Data behind the README results, measured September 7, 2026 UTC on an Apple M4 with 16 GiB RAM.
Download the run · SHA-256 checksum · Per-diagram timings
Extract the archive and open benchmark-run/index.html for the 130 cases, rendered outputs, raw samples, confidence intervals, exact commands, and build metadata. All 2,600 measured invocations and 390 warm-ups exited successfully. Visual review rejected Mermaid's TPMJS PNGs: both observed variants omit lower content, affecting all 20 measurements and 3 warm-ups for that case. Its aggregate is unavailable. The original observations are unchanged; review.json records the rejected hashes and reasons, with a diagnostic reproduction beside the retained outputs.
All PNG cases use 2× density. D2 output hashes were stable throughout. Background activity was uncontrolled.
Both D2 layouts use the same public build; TALA uses seeds 1,2,3. Local repository paths are replaced with ${REPO}; input and rendered bytes retain their original hashes.
To rerun with the pinned tools:
env -u GOMAXPROCS ./make.sh --seed 20260905 --output results/referenceSee setup for prerequisites and options.