saw.
“Has the ganzfeld effect size changed over the last fifty years?”See the answer → Read this answer or ask your own question →How it works: the AI reads every document, and checks every claim against its source
Nothing here is summarized from memory or assembled from commentary about the field. The collected literature is held in full text; frontier AI models read it, and the accuracy of that reading is measured. What they produce is checked against the page it came from — mechanically, on every claim. A random sample of every batch is verified by hand, and decisive rows are re-checked by an independent model of a different family.
- It reads the papers themselves
- Documents are held in full text. The AI reads the paper, not an abstract or a citation record.
- It records the numbers as data
- 6,076 statistical results extracted so far, each marked as reported by the authors or derived here, so the two stay apart.
- A claim it cannot verify is dropped
- If the cited source cannot be shown to support a claim, the claim does not go on the page — it fails closed: not stored, not published, not counted toward coverage. Metadata pulled in from outside is checked against the record it is attached to before anything is stored.
- Its accuracy is measured
- Blind benchmark against an independent published meta-analytic database: 44 of 44 rows exact on trials and hits; on the separate question of derived effect size, mean absolute discrepancy 0.066. This is an initial benchmark inside verification that is still running, and publishing the measured error rate along with the results is planned work.
- Better models mean a better library
- The reading layer is not tied to one model. When a stronger one is released, the literature is read again with it, and the checks around it stay the same. The instrument improves; the standard of proof does not move.
The reference library is curated and held locally. It is not a search index pointed at the web, but a selected library of the field’s primary literature, held in full text, in one place, under one roof. That is what makes everything above possible: only when a whole literature is collated, sorted and analyzable together can it be asked questions that today cannot be answered at all — not because nobody wants the answer, but because the evidence has never been in one place to ask. The field’s primary literature is not absent from the world — it is absent from the machine-readable layer the world now reads through, and this library puts the studies themselves into that layer.
What you are looking at is the partial version — and the numbers below say exactly how partial. The system works end to end today: it reads, extracts, checks, and answers. But acquisition is far from finished. In the one domain where holdings have been measured against the published count, 31% of the known studies are still missing from the shelves — and eleven more domains await that measurement. Of the 9,001 documents acquired, 6% have been through structured extraction. What remains is structured extraction for most of the corpus, coverage measurement for the other domains, acquisition of what those measurements show missing, and recovery of the standard errors that pooling requires. Whether publishing this record changes what the machine-readable layer returns is itself designed as a preregistered test — one that can fail, and will be reported either way.
Counts current as of July 29, 2026.
The introduction to the field — the scientists, the methods, and the evidence — now has its own page: Scientists, Methods, and Evidence →
More from ESP-Nexus: