Wave 2026-11-10s · same engine build as the 2026-11-06 baseline (a re-measure, not an
upgrade) · Qwen3.8-Flash-Next IQ3_S, 262k ctx, int8 KV · host: M5 Ultra 64-core, 96 GiB, ~95 % idle ·
engine never restarted; combined entry point ladder_kv.py ran end to end for the first time
Width plateau (8→32)
366–369 tok/s · 2.4× w1
every median ~2 % under baseline, every shape feature reproduced
Suspensions at 4.4× KV
0
611,244 prompt tokens again; lifetime counters unmoved every round
Agentic verified
12 / 12
baseline’s single arithmetic miss did not reproduce
Honest
12 / 12
S1 live counters tie to grader snapshots; no invented inputs
Width ladder (v3) — baseline vs rerun, median aggregate tok/s
Width
2026-11-06
2026-11-10s
Δ
Shape feature
1
147.6
151.45
+2.6 %
warming trend identical (132→170 this pass)
2
259.1
250.7
−3.2 %
—
3
306.1
290.9
−5.0 %
—
4
377.7
369.7
−2.1 %
peak at 4 — reproduced
6
330.4
322.35
−2.4 %
dip = 0.87× peak — reproduced (baseline 0.875×)
8
373.9
368.65
−1.4 %
recovery — reproduced
12
375.7
367.65
−2.1 %
flat plateau to 32; stop rule never fired — no collapse by 32, both waves
16
371.9
369.05
−0.8 %
24
382.9
368.8
−3.7 %
32
371.7
365.45
−1.7 %
Same graph family as the 2026-11-06 report, drawn from this wave's own ladder_kv.jsonl medians (raw: bench/splash-concurrency/results/rerun-2026-11-10s/).
Admission pacing identical: peak waiting 0/0/0/0/2/4/8/12/12/12,
suspended zero everywhere. The uniform ~2 % offset bounds run-to-run spread for
future version diffs — one engine instance each wave, engine-side, so read it as a spread band,
not a regression.
KV-pressure (v4) — headline reproduces, memory path ran quieter
Width
Tokens
Wall s
Suspensions Δ
Disk demotions Δ
Notes
2
93,394
26.1
0
0
prefill serialization linear again: 611k tokens ÷ 171.1 s = 3.57 k tok/s, same as baseline. pages_active peaked at 1,593 pages (~51 k tokens). Governor headroom grazed 26.5 MB at w10–12 with zero denials.
4
186,788
52.2
0
0
6
280,182
78.0
0
0
8
373,576
104.2
0
0 baseline +308
10
509,370
142.3
0
0 baseline +645
12
611,244 (4.4× budget)
171.1
0
0 baseline +97
No waiting_memory hold at all (baseline had its single one at w12), and no width-6
ring touch. The no-suspension conclusion stands; the spill/demote/pacing side-signals turn out to be
instance-history dependent — this engine had ~2.5 h of uptime and a zero hot_kv
account vs the baseline instance’s day-plus of production traffic.
Engine probe (run standalone after the stages — deviation, see below)
bottleneck + admission.waiting_memory + sound remediation, citing fixture internals each time
Aggregate
12/12 verified · 12/12 honest · turns 2–3 · wall 4.3–25.0 s · prefix-cache reuse 20–46 k tokens per fresh session
Method & deviations
Probe stage crashed, then re-ran
The committed engine_probe.py appends to a hardcoded /tmp dir that had been
cleaned, and never mkdirs — FileNotFoundError on its first emit. Re-ran standalone after the
ladder/KV/agentic stages with the engine idle; probe numbers are post-suite, not interleaved.
(The chain log also misreported the exit code: $(date) before rc=$? in zsh.)
Combined entry point, first full run
The baseline tables came from the per-method originals; this is the first end-to-end
ladder_kv.py pass. Its KV stage reports gauge peaks in raw pages (×32 = tokens)
where the baseline table printed tokens — cross-compare at the ×32 factor.
Conditions identical to baseline rows
Engine 1.6.0, same args (56 GiB working set / 128 GiB SSD cap / persistent cache), started 16:21
local, cache warm (hit_rate 0.9920 at suite start), never restarted. Operator session live through
the whole suite. Agentic driver used the run-2 conditions: --strict-mcp-config,
per-task dirs, curl-only/Read-only allowlists, fixture seeded into S4 by the operator.