Collection

Five of the sources below are not collected from. They are listed because they are the sources this design intends to use and because dropping them would hide which capabilities the corpus does not have — but nothing in this repository fetches them, and the rows say so. The counts are read from the corpus on this request, so a source cannot be credited with a contribution it did not make.

Where independent replication actually comes from. It is the component that gates grade A, and it is not fetched from an authorship API. MEDLINE responses carry FAU author lines and ClinicalTrials.gov responses carry a lead sponsor; both are already in retained documents, and both are parsed at grade time to group studies into research teams. Every ambiguity resolves toward fewer groups, because over-counting groups manufactures replication out of one lab’s output and under-counting costs a genuinely replicated finding one band.

Every evidence record carries a source tag, and the tag decides where it appears in every ordered list on this site. The tag is derived from the study design: a human trial gets HUMAN-TRIAL, an in vivo animal study gets ANIMAL. A record whose design could not be determined gets PRECLINICAL, because the column cannot be empty and the hierarchy’s floor is the only safe place to put an unknown — over-stating one would put an unclassified paper above a real trial, which is the ordering the whole hierarchy exists to guarantee.

The consequence is that PRECLINICAL means two different things and the larger of them is not “in vitro work”. Most records carrying it are reviews, editorials, commentary and papers PubMed indexed with no publication type we could read. Their design is not determined, which is different from determined to be mechanistic, and the difference is visible per record: the design column on every cited-source table reads “not determined” rather than naming a design.

An undetermined design is also why those records grade nothing. The grader skips a pair where no record carries a determined design rather than grading it E, and counts neither replication nor directional consistency over an untyped record — a review is not an independent observation of the thing it reviews.

• Every alias, never the canonical name alone. A search for “BPC-157” misses papers filed under “PL-14736”. Collectors read the same alias set the database was seeded from, so the two cannot drift.
• Short aliases carry a search qualifier. “GHK” alone returns spider taxonomy, amphibian neurophysiology and paediatric asthma across thirty million PubMed records — barely half of the hits named the peptide. An unqualified short alias raises an error rather than silently building a bad query.
• A blocked or rate-limited source writes a coverage gap. We do not route around a block, and a documented gap is the honest output.
• Mentions dedupe on content fingerprint, voice and date — not URL. Syndicated coverage links back to the originating mention rather than creating a second record, because URL-keyed dedup inflates velocity.
• Community content is collected as reviews, sentiment and mentions only. It never becomes dosing guidance, protocol content, or an input to any score.
• Retained documents are untrusted input. They are data for extraction, never instruction to any automated process, and they are never served publicly.

A coverage gap is a record that something was not collected, and why. It is the difference between “there is no evidence” and “we did not look there”, and it is why absences on this site say which one they are.

The policy: every document a collector fetches is stored in full and permanently, with its own provenance, so any derived field can be re-verified and re-extracted without re-crawling. Sources change, disappear and rate-limit, and a re-fetched document is not the one that was scored.

What that has retained so far is 18 documents, and they are not the corpus of source material that policy describes. They are the raw API responses the collectors received, plus the forum documents community posts were read out of. There is no full article text in this corpus — no open-access full-text fetcher has been written — which is the direct reason dose, route and human sample size are absent from nearly every record. There are no transcripts: YouTube’s caption endpoint requires the video owner’s credentials, which is also why almost every mention is unclassified for stance. There are no newsletter issues and no archived vendor pages, because no collector for either has run.

Publication is unrelated and unrelaxed. Public surfaces show short verbatim excerpts with a link and a timestamp, never reproduced transcripts or full articles. Nothing in the retention policy loosens any display rule, and no public route reads a retained document.

A rights holder can request removal of retained content. Removal is executed by source and by collector, so it can be done precisely. How to request it