Method and claim boundary

Compared routes

A standalone Arbace native x86-64 AOT executable, Arbace/WasmGC, and Clojure on HotSpot. The native route uses guarded static specialization. It is not Arbace's unfinished runtime JIT.

Inputs and estimator

Ten fixed, versioned workloads each use one canonical input. Every table cell is the median of five fresh-process medians. Each process contributes twelve hot samples. The statistical unit is the process, not sixty pooled samples.

Warmup and host

Every measured process completed at least ten seconds of directly timed warmup. Engine order rotated. Every process was pinned to logical CPU 8, and host load was logged for the full campaign.

HotSpot

The JVM flags were -Xms4g -Xmx4g -XX:+UseSerialGC. This fixes the heap and collector; it does not disable C1/C2. Other JVMs, flags, collectors, heaps, hosts, or CPUs are different experiments.

Correctness gates

Every process had to return the expected checksum, and the engines had to agree. Native hot-solve allocation counters had to remain zero. The exact frozen source passed 294 compiler tests and 210 runtime tests. All ten PGO artifacts passed their activation and verification gates.

INC101 causal experiment

The transducers-vector result specialization was tested separately as exact OFF/ON artifacts in six fresh process pairs. Both the sample-median and sample-mean tracks produced 6/6 strict wins, exact two-sided sign probability 1/32, paired median ratio above 1.10, and geometric mean ratio above 1.10.

Artifact identity

The build receipt binds the source inventory, toolchain, compiler, runtime archive, full typed IR, ten binaries, activation reports, verification reports, profiles, assembly, and cold/stress checks. The public receipt binds those records to the raw benchmark and test logs by SHA-256.

What follows — and what does not

Wave 19 shows that the exact native artifacts beat the two recorded comparison routes on all ten listed rows under this protocol. It does not establish performance for arbitrary Clojure code, changed inputs, guard misses, startup, compilation, another machine, or another JVM configuration. Nanosecond results are guarded whole-workload static specializations, not generic interpreter throughput.

Primary files

• release receipt
• raw result
• load log
• correctness receipt
• INC101 causal result