Status: open (2026-10-20). Found by Julia Conformance part 2. Related: Julia Name Resolution, SPEC-006 Worlds, Artifacts and Evidence §2, Julia Multiple Dispatch and World Age.
Problem. Which function a Julia method extends depends on bindings in modules Sophia has not ingested. Base.Matrix(x::Tridiagonal) = … adds a constructor to Core.Array, because Base’s Matrix is an alias of Core’s Array{T,2}. Base.cbrt(A::AbstractMatrix) = … extends Base.Math.cbrt, which Base re-exports. Reading LinearAlgebra’s source cannot tell; the answer is in Base’s binding table. On LinearAlgebra this accounts for 86 of the 162 methods whose owner Sophia gets wrong.
The same holds for using (a name brought in by using M belongs to M’s binding, possibly re-exported from elsewhere) and for any package’s dependencies.
Options.
- Ingest Base and Core. Base is written in Julia, so possible in principle. But Base’s bindings are created by bootstrapping (
Coreis built in C), so the source does not give all of them either. - A binding table produced by Julia, stored as an environment artifact: for each module of the environment, each exported or public name and the module that owns the object (
parentmodule,Base.binding_module). Content-addressed, and cited by the facts that used it, like a world. Produced once per (Julia version, Manifest). - Leave it approximate and flag methods whose owner path is not a module Sophia ingested (“owner not verified”).
Proposal. Option 2, with option 3 until it exists. It fits SPEC-006: the binding table is part of the world a fact depends on, and it is exactly the kind of fact Sophia should get from Julia rather than infer (Semantic Boundary Experiments E-1).
Consequence. Generic-function entities for external modules (jl.Base.Matrix) would be renamed to the real owner (jl.Core.Array.new), so all packages extending the same function meet in one entity.