Skip to content
← Software
Proofmark: partial. The evidence exists and something in it is outstanding, named below.
mark
partial
written in
Zig
license
Apache-2.0
version
0.3.0
install
zig build -Doptimize=ReleaseFast

irregex

The-Billy-Company/irregex

The engine underneath: regex, trigram index, corpus walk, ranking.

The library the other three are built on. A regex stack with a linear engine and a vendored PCRE2 escalation path, a persisted trigram index, a corpus walker carrying a wall-clock freshness anchor, and the ranking that knows a definition from a call site.

It ships its own face too, for the two questions that need current bytes rather than a narrowing: provenance attributes a pasted snippet and re-verifies it against what is on disk right now, and blast walks a symbol's live radius with no precomputed graph.

Conformance here is measured against a denominator ripgrep owns: its own documented flag surface, and its own tests/ corpus mined into replayable records driven through both binaries on byte-identical fixtures.

411/411
of ripgrep's mined suite matchesLayer I, ripgrep conformance
99.45%
of 6000 fuzz triples byte-identicalLayer I, ripgrep conformance

What this does not prove

  • Thirteen of the 6000 randomized triples are still unresolved. Nothing in them is an abnormal exit, but a disagreement nobody has explained is a disagreement, and it is the reason this mark is partial rather than struck.
  • Fourteen of the suite’s 446 cases are NA rather than passing. The engine declines a construct outside its guaranteed-linear syntax and exits 2 pointing at -P, and it deliberately does not read two ignore sources. A case that never ran proves nothing either way.
  • The gate that judges this package’s structure, zoning, is public and runs in this repository’s CI, but its own first release is fresh enough that nobody outside has stress-tested it.
  • The Python wheel on PyPI is an alpha placeholder. Native binary wheels have not shipped.

Documents