TypeSafe AI + Swamp / alert funnel

The same alerts. Fewer Claude calls.

Our alert pipeline, before and after the change. Watch how many events reach each stage.

100 workflow runs · 50 before + 50 after · same 50 source alerts

Before

50 runs · Claude-first routing

Before: all 50 events reach Claude The baseline route for the 50-event cohort: 50 alerts ingested, all 50 continue without screening, and all 50 reach Claude triage, including 23 resolution events.

Every event reaches the expensive step.

After Label-affected batch

50 runs · TypeSafe AI + resolution routing

Recorded new route: 50 events, 27 screens, 4 Claude calls Of 50 main-batch events, 23 resolutions bypassed AI and 27 firing alerts were screened. The CSV records 23 dismissals and 4 Claude calls; all 23 dismissals are flagged as confounded by the bench_run input label. One of the four escalated runs subsequently failed. These are batch observations, not a validated production suppression rate.

4 Claude calls; 1 run subsequently failed.