Methodology
Every entry follows the same construction and grading process; two compound pages mean the same thing by the same words.
No single token may merge two orthogonal dimensions. Every claim carries what is claimed, and how well it is evidenced — never one number.
Evidence grade vs effect signal — why both
A large effect in a weak study and a small effect in a strong one are different findings. Averaging them into one score destroys the distinction, so the reference refuses to. Every indication carries two tokens: the grade for how the claim was studied, and the effect signal for what was observed.
An effect signal never renders without its grade. Large alone is a lie by omission; D · Large reads correctly as “a big effect, in cell culture.”
Mixed and No effect are states a single score cannot express. An archive that cannot record a negative finding is not an archive.
How grades are assigned today
References are typed by study design from PubMed itself — MeSH publication types and the Humans/Animals descriptors — for every citation carrying a PMID. References PubMed cannot settle fall back to a text heuristic and are labelled inferred wherever they appear, so the provenance of every label is visible. A narrative review never lifts a grade on its own — secondary summaries are not cited as fact.
Each compound carries the published rubric’s three axes, scored 0–3: preclinical depth (P) and human evidence (H) are computed from the typed references; regulatory standing (R) is always authored, with its authority recorded, never derived. A validator recomputes the axes from the references on every grading run and warns on any mismatch. Per-indication grades appear only where an indication’s own references are linked; everything else renders as ungradedrather than borrowing the compound’s strongest study.
The upstream corpus never tied individual indications to individual studies, so those links are made here: a conservative matcher joins an indication to a compound’s own citations only when they share distinctive domain terms, weighted by rarity, above a strict threshold. Every applied link — and the exact terms that earned it — is committed to the repository as an audit file, reviewable line by line. Indications that do not clear the bar stay honestly ungraded until an editorial pass links them by hand; roughly a third of indications currently carry linked, graded references.
The three dots
Compound cards carry three dots. They measure how complete an entry is— how much a reader can check — not a claim about the strength of the science. Dots never carry meaning alone; every card also states the entry’s depth in words.
A three-dot entry is a well-documented one. It is not a well-evidenced compound — that is what the grade is for.
Sourcing & citations
Every reference resolves to something a reader can open — a PubMed record, a regulatory document or a compendial database. A citation with no reachable source is not shown.
Structured content is compiled from published sources and cross-checked against primary databases where one exists. Where an upstream source disagrees with a primary database, the primary database wins and the override is recorded with the authority that justifies it.
Two examples from this build: two stored PubMed identifiers resolved to unrelated papers and were removed, and BPC-157’s published sequence disagreed with PubChem at residue 2 and was corrected to match.
How suppliers are vetted
A supplier appears on a monograph only where it publishes independent analytics a reader can check: a third-party Certificate of Analysis, dated and matched to the lot, with HPLC purity, mass-spec identity and the stated counterion. The card lists which of those criteria the supplier meets, and lists nothing about price, stock or suitability.
Where the publisher of this reference has an ownership or referral interest in a listed supplier, that is disclosed on the card itself — beside the recommendation, not in a footer. A listing never affects a compound’s grades, citations or recorded data, all of which are compiled before any supplier is considered. Compounds with no qualifying supplier show no card.
Corrections
Errors are fixed in place and logged with a dated note. Nothing is silently rewritten. Corrections that overrule an upstream source are held in version control so a re-import cannot reintroduce the original error.
Each dated correction is recorded in the public changelog — the BPC-157 sequence correction and the removal of two fabricated citations are logged there.