Four hand-written WAT modules. No compiler is involved, deliberately: the question is about the design in doc/design/0004-dispatch-design.md, and a compiler between the design and the number would only add doubt.

Predictions were recorded in doc/design/0002-measure-first.md before the first run. Do not read them until you have your own expectation.

Each runs on both node (V8: speculative inlining) and wasmtime (no adaptive tier). Both numbers are reported; neither is "the" answer.

B1–B3 are compared against JVM Clojure equivalents in bench/s0/jvm/, run on the same machine in the same session. A cross-machine comparison is not one.

bb bench-s0                            # everything, at the sizes the numbers were taken at
bb bench-s0 B1                         # one benchmark
bb bench-s0 --n 2000000 --reps 5       # a quicker, noisier pass while editing

Needs wasmtime, wasm-opt, node and wasm-tools — nix develop has them at the pinned versions. The driver prints the machine, the tool versions and the command to reproduce, which is what gets pasted into doc/design/0002-measure-first.md alongside the numbers.

Every lane's result is checked against the value the benchmark is supposed to compute, and a mismatch stops the run. The driver also refuses an n whose expected answer is the one an empty loop would return — for a ring walk that is any multiple of the ring length, and without the check a benchmark doing no work passes. Both guards were confirmed by breaking the benchmark on purpose.

Controls are not optional here. A benchmark and the variant you subtract from it usually differ in more than one way, and the difference is then not an attribution — see .claude/rules/measurement.md, and the B1 incident in doc/status.md that put the rule there.

bench plus whatever controls it needs, all with the signature (i32) -> i32 — take an iteration count, return a value that depends on every iteration. No imports, so the same file runs unchanged under node and wasmtime run --invoke. The JVM counterpart lives in jvm/, takes <variant> <n> <reps> <warmup>, and prints EDN.

wasmtime has no clock in the guest and no way to loop across invocations, so its lane is timed by process slope: wall time at n and at 2n, difference over n. Startup, compilation and instantiation are identical in both and cancel.

Read .claude/rules/measurement.md. The failure modes that matter — the optimizer deleting unconsumed work, process-spawn overhead swamping short runs, warm-up effects — all produce plausible-looking wrong numbers rather than errors.

All measured — B1–B4 as contracted, B5 and B7 which B1's and B2's findings forced, B6 (the component boundary, measured 2026-07-30), and B8 (the boxed-i64 lane, S3's numeric-representation benchmark). Numbers, controls and what each means are in doc/design/0002-measure-first.md; the S0 verdict is doc/design/0010-*, the B8 decision doc/design/0025-*.

B5 and B7 live inside b2_megamorphic.wat rather than their own files so that their type graph and rings are provably identical to the baselines they are compared against. That is deliberate: comparing across files is how three of this project's controls went wrong.