Each campaign is a documented design effort: models used, generation / selection / filtering methods, and the generated → selected → tested → confirmed funnel. Funnel stages render only where counts are reported in the primary source; missing counts show Unknown and no success rate is invented.

Named development assets (preclinical → clinical) with an AI-linked origin. Each is placed at its verified stage. Clinical stages require a real trial record; the AI role is graded separately and is often weaker than the trial status — both are shown so the distinction stays visible. Discontinued and failed assets remain listed.

Families and descent: starting points, scaffolds, optimization series, and the assets molecules became. Trees are drawn from recorded relationships (developed_as, co_designed_with, scaffold_platform_for, methodological_precursor_to). Where descent is not documented, it is not drawn.

Which AI models produced which molecules, via which campaigns. Focus on a single model to see its downstream molecules and the campaigns that link them. Edges reflect recorded generated / used relationships only.

Academic labs, institutes, and companies, with the molecules, models, campaigns, and assets attributed to each in this registry.

The AI models and methods behind the molecules, with code / weights availability and the molecules and campaigns that used them. Availability is recorded as stated by the source, not assumed.

Assets with registered human trials. Trial IDs link to ClinicalTrials.gov. Trial status is a VERIFIED registry fact; the AI-design contribution is graded independently and may be DERIVED or weak — a high trial phase does not upgrade the AI attribution.

Negative and discontinued outcomes are first-class records. Distinguishing a molecule-level failure from a strategic/portfolio decision matters; each is typed and sourced, and stays visible rather than being quietly dropped.

Publicly reported collaborations, options, and licensing deals tied to these assets. Where only a company press release exists, the record is DERIVED and flagged — never treated as peer-reviewed fact.

Multi-stage design pipelines (chained models and steps) documented as a named workflow in a primary source, distinct from single-model campaigns.

How strong is the record, in aggregate? Distributions of evidence level (E0–E5), AI contribution (A0–A5), attribution strength, structural evidence, and provenance tags — all computed from the embedded data.

The registry auditing itself: field completeness, Unknown coverage, sourced-field counts, recorded conflicts, and the human-review queue. Gaps are shown honestly rather than hidden.

Select 2–4 molecules for a side-by-side field comparison. There is no universal winner: these targets, modalities, and evidence depths are not commensurable, so this is a comparison, not a ranking.