Trust & safety

A conclusion is only as sound as the papers underneath it. Before any source is allowed to shape your brief, the agent screens it for reliability — quietly filtering out the papers the scientific record has already disowned and flagging the ones that have since been amended.

#Screening sources before they count

Retrieval casts a wide net. A single question can pull hundreds of candidate papers from the life-science literature, and not all of them still deserve to be trusted. Some have been formally retracted. Some come from journals that have been hijacked by predatory operators. Some carry corrections that change what the paper can be said to show. Left unchecked, any of these could quietly tilt an evidence base without you ever knowing.

So the agent runs a screening pass over the sources it gathers before they are permitted to influence analysis. This is not a setting you turn on or a filter you configure — it is a standing part of every research run. The intent is simple: the papers that reach your conclusions should be papers you would be comfortable citing yourself.

Screening resolves each source into one of three outcomes:

• Blocked — the source is disqualified and never counts toward a claim. Retracted papers and hijacked-journal sources fall here.
• Flagged — the source still counts, but it carries a visible marker so you know its status. Corrected papers fall here.
• Clear — no integrity signal was found, and the paper is weighed on its scientific merits like any other.

#Retracted papers

A retraction is the strongest signal the scientific record can send about a paper: the publisher or authors have formally withdrawn it, whether for misconduct, unreproducible results, image manipulation, or errors serious enough to invalidate the findings. A retracted result should carry no weight in a research conclusion — yet retracted papers are cited for years after withdrawal, often by authors who never learned the paper was pulled.

The agent checks the retraction status of every candidate source and blocks any paper that has been retracted. Blocked means blocked: a retracted study cannot support a claim, cannot appear as a citation in your brief, and cannot contribute to the Trust Index or the evidence balance for or against a finding.

This closes off the most common ways a retracted result sneaks into an analysis:

• It cannot be counted as supporting evidence for a hypothesis.
• It cannot be cited in a mechanism, method, or synthesis.
• It cannot be laundered through a downstream paper that cited it before the retraction — the agent evaluates each source on its own current status.

#Hijacked-journal sources

A hijacked journal is a fraudulent operation that impersonates a legitimate publication — cloning its name, ISSN, and sometimes its entire website — to sell publication slots and collect fees while performing little or no genuine peer review. To a keyword search, a paper in a hijacked journal can look indistinguishable from one in the real title it is impersonating. The metadata matches; the science underneath may be worthless.

The agent screens sources against known hijacked-journal signals and blocks any paper whose venue is a hijacked or cloned title. As with retractions, these sources are removed from the evidence base entirely rather than merely down-weighted — a paper that never underwent real review is not a matter of degree.

The point is to protect the evidence base from a class of sources that is specifically engineered to defeat surface-level checks. A researcher scanning a reference list has almost no way to tell a hijacked venue from the legitimate journal it mimics; the screening pass does that separation before the paper is ever counted.

#Flagged corrections

Not every integrity signal means a paper should be discarded. A correction — sometimes issued as an erratum or corrigendum — means the published record has been amended: a figure replaced, a statistic fixed, an author or method clarified. The paper still stands as evidence, but a reader deserves to know it was changed after publication, because the correction can bear directly on the exact result being cited.

Rather than block corrected papers, the agent flags them. The source still counts toward the analysis, but it carries a visible marker so you can see that the paper has been amended and judge whether the correction touches the claim you care about. Transparency, not exclusion, is the right response when the underlying finding may well remain valid.

The flag is there so you can:

• See at a glance that a cited paper was corrected rather than published clean.
• Open the source and check whether the correction affects the specific figure or number the brief relies on.
• Weigh a corrected-but-standing result with appropriate care instead of treating it as untouched.

#Where screening fits in a run

Integrity screening sits between search and analysis. The agent gathers candidate sources, screens them, and only the survivors flow into full-text reading, evidence weighing, and citation verification. By the time a claim in your brief points to a source, that source has already cleared retraction and hijacked-journal checks, and any correction on it is surfaced rather than buried.

The result is quiet but load-bearing: you can read a BioSkepsis brief without having to second-guess whether one of its citations was withdrawn last year or planted in a counterfeit journal. The screening pass has already asked those questions on every source, so your attention stays on the science instead of the provenance.