Benchmarks
There are two questions worth measuring, and we keep them separate. Structural: what does Glyph reject by construction that TypeScript accepts? That's provable today, shown below. Reliability: do agents actually ship fewer defects in Glyph than in a well-configured strict-TypeScript setup? That's the number that decides adoption, and we don't have it yet — it's a study we're building, and we won't dress up a proxy as if it were that answer. (The real comparison isn't Glyph vs naive TS; it's Glyph vs strict TS + zod + eslint + ts-pattern.)
Not a statistic — paired demonstrations. Each is a mistake Glyph rejects and
tsc --strict accepts, reproducible from the
benchmarks directory.
Two mistakes agents make constantly — a missing match variant and
an unsafe cast — are rejected by Glyph and accepted by tsc --strict.
Wrong code that compiles is the expensive failure. Glyph won't let it.
Every symbol has exactly one declaration form — no overloads, decorators, or merged namespaces to scatter it.
A one-line Glyph edit produces a one-line TypeScript diff through the transpiler, and
glyph fmt adds no churn. The diff is the change, not the noise around it.
Safety usually costs length — Go and Rust, the other statically typed rows with
explicit error handling, are the longest. Glyph is the only one that is both safe and
lean: ~29% fewer tokens than TypeScript and ~40% fewer than Go, on real
cl100k_base counts. Honest caveat: on trivial one-liners Glyph's
module and import overhead can make it slightly larger than TS
— the lead only appears once there is real logic. Three functions, a small sample.
See the methodology and caveats →
The structural results above are reasons to expect Glyph helps; they are not proof
that it does. The proof is a differential study: the same agent, the same realistic tasks,
written once in Glyph and once in a fully-tooled strict-TypeScript setup
(tsconfig strict + typescript-eslint + ts-pattern + zod), scored on
escaped semantic defects and one-shot green-build rate. That's the number
five different reviewers told us would move them — so we're building it, and until it
exists these bars are a placeholder, not a claim.
The apparatus — a task corpus, a run/replay harness, and the pinned strongest-honest TS baseline — is in the benchmarks directory; the first real pilot (with its n, the model, and the cost stated on the label) publishes here the moment it runs. Whatever it shows.