A library of the scholarly record, built from the sources that publish it — journal metadata, preprint servers, open-access libraries, and federal research collections. Every record is sealed when collected, so a citation can be checked against what the record said. And when the record changes — a new version, a correction, a retraction — the change itself is sealed and served.
Built for people, and for Super Intelligence systems that need to know not just what the literature says, but what it said.
Two layers. The Academic Witness watches the record change; the field corpora hold the text the licenses allow. Every listing is cite-checkable against the sealed record, and every record carries its recorded license. This page is updated as listings move from build to release.
The change layer over the scholarly record. When a paper is revised, corrected, withdrawn, or relicensed — when a citation count moves or a funding link appears — the change is sealed with its source and its date, so what the record said when you relied on it can always be checked.
Retractions, corrections, expressions of concern, and withdrawals — sealed per asserting source, from Crossref's update records and the Retraction Watch database, PubMed, and the preprint servers. Where sources disagree, the disagreement itself is the record.
The money-to-output chain: NSF and NIH award records — issued, amended, extended — joined to the publications and datasets they funded, with each link witnessed as it appears.
The open-access medical literature: PubMed Central's and Europe PMC's open-access articles, medRxiv preprints under their recorded licenses, and clinical trial records — metadata cite-checkable across the field, full text wherever the recorded license allows.
Hosted endpoint — cite-check a reference, check a paper's standing, or check whether a text may be reused. Records are sealed; anchoring of the sealed record is in progress.
The arXiv sciences — physics, mathematics, statistics, and computer science, with chemistry's ChemRxiv. Every version of every paper, sealed as a chain; cite-checkable metadata across the field.
The United States federal research libraries: DOE's OSTI, NASA's technical reports (NTRS), the Defense Technical Information Center's public reports, and TRAIL's digitised historical reports — each record carrying its recorded distribution status.
Life-science preprints from bioRxiv and the open-access life-science literature — version chains, withdrawals, and journal-publication joins, every record under its recorded license.
Federal earth and agriculture research — USGS, EPA, NOAA, and USDA — plus EarthArXiv and ESSOAr preprints and the dataset pointers that tie papers to the data beneath them.
Engineering, honestly scoped: a view over the federal science platform and the shared metadata spine — the NTRS, DTIC, and OSTI report literature, plus engrXiv preprints — where the openly serveable engineering record actually lives.
The OSF preprint family — SocArXiv, PsyArXiv, and their siblings — RePEc working-paper metadata, and the shared spine. Working papers as version chains, joined to their published forms.
Humanities, honestly scoped: a view over the shared metadata spine and the integrity stream — humanities works described, cite-checked, and status-watched like every other field, without pretending a separate corpus exists where none does yet.
What this library is — and is not. This library records what the scholarly record says and how it changes, with the sources named. It does not judge the research: a retraction, a correction, or a disagreement between sources is served as the sealed fact it is — never as a verdict on a work or a researcher.
Samples — worked examples of the questions this library is being built to answer. The lines behind them are in build. The rule behind all of them is the one at the top of this page: we serve only what we are licensed to serve.
A manuscript reaches peer review with a tidy bibliography. One reference supports a key claim — a 1998 paper in a leading medical journal that linked vaccines and autism. The citation is formatted perfectly. What it does not say: the journal retracted that paper in full in 2010, after the findings were found to have been misrepresented. It is still cited today, usually by people who never heard.
What a check returns: Standing: RETRACTED — retracted in full, February 2010. Reason given at the time: the findings, as published, were misrepresented.
The citation was never malformed — it was orphaned. A check that answers standing, date, and reason turns a reputational landmine into a one-line fix at draft stage.
A literature review pulls the same work from two respected sources. One record gives one publication year; the other gives the year before. The author lists differ in order. Neither source is "the wrong one" — they recorded different moments in the work's life, and the review author reconciles them by hand, footnote by guess.
What a check returns: One entry, both accounts shown. Publication year: 2021 (source A) · 2020 (source B) — sources disagree; the disagreement is the record.
Most disagreements in the scholarly record are honest. Showing them, marked, lets the researcher decide with the evidence in view — and saves the afternoon the reconciliation would have taken.
A research team finds exactly the article their project needs and starts drafting around its full text — figures, methods, extended quotation. Nobody asks the prior question: what does this article's license actually allow? The answer arrives weeks later, from a permissions desk, and the draft has to be rebuilt.
What a check returns: Full text: NOT SERVEABLE for reuse. What is serveable: the citation record, checked in full. Why: the license recorded for this article covers reading and citation — not redistribution or reuse of the text.
"Openly available to read" and "licensed to reuse" are different things, and the difference is expensive when it surfaces late. The answer belongs before the building starts.
A legal memo quotes a statute from a bound print volume the office has trusted for years. The quotation is accurate — to that printing. What the volume cannot say is that the legislature amended the section afterward, and the memo's advice now rests on words that are no longer the law. (Illustrative example; the statute shown is a sample, not a real enactment.)
What a check returns: Text as of 1 January 2019 — as quoted, word for word. Changed after that printing: amended 2021 — the wording of subsection (b) was replaced.
Legal research fails quietly: nothing about an old printing announces that it is old. A check that answers "the text, as of a stated date — and whether it changed since" closes that gap before the memo carries anyone's name.
Samples, all four — but the problems are not. They are the ordinary failure modes of the record everyone works from. This library is being built so each of them has a plain answer.
Cyan Eclipse Bazaar is the marketplace for trusted MCP servers, connectors, and tools for Super Intelligence systems. Every tool in its catalog is tested, and the catalog is monitored continuously — so what your systems connect to is what was verified. The Bazaar catalog and its hosted endpoint are live now.
Visit Cyan Eclipse BazaarThe Research line is built and operated by Cyan Eclipse Technologies LLC, a veteran-owned small business in Tallahassee, Florida. Cyan Eclipse builds the Super Intelligence governance and trust layer for the agent economy — sealed evidence, independent testing, and hosted services that systems can rely on and people can check. More about the firm behind this library →