Benchmark figure QA

Evidence and interpretation

• Claim represented: matched single-call point-estimator runtimes for fastphylosig 0.3.0 versus phytools 2.5.2, shown separately for Blomberg’s K and Pagel’s lambda.
• Scope: test = FALSE; K is its point estimator, not the permutation test. The timing starts after tree/trait creation, matching, and preparation. One public call per method was timed with Sys.time; the 10 paired replicates per cell are summarized by the ratio of medians.
• Statistic: median phytools runtime / median fastphylosig runtime. Values above 1 favor fastphylosig; values below 1 remain visible and favor phytools.
• Coverage: 4 species counts (50, 100, 500, 2,000) × 3 signal scenarios × 2 metrics = 24 cells; 10 matched pairs per cell (240 pairs; 480 timed calls).
• This is a descriptive runtime figure only. It does not establish numerical-result equivalence or provide an inferential test/confidence interval.

Benchmark result summary

• K reference/fast ratio range: 0.1434–151.3968. The lowest cell is n=50 (no signal): the reference median is faster; the largest is n=2,000 (moderate signal), 151.4×.
• Lambda reference/fast ratio range: 6.4802–23,591.4444. The largest is n=2,000 (moderate signal), 23,591.4×.
• No cells were filtered for falling below 1. In particular, the K results at n=50 and n=100 show slower fastphylosig calls; the plot preserves this regression visibly.

Data and plot QA

• Matched-input audit: PASS (240 pairs; all 24 cells have 10 pairs; fast/reference species counts equal the target count; input hashes match; zero removed trait values; all timed-call statuses are ok).
• Data grid and ratio guard: PASS (all 24 cells present; finite positive ratios; values below 1 retained).
• Static source validator: 17 PASS, 4 WARN, 0 FAIL. Warnings: generic static syntax check asks for backend parsing (R parse() passed); no TIFF export; raster is 300 dpi rather than the 600-dpi default; static width detector does not infer the explicit 183-mm setting. This is a GitHub preview package, not a journal-raster submission package.
• R rendering and output: PASS; PNG, SVG, and PDF rendered at 183 × 115 mm. Final PNG was reviewed at full resolution; title/subtitle, both panels, legend, and caption are not clipped, and no visible text/mark collision was found. Some scenario points are numerically close and overlap in the panels; the legend explicitly identifies their color, marker, and line-type encodings.
• Facet alignment: PASS (R/ggplot2 gtable measurement; 1 comparison, 0 failures/warnings; tolerance 1.5 pt). Actual panel boxes are recorded in benchmark_reference_speedup.alignment-layout.json; full audit is benchmark_reference_speedup.alignment.json.
• PDF text size: the skill’s audit_pdf_text.py --min-pt 5 reports 9 text operators at 1 pt. Cross-checking the final PDF with PyMuPDF 1.28.2 finds 25 rendered text spans, minimum 6.5 pt, and 0 spans below 5 pt. The discrepancy is consistent with the primary auditor reading the Cairo PDF text matrix/operator size without its scale; therefore the script result is recorded as a limitation, not relabeled as a PASS. Span-level evidence is in benchmark_reference_speedup.pdf-text-audit.json.
• Collision audit: the automated collision tool returns FIX BEFORE DELIVERY (15 FAIL, 1 WARN). Inspection of its JSON findings against individual PyMuPDF text-span boxes and the full-size PNG indicates coarse union text-trace boxes: e.g., separate x ticks, facet titles, subtitle/caption lines and legend labels are grouped into overlapping bounding boxes; the reported legend stroke does not cross the actual “No signal” span. The audit output is preserved as benchmark_reference_speedup.collision-audit.json with its overlay PDF. Because the automated audit does not pass, this figure is suitable as a local/GitHub preview for user review, not certified here as publication-ready; resolve the auditor limitations or obtain an independent final collision review before journal submission.
• Editable text: SVG/PDF retain vector text; verify selection after opening the exports in the target editor.

Provenance

• Source data: benchmark_reference_summary.csv; pair-level records: benchmark_reference_pairs.csv.
• Benchmark metadata: benchmark_reference_metadata.csv; R session details: benchmark_reference_session_info.txt.
• Metadata plot-script SHA-256 captured at benchmark time: 7524130660ae1d3c141d8c0f153b79ea0175584553b7fab07d0d7446d482f297.
• Final plotting script SHA-256 used for this export: 5d6cfeeb615ef41d7f19f70de933f66f6eab4a2ff6e0b07e97f7b1f4baba964c.
• Summary CSV SHA-256: 9e49227b2544d7eefbb2922afb8b4ac95cbd14ad3af60178a44bff4680206387.
• Pair CSV SHA-256: 74df84f9b35081e3fdc7c79232c2ad82c744f65da6f3d8d410f438ba9f131797.