MEASUREMENT CONTRACT
Evidence before headlines
Fresh isolated processes, two warmup rounds, eleven retained alternating observations, exact output checks, and independently reproduced medians.
Read the reproducible benchmark plan ↗This is a dated, complete cross-language snapshot. Its measurement runs only when a specific comparison is needed. See the daily performance page for current status.
REPRODUCIBLE MEASUREMENTS
Small binaries. Light memory. Runtime work remains.
Explore exact Linux arm64 results for three numeric kernels. Every retained observation, toolchain boundary, and source revision remains linked to public evidence. Native now exceeds Pure Go on spectral-norm in the one-core measurements. The other two kernels remain slower. Compare the full results below; these three programs do not establish general language parity.
Included in this snapshot: assignment recovery and verified Array-header reuse in MIR, including individual loops. Safe loops reuse headers while preserving every element bounds check and evaluation order. All seven implementations are measured afresh at Native revision 8980b597. Open the optimization and evidence ↗
Regression recovered: one-core Native n-body is now 7.87414 seconds, down from the previous published 26.2869 seconds and below the pre-regression 10.2644 seconds. Fannkuch-redux is now 61.8302 seconds, down from 202.325 and below the pre-regression 77.1176 seconds. These are comparisons across publication dates; controlled local evidence separately identifies the assignment repair and subsequent improvements. Open the complete comparison ↗
MIR is still a partial migration: verified scalar leaves, selected Integer/Float Array reductions, and bounded Array-header facts support optimization above QBE. General control flow, allocation, and I/O lowering remain incomplete. Open current MIR scope ↗
FORMAL RUNTIME RESULT
Compare runtime medians
PROGRAM
CPU LANE
METRIC
Shorter bars are better · medians of 11 retained processes
Show exact runtime data
| Candidate | Wall | CPU | Peak RSS | Samples | Evidence |
|---|
FORMAL BUILD RESULT
Compare the identical TypeRB sources
These measurements cover exact programs, inputs, host, and toolchains. They do not measure persistent services, package ecosystems, application latency, or developer productivity, and they are not a composite language ranking. Later focused optimization results remain separate until another complete rerun replaces this snapshot.