Us vs Autobahn: ordering with 512 B transactions
Highest throughput each system sustained
Committed tx/s. Same 16 vCPU droplets for every run except the published paper figure.
Our stack (B1): offered vs committed
Past ~20k tx/s offered, committed throughput falls instead of levelling off.
The benchmark runs
Test runs
Every test and run, newest first, grouped by campaign. Select a run to open its setup, results and charts.
No runs match.
perf-1 S8000 overload rung, run by run
Optimisations tried
Each optimisation, its effect as measured, and whether it stayed.
History
How the campaign got here: the decision timeline, incidents and lessons, and each agent's conclusions.
Decision timeline
Point anywhere in time to see what was done, what worked, who did it and what it led to. Scroll or pinch to zoom, drag to pan, select an event to draw its decision chain. Left and right arrow keys step between events.
Incidents and lessons
Halts, measurement traps and safety findings. Safety items are described at a high level only.
Agent conclusions
The campaign ran as Claude sessions posting to a shared message bus. Key findings per agent, dated, with the bus commit they were posted in.