2026-08-31

Parallel Probes

Why more concurrency made verification slower.

Campaign campaign_884b01b822564ab0 ran the same prompt-native pilot-approval-v1 task with parallel live probes. 10/10 accepted. p50 TTPS 25.735s. p95 34.734s.

The immediately prior prompt-native campaign with high-resolution clocks (campaign_8559333f207b412a) was p50 25.181s / p95 30.453s. Parallelism did not buy a faster verified lane. The tail got worse.

The working hypothesis is contention against Next compile-on-demand: live route probes hitting a warming app in parallel force overlapping compiles instead of a single sequential walk that can reuse work already paid.

The campaign is recorded as “parallel probes (failed)” in packages/ttps-bench/README.md — failed as an optimization, not as a P3 verdict. Acceptance stayed 100%. The experiment is evidence, not a buried loss.

All lab notes