LEG — the laptop's 5090, QUEUED rung 95 W
  started 2026-09-21T17:27:07Z on this laptop
  it is WAITING for the card to read 95 W; it never sets a cap
  ceiling 43200s (it reports past this, it does not reap)
  log     /workshop/bench-laptop-5090-2026-09-21/logs/rung-95w.log
receipt: nvidia-powerd is inactive — Dynamic Boost is not moving the cap
waiting for the card to read 95 W. THIS SCRIPT NEVER SETS A CAP — it is waiting
for an operator's paste. The paste is in RUNSHEET.md, Step 4.
the card reads 95 W (field enforced.power.limit) after 0s — the 95 W rung starts now
memory guard: memshed.timer was active at the window's open (remembered in /workshop/bench-laptop-5090-2026-09-21/.memshed-was)
memory guard: memshed.timer is now inactive; MEMSHED_DRY_RUN=1
  (STOPPED, never masked: the unit file is a symlink into ~/estate/systemd and
   'mask --force' would destroy it. glimmer-check.timer is left alone — it next fires
   Tue 14:43Z, outside any daytime window, and masking a timer this bench does not
   need to touch is one more thing to forget to restore.)

================================================================================
RUNG 95 W — LEG 1 (the eleven-arm language bank), then LEG 2 (R1 + R4)
  card   GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff  (NVIDIA GeForce RTX 5090 Laptop GPU)
  class  gpu-5090-laptop-24g
  start  2026-09-21T17:27:07Z
================================================================================
receipt: nvidia-powerd is inactive — Dynamic Boost is not moving the cap
cap read-back [rung 95W open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

---- LEG 1: /workshop/bench-laptop-5090-2026-09-21/harness-llm/driver_block.sh 95 ----
card name read back: 'NVIDIA GeForce RTX 5090 Laptop GPU' (arm one; exact equality, not a substring)
card name read back: 'NVIDIA GeForce RTX 5090 Laptop GPU' (arm one; exact equality, not a substring)

=== the board, before anything is started (this is the 95 W block) === 17:27:07Z

=== the board, before anything is started (this is the 95 W block) === 17:27:07Z
index, uuid, name, power.limit [W], enforced.power.limit [W], power.default_limit [W], power.max_limit [W], persistence_mode, pcie.link.width.current, clocks.current.sm [MHz], clocks.current.memory [MHz], clocks.max.sm [MHz], temperature.gpu, fan.speed [%]
index, uuid, name, power.limit [W], enforced.power.limit [W], power.default_limit [W], power.max_limit [W], persistence_mode, pcie.link.width.current, clocks.current.sm [MHz], clocks.current.memory [MHz], clocks.max.sm [MHz], temperature.gpu, fan.speed [%]
0, GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff, NVIDIA GeForce RTX 5090 Laptop GPU, [N/A], 95.00 W, 95.00 W, 175.00 W, Enabled, 8, 1027 MHz, 9001 MHz, 3090 MHz, 50, [N/A]
0, GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff, NVIDIA GeForce RTX 5090 Laptop GPU, [N/A], 95.00 W, 95.00 W, 175.00 W, Enabled, 8, 1027 MHz, 9001 MHz, 3090 MHz, 50, [N/A]
receipt: nvidia-powerd is inactive — Dynamic Boost is not moving the cap
receipt: nvidia-powerd is inactive — Dynamic Boost is not moving the cap
clock lock: ai-perf.service is active; clocks now 1027,9001,3090 (sm,mem,max.sm MHz)
  ⚠ THE CLOCK LOCK IS STILL ON. This board cannot clock below 1200,2550
    MHz, so a power cap cannot be honoured the way it is on an unlocked card, and these
    figures are NOT like-for-like with the desktop rows. The operator's Step 1 paste
    releases it with 'sudo nvidia-smi -rgc'. The block PROCEEDS and every file records
    the clock floor -- a confound that is measured and printed beats one that is
    avoided by refusing to measure.
clock lock: ai-perf.service is active; clocks now 1027,9001,3090 (sm,mem,max.sm MHz)
  ⚠ THE CLOCK LOCK IS STILL ON. This board cannot clock below 1200,2550
    MHz, so a power cap cannot be honoured the way it is on an unlocked card, and these
    figures are NOT like-for-like with the desktop rows. The operator's Step 1 paste
    releases it with 'sudo nvidia-smi -rgc'. The block PROCEEDS and every file records
    the clock floor -- a confound that is measured and printed beats one that is
    avoided by refusing to measure.
cap read-back [block 95W open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
cap read-back [block 95W open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
memory guard: memshed.timer is inactive
unknown; MEMSHED_DRY_RUN=1
memory guard: memshed.timer is inactive
unknown; MEMSHED_DRY_RUN=1

=== the UPS at the door (expected: absent on this box) === 17:27:07Z

=== the UPS at the door (expected: absent on this box) === 17:27:07Z
UPS UNREADABLE at block start -- whole-box watts and joules-per-1,000-tokens are
WITHHELD with this reason, and board watts are read normally:
  no `upsc` on this box: NUT is not installed and nut-server/nut-monitor are inactive (read 2026-09-21T14:52Z). Whole-box watts and joules-per-1,000-tokens are WITHHELD with this reason; board watts, which is what every table in this ladder prints, are read normally from the card at 2 Hz.
  receipt -> results/gpu-5090-laptop-24g-ups-gate-95w.txt
UPS UNREADABLE at block start -- whole-box watts and joules-per-1,000-tokens are
WITHHELD with this reason, and board watts are read normally:
  no `upsc` on this box: NUT is not installed and nut-server/nut-monitor are inactive (read 2026-09-21T14:52Z). Whole-box watts and joules-per-1,000-tokens are WITHHELD with this reason; board watts, which is what every table in this ladder prints, are read normally from the card at 2 Hz.
  receipt -> results/gpu-5090-laptop-24g-ups-gate-95w.txt

=== instance: one card, kv q8_0, parallel 1 === 17:27:09Z

=== instance: one card, kv q8_0, parallel 1 === 17:27:09Z
  "persistence_mode_at_start": "0, Enabled;",
  "pcie_link_width_at_start": "0, 8;"
}

next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status
pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq8_0-p1.json to every bencher so the result files carry it
  "persistence_mode_at_start": "0, Enabled;",
  "pcie_link_width_at_start": "0, 8;"
}

next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status
pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq8_0-p1.json to every bencher so the result files carry it
cap read-back [card-mapping open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
cap read-back [card-mapping open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clocks at [card-mapping] open: 1590,9001,3090  (sm,mem,max.sm MHz)
clocks at [card-mapping] open: 1590,9001,3090  (sm,mem,max.sm MHz)

=== card mapping: watch which physical card's memory rises === 17:27:51Z

=== card mapping: watch which physical card's memory rises === 17:27:51Z
  one: MATCHES: declared soldered (mobile; no socket — link width is traced under load, never assumed) at 00000000:01:00.0 - vendor unread - GPU-edff232c, and that is the card whose memory grew

every shape checked runs on the card it declares
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-card-mapping-95w.json
  one: MATCHES: declared soldered (mobile; no socket — link width is traced under load, never assumed) at 00000000:01:00.0 - vendor unread - GPU-edff232c, and that is the card whose memory grew

every shape checked runs on the card it declares
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-card-mapping-95w.json
clocks at [card-mapping] close: 1590,9001,3090  (sm,mem,max.sm MHz)
clocks at [card-mapping] close: 1590,9001,3090  (sm,mem,max.sm MHz)
cap read-back [card-mapping close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
cap read-back [card-mapping close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
cap read-back [armA-ladder-gemma4 open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
cap read-back [armA-ladder-gemma4 open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clocks at [armA-ladder-gemma4] open: 1590,9001,3090  (sm,mem,max.sm MHz)
clocks at [armA-ladder-gemma4] open: 1590,9001,3090  (sm,mem,max.sm MHz)

=== armA-ladder-gemma4 === 17:28:03Z

=== armA-ladder-gemma4 === 17:28:03Z
REFUSED: --instance-env-file could not be read: Invalid control character at: line 26 column 38 (char 1389)
REFUSED: --instance-env-file could not be read: Invalid control character at: line 26 column 38 (char 1389)
bench_gpu.py exited 2 for armA-ladder-gemma4
bench_gpu.py exited 2 for armA-ladder-gemma4
ARM REFUSED: armA-ladder-gemma4 exited 2 -- recorded as refused, not done
ARM REFUSED: armA-ladder-gemma4 exited 2 -- recorded as refused, not done
clocks at [armA-ladder-gemma4] close: 1590,9001,3090  (sm,mem,max.sm MHz)
clocks at [armA-ladder-gemma4] close: 1590,9001,3090  (sm,mem,max.sm MHz)
cap read-back [armA-ladder-gemma4 close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
cap read-back [armA-ladder-gemma4 close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
REFUSED: arm A produced no largest_context_that_fits at 95 W. The
         arms that take their window from it do not run on a guess.
REFUSED: arm A produced no largest_context_that_fits at 95 W. The
         arms that take their window from it do not run on a guess.
LEG 1 at 95 W exited 2
cap read-back [between the legs at 95W]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

---- LEG 2: /workshop/bench-laptop-5090-2026-09-21/harness-render/run_rung.sh 95 ----
== render rung 95 W · started 2026-09-21T17:28:03Z ==
   card expected   NVIDIA GeForce RTX 5090 Laptop GPU
   comfy root      /workshop/bench-laptop-5090-2026-09-21/ComfyUI-0.21.1
   comfy python    /workshop/ComfyUI/.venv/bin/python
== render rung 95 W · started 2026-09-21T17:28:03Z ==
   card expected   NVIDIA GeForce RTX 5090 Laptop GPU
   comfy root      /workshop/bench-laptop-5090-2026-09-21/ComfyUI-0.21.1
   comfy python    /workshop/ComfyUI/.venv/bin/python
   comfy version   0.21.1 (26515acd)
   comfy version   0.21.1 (26515acd)
receipt: nvidia-powerd is inactive — Dynamic Boost is not moving the cap
receipt: nvidia-powerd is inactive — Dynamic Boost is not moving the cap
   card reads      0, 00000000:01:00.0, NVIDIA GeForce RTX 5090 Laptop GPU, [N/A], 95.00 W, 24463 MiB, 8, 16, 1590 MHz, 9001 MHz
   card reads      0, 00000000:01:00.0, NVIDIA GeForce RTX 5090 Laptop GPU, [N/A], 95.00 W, 24463 MiB, 8, 16, 1590 MHz, 9001 MHz
cap read-back [render rung 95W open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
cap read-back [render rung 95W open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
   clock lock      active (ai-perf.service; when active this board
                   cannot clock below 1200 MHz and a power cap means something different
                   from what it means on the unlocked desktop cards this leg compares to)
   clock lock      active (ai-perf.service; when active this board
                   cannot clock below 1200 MHz and a power cap means something different
                   from what it means on the unlocked desktop cards this leg compares to)
   port band       18190-18199 free; this box's own comfyui.service is inactive
   port band       18190-18199 free; this box's own comfyui.service is inactive


-- arm inventory · suffix inventory-95w · 2026-09-21T17:28:03Z
-- arm inventory · suffix inventory-95w · 2026-09-21T17:28:03Z
Traceback (most recent call last):
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1527, in <module>
    main()
    ~~~~^^
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1517, in main
    payload = arm_inventory(args.graphs.split(","), name)
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1458, in arm_inventory
    bank = prompt_bank()
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 673, in prompt_bank
    bank = rl.read_manifest_prompts(PROMPT_MANIFEST)
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/renderlib.py", line 850, in read_manifest_prompts
    with open(manifest_path) as fh:
         ~~~~^^^^^^^^^^^^^^^
FileNotFoundError: [Errno 2] No such file or directory: '/workshop/bench-laptop-5090-2026-09-21/harness-render/inputs/set1-exemplars-manifest.json'
Traceback (most recent call last):
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1527, in <module>
    main()
    ~~~~^^
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1517, in main
    payload = arm_inventory(args.graphs.split(","), name)
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1458, in arm_inventory
    bank = prompt_bank()
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 673, in prompt_bank
    bank = rl.read_manifest_prompts(PROMPT_MANIFEST)
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/renderlib.py", line 850, in read_manifest_prompts
    with open(manifest_path) as fh:
         ~~~~^^^^^^^^^^^^^^^
FileNotFoundError: [Errno 2] No such file or directory: '/workshop/bench-laptop-5090-2026-09-21/harness-render/inputs/set1-exemplars-manifest.json'
-- arm inventory exited 1 · 2026-09-21T17:28:03Z
-- arm inventory exited 1 · 2026-09-21T17:28:03Z
cap read-back [after inventory at 95W]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

cap read-back [after inventory at 95W]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

-- arm r1 · suffix r1-95w · 2026-09-21T17:28:03Z
-- arm r1 · suffix r1-95w · 2026-09-21T17:28:03Z
  card available at 2026-09-21T17:28:03Z · the image service holds 0 MiB · quiet: mean 0.1% <= 5.0%
Traceback (most recent call last):
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1527, in <module>
    main()
    ~~~~^^
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1519, in main
    payload = arm_r1(args.graphs.split(","), batches, args.images, name)
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1156, in arm_r1
    "preflight": box_preflight(), "graphs": {}, "refusals": []}
                 ~~~~~~~~~~~~~^^
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1032, in box_preflight
    cap = rl.assert_cap_agrees(expect_cap)
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/renderlib.py", line 814, in assert_cap_agrees
    raise CapRefused(f"the cards report different power limits: {block['per_card_w']}")
renderlib.CapRefused: the cards report different power limits: {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': None}
  card available at 2026-09-21T17:28:03Z · the image service holds 0 MiB · quiet: mean 0.1% <= 5.0%
Traceback (most recent call last):
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1527, in <module>
    main()
    ~~~~^^
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1519, in main
    payload = arm_r1(args.graphs.split(","), batches, args.images, name)
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1156, in arm_r1
    "preflight": box_preflight(), "graphs": {}, "refusals": []}
                 ~~~~~~~~~~~~~^^
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1032, in box_preflight
    cap = rl.assert_cap_agrees(expect_cap)
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/renderlib.py", line 814, in assert_cap_agrees
    raise CapRefused(f"the cards report different power limits: {block['per_card_w']}")
renderlib.CapRefused: the cards report different power limits: {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': None}
-- arm r1 exited 1 · 2026-09-21T17:28:14Z
-- arm r1 exited 1 · 2026-09-21T17:28:14Z
cap read-back [between R1 and R4 at 95W]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

cap read-back [between R1 and R4 at 95W]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

-- arm r4 · suffix r4-95w · 2026-09-21T17:28:14Z
-- arm r4 · suffix r4-95w · 2026-09-21T17:28:14Z
  card available at 2026-09-21T17:28:14Z · the image service holds 0 MiB · quiet: mean 0.0% <= 5.0%
Traceback (most recent call last):
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1527, in <module>
    main()
    ~~~~^^
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1521, in main
    payload = arm_r4(args.graph, batches[0], args.minutes, name)
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1301, in arm_r4
    "preflight": box_preflight(), "cards": {}}
                 ~~~~~~~~~~~~~^^
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1032, in box_preflight
    cap = rl.assert_cap_agrees(expect_cap)
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/renderlib.py", line 814, in assert_cap_agrees
    raise CapRefused(f"the cards report different power limits: {block['per_card_w']}")
renderlib.CapRefused: the cards report different power limits: {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': None}
  card available at 2026-09-21T17:28:14Z · the image service holds 0 MiB · quiet: mean 0.0% <= 5.0%
Traceback (most recent call last):
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1527, in <module>
    main()
    ~~~~^^
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1521, in main
    payload = arm_r4(args.graph, batches[0], args.minutes, name)
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1301, in arm_r4
    "preflight": box_preflight(), "cards": {}}
                 ~~~~~~~~~~~~~^^
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1032, in box_preflight
    cap = rl.assert_cap_agrees(expect_cap)
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/renderlib.py", line 814, in assert_cap_agrees
    raise CapRefused(f"the cards report different power limits: {block['per_card_w']}")
renderlib.CapRefused: the cards report different power limits: {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': None}
-- arm r4 exited 1 · 2026-09-21T17:28:25Z
-- arm r4 exited 1 · 2026-09-21T17:28:25Z
cap read-back [render rung 95W close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

cap read-back [render rung 95W close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

== render rung 95 W · done 2026-09-21T17:28:25Z (R1 exit 1, R4 exit 1) ==
   results/r1-95w.json and results/r4-95w.json
   the UPS is absent on this box: no `upsc` on this box: NUT is not installed and nut-server/nut-monitor are inactive (2026-09-21T14:52Z). Whole-box watts are WITHHELD with this reason; board watts are read normally.
   the fan is unreadable on this board: fan.speed reads [N/A] on this board: the driver reports no fan for it. The fan-drawing cell is a declared non-figure with this reason — arm R4 measures ten minutes of sustained drawing and cannot report a fan curve for it.
== render rung 95 W · done 2026-09-21T17:28:25Z (R1 exit 1, R4 exit 1) ==
   results/r1-95w.json and results/r4-95w.json
   the UPS is absent on this box: no `upsc` on this box: NUT is not installed and nut-server/nut-monitor are inactive (2026-09-21T14:52Z). Whole-box watts are WITHHELD with this reason; board watts are read normally.
   the fan is unreadable on this board: fan.speed reads [N/A] on this board: the driver reports no fan for it. The fan-drawing cell is a declared non-figure with this reason — arm R4 measures ten minutes of sustained drawing and cannot report a fan curve for it.
LEG 2 at 95 W exited 1
cap read-back [rung 95W close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

artifact receipt -> /workshop/bench-laptop-5090-2026-09-21/RUNG-95W-RECEIPT.txt
RUNG 95 W — artifact receipt
  written        2026-09-21T17:28:25Z
  box_class      gpu-5090-laptop-24g
  card           GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff  (NVIDIA GeForce RTX 5090 Laptop GPU)
  cap at close   95 enforced.power.limit
  clocks         180,405,3090  (sm,mem,max.sm MHz)
  clock lock     active
  nvidia-powerd  inactive
  LEG 1 exit     2
  LEG 2 exit     1

  LEG 1 result files carrying 95w in their name:
    ac9b897f61f34351  gpu-5090-laptop-24g-card-mapping-95w.json
    d9d87225ad3a30dd  gpu-5090-laptop-24g-clock-lock-ACTIVE-95w.txt
    6e27d244f8c64a68  gpu-5090-laptop-24g-ups-gate-95w.txt

  LEG 2 result files carrying 95w in their name:

  stated absences at this rung:
    UPS          no `upsc` on this box: NUT is not installed, nut-server and nut-monitor are inactive, the battery exposes no power_now/energy_now, and /sys/class/powercap/*/energy_uj is root-only (all read 2026-09-21T14:52Z). Whole-box watts are WITHHELD with this reason. Board watts — which is what every table in this ladder prints — are read normally at 2 Hz.
    fan          fan.speed reads [N/A] on this board (2026-09-21T14:52Z): the driver reports no fan for it. The fan column is a declared non-figure with this reason, not a gap to be filled later.
    memory temp  temperature.memory reads [N/A], as on every board of this ladder. The core-temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument.
    arm T        the training arm's preprocessed tensor set lives on benchbox, which
                 was unreachable from 2026-09-21T15:03Z. Arm T is a SECOND TRAIN and
                 legs 1 and 2 never depended on it.
rung 95 W finished rc=0 at 2026-09-21T17:28:25Z
memory guard restored: memshed.timer is active (it was active); MEMSHED_DRY_RUN unset
  free receipt: journalctl --user -u memshed --since '2026-09-21T17:27:07Z' | grep SHEDDING
  — every line there is a shed this bench WOULD have taken, with its signature.
LEG — the laptop's 5090, QUEUED rung 95 W
  started 2026-09-21T18:27:50Z on this laptop
  it is WAITING for the card to read 95 W; it never sets a cap
  ceiling 43200s (it reports past this, it does not reap)
  log     /workshop/bench-laptop-5090-2026-09-21/logs/rung-95w.log
receipt: nvidia-powerd is inactive — Dynamic Boost is not moving the cap
waiting for the card to read 95 W. THIS SCRIPT NEVER SETS A CAP — it is waiting
for an operator's paste. The paste is in RUNSHEET.md, Step 4.
the card reads 95 W (field enforced.power.limit) after 0s — the 95 W rung starts now
memory guard: memshed.timer was active at the window's open (remembered in /workshop/bench-laptop-5090-2026-09-21/.memshed-was)
memory guard: memshed.timer is now inactive; MEMSHED_DRY_RUN=1
  (STOPPED, never masked: the unit file is a symlink into ~/estate/systemd and
   'mask --force' would destroy it. glimmer-check.timer is left alone — it next fires
   Tue 14:43Z, outside any daytime window, and masking a timer this bench does not
   need to touch is one more thing to forget to restore.)

================================================================================
PASS 95w — posture fixed — LEG 1 (the eleven-arm language bank), then LEG 2
  card    GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff  (NVIDIA GeForce RTX 5090 Laptop GPU)
  class   gpu-5090-laptop-24g
  cap arg 95
  start   2026-09-21T18:27:50Z
================================================================================
receipt: nvidia-powerd is inactive — Dynamic Boost is not moving the cap
cap read-back [rung 95w open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clock lock [by the CARD]: OFF-BY-CARD
  evidence: clocks.sm was observed at 180 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 180, max 180, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force.
  the unit ai-perf.service reads active — NOT EVIDENCE:
    it is Type=oneshot, so it reads 'active (exited)' for the whole uptime whether or
    not 'nvidia-smi -rgc' has released the lock since. On 2026-09-21 a receipt that
    trusted it published 'ran with the lock ACTIVE' beside clocks that proved otherwise.
clock-lock receipt -> /workshop/bench-laptop-5090-2026-09-21/logs/clock-lock-95w.txt

---- LEG 1: /workshop/bench-laptop-5090-2026-09-21/harness-llm/driver_block.sh 95 ----
/workshop/bench-laptop-5090-2026-09-21/harness-llm/driver_block.sh: line 152: LABEL: unbound variable
LEG 1 at 95w exited 1
cap read-back [between the legs at 95w]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

---- LEG 2: /workshop/bench-laptop-5090-2026-09-21/harness-render/run_rung.sh 95 ----
== render rung 95w · started 2026-09-21T18:27:56Z ==
   card expected   NVIDIA GeForce RTX 5090 Laptop GPU
   comfy root      /workshop/bench-laptop-5090-2026-09-21/ComfyUI-0.21.1
   comfy python    /workshop/ComfyUI/.venv/bin/python
   comfy version   0.21.1 (26515acd)
receipt: nvidia-powerd is inactive — Dynamic Boost is not moving the cap
   card reads      0, 00000000:01:00.0, NVIDIA GeForce RTX 5090 Laptop GPU, [N/A], 95.00 W, 24463 MiB, 8, 16, 180 MHz, 405 MHz
cap read-back [render rung 95w open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
   clock lock [by the CARD]: OFF-BY-CARD
     evidence: clocks.sm was observed at 180 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 180, max 180, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force.
     the unit ai-perf.service reads active — NOT EVIDENCE:
       it is Type=oneshot, so it reads 'active (exited)' for the whole uptime whether or
       not 'nvidia-smi -rgc' has released the lock since. On 2026-09-21 a receipt that
       trusted it published 'ran with the lock ACTIVE' beside clocks that proved otherwise.
   clock lock      unprobed by the CARD (when the lock is in force
                   this board cannot clock below 1200 MHz and a power posture means
                   something different from what it means on the unlocked desktop cards
                   this leg compares to). The unit state is not evidence; see above.
   port band       18190-18199 free; this box's own comfyui.service is inactive

-- arm inventory · suffix inventory-95w · 2026-09-21T18:28:02Z
  -> results/inventory-inventory-95w.json
-- arm inventory exited 0 · 2026-09-21T18:28:02Z
cap read-back [after inventory at 95w]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

-- arm r1 · suffix r1-95w · 2026-09-21T18:28:02Z
  card available at 2026-09-21T18:28:03Z · the image service holds 0 MiB · quiet: mean 0.0% <= 5.0%
  cap 95.0 W per card · UPS load unreadable
  -> results/r1-r1-95w.json
-- arm r1 exited 0 · 2026-09-21T18:28:16Z
cap read-back [between R1 and R4 at 95w]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

-- arm r4 · suffix r4-95w · 2026-09-21T18:28:16Z
  card available at 2026-09-21T18:28:16Z · the image service holds 0 MiB · quiet: mean 0.0% <= 5.0%
  cap 95.0 W per card · UPS load unreadable
Traceback (most recent call last):
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1571, in <module>
    main()
    ~~~~^^
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1565, in main
    payload = arm_r4(args.graph, batches[0], args.minutes, name)
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 1370, in arm_r4
    graph = load_graph(stem)
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/bench_render.py", line 693, in load_graph
    return rl.Graph.load(PRINTLAB_WORKFLOWS, stem)
           ~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/workshop/bench-laptop-5090-2026-09-21/harness-render/renderlib.py", line 94, in load
    with open(path, "rb") as fh:
         ~~~~^^^^^^^^^^^^
FileNotFoundError: [Errno 2] No such file or directory: '/workshop/bench-laptop-5090-2026-09-21/harness-render/inputs/workflows/klein-4b-8st.json'
-- arm r4 exited 1 · 2026-09-21T18:28:30Z
cap read-back [render rung 95w close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

== render rung 95w · done 2026-09-21T18:28:30Z (R1 exit 0, R4 exit 1) ==
   results/r1-95w.json and results/r4-95w.json
   the UPS is absent on this box: no `upsc` on this box: NUT is not installed and nut-server/nut-monitor are inactive (2026-09-21T14:52Z). Whole-box watts are WITHHELD with this reason; board watts are read normally.
   the fan is unreadable on this board: fan.speed reads [N/A] on this board: the driver reports no fan for it. The fan-drawing cell is a declared non-figure with this reason — arm R4 measures ten minutes of sustained drawing and cannot report a fan curve for it.
LEG 2 at 95w exited 1
cap read-back [rung 95w close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

artifact receipt -> /workshop/bench-laptop-5090-2026-09-21/RUNG-95W-FAILED.txt
PASS 95w — FAILED (rc=1)
  posture        fixed
  cap argument   95
  written        2026-09-21T18:28:30Z
  box_class      gpu-5090-laptop-24g
  card           GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff  (NVIDIA GeForce RTX 5090 Laptop GPU)
  cap at close   95 enforced.power.limit
  clocks         180,405,3090  (sm,mem,max.sm MHz)
  clock lock     OFF-BY-CARD, read FROM THE CARD
                 clocks.sm was observed at 180 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 180, max 180, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force.
                 (the unit ai-perf.service reads
                 active, which is NOT evidence: it is a
                 oneshot and stays 'active' for the uptime after -rgc.)
  nvidia-powerd  inactive
  LEG 1 exit     1
  LEG 2 exit     1
  PASS exit      1

  LEG 1 result files carrying 95w in their name:
    ac9b897f61f34351  gpu-5090-laptop-24g-card-mapping-95w.json
    6e27d244f8c64a68  gpu-5090-laptop-24g-ups-gate-95w.txt

  LEG 2 result files carrying 95w in their name:
    04e8b9ea3418de0b  inventory-inventory-95w.json
    510d0154e39b2e3d  r1-r1-95w.json

  stated absences at this pass:
    UPS          no `upsc` on this box: NUT is not installed, nut-server and nut-monitor are inactive, the battery exposes no power_now/energy_now, and /sys/class/powercap/*/energy_uj is root-only (all read 2026-09-21T14:52Z). Whole-box watts are WITHHELD with this reason. Board watts — which is what every table in this ladder prints — are read normally at 2 Hz.
    fan          fan.speed reads [N/A] on this board (2026-09-21T14:52Z): the driver reports no fan for it. The fan column is a declared non-figure with this reason, not a gap to be filled later.
    memory temp  temperature.memory reads [N/A], as on every board of this ladder. The core-temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument.
    arm T        the training arm's preprocessed tensor set lives on benchbox, which
                 was unreachable from 2026-09-21T15:03Z. Arm T is a SECOND TRAIN and
                 legs 1 and 2 never depended on it.
rung 95 W finished rc=1 at 2026-09-21T18:28:30Z
memory guard restored: memshed.timer is active (it was active); MEMSHED_DRY_RUN unset
  free receipt: journalctl --user -u memshed --since '2026-09-21T18:27:50Z' | grep SHEDDING
  — every line there is a shed this bench WOULD have taken, with its signature.
LEG — the laptop's 5090, QUEUED rung 95 W
  started 2026-09-21T18:43:09Z on this laptop
  it is WAITING for the card to read 95 W; it never sets a cap
  ceiling 43200s (it reports past this, it does not reap)
  log     /workshop/bench-laptop-5090-2026-09-21/logs/rung-95w.log
receipt: nvidia-powerd is inactive — Dynamic Boost is not moving the cap
waiting for the card to read 95 W. THIS SCRIPT NEVER SETS A CAP — it is waiting
for an operator's paste. The paste is in RUNSHEET.md, Step 4.
the card reads 95 W (field enforced.power.limit) after 0s — the 95 W rung starts now
memory guard: memshed.timer was active at the window's open (remembered in /workshop/bench-laptop-5090-2026-09-21/.memshed-was)
memory guard: memshed.timer is now inactive; MEMSHED_DRY_RUN=1
  (STOPPED, never masked: the unit file is a symlink into ~/estate/systemd and
   'mask --force' would destroy it. glimmer-check.timer is left alone — it next fires
   Tue 14:43Z, outside any daytime window, and masking a timer this bench does not
   need to touch is one more thing to forget to restore.)

================================================================================
PASS 95w — posture fixed — LEG 1 (the eleven-arm language bank), then LEG 2
  card    GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff  (NVIDIA GeForce RTX 5090 Laptop GPU)
  class   gpu-5090-laptop-24g
  cap arg 95
  start   2026-09-21T18:43:09Z
================================================================================
receipt: nvidia-powerd is inactive — Dynamic Boost is not moving the cap
cap read-back [rung 95w open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clock lock [by the CARD]: OFF-BY-CARD
  evidence: clocks.sm was observed at 180 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 180, max 180, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force.
  the unit ai-perf.service reads active — NOT EVIDENCE:
    it is Type=oneshot, so it reads 'active (exited)' for the whole uptime whether or
    not 'nvidia-smi -rgc' has released the lock since. On 2026-09-21 a receipt that
    trusted it published 'ran with the lock ACTIVE' beside clocks that proved otherwise.
clock-lock receipt -> /workshop/bench-laptop-5090-2026-09-21/logs/clock-lock-95w.txt

---- LEG 1: /workshop/bench-laptop-5090-2026-09-21/harness-llm/driver_block.sh 95 ----
card name read back: 'NVIDIA GeForce RTX 5090 Laptop GPU' (arm one; exact equality, not a substring)

=== the board, before anything is started (this is the 95 W block) === 18:43:15Z
index, uuid, name, power.limit [W], enforced.power.limit [W], power.default_limit [W], power.max_limit [W], persistence_mode, pcie.link.width.current, clocks.current.sm [MHz], clocks.current.memory [MHz], clocks.max.sm [MHz], temperature.gpu, fan.speed [%]
0, GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff, NVIDIA GeForce RTX 5090 Laptop GPU, [N/A], 95.00 W, 95.00 W, 175.00 W, Enabled, 8, 180 MHz, 405 MHz, 3090 MHz, 32, [N/A]
receipt: nvidia-powerd is inactive — Dynamic Boost is not moving the cap
cap read-back [block 95w open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
memory guard: memshed.timer is inactive; MEMSHED_DRY_RUN=1

=== the UPS at the door (expected: absent on this box) === 18:43:15Z
UPS UNREADABLE at block start -- whole-box watts and joules-per-1,000-tokens are
WITHHELD with this reason, and board watts are read normally:
  no `upsc` on this box: NUT is not installed and nut-server/nut-monitor are inactive (read 2026-09-21T14:52Z). Whole-box watts and joules-per-1,000-tokens are WITHHELD with this reason; board watts, which is what every table in this ladder prints, are read normally from the card at 2 Hz.
  receipt -> results/gpu-5090-laptop-24g-ups-gate-95w.txt
clock lock [by the CARD]: OFF-BY-CARD
  evidence: clocks.sm was observed at 180 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 180, max 180, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force.
  the unit ai-perf.service reads active — NOT EVIDENCE:
    it is Type=oneshot, so it reads 'active (exited)' for the whole uptime whether or
    not 'nvidia-smi -rgc' has released the lock since. On 2026-09-21 a receipt that
    trusted it published 'ran with the lock ACTIVE' beside clocks that proved otherwise.
clock-lock receipt -> results/gpu-5090-laptop-24g-clock-lock-OFF-BY-CARD-95w.txt

=== instance: one card, kv q8_0, parallel 1 === 18:43:24Z
  "persistence_mode_at_start": "0, Enabled;",
  "pcie_link_width_at_start": "0, 8;"
}

next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status
pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq8_0-p1.json to every bencher so the result files carry it
cap read-back [card-mapping open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clocks at [card-mapping] open: 1987,9001,3090  (sm,mem,max.sm MHz)

=== card mapping: watch which physical card's memory rises === 18:43:29Z
  one: MATCHES: declared soldered (mobile; no socket — link width is traced under load, never assumed) at 00000000:01:00.0 - vendor unread - GPU-edff232c, and that is the card whose memory grew

every shape checked runs on the card it declares
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-card-mapping-95w.json
clocks at [card-mapping] close: 1590,9001,3090  (sm,mem,max.sm MHz)
cap read-back [card-mapping close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
cap read-back [armA-ladder-gemma4 open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clocks at [armA-ladder-gemma4] open: 1590,9001,3090  (sm,mem,max.sm MHz)

=== armA-ladder-gemma4 === 18:43:38Z
    run 1: decode=118.717 tok/s ttft=357.76 ms W=91.89 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 18845.0}
  contention gate: busy: busiest card mean 95.0% > 5.0% -- waiting (attempt 1/6)
  contention gate: busy: busiest card mean 90.2% > 5.0% -- waiting (attempt 2/6)
    run 2: decode=119.163 tok/s ttft=401.65 ms W=74.97 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 18845.0}
    run 3: decode=118.366 tok/s ttft=407.05 ms W=77.62 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 18845.0}

  --- num_ctx 98,304 ---
    fits: fully resident on the card; per-card delta {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 500.0}
    per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19345.0}; runner RSS 1.48 GiB
    run 1: decode=119.032 tok/s ttft=376.85 ms W=78.35 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19345.0}
    run 2: decode=118.129 tok/s ttft=339.37 ms W=79.93 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19345.0}
    run 3: decode=117.938 tok/s ttft=403.77 ms W=93.41 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19345.0}

  --- num_ctx 131,072 ---
    fits: fully resident on the card; per-card delta {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 500.0}
    per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19845.0}; runner RSS 1.52 GiB
    run 1: decode=119.239 tok/s ttft=402.18 ms W=76.95 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19845.0}
    run 2: decode=118.424 tok/s ttft=372.49 ms W=79.64 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19845.0}
    run 3: decode=117.802 tok/s ttft=355.78 ms W=94.22 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19845.0}

the biggest context this arm holds for gemma4:26b is 131,072 tokens
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m5-one-95w-auto-ladder.json
clocks at [armA-ladder-gemma4] close: 1597,14001,3090  (sm,mem,max.sm MHz)
cap read-back [armA-ladder-gemma4 close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
arm A's headline rung at 95 W, read from its own result file: num_ctx 131072
cap read-back [armA2-kv-axis open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clocks at [armA2-kv-axis] open: 1590,9001,3090  (sm,mem,max.sm MHz)

=== A2 instance kv=f16 === 18:52:00Z
}

next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status
pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvf16-p1.json to every bencher so the result files carry it

=== armA2-kvf16-at131072 === 18:52:35Z
prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101
base=http://127.0.0.1:11470 arm=one live=False runs=3

### gpu-5090-laptop-24g m5 gemma4:26b -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ###
  memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead.
  UPS before arm: load=None% status=None
  keep-alive: bench instance: held for the arm's duration
  /api/ps: size=17477150965 size_vram=17477150965 -> fits: fully resident on the card
  per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20696.0}
  run 1: decode=123.017 tok/s (wall 122.35) ttft=381.96 ms W=76.53 J/1k=622.11 temp=46.0C fan=None% clk=765.0-1927.0MHz mem=14001.0MHz
           NEITHER cap: the SM clock varied with draw at 80.6% of the cap and the card at 46 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it
  contention gate: busy: busiest card mean 83.7% > 5.0% -- waiting (attempt 1/6)
  run 2: decode=121.379 tok/s (wall 120.747) ttft=435.96 ms W=72.76 J/1k=599.44 temp=48.0C fan=None% clk=712.0-2032.0MHz mem=14001.0MHz
           NEITHER cap: the SM clock varied with draw at 76.6% of the cap and the card at 48 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it
  run 3: decode=120.822 tok/s (wall 120.167) ttft=438.79 ms W=91.13 J/1k=754.25 temp=50.0C fan=None% clk=997.0-1695.0MHz mem=14001.0MHz
           the POWER CAP doing its job: the SM clock fell while draw sat at 95.9% of the 95 W cap, at 50 C -- well below any throttling temperature
  spread gate: spread 1.8% of the median within the 15% gate
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m5-one-95w-kvf16-at131072.json

wrote 1 file(s)
kv=f16 holds num_ctx 131072 at 95 W -- that is this KV type's largest window

=== A2 instance kv=q4_0 === 18:53:47Z
}

next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status
pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq4_0-p1.json to every bencher so the result files carry it

=== armA2-kvq4_0-at131072 === 18:54:22Z
prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101
base=http://127.0.0.1:11470 arm=one live=False runs=3

### gpu-5090-laptop-24g m5 gemma4:26b -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ###
  memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead.
  UPS before arm: load=None% status=None
  keep-alive: bench instance: held for the arm's duration
  /api/ps: size=17747851344 size_vram=17747851344 -> fits: fully resident on the card
  per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19116.0}
  run 1: decode=116.716 tok/s (wall 116.111) ttft=326.27 ms W=86.87 J/1k=744.29 temp=55.0C fan=None% clk=840.0-1995.0MHz mem=14001.0MHz
           the POWER CAP doing its job: the SM clock fell while draw sat at 91.4% of the 95 W cap, at 55 C -- well below any throttling temperature
  run 2: decode=115.846 tok/s (wall 115.237) ttft=320.25 ms W=88.05 J/1k=760.06 temp=56.0C fan=None% clk=795.0-1755.0MHz mem=14001.0MHz
           the POWER CAP doing its job: the SM clock fell while draw sat at 92.7% of the 95 W cap, at 56 C -- well below any throttling temperature
  contention gate: busy: busiest card mean 32.5% > 5.0% -- waiting (attempt 1/6)
  run 3: decode=116.125 tok/s (wall 115.503) ttft=386.68 ms W=78.49 J/1k=675.91 temp=56.0C fan=None% clk=1027.0-1905.0MHz mem=14001.0MHz
           NEITHER cap: the SM clock varied with draw at 82.6% of the cap and the card at 56 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it
  spread gate: spread 0.7% of the median within the 15% gate
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m5-one-95w-kvq4_0-at131072.json

wrote 1 file(s)
kv=q4_0 holds num_ctx 131072 at 95 W -- that is this KV type's largest window
clocks at [armA2-kv-axis] close: 1590,14001,3090  (sm,mem,max.sm MHz)
cap read-back [armA2-kv-axis close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

=== restore the q8_0 instance (the seat posture) === 18:55:35Z

next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status
pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq8_0-p1.json to every bencher so the result files carry it
cap read-back [armA3-filled open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clocks at [armA3-filled] open: 1590,9001,3090  (sm,mem,max.sm MHz)

=== armA3-filled === 18:56:10Z
prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101
base=http://127.0.0.1:11470 arm=one live=False runs=3
  memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead.

### gpu-5090-laptop-24g m5 gemma4:26b -- FILLED WINDOW at num_ctx 131,072 (75% target = 98,304 tokens) ###
  fit at the window: fits: fully resident on the card
  per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 19845.0}
    filler round 1: 183 copies -> 97992 tokens (target 98304)
  window filled: 97,992 tokens of 131,072 (74.8% occupied)
  COLD prefill: 1734.777 tok/s over 98025 tokens, ttft 57703.64 ms
  contention gate: busy: busiest card mean 14.2% > 5.0% -- waiting (attempt 1/6)
  run 1: prefill=1128641.028 tok/s (97992 tokens) ttft=807.35 ms decode=56.174 tok/s W=88.38
  run 2: prefill=1430477.497 tok/s (97992 tokens) ttft=744.92 ms decode=56.039 tok/s W=92.16
  run 3: prefill=1421822.403 tok/s (97992 tokens) ttft=708.56 ms decode=56.374 tok/s W=86.48

  prompt tokens: 97992 of 131072 (0.7476)
  prefill: 1421822.403 tok/s   ttft: 744.92 ms   decode at depth: 56.174 tok/s
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m5-one-95w-filled.json
clocks at [armA3-filled] close: 1597,14001,3090  (sm,mem,max.sm MHz)
cap read-back [armA3-filled close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
cap read-back [armD-ladder-mistral open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clocks at [armD-ladder-mistral] open: 1590,14001,3090  (sm,mem,max.sm MHz)

=== armD-ladder-mistral === 18:59:24Z
    run 2: decode=28.815 tok/s ttft=223.65 ms W=89.92 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 18865.0}
  contention gate: busy: busiest card mean 14.8% > 5.0% -- waiting (attempt 1/6)
    run 3: decode=28.851 tok/s ttft=197.14 ms W=89.92 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 18865.0}

  --- num_ctx 65,536 ---
    fits: fully resident on the card; per-card delta {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 1440.0}
    per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20305.0}; runner RSS 0.89 GiB
  contention gate: busy: busiest card mean 10.4% > 5.0% -- waiting (attempt 1/6)
    run 1: decode=28.672 tok/s ttft=210.94 ms W=92.53 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20305.0}
  contention gate: busy: busiest card mean 99.0% > 5.0% -- waiting (attempt 1/6)
  contention gate: busy: busiest card mean 89.1% > 5.0% -- waiting (attempt 2/6)
    run 2: decode=28.874 tok/s ttft=219.98 ms W=90.24 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20305.0}
  contention gate: busy: busiest card mean 79.2% > 5.0% -- waiting (attempt 1/6)
    run 3: decode=28.839 tok/s ttft=199.55 ms W=90.31 mem/card={'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 20305.0}

  --- num_ctx 98,304 ---
    does not fit: only part of the model is on the card (92.2% VRAM / 7.8% RAM); per-card delta {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 2550.0}
    per-card memory.used (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 22855.0}; runner RSS 14.18 GiB
  ladder stops at num_ctx 98,304: does not fit: only part of the model is on the card (92.2% VRAM / 7.8% RAM)

the biggest context this arm holds for mistral-small3.2:24b is 65,536 tokens
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m4-one-95w-auto-ladder.json
clocks at [armD-ladder-mistral] close: 1590,14001,3090  (sm,mem,max.sm MHz)
cap read-back [armD-ladder-mistral close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
cap read-back [armD-forced-probe open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clocks at [armD-forced-probe] open: 1597,14001,3090  (sm,mem,max.sm MHz)

=== D-probe: the planner SPILLED at num_ctx 98304 -- re-running it forced === 19:09:33Z

=== armD-forced-at98304 === 19:09:33Z
prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101
base=http://127.0.0.1:11470 arm=one live=False runs=3

### gpu-5090-laptop-24g m4 mistral-small3.2:24b -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ###
  memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead.
  UPS before arm: load=None% status=None
  keep-alive: bench instance: held for the arm's duration
  /api/ps: size=23109042175 size_vram=23109042175 -> fits: fully resident on the card
  per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 23152.0}
  run 1: decode=28.721 tok/s (wall 28.603) ttft=201.99 ms W=93.25 J/1k=3246.78 temp=46.0C fan=None% clk=367.0-1762.0MHz mem=9001.0MHz
           the POWER CAP doing its job: the SM clock fell while draw sat at 98.2% of the 95 W cap, at 46 C -- well below any throttling temperature
  contention gate: busy: busiest card mean 14.8% > 5.0% -- waiting (attempt 1/6)
  run 2: decode=28.851 tok/s (wall 28.733) ttft=214.37 ms W=93.13 J/1k=3227.99 temp=45.0C fan=None% clk=322.0-1755.0MHz mem=9001.0MHz
           the POWER CAP doing its job: the SM clock fell while draw sat at 98.0% of the 95 W cap, at 45 C -- well below any throttling temperature
  contention gate: busy: busiest card mean 99.0% > 5.0% -- waiting (attempt 1/6)
  contention gate: busy: busiest card mean 79.2% > 5.0% -- waiting (attempt 2/6)
  run 3: decode=28.993 tok/s (wall 28.872) ttft=222.6 ms W=90.14 J/1k=3109.06 temp=44.0C fan=None% clk=300.0-1822.0MHz mem=9001.0MHz
           the POWER CAP doing its job: the SM clock fell while draw sat at 94.9% of the 95 W cap, at 44 C -- well below any throttling temperature
  spread gate: spread 0.9% of the median within the 15% gate
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m4-one-95w-forced-at98304.json

wrote 1 file(s)

=== D-probe: the forced rung FIT -- climbing one more to find the first that does not === 19:11:29Z

=== armD-forced-at131072 === 19:11:29Z
prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101
base=http://127.0.0.1:11470 arm=one live=False runs=3

### gpu-5090-laptop-24g m4 mistral-small3.2:24b -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ###
  memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead.
  UPS before arm: load=None% status=None
  keep-alive: bench instance: held for the arm's duration
  /api/ps: size=None size_vram=None -> refused: /api/ps reported no size for this model
  per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 0.0}
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-m4-one-95w-forced-at131072.json

wrote 1 file(s)
clocks at [armD-forced-probe] close: 1590,9001,3090  (sm,mem,max.sm MHz)
cap read-back [armD-forced-probe close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
cap read-back [armE-concurrency open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clocks at [armE-concurrency] open: 1590,9001,3090  (sm,mem,max.sm MHz)

=== E instance: parallel 4 === 19:11:41Z

next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status
pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq8_0-p4.json to every bencher so the result files carry it

=== E: concurrency 1 / 2 / 4 at 95 W === 19:12:18Z

### concurrency level 2 ###
  level 2: sum 198.006 tok/s, wall 113.446 tok/s over 4.5132s, W=94.6 temp=42.0C fan=None%
    per stream: [101.631, 96.375]
  level 2: sum 200.38 tok/s, wall 167.331 tok/s over 3.0598s, W=84.38 temp=43.0C fan=None%
    per stream: [100.189, 100.191]
  level 2: sum 199.413 tok/s, wall 165.321 tok/s over 3.097s, W=83.14 temp=43.0C fan=None%
    per stream: [99.703, 99.71]

### concurrency level 4 ###
  contention gate: busy: busiest card mean 37.2% > 5.0% -- waiting (attempt 1/6)
  level 4: sum 275.732 tok/s, wall 235.458 tok/s over 4.349s, W=78.34 temp=43.0C fan=None%
    per stream: [69.324, 69.315, 67.784, 69.309]
  level 4: sum 277.023 tok/s, wall 161.574 tok/s over 6.3376s, W=80.75 temp=43.0C fan=None%
    per stream: [69.257, 69.255, 69.255, 69.256]
  level 4: sum 275.117 tok/s, wall 231.239 tok/s over 4.4283s, W=91.77 temp=44.0C fan=None%
    per stream: [68.78, 68.779, 68.779, 68.779]

one-95w takes 4 parallel streams at 275.7 tok/s aggregate (sum of streams) / 231.2 tok/s over wall
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-concurrency-one-95w.json
clocks at [armE-concurrency] close: 1590,14001,3090  (sm,mem,max.sm MHz)
cap read-back [armE-concurrency close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
cap read-back [g2-doorman open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clocks at [g2-doorman] open: 1590,9001,3090  (sm,mem,max.sm MHz)

=== G2: the doorman on mistral at num_ctx 65536 (1 caller, then 4) === 19:15:15Z
  memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead.
  fit: SPLIT and scored as one: 56.2% VRAM / 43.8% RAM
  contention gate: busy: busiest card mean 8.2% > 5.0% -- waiting (attempt 1/6)
  concurrency 1: 10 calls, median 461.64 ms, p95 493.52 ms, max 493.52 ms, W/card {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 27.81}
  concurrency 4: 8 calls, median 928.54 ms, p95 4671.24 ms, max 4671.24 ms, W/card {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 66.59}

a doorman call on mistral-small3.2:24b costs 461.64 ms at the median and 493.52 ms at p95, one caller at a time
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-g2-doorman-95w.json
clocks at [g2-doorman] close: 1590,9001,3090  (sm,mem,max.sm MHz)
cap read-back [g2-doorman close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

=== G3/G4 instance: back to parallel 1 === 19:16:20Z

next: /workshop/bench-laptop-5090-2026-09-21/harness-llm/setup_laptop_instance.sh --shape one --status
pass --instance-env-file /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/instance-env-one-kvq8_0-p1.json to every bencher so the result files carry it
cap read-back [g3-vision open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clocks at [g3-vision] open: 1590,9001,3090  (sm,mem,max.sm MHz)

=== G3: a vision seat (TEXT path only -- no image is read or sent by this bench) === 19:16:48Z

=== g3-vision === 19:16:48Z
prompt sha256=90eedd0c53f9554ae3837674504fcb7090013d9a432a653352183c0f25a7ce5c bytes=2101
base=http://127.0.0.1:11470 arm=one live=False runs=3

### gpu-5090-laptop-24g g3 minicpm-v4.5:latest -- arm one (the one NVIDIA GeForce RTX 5090 Laptop GPU 24 GB soldered to this machine's board, everything ollama's planner will put on it) ###
  memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead.
  UPS before arm: load=None% status=None
  keep-alive: bench instance: held for the arm's duration
  /api/ps: size=5596931685 size_vram=5596931685 -> fits: fully resident on the card
  per-card VRAM delta (MiB): {'GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff': 6696.0}
  run 1: decode=86.808 tok/s (wall 86.021) ttft=147.09 ms W=79.27 J/1k=913.16 temp=41.0C fan=None% clk=892.0-2002.0MHz mem=14001.0MHz
           NEITHER cap: the SM clock varied with draw at 83.4% of the cap and the card at 41 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it
  contention gate: busy: busiest card mean 68.6% > 5.0% -- waiting (attempt 1/6)
  run 2: decode=85.795 tok/s (wall 85.004) ttft=150.97 ms W=57.18 J/1k=666.47 temp=41.0C fan=None% clk=487.0-1912.0MHz mem=14001.0MHz
           NEITHER cap: the SM clock varied with draw at 60.2% of the cap and the card at 41 C. On a memory-bound decode the clock follows the work, and nothing here was limiting it
  run 3: decode=87.364 tok/s (wall 86.539) ttft=149.86 ms W=87.84 J/1k=1005.45 temp=41.0C fan=None% clk=427.0-1447.0MHz mem=14001.0MHz
           the POWER CAP doing its job: the SM clock fell while draw sat at 92.5% of the 95 W cap, at 41 C -- well below any throttling temperature
  spread gate: spread 1.8% of the median within the 15% gate
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-g3-one-95w-g3-vision.json

wrote 1 file(s)
clocks at [g3-vision] close: 1582,14001,3090  (sm,mem,max.sm MHz)
cap read-back [g3-vision close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
cap read-back [g4-embed open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clocks at [g4-embed] open: 1590,9001,3090  (sm,mem,max.sm MHz)

=== G4: embeddings === 19:17:59Z
  memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead.
  keep-alive: bench instance: held for the arm's duration
  batch 1: 251.04 ms, 254.936 texts/s, W=15.69
  batch 2: 257.52 ms, 248.521 texts/s, W=16.81
  batch 3: 253.4 ms, 252.57 texts/s, W=16.86
  batch 4: 254.37 ms, 251.604 texts/s, W=16.52
  batch 5: 247.32 ms, 258.773 texts/s, W=16.09

one-95w: 252.6 texts/s on a batch of 64 (253.4 ms per batch)
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-embed-one-95w.json
clocks at [g4-embed] close: 1867,14001,3090  (sm,mem,max.sm MHz)
cap read-back [g4-embed close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
cap read-back [idle open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
clocks at [idle] open: 1590,14001,3090  (sm,mem,max.sm MHz)

=== idle receipt at 95 W === 19:18:24Z
  memory-temperature sensor: the memory die's temperature is NOT readable on these cards through any of the three paths asked, so the memory-temperature stop condition could not arm and NO memory temperature is reported anywhere in this bench. The core temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument instead.
empty: no model on the cards
  cards empty, nothing loaded: per card soldered (mobile; no socket — link width is traced under load, never assumed) at 00000000:01:00.0 - vendor unread - GPU-edff232c 7.18 W | pair 7.18 W | UPS whole box None W (None %)
loading gemma4:26b (num_ctx=131072, num_gpu=None)
  fit: fits: fully resident on the card
  model resident, not generating: per card soldered (mobile; no socket — link width is traced under load, never assumed) at 00000000:01:00.0 - vendor unread - GPU-edff232c 9.03 W | pair 9.03 W | UPS whole box None W (None %)

with gemma4:26b resident and not generating, the card draws 9.03 W (board power), against 7.18 W with the cards empty -- a resident cost of 1.85 W
  -> /workshop/bench-laptop-5090-2026-09-21/harness-llm/results/gpu-5090-laptop-24g-idle-gemma4-95w.json
clocks at [idle] close: 180,9001,3090  (sm,mem,max.sm MHz)
cap read-back [idle close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

=== 95 W block finished === 19:19:54Z
stopping bench-5090laptop-cpu-box
--- receipt: units ---
--- receipt: ports ---
no bench port listening (1147[01]) -- clean
index, uuid, power.limit [W], enforced.power.limit [W], persistence_mode, temperature.gpu, fan.speed [%], memory.used [MiB], clocks.current.sm [MHz], clocks.current.memory [MHz], clocks.max.sm [MHz]
0, GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff, [N/A], 95.00 W, Enabled, 34, [N/A], 15 MiB, 682 MHz, 810 MHz, 3090 MHz
cap read-back [block 95w close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
the UPS is absent on this box -- see results/gpu-5090-laptop-24g-ups-gate-95w.txt for the
withheld-figure receipt written at the block's open. Board watts are the measured
quantity in every table this bench fills.

LEG 1 at 95w exited 0
cap read-back [between the legs at 95w]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

---- LEG 2: /workshop/bench-laptop-5090-2026-09-21/harness-render/run_rung.sh 95 ----
== render rung 95w · started 2026-09-21T19:19:55Z ==
   card expected   NVIDIA GeForce RTX 5090 Laptop GPU
   comfy root      /workshop/bench-laptop-5090-2026-09-21/ComfyUI-0.21.1
   comfy python    /workshop/ComfyUI/.venv/bin/python
   comfy version   0.21.1 (26515acd)
receipt: nvidia-powerd is inactive — Dynamic Boost is not moving the cap
   card reads      0, 00000000:01:00.0, NVIDIA GeForce RTX 5090 Laptop GPU, [N/A], 95.00 W, 24463 MiB, 8, 16, 1192 MHz, 810 MHz
cap read-back [render rung 95w open]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff
   clock lock [by the CARD]: OFF-BY-CARD
     evidence: clocks.sm was observed at 180 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 180, max 1192, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force.
     the unit ai-perf.service reads active — NOT EVIDENCE:
       it is Type=oneshot, so it reads 'active (exited)' for the whole uptime whether or
       not 'nvidia-smi -rgc' has released the lock since. On 2026-09-21 a receipt that
       trusted it published 'ran with the lock ACTIVE' beside clocks that proved otherwise.
   clock lock      unprobed by the CARD (when the lock is in force
                   this board cannot clock below 1200 MHz and a power posture means
                   something different from what it means on the unlocked desktop cards
                   this leg compares to). The unit state is not evidence; see above.
   port band       18190-18199 free; this box's own comfyui.service is inactive

-- arm inventory · suffix inventory-95w · 2026-09-21T19:20:01Z
  -> results/inventory-inventory-95w.json
-- arm inventory exited 0 · 2026-09-21T19:20:01Z
cap read-back [after inventory at 95w]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

-- arm r1 · suffix r1-95w · 2026-09-21T19:20:01Z
  card available at 2026-09-21T19:20:01Z · the image service holds 0 MiB · quiet: mean 0.0% <= 5.0%
  cap 95.0 W per card · UPS load unreadable
  R1 klein-4b-8st on index 0 (an NVIDIA GeForce RTX 5090 Laptop GPU 24 GB, SOLDERED to this machine — no slot, no partner, and no card that can replace it for a control, soldered (mobile; no socket — link width is traced under load))
    r1-r1-95w-klein-4b-8st-c0-b3-warm[0] 15.588s exec / 3 img = 5.196s per image
    r1-r1-95w-klein-4b-8st-c0-b3-timed[0] 4.750s exec / 3 img = 1.583s per image
    r1-r1-95w-klein-4b-8st-c0-b3-timed[1] 5.093s exec / 3 img = 1.698s per image
  R1 klein-4b-32st on index 0 (an NVIDIA GeForce RTX 5090 Laptop GPU 24 GB, SOLDERED to this machine — no slot, no partner, and no card that can replace it for a control, soldered (mobile; no socket — link width is traced under load))
    r1-r1-95w-klein-4b-32st-c0-b3-warm[0] 37.655s exec / 3 img = 12.552s per image
    r1-r1-95w-klein-4b-32st-c0-b3-timed[0] 18.083s exec / 3 img = 6.028s per image
    r1-r1-95w-klein-4b-32st-c0-b3-timed[1] 18.434s exec / 3 img = 6.145s per image
  R1 klein-4b-8st-pixel4 on index 0 (an NVIDIA GeForce RTX 5090 Laptop GPU 24 GB, SOLDERED to this machine — no slot, no partner, and no card that can replace it for a control, soldered (mobile; no socket — link width is traced under load))
    r1-r1-95w-klein-4b-8st-pixel4-c0-b3-warm[0] 6.676s exec / 3 img = 2.225s per image
    r1-r1-95w-klein-4b-8st-pixel4-c0-b3-timed[0] 4.736s exec / 3 img = 1.579s per image
    r1-r1-95w-klein-4b-8st-pixel4-c0-b3-timed[1] 5.026s exec / 3 img = 1.675s per image
  -> results/r1-r1-95w.json
-- arm r1 exited 0 · 2026-09-21T19:24:03Z
cap read-back [between R1 and R4 at 95w]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

-- arm r4 · suffix r4-95w · 2026-09-21T19:24:03Z
  card available at 2026-09-21T19:24:03Z · the image service holds 0 MiB · quiet: mean 0.0% <= 5.0%
  cap 95.0 W per card · UPS load unreadable
    r4-c0-warm[0] 4.195s exec / 1 img = 4.195s per image
    r4-c0-warm[1] 1.907s exec / 1 img = 1.907s per image
    r4-c0-warm[2] 1.914s exec / 1 img = 1.914s per image
  -> results/r4-r4-95w.json
-- arm r4 exited 0 · 2026-09-21T19:34:29Z
cap read-back [render rung 95w close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

== render rung 95w · done 2026-09-21T19:34:29Z (R1 exit 0, R4 exit 0) ==
   results/r1-95w.json and results/r4-95w.json
   the UPS is absent on this box: no `upsc` on this box: NUT is not installed and nut-server/nut-monitor are inactive (2026-09-21T14:52Z). Whole-box watts are WITHHELD with this reason; board watts are read normally.
   the fan is unreadable on this board: fan.speed reads [N/A] on this board: the driver reports no fan for it. The fan-drawing cell is a declared non-figure with this reason — arm R4 measures ten minutes of sustained drawing and cannot report a fan curve for it.
LEG 2 at 95w exited 0
cap read-back [rung 95w close]: 95 W, read from enforced.power.limit, on GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff

artifact receipt -> /workshop/bench-laptop-5090-2026-09-21/RUNG-95W-RECEIPT.txt
PASS 95w — COMPLETE
  posture        fixed
  cap argument   95
  written        2026-09-21T19:34:30Z
  box_class      gpu-5090-laptop-24g
  card           GPU-edff232c-7dbf-2bac-07fb-921a7c9eecff  (NVIDIA GeForce RTX 5090 Laptop GPU)
  cap at close   95 enforced.power.limit
  clocks         1590,9001,3090  (sm,mem,max.sm MHz)
  clock lock     OFF-BY-CARD, read FROM THE CARD
                 clocks.sm was observed at 180 MHz, BELOW the lock's floor of 1200 MHz, over 12 reads (min 180, max 180, clocks.max.sm 3090). A locked card cannot clock below its floor, so the lock is not in force.
                 (the unit ai-perf.service reads
                 active, which is NOT evidence: it is a
                 oneshot and stays 'active' for the uptime after -rgc.)
  nvidia-powerd  inactive
  LEG 1 exit     0
  LEG 2 exit     0
  PASS exit      0

  LEG 1 result files carrying 95w in their name:
    ad959060e1944079  gpu-5090-laptop-24g-card-mapping-95w.json
    0aeb1f2bd9be3294  gpu-5090-laptop-24g-clock-lock-OFF-BY-CARD-95w.txt
    8dcace86350d584c  gpu-5090-laptop-24g-concurrency-one-95w.json
    ba9a0f2f6fec7b4f  gpu-5090-laptop-24g-embed-one-95w.json
    0547543c54b048f0  gpu-5090-laptop-24g-g2-doorman-95w.json
    704d63148fde16c6  gpu-5090-laptop-24g-g3-one-95w-g3-vision.json
    be924658df8af88f  gpu-5090-laptop-24g-idle-gemma4-95w.json
    50f8194b1117b66f  gpu-5090-laptop-24g-m4-one-95w-auto-ladder.json
    21ad21d27f0b043d  gpu-5090-laptop-24g-m4-one-95w-forced-at131072.json
    521250d4c07e69c6  gpu-5090-laptop-24g-m4-one-95w-forced-at98304.json
    a81bbb7af2e8e82d  gpu-5090-laptop-24g-m5-one-95w-auto-ladder.json
    b20c7b16b17c9eee  gpu-5090-laptop-24g-m5-one-95w-filled.json
    2188c0db39fb316f  gpu-5090-laptop-24g-m5-one-95w-kvf16-at131072.json
    9e33818a11812070  gpu-5090-laptop-24g-m5-one-95w-kvq4_0-at131072.json
    6a95e1ab9188cd1e  gpu-5090-laptop-24g-ups-gate-95w.txt

  LEG 2 result files carrying 95w in their name:
    3334b28179ca3d77  inventory-inventory-95w.json
    131e3c4238f41039  r1-r1-95w.json
    9215364be7285ec8  r4-r4-95w.json

  stated absences at this pass:
    UPS          no `upsc` on this box: NUT is not installed, nut-server and nut-monitor are inactive, the battery exposes no power_now/energy_now, and /sys/class/powercap/*/energy_uj is root-only (all read 2026-09-21T14:52Z). Whole-box watts are WITHHELD with this reason. Board watts — which is what every table in this ladder prints — are read normally at 2 Hz.
    fan          fan.speed reads [N/A] on this board (2026-09-21T14:52Z): the driver reports no fan for it. The fan column is a declared non-figure with this reason, not a gap to be filled later.
    memory temp  temperature.memory reads [N/A], as on every board of this ladder. The core-temperature stop and the driver's own thermal-slowdown reasons are the thermal instrument.
    arm T        the training arm's preprocessed tensor set lives on benchbox, which
                 was unreachable from 2026-09-21T15:03Z. Arm T is a SECOND TRAIN and
                 legs 1 and 2 never depended on it.
rung 95 W finished rc=0 at 2026-09-21T19:34:30Z
memory guard restored: memshed.timer is active (it was active); MEMSHED_DRY_RUN unset
  free receipt: journalctl --user -u memshed --since '2026-09-21T18:43:09Z' | grep SHEDDING
  — every line there is a shed this bench WOULD have taken, with its signature.