NEWS.md

In seqbench: Anytime-Valid Sequential Benchmarking of Algorithms

seqbench 0.1.0

First CRAN release. Development history (0.0.0.9000 to 0.0.0.9002) is kept below.

• comparison_design(), initialize_comparison(), update_comparison(), stopping_decision(), comparison_report(), planning_horizon(); S3 methods print, summary, tidy, glance, autoplot for seqbench_comparison.
• Boundaries: "betting" (default; Waudby-Smith & Ramdas 2024), "empirical_bernstein", "hoeffding", "bernstein_declared" (declared sd_max, assumption A5) and the invalid negative control "naive_fixed", labelled as such in every output.
• Internal kernels boundary_init(), boundary_update(), boundary_interval() exported for sister packages.

seqbench 0.0.0.9002

• New boundary "bernstein_declared": predictable-plug-in Bennett confidence sequence with a declared upper bound sd_max on the standard deviation of the paired difference (assumption A5, recorded in the report). Valid only if the declared bound holds; its gain over betting is modest because the Bennett range term keeps a log(2/alpha)/t floor for bounded observations.

seqbench 0.0.0.9001

• Betting boundary: endpoint refinement by bisection is now lazy inside update_comparison() (only when a decision threshold falls in the boundary grid cell, plus once per call for reporting). Decisions and stopping times are unchanged; per-instance cost is now linear in the grid size instead of linear in the number of observations (a 2000-instance stream at the margin went from minutes to 0.3 s). boundary_interval() gained a thresholds argument.
• Betting boundary: a nonempty confidence set narrower than one grid cell is now located by searching the exact capital instead of being reported as empty (found by external review; constant streams triggered it).
• update_comparison() re-evaluates the decision after the final endpoint refinement, so an empty running intersection is always reported as such.
• planning_horizon() honours max_t; n_max, betting_grid and refine are validated as whole numbers / logical; costs must be finite.

seqbench 0.0.0.9000

• Core implemented: comparison_design(), initialize_comparison(), update_comparison(), stopping_decision(), comparison_report(), planning_horizon(); boundaries betting (default), empirical_bernstein, hoeffding (Waudby-Smith & Ramdas 2024) and the invalid negative control naive_fixed; S3 methods print, summary, tidy, glance, autoplot.
• Internal boundary kernels boundary_init(), boundary_update(), boundary_interval() exported for sister packages.

seqbench documentation built on Oct. 8, 2026, 5:07 p.m.