Citations & verification
Every claim in a BioSkepsis brief traces back to a specific passage in a specific paper. Citations are numbered, linked, and — where the agent could read the paper in full — checked against the source's complete text rather than its abstract alone.
#Every claim traces to a source
The agent does not write prose and attach references afterward. As it reasons, it grounds each assertion in the passages it actually read, and the finished brief carries that grounding forward as inline, numbered citations. A finding without a source behind it does not make it into the brief.
- Numbered citations. Each claim is tagged with one or more numbers that map to entries in the source list, so you can always see how many papers stand behind a statement and which ones.
- One click to the passage. Selecting a citation takes you from the claim in the brief to the exact passage the agent relied on, and onward to the original publication.
- Multi-source claims. When two or more independent papers support the same point, all of them are cited, so agreement across the literature is visible rather than hidden behind a single reference.
#Full-text verification
A citation in a BioSkepsis brief means more than “this paper was in the results.” The agent reads entire papers — methods, results, and discussion — not just abstracts, and it verifies each claim by matching it against that full text before the claim is allowed to stand.
#What “verified” means here
A citation is verified once the agent has located the supporting statement inside the source's full text and confirmed the claim is faithful to it. Abstracts routinely omit sample sizes, effect directions, model systems, and caveats, so a match against the abstract alone is a weaker guarantee than a match against the paper's body.
#Full-text-verified vs. abstract-only
The two are marked distinctly in every brief, because they carry different weight:
- Full-text-verified. The claim was matched against the complete paper — the level of grounding you want for any load-bearing conclusion.
- Abstract-only. The paper's full text was not available to read, so the claim rests on the abstract. It is cited honestly as such rather than presented as if the whole paper had been read.
#Reading and checking citations
Citations are designed to be checked, not merely counted. When a finding matters to your work, follow it back to the source.
- Inline markers. The numbered marker beside a claim links to its source(s). Follow it to see which papers were used.
- Source list. Each numbered entry carries the bibliographic detail you need to find the paper independently, plus a link to the original publication.
- Verification badge. The full-text-verified vs. abstract-only marker tells you, at a glance, how strong the grounding is for that specific citation.
A practical habit: for any surprising, causal, or decision-driving claim, open the source and read the surrounding context. The agent tells you what a paper reports; the paper tells you the conditions under which it reports it.
#Honest limits on hallucination
Grounding every claim in read sources sharply reduces fabrication, but no retrieval or verification system can honestly promise zero hallucinations. BioSkepsis is built to be candid about that boundary rather than to paper over it.
- Reasoning stays inside the evidence. The agent works from the papers it actually read, not from unsourced recall, which is what makes each claim checkable in the first place.
- Insufficient evidence is stated, not invented. When the literature does not support a confident answer, the agent says so — marking a sub-question as limited or not established — instead of manufacturing a citation to fill the gap.
- Verification narrows, but does not eliminate, error. A full-text match confirms a claim is faithful to what a paper says; it cannot confirm the paper itself is correct. Full-text verification is a strong signal, not a guarantee of truth.