implementation experiment

Compares Sophia’s view of a Julia package with Julia’s own: for each method Sophia stores, does Julia say the same generic function owns it, and which of Julia’s methods does Sophia not see? Source: julia_conformance.rs. Results: Julia Conformance.

Sources: Julia Name Resolution; tests/julia-semantics/methods.jl (the ground truth, printed by Julia from its global method table). code: julia_conformance.rs, sophia_ingest::binding

julia crates/sophia-ingest/tests/julia-semantics/methods.jl LinearAlgebra > truth.tsv
cargo run --release -p sophia-ingest --example julia_conformance -- <stdlib>/LinearAlgebra truth.tsv
  • Julia’s side: Base.visit(Core.methodtable) (Julia ≥ 1.12 has one global method table); for each method from the package’s src/, its file, line and owner: parentmodule(f).name, or Module.T.new for a constructor.
  • Sophia’s side: the module-mode ingest and binding, the rule store_items uses.
  • Matching is by (file, line of the definition’s first token). Constructor naming differs by convention, so .new is ignored on both sides.
  • Output: methods with the same owner, a different owner (grouped by module pattern, with examples), or no Julia method at that line; and how many of Julia’s methods Sophia never sees.
  • --by-name reproduces the grouping before 2026-10-20 (bare name under the package), so the effect of the (module, name) key is measurable: 53.7 % → 94.5 % on LinearAlgebra.