# The benchbox 3090 reading 0: the cost-of-intelligence instrument on one GeForce RTX 3090 24 GB at 300 W

**What this is.** The 3090 reading of the cost-of-intelligence instrument (RULINGS-2026-09-23,
R-0 to R-12). One GeForce RTX 3090 24 GB (`<withheld: GPU UUID of board A>`, the only
board on benchbox's bus, x16 at `withheld-2026-10-14`) reads the Jev bench's frozen six-way
(task_c, 108 items, series set S0) and the doorman's planted 48, through the Jev bench's own
`run.py`. It ran the calibration first (K0), then the smoke's 2023 model (K1), then a 24 GB
backfill from 2023 to now, one model at a time, then the reasoning-on arms (R-11). Each model
was pulled, read, then deleted.

**Window (UTC, `date -u` on benchbox).** The instance started at 2026-09-23T16:01:54Z and the
dry run at 16:06:03Z. The first measured arm (K0) began at 2026-09-23T16:09:53Z. The last arm,
`t-gemma4-31b-on` (queue line 154), ended at 2026-09-24T03:21:06Z; its model was deleted by
03:21:08Z, and the runner wrote `QUEUE EMPTY` at 2026-09-24T03:21:09Z (`queue/queue.log`). The
log records 137 step ends: 136 at rc 0 and one at rc 5, the arm `b-nemotron3-nano-30b`, which
did not fit whole on the card (§3 and §5).

**Posture, on every row.** an operator ruled 300 W with persistence mode on (R-12, ~16:0xZ). Every
row carries `x_cap_w` and `x_persistence`. Each task's table row also carries the enforced
limit and persistence mode as read from the card at that task's start. An arm refuses (rc 4)
unless the card reads `300.00 W` and `Enabled`, and no arm refused on posture. Box class `gpu-coi3090-24g`
passed the master harness's shape and prefix preflight (`benchlib.check_box_class`, estate
`51be5b25`).

**Two instruments, never mixed (R-6).**
- **BOARD** is `nvidia-smi power.draw` by UUID at 1 Hz, from run.py's own `Boards` sampler.
  It is the energy of record.
- **WALL** is the rack UPS's `ups.realpower` for the whole box (CyberPower CP1500PFCLCDa via
  NUT `usbhid-ups`), polled at 1 Hz by the driver's `Wall` thread (`upsc`).

The WALL number is coarse. NUT's quick poll runs every 2 s (`pollinterval 2`), but
`ups.realpower` only moves on the driver's full poll (`pollfreq 12`). Across 52 arm
files, the value changed a median of every 14.0 s (p10 6.0 s, p90 28.0 s;
`rows/wall-cadence.json`). So a WALL figure is a step function about 14 s wide that lags the
load. It is honest over a task window of a minute or more, and it is not a per-item figure.
One row shows the lag outright. The doorman window of `b-mistral-7b-v0.3` lasted 13 s, and it
reads a WALL mean of 178.1 W against a BOARD mean of 234.4 W: the whole box below its own card
(`rows/figures.json`). Read no WALL figure from a window under a minute.
The WALL "idle, model resident" reading lags the load by one or two steps, which is why it can
read above the no-model idle by more than the board's own idle delta.

**Verdict.** K0 reproduced the Jev bench's own control in its answers, but not in its joules.
- **The answers.** K0 made the same choice as the 2026-09-21 row on all 108 items, with
  byte-identical reply text on all 108. Both sides score 89 by run.py's count and 88 strict
  (`rows/k0-vs-jev.json`).
- **The joules, which disagree in direction.** BOARD joules per strict-correct answer read
  176.3 today against 194.3 then, which is lower. Net of the resident idle, the same figure
  reads 46.9 against 37.5, which is higher. The card idled at 106.4 W with the model resident
  today, against 120.9 W then (`rows/k0-vs-jev.json`: `board_j_per_correct_strict`,
  `net_j_per_correct_strict`, `idle_w`). This reading does not choose between the two.
- **K1 on a different card and server.** It reproduced the smoke's choices: the same choice on
  108 of 108 items, identical text on 100 of 108, and 10 correct by run.py's count on both.
  Time and energy did not carry over: a median of 0.248 s per decision against 0.135 s, and
  5,384 BOARD joules for the task against 1,534 (`rows/k1-vs-smoke.json`).



**What the report covers** (generated from `rows/figures.json`, one entry per `rows/<arm>/arm.json`): 36 backfill arms `b-*` (think off, or the lowest level a model offers), of which 35 read and 1 refused or failed (`b-nemotron3-nano-30b`); 9 reasoning-on arms `t-*` (R-11), of which 9 read and 0 refused or failed; 6 calibration arms `k0*` / `k1*` (K0, K0a, K1), of which 6 read and 0 refused or failed. That is 91 task rows and 38 probe windows. The dry run (`rows/dryrun-k0`) appears only in the residency table.


## 1. K0, the calibration: the Jev bench's own control, reproduced

The same tag and digest (`gemma4:26b`, `08ae7ec1744b`), the same kit, `run.py`, body, seed, `num_ctx` 131,072 and ollama instance environment (FLASH_ATTENTION on, f16 KV, keep-alive -1, Vulkan blanked). What differs is the posture and the box around the card (below). Joined item by item by index, with the id checked at every position (`rows/k0-vs-jev.json`, generated by `instrument/compare_k0.py`).

| | K0 today (benchbox, 1 card) | Jev arm 1, 2026-09-21 | file |
|---|---:|---:|---|
| started (UTC) | 2026-09-23T16:10:27Z | 2026-09-21T17:56:12Z | reports |
| cap (enforced, read) | 300 W, persistence Enabled | 250 W (`power_cap_w` on every row) | rows |
| run.py's count | 89 / 108 | 89 / 108 | reports |
| strict (parser v2-benchbox-1) | 88 / 108 | 88 / 108 | k0-vs-jev.json |
| unparsed (run.py) | 1 | 1 | reports |
| same choice, item by item | 108 / 108 | | k0-vs-jev.json |
| byte-identical reply text | 108 / 108 | | k0-vs-jev.json |
| median / p95 s per decision | 1.000 / 1.066 | 1.060 / 1.116 | reports |
| task wall seconds | 107.9 | 114.5 | reports |
| BOARD J, task window | 15,514 | 17,094 | watts.csv |
| BOARD mean W, task | 145.0 | 149.9 | reports |
| BOARD idle W, model resident (10 s before) | 106.4 | 120.9 | idle.csv |
| BOARD J per decision (net of resident idle) | 143.6 (38.2) | 158.3 (30.6) | reports |
| BOARD J per correct, strict (net) | 176.3 (46.9) | 194.3 (37.5) | k0-vs-jev.json |


**What differed from the 09-21 row, and only this:**
- **The cap:** 300 W enforced with persistence Enabled, against 250 W then.
- **The box:** one card on the bus. On 09-21 a second 3090 (`<withheld: GPU UUID of board E>`) sat idle beside
  it.
- **The slot:** withheld-2026-10-14, x16 gen 4. The 09-21 slot is disputed. The arm-2 receipt says
  withheld-2026-10-14, and the 09-22 shelf record says withheld-2026-10-14 (memory
  `benchbox-slots-and-clock-2026-09-22`). Every row keys on the UUID.
- **The ollama binary:** today it reads 0.32.13 (`receipts/substrate-2026-09-23.txt`). The Jev
  receipts do not restate their version, and benchbox ran 0.32.13 on 09-20.
- **The store:** this lane pulled the tag fresh into its own store, and its digest matches the
  Jev id.

**Verdict: K0 reproduces the 09-21 row, answer for answer and byte for byte.** Neither the cap,
the slot nor the neighbour card changed an answer. The board energy per correct answer is lower
today, and the net-of-idle figure is higher. That is because this card idled lower with the
model loaded, and it is a posture and box difference, not a model one.

Three instrument facts came out of K0:
1. **`/api/ps` misreports this tag's size on a single-board instance.** It read
   `size = size_vram = 1,105,555,619` B for `gemma4:26b` while the board grew by 21,449 MiB
   (`rows/k0-gemma4-26b-c/preflight.json`). The memory `api-ps-reports-a-moe-active-size` had
   left open whether the tiny ps size was the pair instance's alone. It is not: this instance is
   one board, Vulkan blanked, ollama 0.32.13. The fit rule here is the growth delta (≥ 90 % of
   the store bytes); ps is corroboration only.
2. **ollama charges a `load_duration` on every call to a resident model on benchbox.** The median
   is 778 ms for `gemma4:26b` and 125 ms for `mistral:v0.1`, against 33 ms for `mistral:v0.1` on
   the laptop (the `x_load_ms` field on every row). That is most of the Jev README's "870 ms
   per-request overhead". It dominates latency and the energy per decision for the short
   decision sets, so it is a property of this server and must never be carried to another one.
3. **The probe's 146.6 tok/s decode includes the bare tag's MTP draft model**
   (`draft_num_predict 3`). The anchor build below has no draft and reads 129.3 tok/s.



## 2. K1, the first cross-card point: the smoke's 2023 model on the GeForce RTX 3090 24 GB

`mistral:v0.1` (digest `b17615239298`, Mistral-7B-Instruct-v0.1, weights 2023-09-27), the six-way, `num_ctx` 4,096, against the smoke's pass 1 on the laptop's RTX 5090 Laptop GPU (`rows/k1-vs-smoke.json`, `instrument/compare_k1.py`).

| | benchbox GeForce RTX 3090 24 GB, 300 W | the laptop 5090 Laptop, boost posture (enforced 150 W) | file |
|---|---:|---:|---|
| run.py's count | 10 / 108 | 10 / 108 | reports |
| same choice, item by item | 108 / 108 | | k1-vs-smoke.json |
| byte-identical reply text | 100 / 108 | | k1-vs-smoke.json |
| median s per decision | 0.248 | 0.135 | reports |
| BOARD J, task window | 5,384 | 1,534 | watts.csv |
| BOARD mean W, task | 207.0 | 109.5 | reports |


**Verdict: the 2023 model reads the six-way the same on both cards.** It made the same choice
on all 108 items. The text differed on 8, all tails after the letter (the laptop runs a q8_0 KV
cache; this instance runs f16). So the capability axis of the instrument transfers across cards
and servers. Latency and joules do not: this card spent more time and more energy on the same
answers. Most of the time gap is benchbox's per-call `load_duration` (above), not the decode.



## 3. The curve on one GeForce RTX 3090 24 GB: the best six-way model per half-year (think off; strict)

Half-years come from each model's weights date: the Hugging Face `created` date of the verified repo in the recon's `models-verify-2026-09-23.json` (so the bare `gemma4:26b` tag sits in 2026H1 with its weights, not in the half it was re-pointed). The six-way is strict accuracy over all 108; the doorman column is strict over the 36 in-contract items. BOARD J per correct is the six-way's task window divided by its strict correct count.

| half | best six-way model (tag) | weights date | six-way strict (Wilson 95 %) | BOARD J per correct | WALL J per correct | doorman, 36 in contract (strict) | models read this half | cap / persistence | file |
|---|---|---|---:|---:|---:|---:|---:|---|---|
| 2023H1 | `vicuna:33b-q4_0` | 2023-06-21 | 17/108 = 15.7 % (10.1 %–23.8 %) | 1,133.7 | 1,497.8 | 20/36 | 2 | 300.00 W / Enabled | `b-vicuna-33b/b-vicuna-33b.c.jsonl` |
| 2023H2 | `yi:34b-chat-q4_0` | 2023-11-22 | 21/108 = 19.4 % (13.1 %–27.9 %) | 635.0 | 842.2 | 24/36 | 5 | 300.00 W / Enabled | `b-yi-34b/b-yi-34b.c.jsonl` |
| 2024H1 | `gemma2:27b-instruct-q4_K_M` | 2024-06-24 | 73/108 = 67.6 % (58.3 %–75.7 %) | 243.0 | 353.9 | 29/36 | 6 | 300.00 W / Enabled | `b-gemma2-27b/b-gemma2-27b.c.jsonl` |
| 2024H2 | `qwen2.5:32b-instruct-q4_K_M` | 2024-09-17 | 89/108 = 82.4 % (74.2 %–88.4 %) | 159.2 | 203.5 | 34/36 | 5 | 300.00 W / Enabled | `b-qwen2.5-32b/b-qwen2.5-32b.c.jsonl` |
| 2025H1 | `gemma3:27b-it-q4_K_M` | 2025-03-01 | 89/108 = 82.4 % (74.2 %–88.4 %) | 265.5 | 364.6 | 14/36 | 6 | 300.00 W / Enabled | `b-gemma3-27b/b-gemma3-27b.c.jsonl` |
| 2025H2 | `ministral-3:14b-instruct-2512-q4_K_M` | 2025-10-31 | 71/108 = 65.7 % (56.4 %–74.0 %) | 135.2 | 182.4 | 17/36 | 5 | 300.00 W / Enabled | `b-ministral3-14b/b-ministral3-14b.c.jsonl` |
| 2026H1 | `qwen3.5:27b-q4_K_M` | 2026-02-24 | 97/108 = 89.8 % (82.7 %–94.2 %) | 231.0 | 306.8 | 36/36 | 8 | 300.00 W / Enabled | `b-qwen3.5-27b/b-qwen3.5-27b.c.jsonl` |
| 2026H2 | `qwen3.8:27b-q4_K_M` | 2026-08-05 | 95/108 = 88.0 % (80.5 %–92.8 %) | 224.1 | 321.3 | 9/36 | 1 | 300.00 W / Enabled | `b-qwen3.8-27b/b-qwen3.8-27b.c.jsonl` |


**How to read it.** The strict parser counts a reply as correct only when it opens with the
option letter, so the 2023 chat models lose most of their points to format. Llama 2 writes
"The correct answer is [C]", and falcon writes "The best option is [B]". The lenient column in
§5 reads those letters and names, and it never enters a ratio. The format failure is part of
the measurement, because the product needs the letter. Both views are in the tables, with the
format classes counted.

**One cell has no reading (the rc 5).** `nemotron-3-nano:30b-a3b-q4_K_M` (weights 2025-12-04)
did not fit whole on 24 GB. The board grew by 99 % of the tag's store bytes, which passes the
growth rule, but `/api/ps` put 1,158,960,373 B of it on the CPU. The arm refused on fit 34 s
after it started, before any item was asked (`rows/b-nemotron3-nano-30b/REFUSED.json`; its
figures are in §5's refused-arms table). It is a largecard-only cell (PLAN-v2 §6, "The desktop's larger card's
rows"), and it is not a zero. It is in no accuracy, no energy figure, and not in 2025H2's count
of models read.

**One row is a protocol result, not a capability reading.** `deepseek-r1:32b-qwen-distill`
keeps reasoning when it is sent `think: false`. Its think-off arm returned empty content on all
108 six-way items and all 48 doorman items. Every one of them stopped at the 8-token cap, with
thinking text beside it: 2,387 and 917 characters in total (`rows/figures.json`:
`empty_replies`, `done_reasons`, `thinking_chars`). Its 0/108 in 2025H1's count is that result.
Its reasoning arm in §4 is the model's reading.



## 4. The price of thinking (R-11): each model's reasoning setting side by side

| model | task | setting | num_predict | strict | lenient | budget exhausted (stop: length) | tokens out per item (incl. reasoning) | median s | BOARD J per correct (strict / lenient) | BOARD mean W | cap / persistence | file |
|---|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---|---|
| `deepseek-r1:32b-qwen-distill-q4_K_M` | c | off | 8 | 0/108 | 0/108 | n/a (8-token cap) | 8.0 | 0.812 | — / — | 221.7 | 300.00 W / Enabled | `b-deepseek-r1-32b/b-deepseek-r1-32b.c.jsonl` |
| `deepseek-r1:32b-qwen-distill-q4_K_M` | c | on | 2048 | 1/108 | 76/108 | 3 | 731.3 | 19.836 | 696,094 / 9,159 | 295.4 | 300.00 W / Enabled | `t-deepseek-r1-32b-on/t-deepseek-r1-32b-on.c.jsonl` |
| `gemma4:12b-it-q4_K_M` | c | off | 8 | 33/108 | 79/108 | n/a (8-token cap) | 6.6 | 1.069 | 593 / 248 | 171.5 | 300.00 W / Enabled | `b-gemma4-12b/b-gemma4-12b.c.jsonl` |
| `gemma4:12b-it-q4_K_M` | c | on | 4096 | 32/108 | 46/108 | 59 | 2,616.9 | 59.361 | 38,022 / 26,450 | 294.5 | 300.00 W / Enabled | `t-gemma4-12b-on/t-gemma4-12b-on.c.jsonl` |
| `gemma4:12b-it-q4_K_M` | d | off | 8 | 35/48 | 35/48 | n/a (8-token cap) | 3.1 | 1.190 | 279 / 279 | 171.4 | 300.00 W / Enabled | `b-gemma4-12b/b-gemma4-12b.d.jsonl` |
| `gemma4:12b-it-q4_K_M` | d | on | 4096 | 31/48 | 31/48 | 9 | 1,395.3 | 12.497 | 9,422 / 9,422 | 290.6 | 300.00 W / Enabled | `t-gemma4-12b-on/t-gemma4-12b-on.d.jsonl` |
| `gemma4:26b` | c | off | 8 | 88/108 | 88/108 | n/a (8-token cap) | 3.7 | 1.000 | 176 / 176 | 145.0 | 300.00 W / Enabled | `k0-gemma4-26b-c/k0-gemma4-26b-c.c.jsonl` |
| `gemma4:26b` | c | on | 4096 | 24/108 | 73/108 | 30 | 2,033.9 | 6.346 | 13,819 / 4,543 | 284.9 | 300.00 W / Enabled | `k0-gemma4-26b-think/k0-gemma4-26b-think.c.jsonl` |
| `gemma4:26b` | d | off | 8 | 35/48 | 35/48 | n/a (8-token cap) | 2.9 | 1.108 | 246 / 246 | 162.4 | 300.00 W / Enabled | `k0-gemma4-26b-d/k0-gemma4-26b-d.d.jsonl` |
| `gemma4:26b` | d | on | 4096 | 35/48 | 35/48 | 1 | 677.8 | 3.703 | 1,562 / 1,562 | 261.5 | 300.00 W / Enabled | `k0-gemma4-26b-think/k0-gemma4-26b-think.d.jsonl` |
| `gemma4:31b-it-q4_K_M` | c | off | 8 | 94/108 | 98/108 | n/a (8-token cap) | 4.1 | 1.481 | 309 / 297 | 182.9 | 300.00 W / Enabled | `b-gemma4-31b/b-gemma4-31b.c.jsonl` |
| `gemma4:31b-it-q4_K_M` | c | on | 2048 | 73/108 | 93/108 | 6 | 748.4 | 19.437 | 10,093 / 7,922 | 290.0 | 300.00 W / Enabled | `t-gemma4-31b-on/t-gemma4-31b-on.c.jsonl` |
| `glm-4.7-flash:q4_K_M` | c | off | 8 | 44/108 | 44/108 | n/a (8-token cap) | 3.0 | 0.623 | 255 / 255 | 167.2 | 300.00 W / Enabled | `b-glm4.7-flash/b-glm4.7-flash.c.jsonl` |
| `glm-4.7-flash:q4_K_M` | c | on | 4096 | 67/108 | 67/108 | 23 | 1,924.6 | 10.965 | 7,846 / 7,846 | 290.5 | 300.00 W / Enabled | `t-glm4.7-flash-on/t-glm4.7-flash-on.c.jsonl` |
| `glm-4.7-flash:q4_K_M` | d | off | 8 | 23/48 | 25/48 | n/a (8-token cap) | 4.7 | 0.697 | 258 / 237 | 179.5 | 300.00 W / Enabled | `b-glm4.7-flash/b-glm4.7-flash.d.jsonl` |
| `glm-4.7-flash:q4_K_M` | d | on | 4096 | 3/48 | 20/48 | 16 | 2,319.8 | 15.895 | 95,092 / 14,264 | 292.5 | 300.00 W / Enabled | `t-glm4.7-flash-on/t-glm4.7-flash-on.d.jsonl` |
| `gpt-oss:20b` | c | low | 4096 | 68/108 | 68/108 | 0 | 120.1 | 1.351 | 604 / 604 | 229.2 | 300.00 W / Enabled | `b-gpt-oss-20b-low/b-gpt-oss-20b-low.c.jsonl` |
| `gpt-oss:20b` | c | medium | 4096 | 72/108 | 73/108 | 3 | 888.9 | 4.721 | 3,077 / 3,035 | 282.9 | 300.00 W / Enabled | `t-gpt-oss-20b-medium/t-gpt-oss-20b-medium.c.jsonl` |
| `gpt-oss:20b` | d | low | 4096 | 29/48 | 29/48 | 0 | 78.4 | 1.359 | 498 / 498 | 215.3 | 300.00 W / Enabled | `b-gpt-oss-20b-low/b-gpt-oss-20b-low.d.jsonl` |
| `gpt-oss:20b` | d | medium | 4096 | 35/48 | 35/48 | 0 | 367.1 | 2.955 | 1,264 / 1,264 | 264.8 | 300.00 W / Enabled | `t-gpt-oss-20b-medium/t-gpt-oss-20b-medium.d.jsonl` |
| `qwen3.5:27b-q4_K_M` | c | off | 8 | 97/108 | 97/108 | n/a (8-token cap) | 2.1 | 1.123 | 231 / 231 | 186.7 | 300.00 W / Enabled | `b-qwen3.5-27b/b-qwen3.5-27b.c.jsonl` |
| `qwen3.5:27b-q4_K_M` | c | on | 2048 | 79/108 | 79/108 | 28 | 1,301.3 | 31.558 | 14,346 / 14,346 | 294.6 | 300.00 W / Enabled | `t-qwen3.5-27b-on/t-qwen3.5-27b-on.c.jsonl` |
| `qwen3.8:27b-q4_K_M` | c | off | 8 | 95/108 | 95/108 | n/a (8-token cap) | 2.0 | 1.029 | 224 / 224 | 191.7 | 300.00 W / Enabled | `b-qwen3.8-27b/b-qwen3.8-27b.c.jsonl` |
| `qwen3.8:27b-q4_K_M` | c | on | 2048 | 98/108 | 98/108 | 0 | 476.1 | 12.481 | 4,228 / 4,228 | 290.1 | 300.00 W / Enabled | `t-qwen3.8-27b-on/t-qwen3.8-27b-on.c.jsonl` |
| `qwen3:30b-a3b-q4_K_M` | c | off | 8 | 67/108 | 67/108 | n/a (8-token cap) | 6.8 | 0.510 | 142 / 142 | 176.3 | 300.00 W / Enabled | `b-qwen3-30b-a3b/b-qwen3-30b-a3b.c.jsonl` |
| `qwen3:30b-a3b-q4_K_M` | c | on | 4096 | 75/108 | 76/108 | 1 | 1,179.2 | 6.240 | 3,105 / 3,064 | 287.4 | 300.00 W / Enabled | `t-qwen3-30b-a3b-on/t-qwen3-30b-a3b-on.c.jsonl` |
| `qwen3:30b-a3b-q4_K_M` | d | off | 8 | 2/48 | 39/48 | n/a (8-token cap) | 7.8 | 0.563 | 2,493 / 128 | 184.6 | 300.00 W / Enabled | `b-qwen3-30b-a3b/b-qwen3-30b-a3b.d.jsonl` |
| `qwen3:30b-a3b-q4_K_M` | d | on | 4096 | 0/48 | 30/48 | 0 | 501.9 | 2.602 | — / 1,550 | 276.7 | 300.00 W / Enabled | `t-qwen3-30b-a3b-on/t-qwen3-30b-a3b-on.d.jsonl` |
| `qwen3:32b-q4_K_M` | c | off | 8 | 80/108 | 80/108 | n/a (8-token cap) | 7.0 | 0.799 | 237 / 237 | 220.4 | 300.00 W / Enabled | `b-qwen3-32b/b-qwen3-32b.c.jsonl` |
| `qwen3:32b-q4_K_M` | c | on | 2048 | 66/108 | 67/108 | 4 | 718.7 | 16.806 | 10,393 / 10,238 | 295.4 | 300.00 W / Enabled | `t-qwen3-32b-on/t-qwen3-32b-on.c.jsonl` |

**The reasoning arms and their pairs** (generated from `rows/figures.json`; `think` is the request field exactly as sent, read off every row's `x_think`):

| reasoning arm | task | `think` sent | num_predict | num_ctx | paired arm (`think` sent) | stopped at the budget (stop: length) | empty reply | arm rc | ended (UTC) | file |
|---|---|---|---:|---:|---|---:|---:|---:|---|---|
| `t-deepseek-r1-32b-on` | c | `true` | 2048 | 8,192 | `b-deepseek-r1-32b` (`false`) | 3/108 | 3/108 | 0 | 2026-09-24T01:27:01Z | `t-deepseek-r1-32b-on/t-deepseek-r1-32b-on.c.jsonl` |
| `t-gemma4-12b-on` | c | `true` | 4096 | 8,192 | `b-gemma4-12b` (`false`) | 59/108 | 59/108 | 0 | 2026-09-23T22:22:47Z | `t-gemma4-12b-on/t-gemma4-12b-on.c.jsonl` |
| `t-gemma4-12b-on` | d | `true` | 4096 | 8,192 | `b-gemma4-12b` (`false`) | 9/48 | 9/48 | 0 | 2026-09-23T22:22:47Z | `t-gemma4-12b-on/t-gemma4-12b-on.d.jsonl` |
| `k0-gemma4-26b-think` | c | `true` | 4096 | 131,072 | `k0-gemma4-26b-c` (`false`) | 30/108 | 29/108 | 0 | 2026-09-23T16:41:31Z | `k0-gemma4-26b-think/k0-gemma4-26b-think.c.jsonl` |
| `k0-gemma4-26b-think` | d | `true` | 4096 | 131,072 | `k0-gemma4-26b-d` (`false`) | 1/48 | 1/48 | 0 | 2026-09-23T16:41:31Z | `k0-gemma4-26b-think/k0-gemma4-26b-think.d.jsonl` |
| `t-gemma4-31b-on` | c | `true` | 2048 | 8,192 | `b-gemma4-31b` (`false`) | 6/108 | 6/108 | 0 | 2026-09-24T03:21:05Z | `t-gemma4-31b-on/t-gemma4-31b-on.c.jsonl` |
| `t-glm4.7-flash-on` | c | `true` | 4096 | 8,192 | `b-glm4.7-flash` (`false`) | 23/108 | 23/108 | 0 | 2026-09-23T23:16:28Z | `t-glm4.7-flash-on/t-glm4.7-flash-on.c.jsonl` |
| `t-glm4.7-flash-on` | d | `true` | 4096 | 8,192 | `b-glm4.7-flash` (`false`) | 16/48 | 16/48 | 0 | 2026-09-23T23:16:28Z | `t-glm4.7-flash-on/t-glm4.7-flash-on.d.jsonl` |
| `t-gpt-oss-20b-medium` | c | `"medium"` | 4096 | 8,192 | `b-gpt-oss-20b-low` (`"low"`) | 3/108 | 3/108 | 0 | 2026-09-23T19:27:44Z | `t-gpt-oss-20b-medium/t-gpt-oss-20b-medium.c.jsonl` |
| `t-gpt-oss-20b-medium` | d | `"medium"` | 4096 | 8,192 | `b-gpt-oss-20b-low` (`"low"`) | 0/48 | 0/48 | 0 | 2026-09-23T19:27:44Z | `t-gpt-oss-20b-medium/t-gpt-oss-20b-medium.d.jsonl` |
| `t-qwen3.5-27b-on` | c | `true` | 2048 | 8,192 | `b-qwen3.5-27b` (`false`) | 28/108 | 28/108 | 0 | 2026-09-24T02:34:46Z | `t-qwen3.5-27b-on/t-qwen3.5-27b-on.c.jsonl` |
| `t-qwen3.8-27b-on` | c | `true` | 2048 | 8,192 | `b-qwen3.8-27b` (`false`) | 0/108 | 0/108 | 0 | 2026-09-24T00:01:00Z | `t-qwen3.8-27b-on/t-qwen3.8-27b-on.c.jsonl` |
| `t-qwen3-30b-a3b-on` | c | `true` | 4096 | 8,192 | `b-qwen3-30b-a3b` (`false`) | 1/108 | 1/108 | 0 | 2026-09-23T18:38:34Z | `t-qwen3-30b-a3b-on/t-qwen3-30b-a3b-on.c.jsonl` |
| `t-qwen3-30b-a3b-on` | d | `true` | 4096 | 8,192 | `b-qwen3-30b-a3b` (`false`) | 0/48 | 0/48 | 0 | 2026-09-23T18:38:34Z | `t-qwen3-30b-a3b-on/t-qwen3-30b-a3b-on.d.jsonl` |
| `t-qwen3-32b-on` | c | `true` | 2048 | 8,192 | `b-qwen3-32b` (`false`) | 4/108 | 1/108 | 0 | 2026-09-24T00:43:45Z | `t-qwen3-32b-on/t-qwen3-32b-on.c.jsonl` |



**How to read it.** With reasoning on, `num_predict` rises from 8 to the budget named on the
row, so the reasoning has room. The letter is still parsed from `message.content`, never from
the thinking. A reply whose reasoning used up the budget has empty content and counts as a
format failure; the "budget exhausted" column counts them. Reasoning models also tend to write
an explanation after thinking instead of a bare letter, so strict accuracy falls on the format
alone. Lenient (the one bracketed letter, or the one option named) is the fairer reading of
what the reasoning bought, and it is labelled as lenient everywhere. The tokens-out column
includes the reasoning tokens (ollama's `eval_count`); the thinking characters are on each row
(`x_thinking_chars`, and the text itself in `x_thinking`).

**What "on" means, per family (R-11).** The driver sends run.py's body with the request's
`think` field set (`instrument/coi3090.py`, `--think`). The "`think` sent" column above is read
off every row.
- **`think: true`** asks for the model's default reasoning. It is sent to qwen3 (30b-a3b and
  32b), qwen3.5:27b, qwen3.8:27b, deepseek-r1:32b (the Qwen distill), gemma4 (12b, 26b and
  31b) and glm-4.7-flash. Each is paired with the same tag's think-off arm (`think: false`,
  8 tokens): its `b-*` arm, or K0's arms for `gemma4:26b`.
- **`think: "medium"`** goes to gpt-oss:20b, which takes a level rather than a switch. Its pair
  is its lowest level, `"low"` (`b-gpt-oss-20b-low`), at the same 4,096-token budget. At
  `think: false` and 8 tokens it answered nothing: 108/108 unparsed on the laptop (PLAN-v2 §3's
  table, `cloud-reading-0/calibration/`).
- **Two budgets.** Five arms read both sets at 4,096 tokens: K0's `gemma4:26b` (queue line 7),
  `qwen3:30b-a3b` (69), `gpt-oss:20b` (82), `gemma4:12b` (110) and `glm-4.7-flash` (114). The
  five dense arms appended by the ~18:00Z ruling read the six-way only, at 2,048 tokens (lines
  142–154): `qwen3.8:27b`, `qwen3:32b`, `deepseek-r1:32b`, `qwen3.5:27b` and `gemma4:31b`. Their
  doorman has no reasoning pair. Every reasoning arm ran at `num_ctx` 8,192, except K0's at
  131,072, the Jev control's (`instrument/queue.txt`).

**No reasoning arm failed, but several ran out of budget.** Every reasoning arm ended at rc 0
(`queue/queue.log`). A reply stopped at the budget almost always has no content (179 of the 183
such replies across the ten reasoning arms, read off their row files), and it counts as a format
failure. The pairs table above counts both. The heaviest cases:
- `t-gemma4-12b-on`: 59 of 108 on the six-way, at 4,096 tokens.
- `k0-gemma4-26b-think`: 30 of 108.
- `t-qwen3.5-27b-on`: 28 of 108, at 2,048 tokens.
- `t-glm4.7-flash-on`: 23 of 108 on the six-way and 16 of 48 on the doorman.

These rows are a floor set by the budget, not the model's ceiling (`rows/figures.json`:
`done_reasons`, `empty_replies`). The two counts part on two arms. On `t-qwen3-32b-on`, 4 replies
stopped at the budget and 1 is empty, so three were cut off mid-answer. On
`k0-gemma4-26b-think`'s six-way, 30 stopped and 29 are empty.

**The doorman in JSON.** With reasoning on, `qwen3:30b-a3b` answered 47 of 48 doorman items with
a JSON verdict (`{"pass": …}`) instead of a letter, so it reads 0/48 strict and 30/48 lenient.
`glm-4.7-flash` did the same on 29 of 48 (the format classes, §5).



## 5. Every row (generated: `rows/FIGURES.md`, by `instrument/tables.py` from `rows/figures.json`)


Card: GeForce RTX 3090 24 GB `<withheld: GPU UUID of board A>`, benchbox, x16 at withheld-2026-10-14. Every row carries the enforced power limit and persistence mode as READ from the card at the start of that task (`cap`). BOARD = nvidia-smi power.draw, 1 Hz; WALL = the UPS's whole-box ups.realpower. Strict accuracy is the headline; v1 is run.py's own count, for continuity. Figures generated 2026-09-24T04:04:53Z.


### The six-way (task_c, 108 items, chance 16.7 %)

| weights date | model (tag) | think | strict acc (Wilson 95 %) | lenient acc (late letter / name read; never in a ratio) | v1 acc | format failures | median s/decision | board J per correct (net of resident idle) | wall J per correct | cap / persistence | contamination (R-1) | file |
|---|---|---|---:|---:|---:|---|---:|---:|---:|---|---|---|
| 2023-04-25 | `falcon:7b-instruct-q4_0` | off | 3/108 = 2.8 % (0.9 %–7.8 %) | 21.3 % | 23.2 % | article_A 1, late_letter 80, name_only 5, unparsed 10 | 0.300 | 2,026.3 (839.3) | 2,661.0 | 300.00 W / Enabled | clean (weights 2023-04-25 < 2026-09-19) | `b-falcon-7b/b-falcon-7b.c.jsonl` |
| 2023-06-21 | `vicuna:33b-q4_0` | off | 17/108 = 15.7 % (10.1 %–23.8 %) | 15.7 % | 16.7 % | contradictory 3, unparsed 1 | 0.813 | 1,133.7 (585.0) | 1,497.8 | 300.00 W / Enabled | clean (weights 2023-06-21 < 2026-09-19) | `b-vicuna-33b/b-vicuna-33b.c.jsonl` |
| 2023-07-13 | `llama2:13b-chat-q4_K_M` | off | 0/108 = 0.0 % (0.0 %–3.4 %) | 8.3 % | 8.3 % | late_letter 107, unparsed 1 | 0.474 | — (—) | — | 300.00 W / Enabled | clean (weights 2023-07-13 < 2026-09-19) | `b-llama2-13b/b-llama2-13b.c.jsonl` |
| 2023-07-13 | `llama2:7b-chat-q4_K_M` | off | 8/108 = 7.4 % (3.8 %–13.9 %) | 8.3 % | 14.8 % | late_letter 7, unparsed 27, word_letter 21 | 0.339 | 881.1 (380.5) | 1,130.6 | 300.00 W / Enabled | clean (weights 2023-07-13 < 2026-09-19) | `b-llama2-7b/b-llama2-7b.c.jsonl` |
| 2023-09-27 | `mistral:v0.1` | off | 10/108 = 9.3 % (5.1 %–16.2 %) | 9.3 % | 9.3 % | 0 | 0.248 | 538.4 (246.8) | 690.1 | 300.00 W / Enabled | clean (weights 2023-09-27 < 2026-09-19) | `k1-mistral-v0.1-c/k1-mistral-v0.1-c.c.jsonl` |
| 2023-11-22 | `yi:34b-chat-q4_0` | off | 21/108 = 19.4 % (13.1 %–27.9 %) | 26.9 % | 26.9 % | contradictory 3, late_letter 36 | 0.516 | 635.0 (359.3) | 842.2 | 300.00 W / Enabled | clean (weights 2023-11-22 < 2026-09-19) | `b-yi-34b/b-yi-34b.c.jsonl` |
| 2023-12-13 | `phi:2.7b-chat-v2-q4_K_M` | off | 9/108 = 8.3 % (4.5 %–15.1 %) | 8.3 % | 8.3 % | late_letter 1, name_only 1 | 0.254 | 459.4 (126.8) | 674.4 | 300.00 W / Enabled | clean (weights 2023-12-13 < 2026-09-19) | `b-phi2-2.7b/b-phi2-2.7b.c.jsonl` |
| 2024-03-26 | `gemma:7b-instruct-v1.1-q4_K_M` | off | 21/108 = 19.4 % (13.1 %–27.9 %) | 19.4 % | 19.4 % | 0 | 0.565 | 397.4 (79.8) | 627.8 | 300.00 W / Enabled | clean (weights 2024-03-26 < 2026-09-19) | `b-gemma1.1-7b/b-gemma1.1-7b.c.jsonl` |
| 2024-04-03 | `qwen:32b-chat-v1.5-q4_K_M` | off | 62/108 = 57.4 % (48.0 %–66.3 %) | 57.4 % | 57.4 % | 0 | 0.639 | 228.1 (107.2) | 317.7 | 300.00 W / Enabled | clean (weights 2024-04-03 < 2026-09-19) | `b-qwen1.5-32b/b-qwen1.5-32b.c.jsonl` |
| 2024-04-17 | `llama3:8b-instruct-q4_K_M` | off | 45/108 = 41.7 % (32.8 %–51.1 %) | 41.7 % | 41.7 % | 0 | 0.530 | 214.4 (76.7) | 310.2 | 300.00 W / Enabled | clean (weights 2024-04-17 < 2026-09-19) | `b-llama3-8b/b-llama3-8b.c.jsonl` |
| 2024-05-07 | `phi3:14b-medium-128k-instruct-q4_0` | off | 46/108 = 42.6 % (33.7 %–52.0 %) | 42.6 % | 42.6 % | 0 | 0.268 | 148.0 (81.0) | 174.2 | 300.00 W / Enabled | clean (weights 2024-05-07 < 2026-09-19) | `b-phi3-14b/b-phi3-14b.c.jsonl` |
| 2024-05-22 | `mistral:7b-instruct-v0.3-q4_K_M` | off | 20/108 = 18.5 % (12.3 %–26.9 %) | 18.5 % | 18.5 % | 0 | 0.195 | 243.3 (126.6) | 320.4 | 300.00 W / Enabled | clean (weights 2024-05-22 < 2026-09-19) | `b-mistral-7b-v0.3/b-mistral-7b-v0.3.c.jsonl` |
| 2024-06-24 | `gemma2:27b-instruct-q4_K_M` | off | 73/108 = 67.6 % (58.3 %–75.7 %) | 67.6 % | 67.6 % | 0 | 0.833 | 243.0 (111.7) | 353.9 | 300.00 W / Enabled | clean (weights 2024-06-24 < 2026-09-19) | `b-gemma2-27b/b-gemma2-27b.c.jsonl` |
| 2024-07-17 | `mistral-nemo:12b-instruct-2407-q4_K_M` | off | 47/108 = 43.5 % (34.5 %–52.9 %) | 43.5 % | 43.5 % | 0 | 0.599 | 246.8 (97.0) | 325.7 | 300.00 W / Enabled | clean (weights 2024-07-17 < 2026-09-19) | `b-mistral-nemo-12b/b-mistral-nemo-12b.c.jsonl` |
| 2024-07-18 | `llama3.1:8b-instruct-q4_K_M` | off | 43/108 = 39.8 % (31.1 %–49.2 %) | 39.8 % | 39.8 % | 0 | 0.505 | 171.4 (31.8) | 270.4 | 300.00 W / Enabled | clean (weights 2024-07-18 < 2026-09-19) | `b-llama3.1-8b/b-llama3.1-8b.c.jsonl` |
| 2024-09-17 | `mistral-small:22b-instruct-2409-q4_K_M` | off | 53/108 = 49.1 % (39.8 %–58.4 %) | 49.1 % | 49.1 % | 0 | 0.315 | 160.0 (91.5) | 195.6 | 300.00 W / Enabled | clean (weights 2024-09-17 < 2026-09-19) | `b-mistral-small-22b/b-mistral-small-22b.c.jsonl` |
| 2024-09-17 | `qwen2.5:32b-instruct-q4_K_M` | off | 89/108 = 82.4 % (74.2 %–88.4 %) | 82.4 % | 82.4 % | 0 | 0.655 | 159.2 (74.8) | 203.5 | 300.00 W / Enabled | clean (weights 2024-09-17 < 2026-09-19) | `b-qwen2.5-32b/b-qwen2.5-32b.c.jsonl` |
| 2024-12-11 | `phi4:14b-q4_K_M` | off | 82/108 = 75.9 % (67.1 %–83.0 %) | 76.8 % | 76.8 % | name_only 2, unparsed 3 | 0.473 | 128.0 (59.5) | 175.3 | 300.00 W / Enabled | clean (weights 2024-12-11 < 2026-09-19) | `b-phi4-14b/b-phi4-14b.c.jsonl` |
| 2025-01-20 | `deepseek-r1:32b-qwen-distill-q4_K_M` | off | 0/108 = 0.0 % (0.0 %–3.4 %) | 0.0 % | 16.7 % | unparsed 108 | 0.812 | — (—) | — | 300.00 W / Enabled | clean (weights 2025-01-20 < 2026-09-19) | `b-deepseek-r1-32b/b-deepseek-r1-32b.c.jsonl` |
| 2025-01-20 | `deepseek-r1:32b-qwen-distill-q4_K_M` | on | 1/108 = 0.9 % (0.2 %–5.1 %) | 70.4 % | 10.2 % | contradictory 8, late_letter 84, name_only 12, unparsed 3 | 19.836 | 696,093.8 (440,798.4) | 891,820.7 | 300.00 W / Enabled | clean (weights 2025-01-20 < 2026-09-19) | `t-deepseek-r1-32b-on/t-deepseek-r1-32b-on.c.jsonl` |
| 2025-01-28 | `mistral-small:24b-instruct-2501-q4_K_M` | off | 81/108 = 75.0 % (66.1 %–82.2 %) | 75.0 % | 75.0 % | 0 | 0.631 | 153.9 (62.7) | 207.2 | 300.00 W / Enabled | clean (weights 2025-01-28 < 2026-09-19) | `b-mistral-small-24b-2501/b-mistral-small-24b-2501.c.jsonl` |
| 2025-03-01 | `gemma3:27b-it-q4_K_M` | off | 89/108 = 82.4 % (74.2 %–88.4 %) | 82.4 % | 82.4 % | 0 | 1.225 | 265.5 (102.6) | 364.6 | 300.00 W / Enabled | clean (weights 2025-03-01 < 2026-09-19) | `b-gemma3-27b/b-gemma3-27b.c.jsonl` |
| 2025-04-27 | `qwen3:30b-a3b-q4_K_M` | off | 67/108 = 62.0 % (52.6 %–70.6 %) | 62.0 % | 62.0 % | 0 | 0.510 | 142.1 (54.8) | 189.5 | 300.00 W / Enabled | clean (weights 2025-04-27 < 2026-09-19) | `b-qwen3-30b-a3b/b-qwen3-30b-a3b.c.jsonl` |
| 2025-04-27 | `qwen3:32b-q4_K_M` | off | 80/108 = 74.1 % (65.1 %–81.4 %) | 74.1 % | 74.1 % | 0 | 0.799 | 237.0 (121.2) | 289.6 | 300.00 W / Enabled | clean (weights 2025-04-27 < 2026-09-19) | `b-qwen3-32b/b-qwen3-32b.c.jsonl` |
| 2025-04-27 | `qwen3:30b-a3b-q4_K_M` | on | 75/108 = 69.4 % (60.2 %–77.3 %) | 70.4 % | 69.4 % | late_letter 1, unparsed 2 | 6.240 | 3,105.1 (1,897.5) | 4,069.5 | 300.00 W / Enabled | clean (weights 2025-04-27 < 2026-09-19) | `t-qwen3-30b-a3b-on/t-qwen3-30b-a3b-on.c.jsonl` |
| 2025-04-27 | `qwen3:32b-q4_K_M` | on | 66/108 = 61.1 % (51.7 %–69.8 %) | 62.0 % | 64.8 % | contradictory 29, late_letter 2, unparsed 3 | 16.806 | 10,392.9 (6,591.8) | 13,387.5 | 300.00 W / Enabled | clean (weights 2025-04-27 < 2026-09-19) | `t-qwen3-32b-on/t-qwen3-32b-on.c.jsonl` |
| 2025-06-19 | `mistral-small3.2:24b-instruct-2506-q4_K_M` | off | 73/108 = 67.6 % (58.3 %–75.7 %) | 67.6 % | 67.6 % | 0 | 0.619 | 170.1 (69.0) | 238.4 | 300.00 W / Enabled | clean (weights 2025-06-19 < 2026-09-19) | `b-mistral-small3.2-24b/b-mistral-small3.2-24b.c.jsonl` |
| 2025-07-28 | `qwen3:30b-a3b-instruct-2507-q4_K_M` | off | 69/108 = 63.9 % (54.5 %–72.3 %) | 63.9 % | 63.9 % | 0 | 0.510 | 133.1 (49.7) | 172.9 | 300.00 W / Enabled | clean (weights 2025-07-28 < 2026-09-19) | `b-qwen3-30b-a3b-2507/b-qwen3-30b-a3b-2507.c.jsonl` |
| 2025-08-04 | `gpt-oss:20b` | low | 68/108 = 63.0 % (53.6 %–71.5 %) | 63.0 % | 63.0 % | 0 | 1.351 | 603.6 (322.4) | 802.5 | 300.00 W / Enabled | clean (weights 2025-08-04 < 2026-09-19) | `b-gpt-oss-20b-low/b-gpt-oss-20b-low.c.jsonl` |
| 2025-08-04 | `gpt-oss:20b` | medium | 72/108 = 66.7 % (57.3 %–74.9 %) | 67.6 % | 67.6 % | late_letter 1, unparsed 4 | 4.721 | 3,077.1 (1,860.4) | 4,091.6 | 300.00 W / Enabled | clean (weights 2025-08-04 < 2026-09-19) | `t-gpt-oss-20b-medium/t-gpt-oss-20b-medium.c.jsonl` |
| 2025-09-16 | `granite4:32b-a9b-h` | off | 44/108 = 40.7 % (31.9 %–50.2 %) | 40.7 % | 40.7 % | 0 | 0.662 | 307.7 (133.9) | 422.3 | 300.00 W / Enabled | clean (weights 2025-09-16 < 2026-09-19) | `b-granite4-32b-a9b/b-granite4-32b-a9b.c.jsonl` |
| 2025-10-31 | `ministral-3:14b-instruct-2512-q4_K_M` | off | 71/108 = 65.7 % (56.4 %–74.0 %) | 65.7 % | 65.7 % | 0 | 0.554 | 135.2 (45.2) | 182.4 | 300.00 W / Enabled | clean (weights 2025-10-31 < 2026-09-19) | `b-ministral3-14b/b-ministral3-14b.c.jsonl` |
| 2025-12-10 | `olmo-3.1:32b-instruct-q4_K_M` | off | 61/108 = 56.5 % (47.1 %–65.5 %) | 56.5 % | 56.5 % | 0 | 0.812 | 317.9 (162.5) | 420.9 | 300.00 W / Enabled | clean (weights 2025-12-10 < 2026-09-19) | `b-olmo3.1-32b/b-olmo3.1-32b.c.jsonl` |
| 2026-01-19 | `glm-4.7-flash:q4_K_M` | off | 44/108 = 40.7 % (31.9 %–50.2 %) | 40.7 % | 40.7 % | 0 | 0.623 | 254.7 (89.4) | 370.0 | 300.00 W / Enabled | clean (weights 2026-01-19 < 2026-09-19) | `b-glm4.7-flash/b-glm4.7-flash.c.jsonl` |
| 2026-01-19 | `glm-4.7-flash:q4_K_M` | on | 67/108 = 62.0 % (52.6 %–70.6 %) | 62.0 % | 65.7 % | unparsed 23 | 10.965 | 7,846.2 (4,828.7) | 10,202.6 | 300.00 W / Enabled | clean (weights 2026-01-19 < 2026-09-19) | `t-glm4.7-flash-on/t-glm4.7-flash-on.c.jsonl` |
| 2026-02-24 | `qwen3.5:27b-q4_K_M` | off | 97/108 = 89.8 % (82.7 %–94.2 %) | 89.8 % | 89.8 % | 0 | 1.123 | 231.0 (97.3) | 306.8 | 300.00 W / Enabled | clean (weights 2026-02-24 < 2026-09-19) | `b-qwen3.5-27b/b-qwen3.5-27b.c.jsonl` |
| 2026-02-24 | `qwen3.5:27b-q4_K_M` | on | 79/108 = 73.2 % (64.1 %–80.6 %) | 73.2 % | 77.8 % | unparsed 28 | 31.558 | 14,346.2 (9,032.9) | 18,463.6 | 300.00 W / Enabled | clean (weights 2026-02-24 < 2026-09-19) | `t-qwen3.5-27b-on/t-qwen3.5-27b-on.c.jsonl` |
| 2026-03-11 | `gemma4:31b-it-q4_K_M` | off | 94/108 = 87.0 % (79.4 %–92.1 %) | 90.7 % | 90.7 % | late_letter 4 | 1.481 | 309.4 (126.7) | 401.3 | 300.00 W / Enabled | clean (weights 2026-03-11 < 2026-09-19) | `b-gemma4-31b/b-gemma4-31b.c.jsonl` |
| 2026-03-11 | `gemma4:26b` | off | 88/108 = 81.5 % (73.1 %–87.7 %) | 81.5 % | 82.4 % | name_only 1 | 1.000 | 176.3 (46.9) | 275.6 | 300.00 W / Enabled | clean (weights 2026-03-11 < 2026-09-19) | `k0-gemma4-26b-c/k0-gemma4-26b-c.c.jsonl` |
| 2026-03-11 | `gemma4:26b` | on | 24/108 = 22.2 % (15.4 %–30.9 %) | 67.6 % | 30.6 % | contradictory 2, late_letter 51, name_only 1, unparsed 29 | 6.346 | 13,819.3 (8,382.4) | 18,067.2 | 300.00 W / Enabled | clean (weights 2026-03-11 < 2026-09-19) | `k0-gemma4-26b-think/k0-gemma4-26b-think.c.jsonl` |
| 2026-03-11 | `gemma4:26b-a4b-it-q4_K_M` | off | 88/108 = 81.5 % (73.1 %–87.7 %) | 81.5 % | 81.5 % | 0 | 0.982 | 183.4 (56.2) | 279.6 | 300.00 W / Enabled | clean (weights 2026-03-11 < 2026-09-19) | `k0a-gemma4-anchor-cd/k0a-gemma4-anchor-cd.c.jsonl` |
| 2026-03-11 | `gemma4:31b-it-q4_K_M` | on | 73/108 = 67.6 % (58.3 %–75.7 %) | 86.1 % | 74.1 % | contradictory 2, late_letter 25, unparsed 6 | 19.437 | 10,092.8 (6,305.6) | 12,977.8 | 300.00 W / Enabled | clean (weights 2026-03-11 < 2026-09-19) | `t-gemma4-31b-on/t-gemma4-31b-on.c.jsonl` |
| 2026-04-06 | `granite4.1:30b-q4_K_M` | off | 78/108 = 72.2 % (63.1 %–79.8 %) | 72.2 % | 72.2 % | 0 | 0.549 | 163.4 (81.8) | 207.7 | 300.00 W / Enabled | clean (weights 2026-04-06 < 2026-09-19) | `b-granite4.1-30b/b-granite4.1-30b.c.jsonl` |
| 2026-04-21 | `qwen3.6:27b-q4_K_M` | off | 96/108 = 88.9 % (81.6 %–93.5 %) | 88.9 % | 88.9 % | 0 | 1.121 | 233.7 (99.9) | 291.5 | 300.00 W / Enabled | clean (weights 2026-04-21 < 2026-09-19) | `b-qwen3.6-27b/b-qwen3.6-27b.c.jsonl` |
| 2026-05-23 | `gemma4:12b-it-q4_K_M` | off | 33/108 = 30.6 % (22.7 %–39.8 %) | 73.2 % | 73.2 % | late_letter 70 | 1.069 | 592.6 (214.8) | 826.8 | 300.00 W / Enabled | clean (weights 2026-05-23 < 2026-09-19) | `b-gemma4-12b/b-gemma4-12b.c.jsonl` |
| 2026-05-23 | `gemma4:12b-it-q4_K_M` | on | 32/108 = 29.6 % (21.8 %–38.8 %) | 42.6 % | 44.4 % | late_letter 15, unparsed 59 | 59.361 | 38,022.0 (23,599.8) | 48,869.5 | 300.00 W / Enabled | clean (weights 2026-05-23 < 2026-09-19) | `t-gemma4-12b-on/t-gemma4-12b-on.c.jsonl` |
| 2026-08-05 | `qwen3.8:27b-q4_K_M` | off | 95/108 = 88.0 % (80.5 %–92.8 %) | 88.0 % | 88.0 % | 0 | 1.029 | 224.1 (96.4) | 321.3 | 300.00 W / Enabled | clean (weights 2026-08-05 < 2026-09-19) | `b-qwen3.8-27b/b-qwen3.8-27b.c.jsonl` |
| 2026-08-05 | `qwen3.8:27b-q4_K_M` | on | 98/108 = 90.7 % (83.8 %–94.9 %) | 90.7 % | 90.7 % | 0 | 12.481 | 4,227.7 (2,655.5) | 5,360.4 | 300.00 W / Enabled | clean (weights 2026-08-05 < 2026-09-19) | `t-qwen3.8-27b-on/t-qwen3.8-27b-on.c.jsonl` |

### The doorman's planted set (48 = 36 in contract + 12 C/D)

| weights date | model (tag) | think | strict, all 48 | strict, 36 in contract | lenient (JSON verdict / name read), 36 in contract | C/D: refused / admitted of 12 (strict) | v1, all 48 | format classes | median s | board J per correct (all 48, strict) | wall J per correct | cap / persistence | file |
|---|---|---|---:|---:|---:|---:|---:|---|---:|---:|---:|---|---|
| 2023-04-25 | `falcon:7b-instruct-q4_0` | off | 1/48 = 2.1 % | 1/36 = 2.8 % | 9/36 | 0 / 0 | 25.0 % | json_verdict 24, letter+name 1, name_only 9, unparsed 14 | 0.355 | 3,228.7 | 4,628.1 | 300.00 W / Enabled | `b-falcon-7b/b-falcon-7b.d.jsonl` |
| 2023-06-21 | `vicuna:33b-q4_0` | off | 29/48 = 60.4 % | 20/36 = 55.6 % | 20/36 | 9 / 2 | 72.9 % | clean 30, letter+name 8, unparsed 10 | 1.246 | 444.5 | 708.7 | 300.00 W / Enabled | `b-vicuna-33b/b-vicuna-33b.d.jsonl` |
| 2023-07-13 | `llama2:13b-chat-q4_K_M` | off | 20/48 = 41.7 % | 10/36 = 27.8 % | 11/36 | 10 / 0 | 45.8 % | json_verdict 1, letter+name 30, unparsed 12, word_letter 5 | 0.663 | 321.1 | 441.5 | 300.00 W / Enabled | `b-llama2-13b/b-llama2-13b.d.jsonl` |
| 2023-07-13 | `llama2:7b-chat-q4_K_M` | off | 5/48 = 10.4 % | 3/36 = 8.3 % | 12/36 | 2 / 0 | 22.9 % | json_verdict 20, letter+name 11, unparsed 15, word_letter 2 | 0.469 | 869.0 | 953.6 | 300.00 W / Enabled | `b-llama2-7b/b-llama2-7b.d.jsonl` |
| 2023-09-27 | `mistral:v0.1` | off | 0/48 = 0.0 % | 0/36 = 0.0 % | 17/36 | 0 / 0 | 25.0 % | json_verdict 48 | 0.321 | — | — | 300.00 W / Enabled | `k1-mistral-v0.1-d/k1-mistral-v0.1-d.d.jsonl` |
| 2023-11-22 | `yi:34b-chat-q4_0` | off | 36/48 = 75.0 % | 24/36 = 66.7 % | 24/36 | 12 / 0 | 75.0 % | clean 48 | 0.680 | 223.1 | 319.5 | 300.00 W / Enabled | `b-yi-34b/b-yi-34b.d.jsonl` |
| 2023-12-13 | `phi:2.7b-chat-v2-q4_K_M` | off | 18/48 = 37.5 % | 13/36 = 36.1 % | 13/36 | 5 / 7 | 37.5 % | clean 44, letter+name 4 | 0.322 | 141.5 | 183.2 | 300.00 W / Enabled | `b-phi2-2.7b/b-phi2-2.7b.d.jsonl` |
| 2024-03-26 | `gemma:7b-instruct-v1.1-q4_K_M` | off | 15/48 = 31.2 % | 14/36 = 38.9 % | 14/36 | 1 / 11 | 31.2 % | clean 48 | 0.754 | 431.3 | 559.6 | 300.00 W / Enabled | `b-gemma1.1-7b/b-gemma1.1-7b.d.jsonl` |
| 2024-04-03 | `qwen:32b-chat-v1.5-q4_K_M` | off | 23/48 = 47.9 % | 20/36 = 55.6 % | 20/36 | 3 / 9 | 47.9 % | clean 43, unparsed 1, word_letter 4 | 0.930 | 440.7 | 631.9 | 300.00 W / Enabled | `b-qwen1.5-32b/b-qwen1.5-32b.d.jsonl` |
| 2024-04-17 | `llama3:8b-instruct-q4_K_M` | off | 29/48 = 60.4 % | 18/36 = 50.0 % | 18/36 | 11 / 0 | 60.4 % | json_verdict 1, letter+name 41, unparsed 6 | 0.562 | 159.2 | 208.8 | 300.00 W / Enabled | `b-llama3-8b/b-llama3-8b.d.jsonl` |
| 2024-05-07 | `phi3:14b-medium-128k-instruct-q4_0` | off | 36/48 = 75.0 % | 24/36 = 66.7 % | 24/36 | 12 / 0 | 75.0 % | clean 48 | 0.329 | 96.7 | 94.1 | 300.00 W / Enabled | `b-phi3-14b/b-phi3-14b.d.jsonl` |
| 2024-05-22 | `mistral:7b-instruct-v0.3-q4_K_M` | off | 0/48 = 0.0 % | 0/36 = 0.0 % | 13/36 | 0 / 0 | 25.0 % | json_verdict 46, unparsed 2 | 0.274 | — | — | 300.00 W / Enabled | `b-mistral-7b-v0.3/b-mistral-7b-v0.3.d.jsonl` |
| 2024-06-24 | `gemma2:27b-instruct-q4_K_M` | off | 40/48 = 83.3 % | 29/36 = 80.6 % | 29/36 | 11 / 1 | 83.3 % | clean 47, json_verdict 1 | 1.011 | 248.4 | 386.5 | 300.00 W / Enabled | `b-gemma2-27b/b-gemma2-27b.d.jsonl` |
| 2024-07-17 | `mistral-nemo:12b-instruct-2407-q4_K_M` | off | 33/48 = 68.8 % | 21/36 = 58.3 % | 23/36 | 12 / 0 | 68.8 % | clean 46, json_verdict 2 | 0.639 | 177.8 | 241.9 | 300.00 W / Enabled | `b-mistral-nemo-12b/b-mistral-nemo-12b.d.jsonl` |
| 2024-07-18 | `llama3.1:8b-instruct-q4_K_M` | off | 23/48 = 47.9 % | 15/36 = 41.7 % | 20/36 | 8 / 0 | 66.7 % | clean 26, json_verdict 18, unparsed 4 | 0.589 | 214.9 | 294.9 | 300.00 W / Enabled | `b-llama3.1-8b/b-llama3.1-8b.d.jsonl` |
| 2024-09-17 | `mistral-small:22b-instruct-2409-q4_K_M` | off | 30/48 = 62.5 % | 24/36 = 66.7 % | 24/36 | 6 / 6 | 62.5 % | clean 48 | 0.547 | 225.5 | 282.3 | 300.00 W / Enabled | `b-mistral-small-22b/b-mistral-small-22b.d.jsonl` |
| 2024-09-17 | `qwen2.5:32b-instruct-q4_K_M` | off | 42/48 = 87.5 % | 34/36 = 94.4 % | 34/36 | 8 / 4 | 87.5 % | clean 48 | 0.873 | 222.0 | 287.7 | 300.00 W / Enabled | `b-qwen2.5-32b/b-qwen2.5-32b.d.jsonl` |
| 2024-12-11 | `phi4:14b-q4_K_M` | off | 0/48 = 0.0 % | 0/36 = 0.0 % | 0/36 | 0 / 0 | 25.0 % | unparsed 47, word_letter 1 | 0.589 | — | — | 300.00 W / Enabled | `b-phi4-14b/b-phi4-14b.d.jsonl` |
| 2025-01-20 | `deepseek-r1:32b-qwen-distill-q4_K_M` | off | 0/48 = 0.0 % | 0/36 = 0.0 % | 0/36 | 0 / 0 | 25.0 % | unparsed 48 | 1.052 | — | — | 300.00 W / Enabled | `b-deepseek-r1-32b/b-deepseek-r1-32b.d.jsonl` |
| 2025-01-28 | `mistral-small:24b-instruct-2501-q4_K_M` | off | 35/48 = 72.9 % | 28/36 = 77.8 % | 28/36 | 7 / 5 | 72.9 % | clean 48 | 0.805 | 219.6 | 333.5 | 300.00 W / Enabled | `b-mistral-small-24b-2501/b-mistral-small-24b-2501.d.jsonl` |
| 2025-03-01 | `gemma3:27b-it-q4_K_M` | off | 21/48 = 43.8 % | 14/36 = 38.9 % | 32/36 | 7 / 0 | 66.7 % | clean 22, json_verdict 24, unparsed 2 | 1.656 | 743.9 | 987.4 | 300.00 W / Enabled | `b-gemma3-27b/b-gemma3-27b.d.jsonl` |
| 2025-04-27 | `qwen3:30b-a3b-q4_K_M` | off | 2/48 = 4.2 % | 2/36 = 5.6 % | 27/36 | 0 / 0 | 22.9 % | clean 2, json_verdict 43, letter+name 3 | 0.563 | 2,493.1 | 2,997.6 | 300.00 W / Enabled | `b-qwen3-30b-a3b/b-qwen3-30b-a3b.d.jsonl` |
| 2025-04-27 | `qwen3:32b-q4_K_M` | off | 43/48 = 89.6 % | 31/36 = 86.1 % | 31/36 | 12 / 0 | 89.6 % | clean 48 | 0.917 | 224.4 | 287.4 | 300.00 W / Enabled | `b-qwen3-32b/b-qwen3-32b.d.jsonl` |
| 2025-04-27 | `qwen3:30b-a3b-q4_K_M` | on | 0/48 = 0.0 % | 0/36 = 0.0 % | 30/36 | 0 / 1 | 25.0 % | clean 1, json_verdict 47 | 2.602 | — | — | 300.00 W / Enabled | `t-qwen3-30b-a3b-on/t-qwen3-30b-a3b-on.d.jsonl` |
| 2025-06-19 | `mistral-small3.2:24b-instruct-2506-q4_K_M` | off | 26/48 = 54.2 % | 24/36 = 66.7 % | 24/36 | 2 / 8 | 54.2 % | clean 45, json_verdict 2, unparsed 1 | 0.750 | 287.1 | 313.7 | 300.00 W / Enabled | `b-mistral-small3.2-24b/b-mistral-small3.2-24b.d.jsonl` |
| 2025-07-28 | `qwen3:30b-a3b-instruct-2507-q4_K_M` | off | 9/48 = 18.8 % | 5/36 = 13.9 % | 27/36 | 4 / 0 | 43.8 % | clean 1, json_verdict 39, letter+name 8 | 0.566 | 539.9 | 725.2 | 300.00 W / Enabled | `b-qwen3-30b-a3b-2507/b-qwen3-30b-a3b-2507.d.jsonl` |
| 2025-08-04 | `gpt-oss:20b` | low | 29/48 = 60.4 % | 29/36 = 80.6 % | 29/36 | 0 / 12 | 60.4 % | clean 48 | 1.359 | 497.6 | 696.5 | 300.00 W / Enabled | `b-gpt-oss-20b-low/b-gpt-oss-20b-low.d.jsonl` |
| 2025-08-04 | `gpt-oss:20b` | medium | 35/48 = 72.9 % | 32/36 = 88.9 % | 32/36 | 3 / 9 | 75.0 % | clean 46, unparsed 2 | 2.955 | 1,264.0 | 1,583.5 | 300.00 W / Enabled | `t-gpt-oss-20b-medium/t-gpt-oss-20b-medium.d.jsonl` |
| 2025-09-16 | `granite4:32b-a9b-h` | off | 15/48 = 31.2 % | 11/36 = 30.6 % | 25/36 | 4 / 0 | 41.7 % | clean 19, json_verdict 28, unparsed 1 | 1.059 | 632.6 | 1,043.7 | 300.00 W / Enabled | `b-granite4-32b-a9b/b-granite4-32b-a9b.d.jsonl` |
| 2025-10-31 | `ministral-3:14b-instruct-2512-q4_K_M` | off | 17/48 = 35.4 % | 17/36 = 47.2 % | 17/36 | 0 / 12 | 35.4 % | clean 48 | 0.655 | 339.1 | 398.9 | 300.00 W / Enabled | `b-ministral3-14b/b-ministral3-14b.d.jsonl` |
| 2025-12-10 | `olmo-3.1:32b-instruct-q4_K_M` | off | 0/48 = 0.0 % | 0/36 = 0.0 % | 19/36 | 0 / 0 | 25.0 % | json_verdict 48 | 1.133 | — | — | 300.00 W / Enabled | `b-olmo3.1-32b/b-olmo3.1-32b.d.jsonl` |
| 2026-01-19 | `glm-4.7-flash:q4_K_M` | off | 23/48 = 47.9 % | 18/36 = 50.0 % | 20/36 | 5 / 0 | 52.1 % | clean 32, json_verdict 16 | 0.697 | 257.6 | 335.1 | 300.00 W / Enabled | `b-glm4.7-flash/b-glm4.7-flash.d.jsonl` |
| 2026-01-19 | `glm-4.7-flash:q4_K_M` | on | 3/48 = 6.2 % | 3/36 = 8.3 % | 20/36 | 0 / 0 | 50.0 % | clean 3, json_verdict 29, unparsed 16 | 15.895 | 95,092.1 | 121,935.8 | 300.00 W / Enabled | `t-glm4.7-flash-on/t-glm4.7-flash-on.d.jsonl` |
| 2026-02-24 | `qwen3.5:27b-q4_K_M` | off | 47/48 = 97.9 % | 36/36 = 100.0 % | 36/36 | 11 / 1 | 97.9 % | clean 48 | 1.545 | 294.2 | 429.6 | 300.00 W / Enabled | `b-qwen3.5-27b/b-qwen3.5-27b.d.jsonl` |
| 2026-03-11 | `gemma4:31b-it-q4_K_M` | off | 41/48 = 85.4 % | 35/36 = 97.2 % | 35/36 | 6 / 6 | 85.4 % | clean 48 | 1.872 | 410.3 | 565.3 | 300.00 W / Enabled | `b-gemma4-31b/b-gemma4-31b.d.jsonl` |
| 2026-03-11 | `gemma4:26b` | off | 35/48 = 72.9 % | 34/36 = 94.4 % | 34/36 | 1 / 11 | 72.9 % | clean 46, unparsed 2 | 1.108 | 246.0 | 343.8 | 300.00 W / Enabled | `k0-gemma4-26b-d/k0-gemma4-26b-d.d.jsonl` |
| 2026-03-11 | `gemma4:26b` | on | 35/48 = 72.9 % | 35/36 = 97.2 % | 35/36 | 0 / 12 | 72.9 % | clean 47, unparsed 1 | 3.703 | 1,562.0 | 2,177.2 | 300.00 W / Enabled | `k0-gemma4-26b-think/k0-gemma4-26b-think.d.jsonl` |
| 2026-03-11 | `gemma4:26b-a4b-it-q4_K_M` | off | 35/48 = 72.9 % | 34/36 = 94.4 % | 34/36 | 1 / 11 | 72.9 % | clean 47, unparsed 1 | 1.076 | 239.4 | 350.6 | 300.00 W / Enabled | `k0a-gemma4-anchor-cd/k0a-gemma4-anchor-cd.d.jsonl` |
| 2026-04-06 | `granite4.1:30b-q4_K_M` | off | 38/48 = 79.2 % | 30/36 = 83.3 % | 30/36 | 8 / 4 | 79.2 % | clean 46, json_verdict 2 | 0.749 | 226.4 | 296.3 | 300.00 W / Enabled | `b-granite4.1-30b/b-granite4.1-30b.d.jsonl` |
| 2026-04-21 | `qwen3.6:27b-q4_K_M` | off | 31/48 = 64.6 % | 28/36 = 77.8 % | 28/36 | 3 / 9 | 64.6 % | clean 48 | 1.555 | 455.8 | 704.8 | 300.00 W / Enabled | `b-qwen3.6-27b/b-qwen3.6-27b.d.jsonl` |
| 2026-05-23 | `gemma4:12b-it-q4_K_M` | off | 35/48 = 72.9 % | 30/36 = 83.3 % | 30/36 | 5 / 7 | 72.9 % | clean 45, unparsed 3 | 1.190 | 279.3 | 355.4 | 300.00 W / Enabled | `b-gemma4-12b/b-gemma4-12b.d.jsonl` |
| 2026-05-23 | `gemma4:12b-it-q4_K_M` | on | 31/48 = 64.6 % | 31/36 = 86.1 % | 31/36 | 0 / 8 | 64.6 % | clean 39, unparsed 9 | 12.497 | 9,422.1 | 12,071.6 | 300.00 W / Enabled | `t-gemma4-12b-on/t-gemma4-12b-on.d.jsonl` |
| 2026-08-05 | `qwen3.8:27b-q4_K_M` | off | 9/48 = 18.8 % | 9/36 = 25.0 % | 35/36 | 0 / 0 | 43.8 % | clean 9, json_verdict 39 | 1.609 | 1,694.9 | 2,278.1 | 300.00 W / Enabled | `b-qwen3.8-27b/b-qwen3.8-27b.d.jsonl` |

### Timings and tokens per task (server fields; R-10c)

| model | think | task | num_predict | stop reasons | prompt tokens in | tokens out (incl. reasoning) | thinking chars | decode tok/s (task) | prefill tok/s as reported (cache-inflated) | server time before decode, median / p95 ms | wall-clock median / p95 s | temp start → end °C (trace max) | file |
|---|---|---|---:|---|---:|---:|---:|---:|---:|---:|---:|---|---|
| `falcon:7b-instruct-q4_0` | off | c | 8 | length 98, stop 10 | 32,206 | 820 | 0 | 170.8 | 6,238 | 251.6 / 274.6 | 0.300 / 0.323 | 51 → 58 (60.0) | `b-falcon-7b/b-falcon-7b.c.report.json` |
| `falcon:7b-instruct-q4_0` | off | d | 8 | length 48 | 25,134 | 384 | 0 | 171.5 | 5,568 | 307.4 / 314.2 | 0.355 / 0.362 | 58 → 61 (64.0) | `b-falcon-7b/b-falcon-7b.d.report.json` |
| `vicuna:33b-q4_0` | off | c | 8 | length 108 | 38,477 | 864 | 0 | 44.9 | 1,595 | 633.4 / 762.5 | 0.813 / 0.942 | 40 → 62 (65.0) | `b-vicuna-33b/b-vicuna-33b.c.report.json` |
| `vicuna:33b-q4_0` | off | d | 8 | length 40, stop 8 | 27,817 | 344 | 0 | 44.7 | 1,263 | 1,065.6 / 1,088.6 | 1.246 / 1.271 | 61 → 63 (66.0) | `b-vicuna-33b/b-vicuna-33b.d.report.json` |
| `llama2:13b-chat-q4_K_M` | off | c | 8 | length 108 | 36,317 | 864 | 0 | 93.1 | 3,422 | 386.6 / 448.1 | 0.474 / 0.535 | 46 → 59 (61.0) | `b-llama2-13b/b-llama2-13b.c.report.json` |
| `llama2:13b-chat-q4_K_M` | off | d | 8 | length 18, stop 30 | 26,857 | 324 | 0 | 93.6 | 2,822 | 596.5 / 609.5 | 0.663 / 0.698 | 59 → 62 (65.0) | `b-llama2-13b/b-llama2-13b.d.report.json` |
| `llama2:7b-chat-q4_K_M` | off | c | 8 | length 60, stop 48 | 36,317 | 854 | 0 | 162.7 | 6,189 | 288.3 / 319.3 | 0.339 / 0.370 | 50 → 58 (60.0) | `b-llama2-7b/b-llama2-7b.c.report.json` |
| `llama2:7b-chat-q4_K_M` | off | d | 8 | length 48 | 26,857 | 384 | 0 | 160.5 | 4,871 | 417.9 / 429.5 | 0.469 / 0.481 | 58 → 62 (64.0) | `b-llama2-7b/b-llama2-7b.d.report.json` |
| `mistral:v0.1` | off | c | 8 | length 39, stop 69 | 35,035 | 781 | 0 | 172.5 | 6,325 | 203.7 / 222.1 | 0.248 / 0.268 | 53 → 60 (61.0) | `k1-mistral-v0.1-c/k1-mistral-v0.1-c.c.report.json` |
| `mistral:v0.1` | off | d | 8 | length 48 | 25,733 | 384 | 0 | 171.3 | 5,058 | 273.2 / 279.2 | 0.321 / 0.327 | 56 → 61 (62.0) | `k1-mistral-v0.1-d/k1-mistral-v0.1-d.d.report.json` |
| `yi:34b-chat-q4_0` | off | c | 8 | length 45, stop 63 | 35,593 | 571 | 0 | 47.1 | 1,492 | 390.1 / 456.7 | 0.516 / 0.637 | 40 → 60 (62.0) | `b-yi-34b/b-yi-34b.c.report.json` |
| `yi:34b-chat-q4_0` | off | d | 8 | stop 48 | 26,322 | 96 | 0 | 72.8 | 1,241 | 651.3 / 665.3 | 0.680 / 0.693 | 60 → 64 (67.0) | `b-yi-34b/b-yi-34b.d.report.json` |
| `phi:2.7b-chat-v2-q4_K_M` | off | c | 8 | length 97, stop 11 | 34,288 | 815 | 0 | 249.0 | 12,278 | 222.7 / 243.0 | 0.254 / 0.276 | 52 → 56 (57.0) | `b-phi2-2.7b/b-phi2-2.7b.c.report.json` |
| `phi:2.7b-chat-v2-q4_K_M` | off | d | 8 | length 21, stop 27 | 26,600 | 249 | 0 | 239.8 | 10,606 | 301.0 / 313.2 | 0.322 / 0.345 | 56 → 59 (61.0) | `b-phi2-2.7b/b-phi2-2.7b.d.report.json` |
| `gemma:7b-instruct-v1.1-q4_K_M` | off | c | 8 | stop 108 | 31,959 | 216 | 0 | 186.1 | 4,906 | 552.7 / 603.4 | 0.565 / 0.615 | 50 → 56 (57.0) | `b-gemma1.1-7b/b-gemma1.1-7b.c.report.json` |
| `gemma:7b-instruct-v1.1-q4_K_M` | off | d | 8 | length 48 | 25,094 | 384 | 0 | 114.1 | 4,496 | 682.7 / 715.0 | 0.754 / 0.787 | 56 → 61 (63.0) | `b-gemma1.1-7b/b-gemma1.1-7b.d.report.json` |
| `qwen:32b-chat-v1.5-q4_K_M` | off | c | 8 | length 1, stop 107 | 32,428 | 222 | 0 | 66.4 | 1,384 | 609.0 / 694.5 | 0.639 / 0.727 | 40 → 58 (60.0) | `b-qwen1.5-32b/b-qwen1.5-32b.c.report.json` |
| `qwen:32b-chat-v1.5-q4_K_M` | off | d | 8 | length 5, stop 43 | 25,478 | 126 | 0 | 55.9 | 1,168 | 898.7 / 933.7 | 0.930 / 1.077 | 58 → 63 (66.0) | `b-qwen1.5-32b/b-qwen1.5-32b.d.report.json` |
| `llama3:8b-instruct-q4_K_M` | off | c | 8 | length 48, stop 60 | 30,868 | 813 | 0 | 142.4 | 5,715 | 474.9 / 510.8 | 0.530 / 0.565 | 49 → 58 (60.0) | `b-llama3-8b/b-llama3-8b.c.report.json` |
| `llama3:8b-instruct-q4_K_M` | off | d | 8 | length 10, stop 38 | 23,784 | 232 | 0 | 150.2 | 4,944 | 525.9 / 564.5 | 0.562 / 0.604 | 58 → 61 (63.0) | `b-llama3-8b/b-llama3-8b.d.report.json` |
| `phi3:14b-medium-128k-instruct-q4_0` | off | c | 8 | length 67, stop 41 | 35,129 | 783 | 0 | 101.7 | 3,510 | 196.5 / 224.6 | 0.268 / 0.303 | 47 → 58 (59.0) | `b-phi3-14b/b-phi3-14b.c.report.json` |
| `phi3:14b-medium-128k-instruct-q4_0` | off | d | 8 | stop 48 | 26,329 | 96 | 0 | 163.5 | 2,998 | 315.6 / 327.4 | 0.329 / 0.340 | 58 → 61 (64.0) | `b-phi3-14b/b-phi3-14b.d.report.json` |
| `mistral:7b-instruct-v0.3-q4_K_M` | off | c | 8 | length 68, stop 40 | 34,387 | 817 | 0 | 160.5 | 5,784 | 147.6 / 168.6 | 0.195 / 0.217 | 49 → 57 (60.0) | `b-mistral-7b-v0.3/b-mistral-7b-v0.3.c.report.json` |
| `mistral:7b-instruct-v0.3-q4_K_M` | off | d | 8 | length 48 | 25,445 | 384 | 0 | 159.8 | 4,635 | 222.7 / 229.9 | 0.274 / 0.281 | 57 → 61 (63.0) | `b-mistral-7b-v0.3/b-mistral-7b-v0.3.d.report.json` |
| `gemma2:27b-instruct-q4_K_M` | off | c | 8 | stop 108 | 30,555 | 340 | 0 | 56.2 | 1,106 | 777.6 / 840.4 | 0.833 / 0.896 | 41 → 61 (63.0) | `b-gemma2-27b/b-gemma2-27b.c.report.json` |
| `gemma2:27b-instruct-q4_K_M` | off | d | 8 | length 1, stop 47 | 24,470 | 102 | 0 | 70.7 | 1,256 | 983.1 / 1,044.2 | 1.011 / 1.078 | 61 → 62 (65.0) | `b-gemma2-27b/b-gemma2-27b.d.report.json` |
| `mistral-nemo:12b-instruct-2407-q4_K_M` | off | c | 8 | length 99, stop 9 | 30,074 | 850 | 0 | 98.5 | 3,728 | 518.5 / 558.6 | 0.599 / 0.641 | 47 → 59 (61.0) | `b-mistral-nemo-12b/b-mistral-nemo-12b.c.report.json` |
| `mistral-nemo:12b-instruct-2407-q4_K_M` | off | d | 8 | length 21, stop 27 | 23,588 | 222 | 0 | 107.9 | 3,318 | 609.0 / 643.8 | 0.639 / 0.725 | 59 → 62 (65.0) | `b-mistral-nemo-12b/b-mistral-nemo-12b.d.report.json` |
| `llama3.1:8b-instruct-q4_K_M` | off | c | 8 | length 1, stop 107 | 30,868 | 305 | 0 | 175.1 | 5,466 | 486.9 / 524.5 | 0.505 / 0.543 | 49 → 55 (56.0) | `b-llama3.1-8b/b-llama3.1-8b.c.report.json` |
| `llama3.1:8b-instruct-q4_K_M` | off | d | 8 | length 22, stop 26 | 23,784 | 254 | 0 | 148.1 | 4,955 | 547.5 / 575.5 | 0.589 / 0.626 | 55 → 59 (62.0) | `b-llama3.1-8b/b-llama3.1-8b.d.report.json` |
| `mistral-small:22b-instruct-2409-q4_K_M` | off | c | 8 | length 10, stop 98 | 34,495 | 364 | 0 | 71.9 | 1,918 | 274.6 / 317.1 | 0.315 / 0.413 | 37 → 53 (56.0) | `b-mistral-small-22b/b-mistral-small-22b.c.report.json` |
| `mistral-small:22b-instruct-2409-q4_K_M` | off | d | 8 | stop 48 | 25,493 | 192 | 0 | 67.1 | 1,528 | 486.9 / 507.0 | 0.547 / 0.568 | 53 → 61 (64.0) | `b-mistral-small-22b/b-mistral-small-22b.d.report.json` |
| `qwen2.5:32b-instruct-q4_K_M` | off | c | 8 | length 1, stop 107 | 32,924 | 274 | 0 | 55.7 | 1,493 | 602.6 / 685.7 | 0.655 / 0.730 | 40 → 58 (61.0) | `b-qwen2.5-32b/b-qwen2.5-32b.c.report.json` |
| `qwen2.5:32b-instruct-q4_K_M` | off | d | 8 | length 1, stop 47 | 24,751 | 102 | 0 | 64.0 | 1,284 | 842.8 / 869.1 | 0.873 / 0.918 | 58 → 63 (66.0) | `b-qwen2.5-32b/b-qwen2.5-32b.d.report.json` |
| `phi4:14b-q4_K_M` | off | c | 8 | length 108 | 30,868 | 864 | 0 | 86.9 | 3,266 | 379.3 / 421.1 | 0.473 / 0.515 | 47 → 60 (62.0) | `b-phi4-14b/b-phi4-14b.c.report.json` |
| `phi4:14b-q4_K_M` | off | d | 8 | length 48 | 23,832 | 384 | 0 | 86.6 | 2,890 | 495.3 / 506.8 | 0.589 / 0.601 | 60 → 63 (66.0) | `b-phi4-14b/b-phi4-14b.d.report.json` |
| `deepseek-r1:32b-qwen-distill-q4_K_M` | off | c | 8 | length 108 | 30,116 | 864 | 2,387 | 40.1 | 1,375 | 611.4 / 681.4 | 0.812 / 0.883 | 42 → 63 (66.0) | `b-deepseek-r1-32b/b-deepseek-r1-32b.c.report.json` |
| `deepseek-r1:32b-qwen-distill-q4_K_M` | off | d | 8 | length 48 | 23,503 | 384 | 917 | 40.0 | 1,226 | 850.7 / 874.3 | 1.052 / 1.075 | 63 → 64 (67.0) | `b-deepseek-r1-32b/b-deepseek-r1-32b.d.report.json` |
| `deepseek-r1:32b-qwen-distill-q4_K_M` | on | c | 2048 | length 3, stop 105 | 30,116 | 78,981 | 358,767 | 34.7 | 1,347 | 726.1 / 828.7 | 19.836 / 33.868 | 43 → 67 (71.0) | `t-deepseek-r1-32b-on/t-deepseek-r1-32b-on.c.report.json` |
| `mistral-small:24b-instruct-2501-q4_K_M` | off | c | 8 | stop 108 | 47,138 | 303 | 0 | 76.1 | 3,091 | 590.8 / 656.3 | 0.631 / 0.700 | 44 → 60 (61.0) | `b-mistral-small-24b-2501/b-mistral-small-24b-2501.c.report.json` |
| `mistral-small:24b-instruct-2501-q4_K_M` | off | d | 8 | stop 48 | 31,172 | 131 | 0 | 76.2 | 2,332 | 768.9 / 798.5 | 0.805 / 0.842 | 60 → 63 (66.0) | `b-mistral-small-24b-2501/b-mistral-small-24b-2501.d.report.json` |
| `gemma3:27b-it-q4_K_M` | off | c | 8 | stop 108 | 30,581 | 250 | 0 | 62.5 | 1,025 | 1,190.5 / 1,273.7 | 1.225 / 1.313 | 43 → 61 (66.0) | `b-gemma3-27b/b-gemma3-27b.c.report.json` |
| `gemma3:27b-it-q4_K_M` | off | d | 8 | length 47, stop 1 | 24,616 | 379 | 0 | 42.7 | 1,165 | 1,468.4 / 1,558.2 | 1.656 / 1.746 | 61 → 63 (66.0) | `b-gemma3-27b/b-gemma3-27b.d.report.json` |
| `qwen3:30b-a3b-q4_K_M` | off | c | 8 | length 58, stop 50 | 31,520 | 729 | 0 | 169.9 | 3,565 | 467.1 / 501.6 | 0.510 / 0.546 | 42 → 56 (56.0) | `b-qwen3-30b-a3b/b-qwen3-30b-a3b.c.report.json` |
| `qwen3:30b-a3b-q4_K_M` | off | d | 8 | length 46, stop 2 | 24,127 | 374 | 0 | 166.7 | 4,143 | 514.5 / 541.5 | 0.563 / 0.593 | 56 → 60 (62.0) | `b-qwen3-30b-a3b/b-qwen3-30b-a3b.d.report.json` |
| `qwen3:32b-q4_K_M` | off | c | 8 | length 52, stop 56 | 31,520 | 760 | 0 | 40.8 | 1,354 | 613.5 / 705.7 | 0.799 / 0.901 | 41 → 62 (65.0) | `b-qwen3-32b/b-qwen3-32b.c.report.json` |
| `qwen3:32b-q4_K_M` | off | d | 8 | stop 48 | 24,127 | 144 | 0 | 50.3 | 1,247 | 856.8 / 877.9 | 0.917 / 0.939 | 62 → 63 (67.0) | `b-qwen3-32b/b-qwen3-32b.d.report.json` |
| `qwen3:30b-a3b-q4_K_M` | on | c | 4096 | length 1, stop 107 | 30,872 | 127,355 | 575,364 | 170.7 | 3,429 | 522.2 / 593.2 | 6.240 / 15.977 | 54 → 66 (71.0) | `t-qwen3-30b-a3b-on/t-qwen3-30b-a3b-on.c.report.json` |
| `qwen3:30b-a3b-q4_K_M` | on | d | 4096 | stop 48 | 23,839 | 24,091 | 99,486 | 171.7 | 4,075 | 553.6 / 592.5 | 2.602 / 8.957 | 66 → 66 (70.0) | `t-qwen3-30b-a3b-on/t-qwen3-30b-a3b-on.d.report.json` |
| `qwen3:32b-q4_K_M` | on | c | 2048 | length 4, stop 104 | 30,872 | 77,623 | 307,598 | 34.7 | 1,341 | 722.4 / 905.5 | 16.806 / 49.204 | 43 → 67 (71.0) | `t-qwen3-32b-on/t-qwen3-32b-on.c.report.json` |
| `mistral-small3.2:24b-instruct-2506-q4_K_M` | off | c | 8 | length 8, stop 100 | 84,506 | 294 | 0 | 77.6 | 5,503 | 590.5 / 649.6 | 0.619 / 0.752 | 44 → 59 (61.0) | `b-mistral-small3.2-24b/b-mistral-small3.2-24b.c.report.json` |
| `mistral-small3.2:24b-instruct-2506-q4_K_M` | off | d | 8 | length 3, stop 45 | 47,780 | 115 | 0 | 83.2 | 3,551 | 728.7 / 760.8 | 0.750 / 0.884 | 59 → 63 (66.0) | `b-mistral-small3.2-24b/b-mistral-small3.2-24b.d.report.json` |
| `qwen3:30b-a3b-instruct-2507-q4_K_M` | off | c | 8 | length 60, stop 48 | 30,656 | 723 | 0 | 171.1 | 3,574 | 468.7 / 494.9 | 0.510 / 0.538 | 35 → 50 (50.0) | `b-qwen3-30b-a3b-2507/b-qwen3-30b-a3b-2507.c.report.json` |
| `qwen3:30b-a3b-instruct-2507-q4_K_M` | off | d | 8 | length 47, stop 1 | 23,743 | 379 | 0 | 170.3 | 4,158 | 518.7 / 548.5 | 0.566 / 0.597 | 50 → 56 (58.0) | `b-qwen3-30b-a3b-2507/b-qwen3-30b-a3b-2507.d.report.json` |
| `gpt-oss:20b` | low | c | 4096 | stop 108 | 36,181 | 12,969 | 51,505 | 141.1 | 3,901 | 802.6 / 856.8 | 1.351 / 2.795 | 38 → 64 (68.0) | `b-gpt-oss-20b-low/b-gpt-oss-20b-low.c.report.json` |
| `gpt-oss:20b` | low | d | 4096 | stop 48 | 26,499 | 3,761 | 13,887 | 141.1 | 4,765 | 841.3 / 895.1 | 1.359 / 1.583 | 64 → 62 (65.0) | `b-gpt-oss-20b-low/b-gpt-oss-20b-low.d.report.json` |
| `gpt-oss:20b` | medium | c | 4096 | length 3, stop 105 | 36,181 | 95,999 | 417,997 | 138.8 | 3,718 | 806.5 / 871.5 | 4.721 / 28.630 | 54 → 66 (71.0) | `t-gpt-oss-20b-medium/t-gpt-oss-20b-medium.c.report.json` |
| `gpt-oss:20b` | medium | d | 4096 | stop 48 | 26,499 | 17,620 | 74,078 | 139.4 | 4,649 | 839.6 / 887.0 | 2.955 / 7.750 | 66 → 65 (69.0) | `t-gpt-oss-20b-medium/t-gpt-oss-20b-medium.d.report.json` |
| `granite4:32b-a9b-h` | off | c | 8 | stop 108 | 33,028 | 225 | 0 | 136.1 | 1,385 | 647.2 / 702.4 | 0.662 / 0.721 | 33 → 54 (56.0) | `b-granite4-32b-a9b/b-granite4-32b-a9b.c.report.json` |
| `granite4:32b-a9b-h` | off | d | 8 | length 30, stop 18 | 24,792 | 276 | 0 | 98.5 | 1,321 | 987.3 / 1,008.2 | 1.059 / 1.089 | 53 → 61 (63.0) | `b-granite4-32b-a9b/b-granite4-32b-a9b.d.report.json` |
| `ministral-3:14b-instruct-2512-q4_K_M` | off | c | 8 | stop 108 | 89,582 | 218 | 0 | 144.3 | 9,860 | 539.6 / 569.6 | 0.554 / 0.584 | 37 → 52 (54.0) | `b-ministral3-14b/b-ministral3-14b.c.report.json` |
| `ministral-3:14b-instruct-2512-q4_K_M` | off | d | 8 | stop 48 | 50,036 | 115 | 0 | 121.9 | 6,188 | 632.2 / 672.5 | 0.655 / 0.692 | 52 → 58 (60.0) | `b-ministral3-14b/b-ministral3-14b.d.report.json` |
| `olmo-3.1:32b-instruct-q4_K_M` | off | c | 8 | stop 108 | 36,916 | 324 | 0 | 50.1 | 957 | 751.5 / 813.9 | 0.812 / 0.875 | 40 → 63 (66.0) | `b-olmo3.1-32b/b-olmo3.1-32b.c.report.json` |
| `olmo-3.1:32b-instruct-q4_K_M` | off | d | 8 | length 48 | 26,520 | 384 | 0 | 40.3 | 1,136 | 933.4 / 986.6 | 1.133 / 1.186 | 63 → 64 (68.0) | `b-olmo3.1-32b/b-olmo3.1-32b.d.report.json` |
| `glm-4.7-flash:q4_K_M` | off | c | 8 | stop 108 | 30,328 | 324 | 0 | 124.0 | 2,777 | 598.5 / 647.0 | 0.623 / 0.672 | 44 → 58 (60.0) | `b-glm4.7-flash/b-glm4.7-flash.c.report.json` |
| `glm-4.7-flash:q4_K_M` | off | d | 8 | length 16, stop 32 | 23,544 | 224 | 0 | 122.3 | 3,312 | 665.0 / 701.0 | 0.697 / 0.767 | 58 → 62 (65.0) | `b-glm4.7-flash/b-glm4.7-flash.d.report.json` |
| `glm-4.7-flash:q4_K_M` | on | c | 4096 | length 23, stop 85 | 30,328 | 207,852 | 833,504 | 120.4 | 2,708 | 668.0 / 758.3 | 10.965 / 35.461 | 54 → 67 (71.0) | `t-glm4.7-flash-on/t-glm4.7-flash-on.c.report.json` |
| `glm-4.7-flash:q4_K_M` | on | d | 4096 | length 16, stop 32 | 23,544 | 111,351 | 439,686 | 119.2 | 3,261 | 747.2 / 834.5 | 15.895 / 35.721 | 67 → 67 (71.0) | `t-glm4.7-flash-on/t-glm4.7-flash-on.d.report.json` |
| `qwen3.5:27b-q4_K_M` | off | c | 8 | stop 108 | 31,303 | 229 | 0 | 62.8 | 937 | 1,090.7 / 1,153.7 | 1.123 / 1.197 | 42 → 62 (65.0) | `b-qwen3.5-27b/b-qwen3.5-27b.c.report.json` |
| `qwen3.5:27b-q4_K_M` | off | d | 8 | stop 48 | 24,550 | 96 | 0 | 66.3 | 916 | 1,513.7 / 1,563.9 | 1.545 / 1.595 | 62 → 62 (65.0) | `b-qwen3.5-27b/b-qwen3.5-27b.d.report.json` |
| `qwen3.5:27b-q4_K_M` | on | c | 2048 | length 28, stop 80 | 31,087 | 140,545 | 594,024 | 37.8 | 916 | 1,134.8 / 1,205.6 | 31.558 / 55.402 | 44 → 67 (71.0) | `t-qwen3.5-27b-on/t-qwen3.5-27b-on.c.report.json` |
| `gemma4:31b-it-q4_K_M` | off | c | 8 | length 4, stop 104 | 30,905 | 444 | 0 | 43.8 | 852 | 1,387.3 / 1,490.3 | 1.481 / 1.596 | 39 → 62 (65.0) | `b-gemma4-31b/b-gemma4-31b.c.report.json` |
| `gemma4:31b-it-q4_K_M` | off | d | 8 | stop 48 | 24,760 | 136 | 0 | 49.6 | 944 | 1,811.8 / 1,897.2 | 1.872 / 1.979 | 62 → 62 (65.0) | `b-gemma4-31b/b-gemma4-31b.d.report.json` |
| `gemma4:26b` | off | c | 8 | length 1, stop 107 | 30,905 | 398 | 0 | 189.3 | 2,143 | 980.9 / 1,047.5 | 1.000 / 1.066 | 34 → 53 (52.0) | `k0-gemma4-26b-c/k0-gemma4-26b-c.c.report.json` |
| `gemma4:26b` | off | d | 8 | length 3, stop 45 | 24,760 | 140 | 0 | 135.5 | 2,572 | 1,089.2 / 1,147.4 | 1.108 / 1.173 | 54 → 61 (62.0) | `k0-gemma4-26b-d/k0-gemma4-26b-d.d.report.json` |
| `gemma4:26b` | on | c | 4096 | length 30, stop 78 | 31,229 | 219,663 | 844,314 | 210.1 | 2,117 | 1,105.6 / 1,166.8 | 6.346 / 21.545 | 56 → 66 (71.0) | `k0-gemma4-26b-think/k0-gemma4-26b-think.c.report.json` |
| `gemma4:26b` | on | d | 4096 | length 1, stop 47 | 24,904 | 32,535 | 123,454 | 212.6 | 2,600 | 1,186.2 / 1,239.1 | 3.703 / 7.903 | 66 → 65 (70.0) | `k0-gemma4-26b-think/k0-gemma4-26b-think.d.report.json` |
| `gemma4:26b-a4b-it-q4_K_M` | off | c | 8 | stop 108 | 30,905 | 428 | 0 | 134.9 | 2,632 | 950.7 / 993.6 | 0.982 / 1.024 | 40 → 58 (58.0) | `k0a-gemma4-anchor-cd/k0a-gemma4-anchor-cd.c.report.json` |
| `gemma4:26b-a4b-it-q4_K_M` | off | d | 8 | length 2, stop 46 | 24,760 | 172 | 0 | 137.6 | 3,342 | 1,046.3 / 1,087.3 | 1.076 / 1.120 | 58 → 62 (64.0) | `k0a-gemma4-anchor-cd/k0a-gemma4-anchor-cd.d.report.json` |
| `gemma4:31b-it-q4_K_M` | on | c | 2048 | length 6, stop 102 | 31,229 | 80,824 | 319,286 | 34.3 | 840 | 1,691.6 / 1,839.7 | 19.437 / 62.029 | 43 → 67 (70.0) | `t-gemma4-31b-on/t-gemma4-31b-on.c.report.json` |
| `granite4.1:30b-q4_K_M` | off | c | 8 | length 5, stop 103 | 30,652 | 362 | 0 | 53.6 | 1,460 | 493.1 / 555.9 | 0.549 / 0.675 | 42 → 60 (62.0) | `b-granite4.1-30b/b-granite4.1-30b.c.report.json` |
| `granite4.1:30b-q4_K_M` | off | d | 8 | length 8, stop 40 | 23,736 | 145 | 0 | 57.2 | 1,289 | 720.0 / 733.6 | 0.749 / 0.902 | 60 → 63 (67.0) | `b-granite4.1-30b/b-granite4.1-30b.d.report.json` |
| `qwen3.6:27b-q4_K_M` | off | c | 8 | stop 108 | 31,303 | 298 | 0 | 55.1 | 944 | 1,068.9 / 1,145.6 | 1.121 / 1.195 | 37 → 62 (65.0) | `b-qwen3.6-27b/b-qwen3.6-27b.c.report.json` |
| `qwen3.6:27b-q4_K_M` | off | d | 8 | stop 48 | 24,550 | 144 | 0 | 52.6 | 918 | 1,496.7 / 1,543.0 | 1.555 / 1.601 | 62 → 62 (65.0) | `b-qwen3.6-27b/b-qwen3.6-27b.d.report.json` |
| `gemma4:12b-it-q4_K_M` | off | c | 8 | length 70, stop 38 | 30,905 | 712 | 0 | 82.8 | 2,045 | 987.2 / 1,041.4 | 1.069 / 1.141 | 46 → 63 (65.0) | `b-gemma4-12b/b-gemma4-12b.c.report.json` |
| `gemma4:12b-it-q4_K_M` | off | d | 8 | length 3, stop 45 | 24,760 | 148 | 0 | 97.1 | 2,327 | 1,162.2 / 1,231.6 | 1.190 / 1.280 | 63 → 61 (63.0) | `b-gemma4-12b/b-gemma4-12b.d.report.json` |
| `gemma4:12b-it-q4_K_M` | on | c | 4096 | length 59, stop 49 | 31,229 | 282,627 | 1,113,825 | 70.8 | 2,015 | 1,162.8 / 1,221.1 | 59.361 / 59.599 | 54 → 66 (71.0) | `t-gemma4-12b-on/t-gemma4-12b-on.c.report.json` |
| `gemma4:12b-it-q4_K_M` | on | d | 4096 | length 9, stop 39 | 24,904 | 66,974 | 252,830 | 71.2 | 2,324 | 1,274.2 / 1,353.0 | 12.497 / 59.744 | 66 → 66 (71.0) | `t-gemma4-12b-on/t-gemma4-12b-on.d.report.json` |
| `qwen3.8:27b-q4_K_M` | off | c | 8 | stop 108 | 31,303 | 216 | 0 | 69.3 | 938 | 999.6 / 1,062.0 | 1.029 / 1.092 | 44 → 62 (66.0) | `b-qwen3.8-27b/b-qwen3.8-27b.c.report.json` |
| `qwen3.8:27b-q4_K_M` | off | d | 8 | length 39, stop 9 | 24,550 | 330 | 0 | 43.8 | 915 | 1,424.7 / 1,457.4 | 1.609 / 1.642 | 62 → 63 (67.0) | `b-qwen3.8-27b/b-qwen3.8-27b.d.report.json` |
| `qwen3.8:27b-q4_K_M` | on | c | 2048 | stop 108 | 31,087 | 51,416 | 224,074 | 39.1 | 922 | 1,045.3 / 1,113.9 | 12.481 / 26.185 | 43 → 66 (70.0) | `t-qwen3.8-27b-on/t-qwen3.8-27b-on.c.report.json` |

### Energy per task window (R-10; board vs wall labelled)

| model | think | task | BOARD J (s) | BOARD mean / peak W | BOARD resident idle W | BOARD J per decision (net) | BOARD J per 1k tokens (in+out) | BOARD J per prompt token, apportioned | BOARD J per generated token, apportioned | WALL J (s) | WALL mean / peak W | WALL box idle, resident / no model W | WALL J per decision | files |
|---|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|
| `falcon:7b-instruct-q4_0` | off | c | 6,079 (32.0) | 189.9 / 204.8 | 111.3 | 56.29 (23.31) | 184.1 | 0.0305 | 1.112 | 7,983 (31.0) | 257.5 / 272.0 | 166.0 / 92.1 | 73.92 | `b-falcon-7b/b-falcon-7b.c.watts.csv`, `b-falcon-7b/wall.csv` |
| `falcon:7b-instruct-q4_0` | off | d | 3,229 (16.0) | 201.7 / 216.2 | 114.5 | 67.26 (29.10) | 126.5 | 0.0362 | 1.176 | 4,628 (15.0) | 308.5 / 316.0 | 271.0 / 92.1 | 96.42 | `b-falcon-7b/b-falcon-7b.d.watts.csv`, `b-falcon-7b/wall.csv` |
| `vicuna:33b-q4_0` | off | c | 19,273 (87.0) | 221.5 / 253.8 | 107.2 | 178.46 (92.09) | 489.9 | 0.1389 | 4.934 | 25,462 (86.0) | 296.0 / 356.0 | 182.0 / 97.3 | 235.76 | `b-vicuna-33b/b-vicuna-33b.c.watts.csv`, `b-vicuna-33b/wall.csv` |
| `vicuna:33b-q4_0` | off | d | 12,891 (58.0) | 222.2 / 240.1 | 119.4 | 268.56 (124.23) | 457.8 | 0.1759 | 4.968 | 20,552 (57.0) | 360.5 / 404.0 | 315.0 / 97.3 | 428.17 | `b-vicuna-33b/b-vicuna-33b.d.watts.csv`, `b-vicuna-33b/wall.csv` |
| `llama2:13b-chat-q4_K_M` | off | c | 10,405 (51.0) | 204.0 / 232.1 | 109.9 | 96.34 (44.41) | 279.8 | 0.0596 | 2.191 | 13,463 (50.0) | 269.2 / 294.0 | 107.1 / 80.0 | 124.66 | `b-llama2-13b/b-llama2-13b.c.watts.csv`, `b-llama2-13b/wall.csv` |
| `llama2:13b-chat-q4_K_M` | off | d | 6,421 (31.0) | 207.1 / 231.0 | 117.9 | 133.77 (57.58) | 236.2 | 0.0734 | 2.213 | 8,830 (30.0) | 294.3 / 319.0 | 206.0 / 80.0 | 183.95 | `b-llama2-13b/b-llama2-13b.d.watts.csv`, `b-llama2-13b/wall.csv` |
| `llama2:7b-chat-q4_K_M` | off | c | 7,049 (36.0) | 195.7 / 210.6 | 111.2 | 65.27 (28.18) | 189.6 | 0.0316 | 1.203 | 9,045 (35.0) | 258.4 / 298.0 | 170.0 / 88.5 | 83.75 | `b-llama2-7b/b-llama2-7b.c.watts.csv`, `b-llama2-7b/wall.csv` |
| `llama2:7b-chat-q4_K_M` | off | d | 4,345 (22.0) | 197.4 / 219.8 | 113.7 | 90.52 (38.37) | 159.5 | 0.0405 | 1.230 | 4,768 (21.0) | 227.0 / 290.0 | 290.3 / 88.5 | 99.33 | `b-llama2-7b/b-llama2-7b.d.watts.csv`, `b-llama2-7b/wall.csv` |
| `mistral:v0.1` | off | c | 5,384 (26.0) | 207.0 / 224.0 | 112.1 | 49.85 (22.85) | 150.3 | 0.0327 | 1.200 | 6,901 (25.0) | 276.0 / 325.0 | 101.6 / 125.2 | 63.90 | `k1-mistral-v0.1-c/k1-mistral-v0.1-c.c.watts.csv`, `k1-mistral-v0.1-c/wall.csv` |
| `mistral:v0.1` | off | d | 3,227 (15.0) | 215.1 / 236.1 | 112.9 | 67.23 (31.93) | 123.6 | 0.0425 | 1.255 | 3,347 (14.0) | 239.0 / 298.0 | 141.6 / 98.9 | 69.72 | `k1-mistral-v0.1-d/k1-mistral-v0.1-d.d.watts.csv`, `k1-mistral-v0.1-d/wall.csv` |
| `yi:34b-chat-q4_0` | off | c | 13,336 (54.0) | 246.9 / 269.1 | 107.2 | 123.48 (69.86) | 368.8 | 0.1655 | 5.240 | 17,686 (53.0) | 333.7 / 348.0 | 172.4 / 86.2 | 163.75 | `b-yi-34b/b-yi-34b.c.watts.csv`, `b-yi-34b/wall.csv` |
| `yi:34b-chat-q4_0` | off | d | 8,031 (32.0) | 250.9 / 278.6 | 124.7 | 167.31 (84.14) | 304.0 | 0.2021 | 3.445 | 11,503 (31.0) | 371.0 / 380.0 | 286.0 / 86.2 | 239.65 | `b-yi-34b/b-yi-34b.d.watts.csv`, `b-yi-34b/wall.csv` |
| `phi:2.7b-chat-v2-q4_K_M` | off | c | 4,134 (27.0) | 153.1 / 159.8 | 110.8 | 38.28 (10.56) | 117.8 | 0.0125 | 0.615 | 6,070 (26.0) | 233.4 / 237.0 | 198.0 / 84.8 | 56.20 | `b-phi2-2.7b/b-phi2-2.7b.c.watts.csv`, `b-phi2-2.7b/wall.csv` |
| `phi:2.7b-chat-v2-q4_K_M` | off | d | 2,548 (15.0) | 169.8 / 190.5 | 112.3 | 53.08 (17.98) | 94.9 | 0.0160 | 0.708 | 3,297 (14.0) | 235.5 / 254.0 | 180.0 / 84.8 | 68.69 | `b-phi2-2.7b/b-phi2-2.7b.d.watts.csv`, `b-phi2-2.7b/wall.csv` |
| `gemma:7b-instruct-v1.1-q4_K_M` | off | c | 8,344 (60.0) | 139.0 / 157.8 | 111.1 | 77.26 (15.51) | 259.4 | 0.0283 | 0.747 | 13,183 (59.0) | 223.4 / 235.0 | 194.0 / 89.3 | 122.07 | `b-gemma1.1-7b/b-gemma1.1-7b.c.watts.csv`, `b-gemma1.1-7b/wall.csv` |
| `gemma:7b-instruct-v1.1-q4_K_M` | off | d | 6,470 (36.0) | 179.7 / 200.1 | 112.7 | 134.78 (50.20) | 253.9 | 0.0400 | 1.575 | 8,395 (35.0) | 239.8 / 284.0 | 212.0 / 89.3 | 174.89 | `b-gemma1.1-7b/b-gemma1.1-7b.d.watts.csv`, `b-gemma1.1-7b/wall.csv` |
| `qwen:32b-chat-v1.5-q4_K_M` | off | c | 14,142 (70.0) | 202.0 / 235.7 | 107.0 | 130.95 (61.56) | 433.1 | 0.1459 | 3.040 | 19,695 (69.0) | 285.4 / 325.0 | 169.1 / 78.9 | 182.36 | `b-qwen1.5-32b/b-qwen1.5-32b.c.watts.csv`, `b-qwen1.5-32b/wall.csv` |
| `qwen:32b-chat-v1.5-q4_K_M` | off | d | 10,137 (45.0) | 225.2 / 253.0 | 119.2 | 211.19 (99.43) | 395.9 | 0.1927 | 4.031 | 14,534 (44.0) | 330.3 / 383.0 | 288.0 / 78.9 | 302.78 | `b-qwen1.5-32b/b-qwen1.5-32b.d.watts.csv`, `b-qwen1.5-32b/wall.csv` |
| `llama3:8b-instruct-q4_K_M` | off | c | 9,648 (56.0) | 172.2 / 191.6 | 110.6 | 89.33 (31.97) | 304.5 | 0.0301 | 1.210 | 13,960 (55.0) | 253.8 / 270.0 | 168.3 / 106.5 | 129.26 | `b-llama3-8b/b-llama3-8b.c.watts.csv`, `b-llama3-8b/wall.csv` |
| `llama3:8b-instruct-q4_K_M` | off | d | 4,616 (26.0) | 177.5 / 191.9 | 117.2 | 96.16 (32.64) | 192.2 | 0.0359 | 1.182 | 6,056 (25.0) | 242.2 / 269.0 | 197.1 / 106.5 | 126.18 | `b-llama3-8b/b-llama3-8b.d.watts.csv`, `b-llama3-8b/wall.csv` |
| `phi3:14b-medium-128k-instruct-q4_0` | off | c | 6,808 (28.0) | 243.1 / 258.2 | 110.0 | 63.04 (34.51) | 189.6 | 0.0693 | 2.390 | 8,012 (27.0) | 296.7 / 341.0 | 139.6 / 78.5 | 74.18 | `b-phi3-14b/b-phi3-14b.c.watts.csv`, `b-phi3-14b/wall.csv` |
| `phi3:14b-medium-128k-instruct-q4_0` | off | d | 3,481 (15.0) | 232.0 / 258.9 | 125.0 | 72.51 (33.43) | 131.7 | 0.0774 | 1.419 | 3,388 (14.0) | 242.0 / 310.0 | 247.3 / 78.5 | 70.58 | `b-phi3-14b/b-phi3-14b.d.watts.csv`, `b-phi3-14b/wall.csv` |
| `mistral:7b-instruct-v0.3-q4_K_M` | off | c | 4,866 (21.0) | 231.6 / 247.0 | 111.1 | 45.05 (23.44) | 138.2 | 0.0401 | 1.443 | 6,407 (20.0) | 320.3 / 342.0 | 177.0 / 88.7 | 59.33 | `b-mistral-7b-v0.3/b-mistral-7b-v0.3.c.watts.csv`, `b-mistral-7b-v0.3/wall.csv` |
| `mistral:7b-instruct-v0.3-q4_K_M` | off | d | 3,049 (13.0) | 234.4 / 250.1 | 117.1 | 63.52 (31.80) | 118.0 | 0.0506 | 1.466 | 2,137 (12.0) | 178.1 / 341.0 | 325.7 / 88.7 | 44.52 | `b-mistral-7b-v0.3/b-mistral-7b-v0.3.d.watts.csv`, `b-mistral-7b-v0.3/wall.csv` |
| `gemma2:27b-instruct-q4_K_M` | off | c | 17,737 (89.0) | 199.3 / 234.0 | 107.7 | 164.23 (75.49) | 574.1 | 0.1801 | 3.546 | 25,838 (88.0) | 293.6 / 345.0 | 162.2 / 76.6 | 239.24 | `b-gemma2-27b/b-gemma2-27b.c.watts.csv`, `b-gemma2-27b/wall.csv` |
| `gemma2:27b-instruct-q4_K_M` | off | d | 9,937 (49.0) | 202.7 / 223.7 | 120.2 | 207.02 (84.26) | 404.4 | 0.1614 | 2.868 | 15,462 (49.0) | 315.5 / 356.0 | 270.0 / 76.6 | 322.12 | `b-gemma2-27b/b-gemma2-27b.d.watts.csv`, `b-gemma2-27b/wall.csv` |
| `mistral-nemo:12b-instruct-2407-q4_K_M` | off | c | 11,599 (64.0) | 181.2 / 193.7 | 109.9 | 107.40 (42.23) | 375.1 | 0.0486 | 1.840 | 15,307 (63.0) | 242.9 / 273.0 | 104.1 / 76.0 | 141.73 | `b-mistral-nemo-12b/b-mistral-nemo-12b.c.watts.csv`, `b-mistral-nemo-12b/wall.csv` |
| `mistral-nemo:12b-instruct-2407-q4_K_M` | off | d | 5,869 (31.0) | 189.3 / 213.8 | 125.8 | 122.27 (40.98) | 246.5 | 0.0570 | 1.754 | 7,982 (30.0) | 266.0 / 277.0 | 275.0 / 76.0 | 166.29 | `b-mistral-nemo-12b/b-mistral-nemo-12b.d.watts.csv`, `b-mistral-nemo-12b/wall.csv` |
| `llama3.1:8b-instruct-q4_K_M` | off | c | 7,369 (54.0) | 136.4 / 149.6 | 111.1 | 68.24 (12.66) | 236.4 | 0.0250 | 0.779 | 11,626 (53.0) | 219.3 / 224.0 | 190.0 / 77.3 | 107.65 | `b-llama3.1-8b/b-llama3.1-8b.c.watts.csv`, `b-llama3.1-8b/wall.csv` |
| `llama3.1:8b-instruct-q4_K_M` | off | d | 4,944 (28.0) | 176.5 / 191.8 | 112.5 | 102.99 (37.37) | 205.7 | 0.0356 | 1.192 | 6,782 (27.0) | 251.1 / 268.0 | 174.0 / 77.3 | 141.29 | `b-llama3.1-8b/b-llama3.1-8b.d.watts.csv`, `b-llama3.1-8b/wall.csv` |
| `mistral-small:22b-instruct-2409-q4_K_M` | off | c | 8,482 (34.0) | 249.4 / 273.8 | 106.8 | 78.54 (44.91) | 243.3 | 0.1300 | 3.470 | 10,365 (33.0) | 314.1 / 343.0 | 176.0 / 68.3 | 95.97 | `b-mistral-small-22b/b-mistral-small-22b.c.watts.csv`, `b-mistral-small-22b/wall.csv` |
| `mistral-small:22b-instruct-2409-q4_K_M` | off | d | 6,764 (26.0) | 260.1 / 271.7 | 122.9 | 140.92 (74.35) | 263.4 | 0.1702 | 3.877 | 8,468 (25.0) | 338.7 / 369.0 | 194.6 / 68.3 | 176.42 | `b-mistral-small-22b/b-mistral-small-22b.d.watts.csv`, `b-mistral-small-22b/wall.csv` |
| `qwen2.5:32b-instruct-q4_K_M` | off | c | 14,170 (70.0) | 202.4 / 231.4 | 107.3 | 131.21 (61.62) | 426.8 | 0.1356 | 3.631 | 18,108 (69.0) | 262.4 / 312.0 | 197.3 / 82.6 | 167.67 | `b-qwen2.5-32b/b-qwen2.5-32b.c.watts.csv`, `b-qwen2.5-32b/wall.csv` |
| `qwen2.5:32b-instruct-q4_K_M` | off | d | 9,322 (42.0) | 221.9 / 244.1 | 119.8 | 194.21 (89.36) | 375.1 | 0.1728 | 3.466 | 12,084 (41.0) | 294.7 / 378.0 | 259.4 / 82.6 | 251.75 | `b-qwen2.5-32b/b-qwen2.5-32b.d.watts.csv`, `b-qwen2.5-32b/wall.csv` |
| `phi4:14b-q4_K_M` | off | c | 10,494 (51.0) | 205.7 / 224.0 | 110.1 | 97.16 (45.14) | 330.7 | 0.0630 | 2.368 | 14,373 (50.0) | 287.4 / 319.0 | 166.0 / 94.3 | 133.08 | `b-phi4-14b/b-phi4-14b.c.watts.csv`, `b-phi4-14b/wall.csv` |
| `phi4:14b-q4_K_M` | off | d | 6,026 (28.0) | 215.1 / 227.6 | 120.7 | 125.53 (55.09) | 248.8 | 0.0745 | 2.484 | 6,873 (27.0) | 254.5 / 315.0 | 278.0 / 94.3 | 143.19 | `b-phi4-14b/b-phi4-14b.d.watts.csv`, `b-phi4-14b/wall.csv` |
| `deepseek-r1:32b-qwen-distill-q4_K_M` | off | c | 19,294 (87.0) | 221.7 / 252.2 | 108.2 | 178.65 (91.48) | 622.8 | 0.1612 | 5.524 | 25,097 (86.0) | 291.8 / 360.0 | 158.3 / 77.0 | 232.38 | `b-deepseek-r1-32b/b-deepseek-r1-32b.c.watts.csv`, `b-deepseek-r1-32b/wall.csv` |
| `deepseek-r1:32b-qwen-distill-q4_K_M` | off | d | 11,774 (50.0) | 235.4 / 251.0 | 121.1 | 245.29 (119.09) | 492.9 | 0.1920 | 5.881 | 15,102 (49.0) | 308.2 / 371.0 | 209.0 / 77.0 | 314.62 | `b-deepseek-r1-32b/b-deepseek-r1-32b.d.watts.csv`, `b-deepseek-r1-32b/wall.csv` |
| `deepseek-r1:32b-qwen-distill-q4_K_M` | on | c | 696,094 (2,356.4) | 295.4 / 302.2 | 108.3 | 6,445.31 (4,081.47) | 6,380.5 | 0.2192 | 8.504 | 891,821 (2,355.3) | 378.6 / 400.0 | 157.1 / 68.1 | 8,257.60 | `t-deepseek-r1-32b-on/t-deepseek-r1-32b-on.c.watts.csv`, `t-deepseek-r1-32b-on/wall.csv` |
| `mistral-small:24b-instruct-2501-q4_K_M` | off | c | 12,463 (67.0) | 186.0 / 211.7 | 110.2 | 115.40 (47.01) | 262.7 | 0.0602 | 2.445 | 16,783 (66.0) | 254.3 / 296.0 | 194.6 / 80.2 | 155.40 | `b-mistral-small-24b-2501/b-mistral-small-24b-2501.c.watts.csv`, `b-mistral-small-24b-2501/wall.csv` |
| `mistral-small:24b-instruct-2501-q4_K_M` | off | d | 7,686 (38.0) | 202.2 / 236.5 | 126.0 | 160.12 (60.31) | 245.5 | 0.0867 | 2.654 | 11,674 (37.0) | 315.5 / 336.0 | 261.0 / 80.2 | 243.21 | `b-mistral-small-24b-2501/b-mistral-small-24b-2501.d.watts.csv`, `b-mistral-small-24b-2501/wall.csv` |
| `gemma3:27b-it-q4_K_M` | off | c | 23,626 (132.0) | 178.9 / 195.1 | 109.8 | 218.76 (84.55) | 766.3 | 0.1746 | 2.862 | 32,454 (131.0) | 247.7 / 311.0 | 200.7 / 86.3 | 300.50 | `b-gemma3-27b/b-gemma3-27b.c.watts.csv`, `b-gemma3-27b/wall.csv` |
| `gemma3:27b-it-q4_K_M` | off | d | 15,621 (79.0) | 197.7 / 241.1 | 115.5 | 325.45 (135.31) | 625.0 | 0.1697 | 4.630 | 20,736 (78.0) | 265.8 / 346.0 | 187.5 / 86.3 | 431.99 | `b-gemma3-27b/b-gemma3-27b.d.watts.csv`, `b-gemma3-27b/wall.csv` |
| `qwen3:30b-a3b-q4_K_M` | off | c | 9,522 (54.0) | 176.3 / 196.0 | 108.4 | 88.17 (33.98) | 295.3 | 0.0495 | 1.038 | 12,695 (53.0) | 239.5 / 277.0 | 127.2 / 73.0 | 117.55 | `b-qwen3-30b-a3b/b-qwen3-30b-a3b.c.watts.csv`, `b-qwen3-30b-a3b/wall.csv` |
| `qwen3:30b-a3b-q4_K_M` | off | d | 4,986 (27.0) | 184.6 / 191.5 | 118.6 | 103.88 (37.16) | 203.5 | 0.0446 | 1.108 | 5,995 (26.0) | 230.6 / 274.0 | 256.9 / 73.0 | 124.90 | `b-qwen3-30b-a3b/b-qwen3-30b-a3b.d.watts.csv`, `b-qwen3-30b-a3b/wall.csv` |
| `qwen3:32b-q4_K_M` | off | c | 18,960 (86.0) | 220.4 / 248.8 | 107.7 | 175.56 (89.80) | 587.4 | 0.1628 | 5.403 | 23,165 (85.0) | 272.5 / 337.0 | 229.0 / 87.0 | 214.49 | `b-qwen3-32b/b-qwen3-32b.c.watts.csv`, `b-qwen3-32b/wall.csv` |
| `qwen3:32b-q4_K_M` | off | d | 9,650 (43.0) | 224.4 / 246.0 | 121.6 | 201.05 (92.09) | 397.6 | 0.1799 | 4.462 | 12,360 (42.0) | 294.2 / 385.0 | 264.0 / 87.0 | 257.49 | `b-qwen3-32b/b-qwen3-32b.d.watts.csv`, `b-qwen3-32b/wall.csv` |
| `qwen3:30b-a3b-q4_K_M` | on | c | 232,881 (810.2) | 287.4 / 299.1 | 111.8 | 2,156.30 (1,317.72) | 1,471.8 | 0.0838 | 1.684 | 305,215 (809.3) | 377.1 / 399.0 | 131.4 / 77.0 | 2,826.06 | `t-qwen3-30b-a3b-on/t-qwen3-30b-a3b-on.c.watts.csv`, `t-qwen3-30b-a3b-on/wall.csv` |
| `qwen3:30b-a3b-q4_K_M` | on | d | 46,498 (168.0) | 276.7 / 299.6 | 123.1 | 968.70 (537.89) | 970.1 | 0.0679 | 1.612 | 61,460 (167.1) | 367.8 / 397.0 | 255.3 / 77.0 | 1,280.41 | `t-qwen3-30b-a3b-on/t-qwen3-30b-a3b-on.d.watts.csv`, `t-qwen3-30b-a3b-on/wall.csv` |
| `qwen3:32b-q4_K_M` | on | c | 685,931 (2,322.4) | 295.4 / 303.1 | 108.0 | 6,351.21 (4,028.35) | 6,322.2 | 0.2203 | 8.520 | 883,576 (2,321.3) | 380.6 / 397.0 | 110.5 / 70.1 | 8,181.26 | `t-qwen3-32b-on/t-qwen3-32b-on.c.watts.csv`, `t-qwen3-32b-on/wall.csv` |
| `mistral-small3.2:24b-instruct-2506-q4_K_M` | off | c | 12,415 (67.0) | 185.3 / 223.9 | 110.1 | 114.95 (46.64) | 146.4 | 0.0337 | 2.388 | 17,405 (66.0) | 263.7 / 316.0 | 159.3 / 74.6 | 161.16 | `b-mistral-small3.2-24b/b-mistral-small3.2-24b.c.watts.csv`, `b-mistral-small3.2-24b/wall.csv` |
| `mistral-small3.2:24b-instruct-2506-q4_K_M` | off | d | 7,464 (36.0) | 207.3 / 241.5 | 123.8 | 155.49 (62.60) | 155.8 | 0.0584 | 2.490 | 8,156 (35.0) | 233.0 / 262.0 | 225.8 / 74.6 | 169.93 | `b-mistral-small3.2-24b/b-mistral-small3.2-24b.d.watts.csv`, `b-mistral-small3.2-24b/wall.csv` |
| `qwen3:30b-a3b-instruct-2507-q4_K_M` | off | c | 9,182 (54.0) | 170.0 / 189.0 | 106.5 | 85.02 (31.74) | 292.6 | 0.0476 | 0.994 | 11,932 (53.0) | 225.1 / 249.0 | 184.0 / 82.3 | 110.48 | `b-qwen3-30b-a3b-2507/b-qwen3-30b-a3b-2507.c.watts.csv`, `b-qwen3-30b-a3b-2507/wall.csv` |
| `qwen3:30b-a3b-instruct-2507-q4_K_M` | off | d | 4,859 (27.0) | 179.9 / 187.9 | 115.6 | 101.22 (36.20) | 201.4 | 0.0433 | 1.056 | 6,526 (26.0) | 251.0 / 268.0 | 244.0 / 82.3 | 135.97 | `b-qwen3-30b-a3b-2507/b-qwen3-30b-a3b-2507.d.watts.csv`, `b-qwen3-30b-a3b-2507/wall.csv` |
| `gpt-oss:20b` | low | c | 41,044 (179.0) | 229.2 / 300.5 | 106.8 | 380.04 (202.97) | 835.1 | 0.0588 | 1.625 | 54,570 (178.0) | 306.5 / 397.0 | 175.6 / 78.2 | 505.28 | `b-gpt-oss-20b-low/b-gpt-oss-20b-low.c.watts.csv`, `b-gpt-oss-20b-low/wall.csv` |
| `gpt-oss:20b` | low | d | 14,430 (67.0) | 215.3 / 298.9 | 120.8 | 300.63 (131.94) | 476.9 | 0.0452 | 1.526 | 20,198 (66.0) | 306.0 / 394.0 | 350.5 / 78.2 | 420.78 | `b-gpt-oss-20b-low/b-gpt-oss-20b-low.d.watts.csv`, `b-gpt-oss-20b-low/wall.csv` |
| `gpt-oss:20b` | medium | c | 221,548 (783.2) | 282.9 / 300.5 | 111.8 | 2,051.37 (1,240.28) | 1,676.1 | 0.0761 | 2.038 | 294,593 (781.5) | 377.0 / 407.0 | 179.3 / 81.2 | 2,727.71 | `t-gpt-oss-20b-medium/t-gpt-oss-20b-medium.c.watts.csv`, `t-gpt-oss-20b-medium/wall.csv` |
| `gpt-oss:20b` | medium | d | 44,240 (167.0) | 264.8 / 302.4 | 128.2 | 921.67 (475.60) | 1,002.7 | 0.0570 | 1.899 | 55,423 (166.1) | 333.8 / 388.0 | 283.0 / 81.2 | 1,154.64 | `t-gpt-oss-20b-medium/t-gpt-oss-20b-medium.d.watts.csv`, `t-gpt-oss-20b-medium/wall.csv` |
| `granite4:32b-a9b-h` | off | c | 13,539 (72.0) | 188.0 / 208.0 | 106.2 | 125.36 (54.55) | 407.2 | 0.1358 | 1.381 | 18,582 (71.0) | 261.7 / 303.0 | 163.0 / 88.3 | 172.05 | `b-granite4-32b-a9b/b-granite4-32b-a9b.c.watts.csv`, `b-granite4-32b-a9b/wall.csv` |
| `granite4:32b-a9b-h` | off | d | 9,488 (50.0) | 189.7 / 202.3 | 116.0 | 197.68 (76.78) | 378.5 | 0.1436 | 1.926 | 15,655 (49.0) | 319.4 / 345.0 | 286.0 / 88.3 | 326.14 | `b-granite4-32b-a9b/b-granite4-32b-a9b.d.watts.csv`, `b-granite4-32b-a9b/wall.csv` |
| `ministral-3:14b-instruct-2512-q4_K_M` | off | c | 9,600 (59.0) | 162.7 / 174.3 | 108.3 | 88.89 (29.70) | 106.9 | 0.0165 | 1.127 | 12,951 (58.0) | 223.3 / 243.0 | 185.0 / 73.4 | 119.92 | `b-ministral3-14b/b-ministral3-14b.c.watts.csv`, `b-ministral3-14b/wall.csv` |
| `ministral-3:14b-instruct-2512-q4_K_M` | off | d | 5,764 (31.0) | 185.9 / 202.5 | 114.3 | 120.09 (46.27) | 114.9 | 0.0300 | 1.525 | 6,782 (30.0) | 226.0 / 270.0 | 231.4 / 73.4 | 141.29 | `b-ministral3-14b/b-ministral3-14b.d.watts.csv`, `b-ministral3-14b/wall.csv` |
| `olmo-3.1:32b-instruct-q4_K_M` | off | c | 19,391 (87.0) | 222.8 / 257.1 | 108.9 | 179.55 (91.77) | 520.7 | 0.2328 | 4.447 | 25,674 (86.0) | 298.5 / 360.0 | 108.2 / 75.0 | 237.72 | `b-olmo3.1-32b/b-olmo3.1-32b.c.watts.csv`, `b-olmo3.1-32b/wall.csv` |
| `olmo-3.1:32b-instruct-q4_K_M` | off | d | 12,767 (54.0) | 236.4 / 248.2 | 121.6 | 265.99 (129.20) | 474.6 | 0.2080 | 5.859 | 15,033 (53.0) | 283.6 / 320.0 | 212.1 / 75.0 | 313.19 | `b-olmo3.1-32b/b-olmo3.1-32b.d.watts.csv`, `b-olmo3.1-32b/wall.csv` |
| `glm-4.7-flash:q4_K_M` | off | c | 11,206 (67.0) | 167.2 / 180.4 | 108.5 | 103.76 (36.42) | 365.6 | 0.0602 | 1.349 | 16,280 (66.0) | 246.6 / 257.0 | 159.0 / 90.2 | 150.74 | `b-glm4.7-flash/b-glm4.7-flash.c.watts.csv`, `b-glm4.7-flash/wall.csv` |
| `glm-4.7-flash:q4_K_M` | off | d | 5,924 (33.0) | 179.5 / 203.3 | 119.4 | 123.41 (41.27) | 249.2 | 0.0542 | 1.468 | 7,707 (32.0) | 240.8 / 273.0 | 218.3 / 90.2 | 160.56 | `b-glm4.7-flash/b-glm4.7-flash.d.watts.csv`, `b-glm4.7-flash/wall.csv` |
| `glm-4.7-flash:q4_K_M` | on | c | 525,697 (1,809.4) | 290.5 / 299.7 | 111.7 | 4,867.56 (2,995.61) | 2,207.1 | 0.1073 | 2.412 | 683,572 (1,808.6) | 377.9 / 390.0 | 123.1 / 87.0 | 6,329.37 | `t-glm4.7-flash-on/t-glm4.7-flash-on.c.watts.csv`, `t-glm4.7-flash-on/wall.csv` |
| `glm-4.7-flash:q4_K_M` | on | d | 285,276 (975.3) | 292.5 / 299.9 | 123.9 | 5,943.26 (3,425.62) | 2,114.8 | 0.0897 | 2.454 | 365,807 (974.4) | 375.4 / 397.0 | 188.6 / 87.0 | 7,620.99 | `t-glm4.7-flash-on/t-glm4.7-flash-on.d.watts.csv`, `t-glm4.7-flash-on/wall.csv` |
| `qwen3.5:27b-q4_K_M` | off | c | 22,405 (120.0) | 186.7 / 200.6 | 108.0 | 207.46 (87.43) | 710.5 | 0.1993 | 2.975 | 29,758 (119.0) | 250.0 / 301.0 | 160.0 / 70.8 | 275.54 | `b-qwen3.5-27b/b-qwen3.5-27b.c.watts.csv`, `b-qwen3.5-27b/wall.csv` |
| `qwen3.5:27b-q4_K_M` | off | d | 13,828 (73.0) | 189.4 / 222.8 | 120.9 | 288.08 (104.09) | 561.1 | 0.2068 | 2.858 | 20,192 (72.0) | 280.4 / 370.0 | 208.3 / 70.8 | 420.67 | `b-qwen3.5-27b/b-qwen3.5-27b.d.watts.csv`, `b-qwen3.5-27b/wall.csv` |
| `qwen3.5:27b-q4_K_M` | on | c | 1,133,348 (3,846.8) | 294.6 / 302.7 | 109.1 | 10,493.96 (6,607.36) | 6,603.4 | 0.3215 | 7.793 | 1,458,624 (3,845.6) | 379.3 / 398.0 | 120.6 / 70.9 | 13,505.78 | `t-qwen3.5-27b-on/t-qwen3.5-27b-on.c.watts.csv`, `t-qwen3.5-27b-on/wall.csv` |
| `gemma4:31b-it-q4_K_M` | off | c | 29,088 (159.0) | 182.9 / 229.3 | 108.0 | 269.34 (110.25) | 927.9 | 0.2147 | 4.174 | 37,725 (158.0) | 238.7 / 320.0 | 122.8 / 74.1 | 349.30 | `b-gemma4-31b/b-gemma4-31b.c.watts.csv`, `b-gemma4-31b/wall.csv` |
| `gemma4:31b-it-q4_K_M` | off | d | 16,821 (89.0) | 189.0 / 237.8 | 121.3 | 350.44 (125.49) | 675.7 | 0.2002 | 3.809 | 23,176 (88.0) | 263.3 / 383.0 | 205.4 / 74.1 | 482.83 | `b-gemma4-31b/b-gemma4-31b.d.watts.csv`, `b-gemma4-31b/wall.csv` |
| `gemma4:26b` | off | c | 15,514 (107.0) | 145.0 / 157.2 | 106.4 | 143.65 (38.24) | 495.6 | 0.0677 | 0.766 | 24,251 (106.0) | 228.8 / 258.0 | 179.3 / 59.7 | 224.54 | `k0-gemma4-26b-c/k0-gemma4-26b-c.c.watts.csv`, `k0-gemma4-26b-c/wall.csv` |
| `gemma4:26b` | off | d | 8,609 (53.0) | 162.4 / 173.2 | 111.5 | 179.35 (56.18) | 345.7 | 0.0631 | 1.198 | 12,033 (52.0) | 231.4 / 279.0 | 149.1 / 114.9 | 250.68 | `k0-gemma4-26b-d/k0-gemma4-26b-d.d.watts.csv`, `k0-gemma4-26b-d/wall.csv` |
| `gemma4:26b` | on | c | 331,664 (1,164.2) | 284.9 / 300.9 | 112.1 | 3,070.96 (1,862.76) | 1,321.9 | 0.1346 | 1.356 | 433,612 (1,163.2) | 372.8 / 406.0 | 186.7 / 113.8 | 4,014.93 | `k0-gemma4-26b-think/k0-gemma4-26b-think.c.watts.csv`, `k0-gemma4-26b-think/wall.csv` |
| `gemma4:26b` | on | d | 54,671 (209.1) | 261.5 / 301.7 | 122.8 | 1,138.98 (604.32) | 951.8 | 0.1006 | 1.230 | 76,203 (208.0) | 366.3 / 398.0 | 200.1 / 113.8 | 1,587.56 | `k0-gemma4-26b-think/k0-gemma4-26b-think.d.watts.csv`, `k0-gemma4-26b-think/wall.csv` |
| `gemma4:26b-a4b-it-q4_K_M` | off | c | 16,138 (104.0) | 155.1 / 177.4 | 107.6 | 149.43 (45.78) | 515.1 | 0.0590 | 1.150 | 24,609 (103.0) | 238.9 / 265.0 | 184.7 / 97.9 | 227.86 | `k0a-gemma4-anchor-cd/k0a-gemma4-anchor-cd.c.watts.csv`, `k0a-gemma4-anchor-cd/wall.csv` |
| `gemma4:26b-a4b-it-q4_K_M` | off | d | 8,380 (51.0) | 164.3 / 176.4 | 114.6 | 174.59 (52.83) | 336.1 | 0.0492 | 1.194 | 12,269 (50.0) | 245.4 / 278.0 | 177.0 / 97.9 | 255.61 | `k0a-gemma4-anchor-cd/k0a-gemma4-anchor-cd.d.watts.csv`, `k0a-gemma4-anchor-cd/wall.csv` |
| `gemma4:31b-it-q4_K_M` | on | c | 736,771 (2,540.5) | 290.0 / 303.3 | 108.8 | 6,821.95 (4,262.08) | 6,575.2 | 0.3453 | 8.447 | 947,382 (2,539.5) | 373.1 / 395.0 | 182.0 / 77.5 | 8,772.05 | `t-gemma4-31b-on/t-gemma4-31b-on.c.watts.csv`, `t-gemma4-31b-on/wall.csv` |
| `granite4.1:30b-q4_K_M` | off | c | 12,745 (59.0) | 216.0 / 239.2 | 107.9 | 118.01 (59.05) | 410.9 | 0.1479 | 4.030 | 16,198 (58.0) | 279.2 / 300.0 | 153.1 / 72.7 | 149.98 | `b-granite4.1-30b/b-granite4.1-30b.c.watts.csv`, `b-granite4.1-30b/wall.csv` |
| `granite4.1:30b-q4_K_M` | off | d | 8,604 (37.0) | 232.5 / 261.0 | 126.0 | 179.25 (82.12) | 360.3 | 0.1803 | 4.066 | 11,260 (36.0) | 312.8 / 338.0 | 211.1 / 72.7 | 234.59 | `b-granite4.1-30b/b-granite4.1-30b.d.watts.csv`, `b-granite4.1-30b/wall.csv` |
| `qwen3.6:27b-q4_K_M` | off | c | 22,430 (120.0) | 186.9 / 201.2 | 107.0 | 207.69 (88.81) | 709.8 | 0.1979 | 3.390 | 27,984 (119.0) | 235.1 / 322.0 | 189.1 / 75.7 | 259.11 | `b-qwen3.6-27b/b-qwen3.6-27b.c.watts.csv`, `b-qwen3.6-27b/wall.csv` |
| `qwen3.6:27b-q4_K_M` | off | d | 14,129 (73.0) | 193.5 / 225.2 | 125.2 | 294.35 (103.93) | 572.1 | 0.2107 | 3.675 | 21,849 (72.0) | 303.4 / 392.0 | 185.1 / 75.7 | 455.18 | `b-qwen3.6-27b/b-qwen3.6-27b.d.watts.csv`, `b-qwen3.6-27b/wall.csv` |
| `gemma4:12b-it-q4_K_M` | off | c | 19,556 (114.0) | 171.5 / 193.4 | 109.3 | 181.08 (65.64) | 618.5 | 0.0839 | 2.072 | 27,286 (113.0) | 241.4 / 303.0 | 98.7 / 69.3 | 252.64 | `b-gemma4-12b/b-gemma4-12b.c.watts.csv`, `b-gemma4-12b/wall.csv` |
| `gemma4:12b-it-q4_K_M` | off | d | 9,774 (57.0) | 171.4 / 185.2 | 121.3 | 203.63 (59.55) | 392.4 | 0.0737 | 1.765 | 12,441 (56.0) | 222.1 / 294.0 | 204.4 / 69.3 | 259.18 | `b-gemma4-12b/b-gemma4-12b.d.watts.csv`, `b-gemma4-12b/wall.csv` |
| `gemma4:12b-it-q4_K_M` | on | c | 1,216,705 (4,131.8) | 294.5 / 303.2 | 111.7 | 11,265.79 (6,992.54) | 3,876.6 | 0.1461 | 4.160 | 1,563,825 (4,130.7) | 378.6 / 402.0 | 217.0 / 78.5 | 14,479.86 | `t-gemma4-12b-on/t-gemma4-12b-on.c.watts.csv`, `t-gemma4-12b-on/wall.csv` |
| `gemma4:12b-it-q4_K_M` | on | d | 292,084 (1,005.2) | 290.6 / 302.7 | 122.6 | 6,085.08 (3,517.40) | 3,179.0 | 0.1250 | 4.084 | 374,218 (1,004.1) | 372.7 / 390.0 | 382.0 / 78.5 | 7,796.21 | `t-gemma4-12b-on/t-gemma4-12b-on.d.watts.csv`, `t-gemma4-12b-on/wall.csv` |
| `qwen3.8:27b-q4_K_M` | off | c | 21,286 (111.0) | 191.7 / 209.5 | 109.2 | 197.10 (84.82) | 675.4 | 0.2045 | 2.768 | 30,525 (110.0) | 277.5 / 332.0 | 213.0 / 88.2 | 282.63 | `b-qwen3.8-27b/b-qwen3.8-27b.c.watts.csv`, `b-qwen3.8-27b/wall.csv` |
| `qwen3.8:27b-q4_K_M` | off | d | 15,254 (75.0) | 203.3 / 245.7 | 119.8 | 317.79 (130.53) | 613.1 | 0.2223 | 4.642 | 20,503 (74.0) | 277.0 / 377.0 | 175.2 / 88.2 | 427.15 | `b-qwen3.8-27b/b-qwen3.8-27b.d.watts.csv`, `b-qwen3.8-27b/wall.csv` |
| `qwen3.8:27b-q4_K_M` | on | c | 414,315 (1,428.3) | 290.1 / 302.2 | 107.9 | 3,836.25 (2,409.59) | 5,021.8 | 0.3146 | 7.409 | 525,323 (1,427.2) | 368.1 / 395.0 | 238.0 / 81.5 | 4,864.10 | `t-qwen3.8-27b-on/t-qwen3.8-27b-on.c.watts.csv`, `t-qwen3.8-27b-on/wall.csv` |

### The probes: decode, prefill, time to first token (one window each, think off / lowest; R-10, R-10c)

| model | decode tok/s (median of 3 × 256) | BOARD mean W (decode) | BOARD J per generated token (phase) | BOARD J per 1k generated tokens | generated tokens per BOARD J | tok/s per BOARD W | WALL J per generated token (phase) | tok/s per WALL W | prefill tok/s (median; nonce-first, no cache) | prompt tokens per run | BOARD J per prompt token (phase / window) | prompt tokens per BOARD J | TTFT median / p95 s (streamed, 12 task_c prompts) | file |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|
| `falcon:7b-instruct-q4_0` | 160.0 | 230.4 | 1.439 | 1,439 | 0.695 | 0.695 | 1.656 | 0.604 | 5,366 | 1,308 | 0.0408 / 0.0838 | 11.93 | 0.222 / 0.252 | `b-falcon-7b/probes.json + b-falcon-7b/probe-decode.board.csv` |
| `vicuna:33b-q4_0` | 40.6 | 282.6 | 6.958 | 6,958 | 0.144 | 0.144 | 8.167 | 0.122 | 1,204 | 1,510 | 0.1790 / 0.3715 | 2.69 | 0.683 / 0.873 | `b-vicuna-33b/probes.json + b-vicuna-33b/probe-decode.board.csv` |
| `llama2:13b-chat-q4_K_M` | 83.3 | 271.8 | 3.265 | 3,265 | 0.306 | 0.306 | 3.700 | 0.270 | 2,725 | 2,543 | 0.0743 / 0.1726 | 5.80 | 0.425 / 0.579 | `b-llama2-13b/probes.json + b-llama2-13b/probe-decode.board.csv` |
| `llama2:7b-chat-q4_K_M` | 149.5 | 255.9 | 1.710 | 1,710 | 0.585 | 0.585 | 2.425 | 0.412 | 4,671 | 2,543 | 0.0433 / 0.1148 | 8.71 | 0.285 / 0.425 | `b-llama2-7b/probes.json + b-llama2-7b/probe-decode.board.csv` |
| `mistral:v0.1` | 159.9 | 261.0 | 1.633 | 1,633 | 0.612 | 0.612 | 2.267 | 0.441 | 5,062 | 2,455 | 0.0449 / 0.0804 | 12.44 | 0.221 / 0.357 | `k1-mistral-v0.1-c/probes.json + k1-mistral-v0.1-c/probe-decode.board.csv` |
| `yi:34b-chat-q4_0` | 39.3 | 286.9 | 7.303 | 7,303 | 0.137 | 0.137 | 9.354 | 0.107 | 1,217 | 2,356 | 0.2147 / 0.2767 | 3.61 | 0.441 / 0.479 | `b-yi-34b/probes.json + b-yi-34b/probe-decode.board.csv` |
| `phi:2.7b-chat-v2-q4_K_M` | 237.6 | 236.6 | 0.996 | 996 | 1.004 | 1.004 | 0.787 | 1.270 | 10,598 | 1,340 | 0.0169 / 0.0756 | 13.24 | 0.224 / 0.276 | `b-phi2-2.7b/probes.json + b-phi2-2.7b/probe-decode.board.csv` |
| `gemma:7b-instruct-v1.1-q4_K_M` | 105.1 | 251.5 | 2.395 | 2,395 | 0.418 | 0.418 | 2.419 | 0.413 | 4,194 | 2,222 | 0.0448 / 0.1374 | 7.28 | 0.588 / 0.639 | `b-gemma1.1-7b/probes.json + b-gemma1.1-7b/probe-decode.board.csv` |
| `qwen:32b-chat-v1.5-q4_K_M` | 35.4 | 283.5 | 8.005 | 8,005 | 0.125 | 0.125 | 10.600 | 0.094 | 1,205 | 2,302 | 0.2043 / 0.2783 | 3.59 | 0.670 / 0.711 | `b-qwen1.5-32b/probes.json + b-qwen1.5-32b/probe-decode.board.csv` |
| `llama3:8b-instruct-q4_K_M` | 128.3 | 253.4 | 1.974 | 1,974 | 0.506 | 0.506 | 2.018 | 0.495 | 4,732 | 2,236 | 0.0435 / 0.0999 | 10.01 | 0.490 / 0.526 | `b-llama3-8b/probes.json + b-llama3-8b/probe-decode.board.csv` |
| `phi3:14b-medium-128k-instruct-q4_0` | 90.7 | 278.9 | 3.074 | 3,074 | 0.325 | 0.325 | 3.774 | 0.265 | 3,021 | 2,532 | 0.0819 / 0.1173 | 8.53 | 0.226 / 0.240 | `b-phi3-14b/probes.json + b-phi3-14b/probe-decode.board.csv` |
| `mistral:7b-instruct-v0.3-q4_K_M` | 144.2 | 264.6 | 1.833 | 1,833 | 0.546 | 0.546 | 2.362 | 0.423 | 4,687 | 2,450 | 0.0526 / 0.0818 | 12.22 | 0.172 / 0.180 | `b-mistral-7b-v0.3/probes.json + b-mistral-7b-v0.3/probe-decode.board.csv` |
| `gemma2:27b-instruct-q4_K_M` | 40.3 | 275.4 | 6.830 | 6,830 | 0.146 | 0.146 | 8.940 | 0.112 | 1,259 | 2,210 | 0.1784 / 0.3068 | 3.26 | 0.814 / 0.903 | `b-gemma2-27b/probes.json + b-gemma2-27b/probe-decode.board.csv` |
| `mistral-nemo:12b-instruct-2407-q4_K_M` | 87.7 | 264.4 | 3.011 | 3,012 | 0.332 | 0.332 | 2.711 | 0.369 | 3,253 | 2,240 | 0.0674 / 0.1371 | 7.29 | 0.528 / 0.589 | `b-mistral-nemo-12b/probes.json + b-mistral-nemo-12b/probe-decode.board.csv` |
| `llama3.1:8b-instruct-q4_K_M` | 128.5 | 251.5 | 1.958 | 1,958 | 0.511 | 0.511 | 2.869 | 0.348 | 4,739 | 2,236 | 0.0439 / 0.1008 | 9.92 | 0.506 / 0.528 | `b-llama3.1-8b/probes.json + b-llama3.1-8b/probe-decode.board.csv` |
| `mistral-small:22b-instruct-2409-q4_K_M` | 52.1 | 285.2 | 5.472 | 5,472 | 0.183 | 0.183 | 6.533 | 0.153 | 1,526 | 2,450 | 0.1726 / 0.2152 | 4.65 | 0.332 / 0.351 | `b-mistral-small-22b/probes.json + b-mistral-small-22b/probe-decode.board.csv` |
| `qwen2.5:32b-instruct-q4_K_M` | 35.0 | 284.4 | 8.120 | 8,120 | 0.123 | 0.123 | 10.640 | 0.094 | 1,207 | 2,264 | 0.2054 / 0.2845 | 3.52 | 0.661 / 0.715 | `b-qwen2.5-32b/probes.json + b-qwen2.5-32b/probe-decode.board.csv` |
| `phi4:14b-q4_K_M` | 75.8 | 271.3 | 3.578 | 3,578 | 0.280 | 0.280 | 4.306 | 0.232 | 2,805 | 2,236 | 0.0821 / 0.1442 | 6.93 | 0.397 / 0.441 | `b-phi4-14b/probes.json + b-phi4-14b/probe-decode.board.csv` |
| `deepseek-r1:32b-qwen-distill-q4_K_M` | 34.9 | 285.1 | 8.155 | 8,155 | 0.123 | 0.123 | 10.909 | 0.092 | 1,192 | 2,238 | 0.2082 / 0.2886 | 3.47 | 0.735 / 0.797 | `b-deepseek-r1-32b/probes.json + b-deepseek-r1-32b/probe-decode.board.csv` |
| `mistral-small:24b-instruct-2501-q4_K_M` | 49.4 | 277.9 | 5.624 | 5,624 | 0.178 | 0.178 | 7.387 | 0.135 | 1,879 | 2,398 | 0.1276 / 0.2001 | 5.00 | 0.581 / 0.676 | `b-mistral-small-24b-2501/probes.json + b-mistral-small-24b-2501/probe-decode.board.csv` |
| `gemma3:27b-it-q4_K_M` | 39.2 | 273.0 | 6.963 | 6,963 | 0.144 | 0.144 | 5.633 | 0.177 | 1,160 | 2,216 | 0.1784 / 0.3296 | 3.03 | 1.226 / 1.345 | `b-gemma3-27b/probes.json + b-gemma3-27b/probe-decode.board.csv` |
| `qwen3:30b-a3b-q4_K_M` | 175.0 | 237.1 | 1.357 | 1,357 | 0.737 | 0.737 | 1.425 | 0.702 | 4,004 | 2,251 | 0.0552 / 0.1063 | 9.41 | 0.397 / 0.414 | `b-qwen3-30b-a3b/probes.json + b-qwen3-30b-a3b/probe-decode.board.csv` |
| `qwen3:32b-q4_K_M` | 35.0 | 285.5 | 8.167 | 8,167 | 0.122 | 0.122 | 9.962 | 0.100 | 1,181 | 2,251 | 0.2111 / 0.2882 | 3.47 | 0.697 / 0.744 | `b-qwen3-32b/probes.json + b-qwen3-32b/probe-decode.board.csv` |
| `mistral-small3.2:24b-instruct-2506-q4_K_M` | 49.2 | 278.0 | 5.649 | 5,650 | 0.177 | 0.177 | 5.425 | 0.184 | 2,139 | 2,744 | 0.1118 / 0.1745 | 5.73 | 0.627 / 0.665 | `b-mistral-small3.2-24b/probes.json + b-mistral-small3.2-24b/probe-decode.board.csv` |
| `qwen3:30b-a3b-instruct-2507-q4_K_M` | 176.9 | 237.5 | 1.345 | 1,345 | 0.743 | 0.743 | 2.104 | 0.475 | 4,044 | 2,243 | 0.0542 / 0.1060 | 9.43 | 0.413 / 0.424 | `b-qwen3-30b-a3b-2507/probes.json + b-qwen3-30b-a3b-2507/probe-decode.board.csv` |
| `gpt-oss:20b` | 140.7 | 234.9 | 1.670 | 1,670 | 0.599 | 0.599 | 2.623 | 0.381 | 5,602 | 2,260 | 0.0333 / 0.1075 | 9.31 | 0.805 / 0.877 | `b-gpt-oss-20b-low/probes.json + b-gpt-oss-20b-low/probe-decode.board.csv` |
| `granite4:32b-a9b-h` | 88.4 | 253.4 | 2.865 | 2,865 | 0.349 | 0.349 | 3.382 | 0.296 | 1,859 | 2,255 | 0.1262 / 0.1935 | 5.17 | 0.937 / 0.943 | `b-granite4-32b-a9b/probes.json + b-granite4-32b-a9b/probe-decode.board.csv` |
| `ministral-3:14b-instruct-2512-q4_K_M` | 77.9 | 267.5 | 3.434 | 3,434 | 0.291 | 0.291 | 3.946 | 0.253 | 3,578 | 2,791 | 0.0622 / 0.1241 | 8.05 | 0.551 / 0.595 | `b-ministral3-14b/probes.json + b-ministral3-14b/probe-decode.board.csv` |
| `olmo-3.1:32b-instruct-q4_K_M` | 35.8 | 284.9 | 7.955 | 7,955 | 0.126 | 0.126 | 9.442 | 0.106 | 1,095 | 2,291 | 0.2167 / 0.3329 | 3.00 | 0.765 / 0.887 | `b-olmo3.1-32b/probes.json + b-olmo3.1-32b/probe-decode.board.csv` |
| `glm-4.7-flash:q4_K_M` | 126.7 | 246.0 | 1.944 | 1,944 | 0.515 | 0.514 | 2.552 | 0.392 | 2,844 | 2,232 | 0.0801 / 0.1475 | 6.78 | 0.515 / 0.582 | `b-glm4.7-flash/probes.json + b-glm4.7-flash/probe-decode.board.csv` |
| `qwen3.5:27b-q4_K_M` | 37.8 | 272.7 | 7.219 | 7,219 | 0.139 | 0.139 | 9.888 | 0.101 | 1,194 | 2,262 | 0.1953 / 0.3098 | 3.23 | 1.388 / 1.417 | `b-qwen3.5-27b/probes.json + b-qwen3.5-27b/probe-decode.board.csv` |
| `gemma4:31b-it-q4_K_M` | 35.3 | 272.8 | 7.734 | 7,734 | 0.129 | 0.129 | 8.504 | 0.118 | 850 | 2,219 | 0.2339 / 0.4228 | 2.37 | 1.437 / 1.519 | `b-gemma4-31b/probes.json + b-gemma4-31b/probe-decode.board.csv` |
| `gemma4:26b` | 146.6 | 226.0 | 1.544 | 1,544 | 0.648 | 0.648 | 1.755 | 0.570 | 2,699 | 2,219 | 0.0663 / 0.1627 | 6.15 | 1.004 / 1.056 | `k0-gemma4-26b-c/probes.json + k0-gemma4-26b-c/probe-decode.board.csv` |
| `gemma4:26b-a4b-it-q4_K_M` | 129.3 | 234.4 | 1.813 | 1,813 | 0.552 | 0.552 | 2.146 | 0.466 | 3,242 | 2,219 | 0.0575 / 0.1573 | 6.36 | 0.973 / 1.018 | `k0a-gemma4-anchor-cd/probes.json + k0a-gemma4-anchor-cd/probe-decode.board.csv` |
| `granite4.1:30b-q4_K_M` | 39.3 | 285.3 | 7.245 | 7,244 | 0.138 | 0.138 | 7.377 | 0.136 | 1,265 | 2,233 | 0.2010 / 0.2657 | 3.76 | 0.546 / 0.590 | `b-granite4.1-30b/probes.json + b-granite4.1-30b/probe-decode.board.csv` |
| `qwen3.6:27b-q4_K_M` | 40.2 | 270.9 | 6.738 | 6,738 | 0.148 | 0.148 | 6.659 | 0.150 | 1,195 | 2,262 | 0.1929 / 0.3062 | 3.27 | 1.376 / 1.419 | `b-qwen3.6-27b/probes.json + b-qwen3.6-27b/probe-decode.board.csv` |
| `gemma4:12b-it-q4_K_M` | 74.5 | 252.1 | 3.381 | 3,382 | 0.296 | 0.296 | 4.587 | 0.218 | 2,226 | 2,218 | 0.0842 / 0.1986 | 5.04 | 1.024 / 1.049 | `b-gemma4-12b/probes.json + b-gemma4-12b/probe-decode.board.csv` |
| `qwen3.8:27b-q4_K_M` | 39.4 | 272.3 | 6.913 | 6,913 | 0.145 | 0.145 | 8.155 | 0.123 | 1,186 | 2,262 | 0.2020 / 0.2972 | 3.36 | 1.304 / 1.325 | `b-qwen3.8-27b/probes.json + b-qwen3.8-27b/probe-decode.board.csv` |

### Arms that were refused or failed (a refusal is a measurement; a failure is the bench not asking)

No row of these arms enters an accuracy, an energy figure or a count of models read: they have no task rows.

| arm | model | arm rc | refused | failed | stamp (UTC) | receipt |
|---|---|---:|---|---|---|---|
| b-nemotron3-nano-30b | `nemotron-3-nano:30b-a3b-q4_K_M` | 5 | did not fit whole on 24 GB — largecard-only (not a score). The board grew 24,032,313,344 B for 24,271,934,866 store bytes; /api/ps held 23,710,201,282 of 24,869,161,655 B in VRAM, so 1,158,960,373 B sat on the CPU; card memory.total 24,576 MiB |  | 2026-09-23T23:28:40Z | `b-nemotron3-nano-30b/REFUSED.json` |

### Residency and fit per arm (the delta rule; /api/ps as corroboration)

| arm | model | board growth MiB | store bytes | ps size_vram / size | whole | BOARD idle, no model W | WALL box idle, no model W | ollama |
|---|---|---:|---:|---:|---|---:|---:|---|
| b-deepseek-r1-32b | `deepseek-r1:32b-qwen-distill-q4_K_M` | 20,041 | 19,851,337,809 | 20,686,684,814 / 20,686,684,814 | True | 24.3 | 77.0 | 0.32.13 |
| b-falcon-7b | `falcon:7b-instruct-q4_0` | 4,241 | 4,210,994,570 | 4,126,272,388 / 4,126,272,388 | True | 24.3 | 92.1 | 0.32.13 |
| b-gemma1.1-7b | `gemma:7b-instruct-v1.1-q4_K_M` | 7,347 | 5,329,768,363 | 7,378,934,168 / 7,378,934,168 | True | 24.4 | 89.3 | 0.32.13 |
| b-gemma2-27b | `gemma2:27b-instruct-q4_K_M` | 17,901 | 16,645,391,492 | 17,673,234,677 / 17,673,234,677 | True | 24.3 | 76.6 | 0.32.13 |
| b-gemma3-27b | `gemma3:27b-it-q4_K_M` | 18,125 | 17,396,936,941 | 17,484,019,138 / 17,484,019,138 | True | 24.2 | 86.3 | 0.32.13 |
| b-gemma4-12b | `gemma4:12b-it-q4_K_M` | 8,353 | 7,556,508,396 | 8,063,046,122 / 8,063,046,122 | True | 24.2 | 69.3 | 0.32.13 |
| b-gemma4-31b | `gemma4:31b-it-q4_K_M` | 21,103 | 19,868,981,791 | 20,253,381,754 / 20,253,381,754 | True | 24.3 | 74.1 | 0.32.13 |
| b-glm4.7-flash | `glm-4.7-flash:q4_K_M` | 18,555 | 19,019,270,897 | 19,138,556,722 / 19,138,556,722 | True | 24.2 | 90.2 | 0.32.13 |
| b-gpt-oss-20b-low | `gpt-oss:20b` | 12,773 | 13,793,441,244 | 12,854,199,582 / 12,854,199,582 | True | 24.3 | 78.2 | 0.32.13 |
| b-granite4-32b-a9b | `granite4:32b-a9b-h` | 19,253 | 19,476,640,854 | 19,847,142,440 / 19,847,142,440 | True | 24.4 | 88.3 | 0.32.13 |
| b-granite4.1-30b | `granite4.1:30b-q4_K_M` | 18,233 | 17,490,259,354 | 18,791,100,579 / 18,791,100,579 | True | 24.2 | 72.7 | 0.32.13 |
| b-llama2-13b | `llama2:13b-chat-q4_K_M` | 11,037 | 7,866,942,896 | 11,249,091,869 / 11,249,091,869 | True | 24.2 | 80.0 | 0.32.13 |
| b-llama2-7b | `llama2:7b-chat-q4_K_M` | 6,269 | 4,081,990,863 | 6,251,033,394 / 6,251,033,394 | True | 24.3 | 88.5 | 0.32.13 |
| b-llama3-8b | `llama3:8b-instruct-q4_K_M` | 5,333 | 4,920,747,238 | 5,271,715,839 / 5,271,715,839 | True | 24.3 | 106.5 | 0.32.13 |
| b-llama3.1-8b | `llama3.1:8b-instruct-q4_K_M` | 5,333 | 4,920,753,328 | 5,271,715,839 / 5,271,715,839 | True | 24.3 | 77.3 | 0.32.13 |
| b-ministral3-14b | `ministral-3:14b-instruct-2512-q4_K_M` | 9,443 | 9,082,537,546 | 8,721,971,281 / 8,721,971,281 | True | 24.4 | 73.4 | 0.32.13 |
| b-mistral-7b-v0.3 | `mistral:7b-instruct-v0.3-q4_K_M` | 5,027 | 4,372,824,384 | 4,950,883,040 / 4,950,883,040 | True | 24.3 | 88.7 | 0.32.13 |
| b-mistral-nemo-12b | `mistral-nemo:12b-instruct-2407-q4_K_M` | 7,827 | 7,477,217,229 | 7,884,568,001 / 7,884,568,001 | True | 24.2 | 76.0 | 0.32.13 |
| b-mistral-small-22b | `mistral-small:22b-instruct-2409-q4_K_M` | 13,957 | 13,341,251,759 | 14,309,360,925 / 14,309,360,925 | True | 24.3 | 68.3 | 0.32.13 |
| b-mistral-small-24b-2501 | `mistral-small:24b-instruct-2501-q4_K_M` | 14,479 | 14,333,921,662 | 14,854,515,588 / 14,854,515,588 | True | 24.2 | 80.2 | 0.32.13 |
| b-mistral-small3.2-24b | `mistral-small3.2:24b-instruct-2506-q4_K_M` | 15,359 | 15,177,384,862 | 14,917,482,577 / 14,917,482,577 | True | 24.3 | 74.6 | 0.32.13 |
| b-nemotron3-nano-30b | `nemotron-3-nano:30b-a3b-q4_K_M` | 22,919 | 24,271,934,866 | 23,710,201,282 / 24,869,161,655 | False | 24.3 | 71.9 | 0.32.13 |
| b-olmo3.1-32b | `olmo-3.1:32b-instruct-q4_K_M` | 19,841 | 19,482,447,875 | 20,204,696,370 / 20,204,696,370 | True | 24.4 | 75.0 | 0.32.13 |
| b-phi2-2.7b | `phi:2.7b-chat-v2-q4_K_M` | 2,625 | 1,789,240,980 | 2,428,848,045 / 2,428,848,045 | True | 24.3 | 84.8 | 0.32.13 |
| b-phi3-14b | `phi3:14b-medium-128k-instruct-q4_0` | 8,677 | 7,897,127,458 | 8,778,174,954 / 8,778,174,954 | True | 24.2 | 78.5 | 0.32.13 |
| b-phi4-14b | `phi4:14b-q4_K_M` | 9,589 | 9,053,116,391 | 9,734,685,981 / 9,734,685,981 | True | 24.3 | 94.3 | 0.32.13 |
| b-qwen1.5-32b | `qwen:32b-chat-v1.5-q4_K_M` | 19,895 | 19,698,969,702 | 20,532,753,857 / 20,532,753,857 | True | 24.2 | 78.9 | 0.32.13 |
| b-qwen2.5-32b | `qwen2.5:32b-instruct-q4_K_M` | 20,045 | 19,851,349,669 | 20,686,684,814 / 20,686,684,814 | True | 24.3 | 82.6 | 0.32.13 |
| b-qwen3-30b-a3b | `qwen3:30b-a3b-q4_K_M` | 18,381 | 18,622,563,179 | 18,936,496,127 / 18,936,496,127 | True | 24.3 | 73.0 | 0.32.13 |
| b-qwen3-30b-a3b-2507 | `qwen3:30b-a3b-instruct-2507-q4_K_M` | 18,307 | 18,556,699,186 | 18,870,635,068 / 18,870,635,068 | True | 24.3 | 82.3 | 0.32.13 |
| b-qwen3-32b | `qwen3:32b-q4_K_M` | 20,381 | 20,201,253,829 | 21,030,670,171 / 21,030,670,171 | True | 24.3 | 87.0 | 0.32.13 |
| b-qwen3.5-27b | `qwen3.5:27b-q4_K_M` | 16,915 | 17,420,432,728 | 16,325,279,742 / 16,325,279,742 | True | 24.3 | 70.8 | 0.32.13 |
| b-qwen3.6-27b | `qwen3.6:27b-q4_K_M` | 16,915 | 17,420,432,475 | 16,325,279,742 / 16,325,279,742 | True | 24.3 | 75.7 | 0.32.13 |
| b-qwen3.8-27b | `qwen3.8:27b-q4_K_M` | 17,079 | 17,741,872,132 | 16,638,866,881 / 16,638,866,881 | True | 24.4 | 88.2 | 0.32.13 |
| b-vicuna-33b | `vicuna:33b-q4_0` | 20,971 | 18,355,968,078 | 21,660,067,429 / 21,660,067,429 | True | 24.2 | 97.3 | 0.32.13 |
| b-yi-34b | `yi:34b-chat-q4_0` | 19,757 | 19,466,547,560 | 20,387,683,368 / 20,387,683,368 | True | 24.3 | 86.2 | 0.32.13 |
| dryrun-k0 | `gemma4:26b` | 21,449 | 18,604,148,513 | 1,105,555,619 / 1,105,555,619 | True | 12.6 | 67.2 | 0.32.13 |
| k0-gemma4-26b-c | `gemma4:26b` | 21,449 | 18,604,148,513 | 1,105,555,619 / 1,105,555,619 | True | 12.8 | 59.7 | 0.32.13 |
| k0-gemma4-26b-d | `gemma4:26b` | 21,449 | 18,604,148,513 | 1,105,555,619 / 1,105,555,619 | True | 24.6 | 114.9 | 0.32.13 |
| k0-gemma4-26b-think | `gemma4:26b` | 21,449 | 18,604,148,513 | 1,105,555,619 / 1,105,555,619 | True | 24.6 | 113.8 | 0.32.13 |
| k0a-gemma4-anchor-cd | `gemma4:26b-a4b-it-q4_K_M` | 18,077 | 17,987,581,215 | 17,347,127,541 / 17,347,127,541 | True | 24.4 | 97.9 | 0.32.13 |
| k1-mistral-v0.1-c | `mistral:v0.1` | 4,777 | 4,109,865,192 | 4,688,770,497 / 4,688,770,497 | True | 24.5 | 125.2 | 0.32.13 |
| k1-mistral-v0.1-d | `mistral:v0.1` | 4,777 | 4,109,865,192 | 4,688,770,497 / 4,688,770,497 | True | 24.5 | 98.9 | 0.32.13 |
| t-deepseek-r1-32b-on | `deepseek-r1:32b-qwen-distill-q4_K_M` | 21,069 | 19,851,337,809 | 21,764,620,942 / 21,764,620,942 | True | 24.3 | 68.1 | 0.32.13 |
| t-gemma4-12b-on | `gemma4:12b-it-q4_K_M` | 8,713 | 7,556,508,396 | 8,372,921,302 / 8,372,921,302 | True | 24.4 | 78.5 | 0.32.13 |
| t-gemma4-31b-on | `gemma4:31b-it-q4_K_M` | 21,427 | 19,868,981,791 | 20,257,576,058 / 20,257,576,058 | True | 24.3 | 77.5 | 0.32.13 |
| t-glm4.7-flash-on | `glm-4.7-flash:q4_K_M` | 18,771 | 19,019,270,897 | 19,364,524,850 / 19,364,524,850 | True | 24.5 | 87.0 | 0.32.13 |
| t-gpt-oss-20b-medium | `gpt-oss:20b` | 12,773 | 13,793,441,244 | 12,854,199,582 / 12,854,199,582 | True | 24.6 | 81.2 | 0.32.13 |
| t-qwen3-30b-a3b-on | `qwen3:30b-a3b-q4_K_M` | 18,757 | 18,622,563,179 | 19,343,343,615 / 19,343,343,615 | True | 24.6 | 77.0 | 0.32.13 |
| t-qwen3-32b-on | `qwen3:32b-q4_K_M` | 21,397 | 20,201,253,829 | 22,108,606,299 / 22,108,606,299 | True | 24.2 | 70.1 | 0.32.13 |
| t-qwen3.5-27b-on | `qwen3.5:27b-q4_K_M` | 17,303 | 17,420,432,728 | 16,734,234,868 / 16,734,234,868 | True | 24.4 | 70.9 | 0.32.13 |
| t-qwen3.8-27b-on | `qwen3.8:27b-q4_K_M` | 17,467 | 17,741,872,132 | 17,047,822,007 / 17,047,822,007 | True | 24.2 | 81.5 | 0.32.13 |



## 6. Instrument, gates and traps

- **The instrument's bytes, sha-checked at every arm start** (`rows/*/preflight.json` →
  `instrument`): `run.py` `withheld-2026-10-14` (the smoke's and the Jev arms'), task_c
  `61a0b1d6…2d7f39` and doorman `4daf059c…22f7` (S0, `../SERIES-SETS.md`). `run_addenda.py`
  (task_d, DOORMAN_SYSTEM v1.2 sha-asserted) and the master harness's `benchlib.py` are copied
  unchanged from the estate and hashed on every arm.
- **The pin, by growth on the named board.** Every arm reads a pre-load baseline, loads with
  the arm's exact options, and requires ≥ 200 MiB of growth on `<withheld: GPU UUID of board A>` with no other
  board present. The instance also runs `OLLAMA_VULKAN=0` + `GGML_VK_VISIBLE_DEVICES=-1` beside
  the CUDA pin, and its discovery log shows one CUDA device at `pci_id=withheld-2026-10-14`
  (`receipts/substrate-2026-09-23.txt`).
- **The gates** (master harness, estate `51be5b25`):
  - the box-class preflight;
  - the disk floor at 30 GiB, on the results and the store before every arm (`disk_floor_gate`)
    and on `/bin/df` before every pull, net of the manifest's bytes (`pull.py`);
  - the UPS readable before every arm;
  - the 80 % UPS load stop, judged by `ups_stop_sample`. It aborts in-process (SIGINT → the
    `finally` unloads) and is announced to the stage's receipt;
  - `stopguard.py` around every arm, which reads the UPS and every board's temperature every
    2 s.

  No stop tripped. The highest UPS load and core temperature per arm are in each arm's
  `stop-receipt.guard.json`.
- **The pinned anchor digest.** `pull.py` pins `gemma4:26b-a4b-it-q4_K_M` to `5571076f3d70`
  (PLAN-v2 item 1). It checks the registry manifest's sha256 before pulling and `/api/tags`
  after, and it deletes on a mismatch. The anchor matched on both (`receipts/pull-gemma4_26b-a4b-it-q4_K_M.json`).
  Every row carries `x_digest` from 17:4xZ on. Earlier rows carry the digest in `arm.json` →
  `tag`.
- **Traps met and closed before a measured row:**
  - run.py's `--limit` filters without stopping. The driver uses `islice` (the smoke's finding).
  - run.py crashes on a task window under two sampler seconds. The dry run used 5 items, and
    every real window is 5 s or longer.
  - The strict parser's article rule first read "A\n\nExplanation" as the English article. It
    was fixed to one space plus a lowercase word on the same line, and self-tested
    (`instrument/analyse.py`, parser `v2-benchbox-1`). Only the generated figures changed.
- **Queue changes, each written atomically (a tmp file, then `mv`):**
  - The anchor swap at lines 55–57 (the lead, ~17:4xZ).
  - The dense reasoning arms: first deferred in place, then appended at lines 140–155 after
    every reasoning-off arm by an operator's ~18:00Z ruling.
  - The 23:55Z halt and the arms' `RuntimeMaxSec` ceiling were lifted by the same ruling. The
    runner stopped at the `STOP` file at 18:09:57Z and restarted as `coi3090-queue2` at
    18:09:59Z; no step was lost or repeated (`queue.log`, `queue.done`).

## 7. Every command, as run (benchbox, `operator@benchbox`, TZ=UTC)

```
# the instance (transient, user scope, no sudo), 16:01:54Z
systemd-run --user --unit=coi3090-ollama -E OLLAMA_HOST=127.0.0.1:withheld-2026-10-14 \
  -E OLLAMA_MODELS=/workshop/bench-coi-2026-09-23/store \
  -E CUDA_VISIBLE_DEVICES=<withheld: GPU UUID of board A> -E OLLAMA_VULKAN=0 \
  -E GGML_VK_VISIBLE_DEVICES=-1 -E HIP_VISIBLE_DEVICES= -E ROCR_VISIBLE_DEVICES= \
  -E OLLAMA_KEEP_ALIVE=-1 -E OLLAMA_FLASH_ATTENTION=1 -E OLLAMA_CONTEXT_LENGTH=131072 -E TZ=UTC \
  /usr/local/bin/ollama serve
# the dry run, 16:06:03Z (5 items of c and d, the probes, the whole path; rc 0)
systemd-run --user --collect --unit=coi3090-arm-dryrun-k0 -E TZ=UTC /bin/bash arm.sh dryrun-k0 \
  --model gemma4:26b --num-ctx 131072 --think off --tasks c,d --probes --limit 5
# the queue, 16:09:53Z (then coi3090-queue2 from 18:09:59Z); each arm is its own unit:
systemd-run --user --collect --unit=coi3090-queue -E TZ=UTC /bin/bash queue.sh
#   arm  -> systemd-run --user --wait --collect --unit=coi3090-arm-<id> -E TZ=UTC /bin/bash arm.sh <id> ...
#   arm.sh -> BENCH_STOP_RECEIPT=rows/<id>/stop-receipt.jsonl BENCH_BOX_CLASS=gpu-coi3090-24g \
#             python3 -B harness/stopguard.py -- python3 -B coi3090.py --arm-id <id> --out rows/<id> ...
#   pull / rm -> python3 -B pull.py pull|rm <tag>
```
The queue as run is `instrument/queue.txt` (line numbers are the `queue.done` keys), and
`queue.log` has every step's start, end and rc.

## 8. Files (this directory, committed to the estate repo; `receipts/pull-*.log`, the pulls' progress output, is git-ignored)

```
RTX-3090-READING-0.md   this report (built by instrument/build_report.py)
instrument/                 coi3090.py (the driver), arm.sh, queue.sh, queue.txt, pull.py, params.sh,
                            requeue.py, analyse.py, tables.py, compare_k0.py, compare_k1.py,
                            wall_cadence.py, build_report.py, NOTES.md (this prose)
rows/<arm>/                 <arm>.<task>.jsonl (every row, raw reply, thinking text), .report.json (run.py's),
                            .watts.csv (BOARD, run.py) .idle.csv .ext.csv (wide 1 Hz trace), wall.csv (WALL),
                            idle-preload.board.csv, preflight.json, arm.json, probes.json,
                            probe-{decode,prefill}.{board,ext}.csv, run.log, stop-receipt.guard.json
rows/figures.json, rows/FIGURES.md, rows/k0-vs-jev.json, rows/k1-vs-smoke.json, rows/wall-cadence.json
queue/                      queue.txt, queue.log, queue.done as the runner left them (QUEUE EMPTY 2026-09-24T03:21:09Z)
receipts/                   pull-<tag>.json and rm-<tag>.json per tag, the substrate receipt
rows/dryrun-k0/             the 16:06Z dry run's files
```