31 August 2026 · Benchmarks
A fifth benchmark, and the claim it would not support
The four benchmarks all built a program small enough to fit on one screen, which measures how quickly a toolchain starts rather than the number a developer waits on. The new build-scaling suite generates the same program at eight sizes, builds it through every toolchain on the host, and then executes what each one produced.
Executing the artifact is what makes it a benchmark rather than a stopwatch. The released Kotoba CLI emits an invalid WebAssembly module above 128 functions and exits successfully, so an unvalidated harness would have recorded its best numbers exactly where it had stopped working. Two further Kotoba lanes stop at declared bounds — a call-fuel budget and a function-count admission limit — which are the opposite result and had to be reported as such.
The suite was added to show build speed as a strength. It shows the released binary with the lowest cold-start build cost of any lane measured, and it shows that at large sizes build speed is not currently a Kotoba strength. Both halves are in the same table.
Read the build-scaling results
29 August 2026 · Benchmarks
Four benchmarks answer four different questions
Compiler startup measures one tiny source-to-artifact path. The developer-loop suite separates resolution, checking, clean and no-change builds, and first result across eleven toolchain paths. Native runtime measures already-built programs. A fourth suite compares strings, collections, allocation, I/O, concurrency, and a small real application across six representative runtime paths. Kotoba publishes them separately so one fast phase cannot be mistaken for universal speed.
Every emitted result is checked, but all speed rankings remain withheld because the recorded host-load gates failed.
Read the benchmark and inspect its evidence
28 August 2026 · Language design
No ambient authority is a language boundary
Kotoba programs do not begin with implicit filesystem, network, process, clock, model, or secret access. Source declares effects, admission intersects grants and policy, and the host binds only the resulting capabilities.
That design complements operating-system isolation; it does not replace the compiler, verifier, runtime, provider, key custody, or host policy in the trusted computing base.
See the computation boundary
