diff --git a/issues/cyclictests-worst-t14-wake-may-fall-in-acpiservers-startup.md b/issues/cyclictests-worst-t14-wake-may-fall-in-acpiservers-startup.md
new file mode 100644
index 00000000000..ee8bb494d8a
--- /dev/null
+++ b/issues/cyclictests-worst-t14-wake-may-fall-in-acpiservers-startup.md
@@ -0,0 +1,38 @@
+---
+status: open
+kind: finding
+opened: 2026-10-22
+---
+
+# cyclictest's worst T14 wake may fall in acpiserver's startup
+
+One T14 `latencycase` boot at `e6d6170d8` (#1009), the first with
+`/system/bin/acpiserver` started, read `cyclictest: 10000 wakes at 200us:
+p50=4us p90=6us p99=16us p99.9=74us max=1874us`. The worst wake is stamped
+638,179,441 ns on `nanos_since_boot`, and cyclictest ran from 12.082 s to
+14.103 s on the log's clock.
+
+acpiserver's startup overlaps the run's first 90 ms:
+- `ACPI_ENABLE` at 12.075 s;
+- the namespace initialised at 12.160 s;
+- power sources read at 12.171 s.
+
+In that window the tables' bytes took 42,906 reads of firmware memory, and
+their AML read SystemMemory 385 times. The boot logged firmware call `0xf5` at
+12.151 s, which held cpu0 297,820 ns. The SMI count on every CPU moved from
+4819 to 4821 across cyclictest's run. One step is that call. The other is not
+attributed: the kernel logs only the first call of each byte.
+
+The kernel's own lines put power-on to the kernel at 11.463 s, so the worst
+wake would be near 12.10 s, inside the window. That is an inference: the two
+clocks have not been aligned.
+
+**Owner**: `issues/toyos-beats-linuxs-latency-on-the-t14.md`.
+
+**Next step**: align cyclictest's worst-wake stamp with the log's clock. Then
+read the kernel's firmware-call count and time at the stop, to attribute the
+second SMI.
+
+**Exit**: the worst wake's cause is named, either as acpiserver's startup or
+as something else. If it is the startup, the owner decides whether timing rows
+start after it.
diff --git a/issues/the-t14s-firmware-interrupts-every-cpu-every-2-2-s-under-toyos.md b/issues/the-t14s-firmware-interrupts-every-cpu-every-2-2-s-under-toyos.md
index dbf3c30c928..e1ad007f983 100644
--- a/issues/the-t14s-firmware-interrupts-every-cpu-every-2-2-s-under-toyos.md
+++ b/issues/the-t14s-firmware-interrupts-every-cpu-every-2-2-s-under-toyos.md
@@ -81,6 +81,16 @@ The `mask-windows` kernel prints such a window as its CPU's own: cpu4's
 (`issues/a-process-lengthens-an-interrupts-off-walk-by-the-threads-it-parks-on-one-ring.md`),
 are this by reading, those kernels not reading the count.
 
+**An image that starts no `/system/bin/acpiserver` stays in legacy mode**, by
+design: the kernel enables ACPI mode only for the `acpi` claim's holder. The
+`osbench` and `latency_wake` images started none before 2026-10-22, so their T14
+readings carried this stop. Three `osbench` boots at `588c821ee` (#989), each
+`legacy mode` in its log, read 27 or 28 SMIs per 60 s on every CPU and 27
+wakes past 1 ms late, 4.1 to 4.6 ms, spaced 2.2165 to 2.248 s; each boot's
+30 s `tail.diag` run read 14 of each. With the server started (#1009), one `osbench`
+boot at `e6d6170d8` read no SMI on any CPU over either run and no wake later
+than 37 µs.
+
 **Owner**: stage 1 of
 `issues/toyos-runs-the-machine-in-acpi-mode-and-interprets-its-aml.md`,
 for the fix and for the reading: its exit holds the count on the T14 to what
diff --git a/tests/images/benchcase.toml b/tests/images/benchcase.toml
index 15141bda82e..43a35c543b7 100644
--- a/tests/images/benchcase.toml
+++ b/tests/images/benchcase.toml
@@ -1,10 +1,11 @@
 # The boot `osbench` measures ToyOS on, for a comparison with the same program
 # on the bench's own Linux (`tests/guest/osbench`).
 #
-# Only what a job list needs to run and to reach the log on the boot stick: no
-# compositor, sound, ACPI server or network, so little but the measurement asks
-# for a CPU. DATA is memory here, the stick having no DATA partition, and the
-# log volume is the one persistent volume a job can write (`osbench persist`).
+# Only what a job list needs to run and to reach the log on the boot stick, and
+# the ACPI server: no compositor, sound or network, so little but the
+# measurement asks for a CPU. DATA is memory here, the stick having no DATA
+# partition, and the log volume is the one persistent volume a job can write
+# (`osbench persist`).
 #
 # The runner's capability carries `jobs`, for the fairness run that starts each
 # workload as a job of its own; `counters`, `trace` and `roster`, so `osbench
@@ -13,12 +14,21 @@
 # handed a duplicate. No other benchmark's path uses them. An unprivileged
 # Linux program has no observation rights to match, which the report says.
 [boot]
-start = ["logkeeper", "diskserver", "fileserver", "test-runner"]
+start = ["logkeeper", "diskserver", "fileserver", "acpiserver", "test-runner"]
 
 [programs.logkeeper]
 service = true
 syscap = ["logread"]
 
+# ACPI mode, as a shipped image and the bench's Linux run it. In the legacy
+# mode the T14's firmware hands over, it stops every CPU for 4.5 ms every 2.2 s
+# (`issues/the-t14s-firmware-interrupts-every-cpu-every-2-2-s-under-toyos.md`).
+# Started before `test-runner`, so its claim's enable precedes every job.
+[programs.acpiserver]
+service = true
+devices = ["acpi"]
+receives = ["power"]
+
 # The order `osbench all` runs the groups in. The bound is the metal row's, its
 # 240 s on the 60 s every list has (`tests/toyos.rs`), restated for a boot of
 # this config built without the row; a derived metal config states its own.
diff --git a/tests/images/latencycase.toml b/tests/images/latencycase.toml
index b813f0c99dd..e6c1c8d1661 100644
--- a/tests/images/latencycase.toml
+++ b/tests/images/latencycase.toml
@@ -11,12 +11,21 @@
 # `power` because the job list is what ends the boot: there is no host on the
 # other end of the console on the machine this config exists for.
 [boot]
-start = ["logkeeper", "diskserver", "fileserver", "test-runner"]
+start = ["logkeeper", "diskserver", "fileserver", "acpiserver", "test-runner"]
 
 [programs.logkeeper]
 service = true
 syscap = ["logread"]
 
+# ACPI mode, as a shipped image and the bench's Linux run it. In the legacy
+# mode the T14's firmware hands over, it stops every CPU for 4.5 ms every 2.2 s
+# (`issues/the-t14s-firmware-interrupts-every-cpu-every-2-2-s-under-toyos.md`).
+# Started before `test-runner`, so its claim's enable precedes every job.
+[programs.acpiserver]
+service = true
+devices = ["acpi"]
+receives = ["power"]
+
 [programs.test-runner]
 receives = ["power"]
 args = ["test_rs_cyclictest", "test_rs_sched_stress", "reboot"]
diff --git a/tests/toyos.rs b/tests/toyos.rs
index ecef65940ca..173c7acf331 100644
--- a/tests/toyos.rs
+++ b/tests/toyos.rs
@@ -974,6 +974,7 @@ const METAL: &[(&str, metal::Metal)] = &[
         metal::Metal {
             arms: LATENCYCASE,
             judge: |b| {
+                in_acpi_mode_before_any_job(b[0])?;
                 b[0].job_passed("test_rs_sched_stress")?;
                 wake_latency_recorded(b[0])
             },
@@ -1457,6 +1458,7 @@ const BENCHCASE: &[metal::Arm] = &[metal::Arm {
 /// Every group exited 0, which `osbench` does only once each of its
 /// benchmarks measured, and said it was done.
 fn osbench_ran(boot: &metal::Readback) -> Result<(), String> {
+    in_acpi_mode_before_any_job(boot)?;
     for job in OSBENCH_GROUPS {
         boot.job_passed(job)?;
         let group = job.strip_prefix("osbench_").expect("a group's job");
@@ -1465,6 +1467,27 @@ fn osbench_ran(boot: &metal::Readback) -> Result<(), String> {
     Ok(())
 }
 
+/// A timing boot was in ACPI mode before its first job and stayed there, as a
+/// shipped image and the bench's Linux are. In legacy mode the T14's firmware
+/// stops every CPU every 2.2 s, and its time would be measured as ToyOS's
+/// (`issues/the-t14s-firmware-interrupts-every-cpu-every-2-2-s-under-toyos.md`).
+fn in_acpi_mode_before_any_job(boot: &metal::Readback) -> Result<(), String> {
+    let log = boot.log();
+    let text = log.text();
+    let first_job = text.find("===TEST_START").ok_or("the boot started no job")?;
+    let in_acpi_mode = ["the firmware handed over in ACPI mode", "acpi: ACPI mode: ACPI_ENABLE "]
+        .iter()
+        .filter_map(|said| text.find(said))
+        .min();
+    if !in_acpi_mode.is_some_and(|at| at < first_job) {
+        return Err("the machine was not in ACPI mode before the first job: the boot measured the firmware's legacy mode".to_string());
+    }
+    if let Ok(left) = log.must_say("acpi: legacy mode again") {
+        return Err(format!("the machine left ACPI mode inside the boot: {left}"));
+    }
+    Ok(())
+}
+
 /// The same boot, its rows ended on the hold, which waits for the ACPI
 /// server's power sources, the last line of its start. The members run
 /// behind it: the hold gives up at a time since boot, which a list of