What would make the WorkPaper benchmark unfair?

[gregkonush]

I maintain Bilig. The current WorkPaper vs HyperFormula artifact is good enough to publish, but I want the critique to be public too.

Current claim:

• WorkPaper has lower mean latency on 46/46 scorecard-eligible comparable workloads.
• The p95 aggregate is still ahead, but lookup-approximate-duplicates is slower at p95 by 1.043x.
• This is a headless formula benchmark. It is not an Excel compatibility claim, a browser grid claim, or proof that every customer workbook is faster.

Links:

• Benchmark explainer: https://proompteng.github.io/bilig/what-workpaper-benchmark-proves.html
• Evidence note: https://proompteng.github.io/bilig/headless-workpaper-benchmark-evidence.html
• Local walkthrough: https://proompteng.github.io/bilig/local-workpaper-benchmark-walkthrough.html
• Checked artifact: https://github.com/proompteng/bilig/blob/main/packages/benchmarks/baselines/workpaper-vs-hyperformula.json
• Share card: https://github.com/proompteng/bilig/blob/main/docs/assets/workpaper-benchmark-card.png

What I want feedback on:

1. Which workload family is missing before this benchmark is useful for your spreadsheet automation work?
2. Which row looks suspicious, too synthetic, or too favorable to WorkPaper?
3. Is the p95 caveat explained clearly enough, or should the public claim be narrower?
4. What would make the comparison unfair against HyperFormula or another headless formula engine?
5. If you maintain or use a spreadsheet engine, what command or artifact would make the comparison easier to audit?

Useful criticism here should turn into benchmark issues, docs fixes, or new fixtures. I am not looking for generic praise; I want the benchmark to survive skeptical review.

[gregkonush]

@/tmp/bilig-discussion-340-comment.md