Skip to content

MISWIRE-AUDIT-EXAMPLES.md

Latest commit

 

History

History
52 lines (42 loc) · 3.52 KB

MISWIRE-AUDIT-EXAMPLES.md

File metadata and controls

52 lines (42 loc) · 3.52 KB
Raw

Mis-Wire Audit — pre-seeded results on real repos

Real miswire-audit runs on public repos (no CodeGraph install — TSA models the name-only design from its own index). Reproduce on any tree: uvx --from "git+https://github.com/aimasteracc/tree-sitter-analyzer@develop" miswire-audit <path>.

Summary

Column definitions:

  • name-only worst case — every call whose name has a definition only in another language, including each language's own builtins (print, range, Ok, …). Absolute ceiling.
  • name-only genuine floor — same, but excluding each caller language's own builtins. The skeptic-resistant lower bound: a smarter name-only index that special-cased builtins would still get all of these wrong. This is the column reported by miswire-audit by default.
  • TSA mis-wires — edges in TSA's live index where caller and callee languages are incompatible. A real measurement, not a model.
repo languages call edges name-only worst case name-only genuine floor TSA mis-wires measured at
huggingface/tokenizers Rust+Py+JS+TS 16,329 — 1,259 (7.71%) 0 v1.21.0 (2026-06-07) — re-run gauntlet_runner.py --all before publication
astral-sh/ruff Rust+Py+TS 187,418 — 7,557 (4.03%) 0 v1.21.0 (2026-06-07) — re-run gauntlet_runner.py --all before publication
pola-rs/polars Rust+Py 267,066 — 9,016 (3.38%) 0 v1.21.0 (2026-06-07) — re-run gauntlet_runner.py --all before publication
tree-sitter-analyzer (this repo) 14 langs 116,672 3,946 (3.38%) 680 (0.58%) 1 measured 2026-06-10 at v1.22.0 (g6b2a266d)
gin-gonic/gin Go (single) 9,134 0 0 0 v1.21.0 (2026-06-07) — re-run gauntlet_runner.py --all before publication

Across 5 repos TSA resolves 0 or 1 cross-language mis-wires on the four polyglot ones (the 1 on its own repo is a single genuine collision — see live head-to-head in REPORT-v1.21.0.md for the CodeGraph comparison). A single-language repo correctly reports 0 and 0 — no false positives.

What the name-only counts include.

Worst case (name-only worst case column): every call whose name has a definition only in another language, including language builtins (print, range, Ok, …) that a smarter index could special-case.

Genuine floor (name-only genuine floor column, default reported by miswire-audit): same set but excluding each caller language's own builtins — leaving only genuine cross-language name collisions no naive index could easily skip. This is the column TSA leads with, as it survives a skeptic's "but you could just exclude builtins" objection.

For this repo (measured 2026-06-10 at v1.22.0): 3,946 worst-case but 680 genuine (sleep()→Java, connect()→Kotlin, Counter()→TS, find()→PHP, draw()→Kotlin). TSA resolves 1 of these — a single genuine collision on its own test-corpus files.

Sample offenders (a name-only resolver would make these; TSA does not)

huggingface/tokenizers — javascript tokenize() → rust def, rust format() → python def, javascript resolve() → rust def.

astral-sh/ruff — rust create() → python def, rust split() → python def, python range() → rust def.

pola-rs/polars — python iter() → rust def, rust f() → python def, python wait() → rust def.