Designing an evidence-first AI catalog

An AI catalog becomes unreliable when the same project is copied into several categories and updated independently. This catalog uses one canonical record that can carry several types. An agent framework may also be a skill; an evaluation application may also publish benchmark evidence.

Every entry includes its sources and review date. Benchmark claims require methodology and result links, an evaluated target, a named metric, and an explicit limitation. The catalog records evidence. It does not turn incomparable results into an invented ranking.

Git review is the publishing system for the first release. Schema validation catches structural errors, while maintainers remain responsible for verifying meaning and provenance.