Read-only copy. This public copy of trans-bench is read-only: it shows the published campaigns. Clone the repository to run benchmarks or to edit systems. github.com/Sirneij/trans-bench, branch verified-rerun-2026
trans-bench · 11 systems · 15 topologies

Every path,
every system,
verified.

The same recursive query on logic engines and databases, over graphs that grow until something breaks. Every answer is checked against the true closure.

hover a node
Systems
11
Topologies
15
Verified runs
12,215
Campaigns
2

Fastest at the largest graphs

verified_2026_v2 →
#SystemWins · podiumsTime as n growsDone
A contest is one structured topology with left or right recursion at its largest n; the fastest completed system wins it. The line is the geometric mean time over topologies as n grows.

Latest campaign

verified_2026_v2 →
99.4%
verified
Executed runs12,289
Timeout23
Out of memory13
Unsupported36
Iteration limit2
Skipped after a failure528

Topologies

hover a card to see what its closure adds · all 15 →

Systems

Manage

From the command line

Linux 6.18.15+deb13-cloud-amd64 · 48 CPUs · Python 3.12.15
$./virtualenv/bin/python benchmark.py --systems duckdb postgres --sizes 100 200 300 --campaign results/my_campaign
$./virtualenv/bin/python transitive.py --systems duckdb --sizes 100 301 100 --campaign results/my_campaign

Both commands, and the Start button of the wizard, run the same engine. A campaign directory holds one runs.jsonl per system, and its tables and figures appear under Campaigns when the run ends.