Correctness

Wakaru tests both JavaScript behavior and syntax recovery against real toolchain output. The two measurements answer different questions: does the result preserve the expected behavior, and did Wakaru recover the original construct?

Test262 round trips

Test262 is the official ECMAScript conformance suite. The harness checks each test's expected result, including expected errors. For executable tests, it compares the original source, compiled or minified source, and Wakaru's output. Syntax-error tests check rejection at parsing.

Current state: 62,171 passing round trips, with zero Wakaru correctness failures across 66,729 runnable inputs. The remaining 4,668 are classified as unsupported or rejected, including test-environment and toolchain limitations. They are not counted as passes. The results and source revision are recorded in test262-stats.json.

CI reruns the tests and compares them against the baseline to catch regressions.

Pattern recovery

Reproduction matrices compile known inputs through real Babel, TypeScript, SWC, esbuild, rollup, sucrase, webpack, and Terser version combinations, then verify that Wakaru recovers the original construct. Current rate: 94.6% across 2,991 test shapes. Per-matrix results are recorded in stats.json.

What this promises, and what it does not

The measurements cover the constructs and toolchains in the matrices. That is strong evidence, not a formal equivalence proof. When behavioral fidelity matters most, use --level minimal, and treat decompiled output as something to verify rather than something verified.

If you want the exact rules, the full semantic contract lives in rewrite-assumptions.md: every named assumption, and which levels may rely on it.

And if you ever observe decompiled code behaving differently from its input, that is a bug we want. Open an issue with the input, reduced if you can.