Re-ran the 20-model subset with --repeats 3 and a warmed edge for every
run: 12/12 runs cache-clean, 0 rate-limited, and the same 182 files /
19.6 MB for every new-tree mode.
- serial -> new-serial: 4.4x less wall time with 5.0x fewer bytes
(97.6 -> 19.6 MB).
- pipeline vs new-serial: -32.8% (1.49x), measured against a warm
baseline, so the pipeline gain is not a cold-start artifact.
- aimd vs pipeline: -16.0%; aimd vs new-serial: -43.5% (1.77x), each run
probed 6 -> 8 with no 429s.
Absolute times were about 2.8x slower than the 2026-11 session (a
higher-latency CDN path), which explains why the relative gains are
larger. The old tables stay marked superseded; the re-run table is the
one to quote.
Refs #1203
Will Miao
cd03d0c114