Cubicle / experiments

Benchmark explorer

Measured performance, verification outcomes, and the evidence behind each run.

Results at a glance

Median wall time in seconds. Select a cell to inspect repetitions, counters, and logs.

Pilot and type-check runs are excluded from medians. “Not scheduled” is not a failure: transaction models were tested only with the two -tx all configurations. Input errors and pilot timeouts are not timings.

Pairwise runtime comparison

Same-model ratios on eligible results with matching verdicts. The current model filters apply here too.

Ratio = reference median ÷ comparison median. Above 1× means the comparison is faster. Bar length shows this ratio on a linear scale; no cross-model average is used.

Provenance & interpretation

These are descriptive measurements, not significance tests. Small runtimes include process startup overhead; related model variants are not independent protocols. Source expectations are audit annotations, not independently established ground truth.
Environment and run policy