# CB-7 fold: reading 0.1 on the laptop's RTX 5090, seven cells read whole under the quiet hold

**Written:** 2026-10-02, 14:53Z → 15:31Z (UTC, `date -u` on the laptop), by the CB-7 fold lane of the cost-of-intelligence series, under an operator's D-20261001-039 ("any additional \"cost of intelligence\" benches we still need to run, let's do those too"), D-20261002-003 and D-20261002-007. The block's record is ledger D-20261001-059 (started) and D-20261001-063 (done). The gaps plan books this fold as its item A-12 (`~/bench-archive/plans-2026-10-02/gaps/GAPS-PLAN-v2.md`). **Read-only:** no repo, worktree, unit, card or model was touched; nothing was re-run, moved, committed, pushed or posted. This lane wrote only into `~/bench-archive/coi-cb7-2026-10-01/fold/` and its session scratch folder (§10). **What it folds:** CB-7, the laptop block of reading 0.1, committed on estate branch `bench/coi-cb7-laptop-0.1` at `483eeafb` (172 files; not on master). **Read with:** the series' own `v2score.score_pass()` (the frozen `parse_v2.py`, sha `689254c5…`, its 48-case selftest run at import), `series/stats.py` and `series/kitlock.py` from estate origin/master `99f676ec`, and the laptop kit's own `analyse_laptop.py` (sha `24da34be…`, from `483eeafb`), each on copies (§10).

**What this is, for a reader who scrolled straight here.** This page reads the seven cost-of-intelligence cells that the laptop's NVIDIA GeForce RTX 5090 Laptop GPU ran on **2026-10-01 from 19:09:33Z to 20:06:45Z** (UTC), the laptop's reading 0.1 block (CB-7). For each model it counts how many six-way questions (`c`, 108 items) and doorman questions (`d`, 48 items) it answered correctly under parser v2's strict rule, how many joules the GPU board drew per correct answer, how fast it decoded, and whether every gate held. It then sets each cell beside the same model's earlier readings, item by item. Against the reference RTX 3090's reading 0 (2026-09-23/24, 300 W) the laptop's strict count lies within one answer on 12 of 13 tasks and four answers below on the gemma4 anchor's six-way, and no paired interval excludes zero. On this same card, the two anchors' 312 replies are byte-identical to the card's own reading 0 of 2026-09-23.

---

## §0 · One screen

**Seven of seven cells read whole, rc 0, in 57 minutes. Every gate that can refuse an arm held; the R-20 receipt, which decides afterwards whether an arm is a row of record, needs the lead's ruling on one arm.** The queue started at 19:09:33Z and ended `QUEUE EMPTY` at 20:06:45Z, against the 105–165 minutes planned. Dynamic Boost held the board's enforced limit at 150.00 W in all 2,846 one-second samples inside the 13 task windows, inside A-L04's (95, 175] W envelope. The memory gate read `ok` on every arm, and AC read online at all 40 posture reads. The one finding is the **R-20 receipt on `t-deepseek-r1-32b-on`**: the bench serve unit's `memory.events max` grew by 337 during that arm (0 on the other six), and by the receipt's letter that arm is superseded (§3.2).

| Role | Cell | Verdict | v2 strict, `c` (of 108) | v2 strict, `d` (of 48) | Board J per correct, gross, `c` | Decode tok/s |
|---|---|---|---:|---:|---:|---:|
| 2024H1 ★ (A-11); a DNS-lost cell | `b-gemma2-27b` | DONE, rc 0 | 74 (68.5 %) | 39 (81.3 %) | 70.68 | 37.16 |
| DNS-lost cell | `b-llama3.1-8b` | DONE, rc 0 | 43 (39.8 %) | 24 (50.0 %) | 31.18 | 126.33 |
| DNS-lost cell | `b-mistral-7b-v0.3` | DONE, rc 0 | 19 (17.6 %) | 0 (0.0 %) | 83.29 | 138.12 |
| DNS-lost cell | `b-mistral-nemo-12b` | DONE, rc 0 | 46 (42.6 %) | 34 (70.8 %) | 63.86 | 84.68 |
| 2025H1 ★ (A-11), think on | `t-deepseek-r1-32b-on` | DONE, rc 0; the R-20 receipt reads +337 (§3.2) | 1 (0.9 %) | not on its line (`c` only) | 369,348.48 (divides by 1 answer) | 31.61 (the task's own; no probes registered) |
| anchor (A-14) | `k0a-gemma4-anchor-cd` | DONE, rc 0 | 84 (77.8 %) | 36 (75.0 %) | 29.32 | 148.6 |
| anchor (A-14) | `k1-mistral-v0.1-cd` | DONE, rc 0 | 10 (9.3 %) | 0 (0.0 %) | 148.13 | 153.63 |

- **The gates held on all 13 tasks.** Every pass is complete with its registered fingerprint (`c` `1feac3433cb4217b`, `d` `617abd299260cee0`), with 0 failed, 0 template-suspect and 0 excluded items, so each prints a rate. Every model loaded whole on the card at its pinned digest, from a line byte-identical to the line that registered it. The temperature guard never tripped (its maximum was 67 °C), and the system daemon served 0 requests inside any task window. The laptop instrument writes no `voided_by_gate` field (that is the desktop's larger card's A-17 (5) cap gate), and no `VOID`, `REFUSED`, `FAILED` or `STOP-HOT` file exists.
- **The R-20 receipt on deepseek.** PREREG-S0 §2.2 registers it as mechanical: "an arm is superseded when the serve unit's `memory.events max` grew". It read 88,920 → 89,257 across that 43 m 43 s arm. In reading 0 on this card, all 37 arms at 32G read +0, the six reasoning-on arms (4.3 to 82.4 minutes) among them. The rows show no stall: prompt eval at most 233.7 ms, load at most 118.9 ms, decode 27.71 to 31.99 tok/s per item. So the replies stand. Only the row-of-record status and the energy wait on the lead's ruling, and this fold moves nothing.
- **Against the reference RTX 3090** (reading 0, same digests, item by item). The strict counts differ by at most one answer on 12 of 13 tasks and by four on `k0a` `c` (84 against 88; exact McNemar p 0.219; paired −3.7 points [−7.7, 0.0]). No paired interval excludes zero. S-7's 97 % bridge carries 8 of the 13 tasks, and 3 of those 8 carry only a shared format failure (§6.1).
- **Against the desktop's larger card.** gemma2 sits beside CB-5b's row of record: `c` 74 against 72 (carried, 106 of 108), `d` 39 against 41 (not carried, 46 of 48). The anchors sit beside CB-5 at 300 W, which is byte-identical to the desktop's larger card's reading 0 at 420 W (§6.2). deepseek is compared with the desktop's larger card's 8 stopped items as counts only.
- **The anchors A-14 needs are re-read, and this card did not move.** `k0a` (gemma4 `5571076f`) and `k1` (mistral:v0.1 `b17615239298`) give replies byte-identical to this card's reading 0 on all 312 items (`c` 108 + `d` 48 each). The paired change is 0.0 points on every task, so there is no detectable change at n = 108 or 48. So A-14's completeness condition is met for the laptop on `c` and `d48` once the series PR carries these rows (§7).
- **deepseek-r1 32B, think on, budget 2,048:** 1 of 108 strict (`late_letter` 103, `budget_exhausted` 3, `clean` 2), lenient 86. 3 items hit the 2,048-token budget (kit 2, 20 and 50), each with an empty visible reply. **D-20261002-014's parenthesis, checked:** the row stands whole as **of record only on the 3090** (series row `withheld-2026-10-14`, 1 of 108). The laptop's 108 replies agree in kind: 1 of 108, and the same class and lenient choice as the desktop's larger card's 8 items. But the laptop row is not on master, not in `series.jsonl`, and, by the R-20 receipt's letter, superseded (§5).
- **Owed** (§9): the lead's ruling on the R-20 receipt; the data PR of `483eeafb` with one audit (R-4); the series PR (the gaps plan's A-4) before any 0.1 row enters `series.jsonl` (T-12); an operator's word on the desktop's larger card's think-on cell, with §5's facts.

---

## §1 · What ran, cell by cell

The sources are each cell's `arm.json`, its `run.log` (`rc=N end` line), `laptop-reading-0.1/queue.log` and `queue.done`, all at `483eeafb`. The registering files are `laptop-reading-0/queue.txt` (sha `a17a1146…`) and benchbox's `instrument/queue.txt` (sha `d844bf2a…`), the shas PREREG-S0 §9.2 registers.

| # | Cell | Its line, byte for byte | Digest = pin | Started (UTC) | Ended (UTC) | Took | Rc | Whole (growth MiB) | Warm load (s) | `memory.events max`, before → after |
|---:|---|---|---|---|---|---|---:|---|---:|---|
| 1 | `b-gemma2-27b` | `laptop-reading-0/queue.txt` line 49 (yes) | `cfd95e5696e9` (yes) | 19:09:33 | 19:12:38 | 3 m 05 s | 0 | True (17,896) | 9.4 | 88,920 → 88,920 (+0) |
| 2 | `b-llama3.1-8b` | `laptop-reading-0/queue.txt` line 52 (yes) | `46e0c10c039e` (yes) | 19:12:38 | 19:14:34 | 1 m 56 s | 0 | True (5,328) | 3.5 | 88,920 → 88,920 (+0) |
| 3 | `b-mistral-7b-v0.3` | `laptop-reading-0/queue.txt` line 46 (yes) | `6577803aa9a0` (yes) | 19:14:34 | 19:16:17 | 1 m 43 s | 0 | True (5,022) | 3.5 | 88,920 → 88,920 (+0) |
| 4 | `b-mistral-nemo-12b` | `laptop-reading-0/queue.txt` line 55 (yes) | `daf673741712` (yes) | 19:16:17 | 19:18:36 | 2 m 19 s | 0 | True (7,822) | 4.9 | 88,920 → 88,920 (+0) |
| 5 | `t-deepseek-r1-32b-on` | benchbox `instrument/queue.txt` line 148 (yes) | `edba8017331d` (yes) | 19:18:36 | 20:02:19 | 43 m 43 s | 0 | True (21,062) | 9.5 | 88,920 → 89,257 (**+337**) |
| 6 | `k0a-gemma4-anchor-cd` | `laptop-reading-0/queue.txt` line 125 (yes) | `5571076f3d70` (yes) | 20:02:19 | 20:05:01 | 2 m 42 s | 0 | True (18,074) | 12.1 | 89,257 → 89,257 (+0) |
| 7 | `k1-mistral-v0.1-cd` | `laptop-reading-0/queue.txt` line 12 (yes) | `b17615239298` (yes) | 20:05:01 | 20:06:45 | 1 m 44 s | 0 | True (4,772) | 3.0 | 89,257 → 89,257 (+0) |

- **The queue.** `QUEUE START lane=laptop-reading-0.1 … queue.sh=cd65804870048ec6 arm.sh=fe1466a42e5274ca pull.py=19dfaca4f8e47ae4` at 19:09:33Z. Each `STEP end` read rc 0 and was followed by `POSTARM api_ps={"models":[]} board_mib=15`. `QUEUE EMPTY` came at 20:06:45Z, and `queue.done` holds the seven keys. No `HALT` or `UNDONE` line exists. Against the plan's estimates (`CB7-DRY-RUN-PLAN.md` B-1): the four DNS-lost cells took 9 m 03 s (planned about 35), deepseek 43 m 43 s (60–120), and the two anchors 4 m 26 s (about 8).
- **The instrument.** It equals A-25 (3) and (4) by sha256: `queue.sh` `cd658048…`, `arm.sh` `fe1466a4…`, `pull.py` `19dfaca4…`, `instance.sh` `095fa3be…`, `coilaptop.py` `0b5dcb41…` and `pins.json` `f644ae95…`, plus the three queues (`prep/` `b196f11c…`, `dry/` `acbee17a…`, cells `008e53f0…`). Every `arm.json` `instrument` block equals the pins: `coilaptop.py` `0b5dcb41…`, the Jev `run.py` `withheld-2026-10-14` and `run_addenda.py` `a34aca61…`, `task_c.json` `61a0b1d6…` and `doorman_planted.json` `4daf059c…`. Each carries A-L01 to A-L04. Each task's rows came from one runner process, so no model reloaded mid-arm.
- **The bench instance** (`receipts/instance-start.txt`, written 18:35:45Z): `/api/version` 0.32.13, binary sha `39c8dea5…` (PREREG-S0 §10's pin), `MemoryMax` 34,359,738,368 bytes (32 GiB, A-L03), KV cache f16, flash attention on. It ran on driver 595.91.07, CUDA 13.2 and kernel 7.0.0-34-generic, the same three as reading 0's receipt on this card.
- **The prep and the dry run, read as such.** `prep/` pulled the seven tags from 18:35:57Z to 18:49:19Z, each rc 0, and each digest equal to its `expected_digest_prefix`. Five of the seven pins cite "benchbox pull receipt (the bytes the 3090 read)"; deepseek and mistral:v0.1 cite the recon manifest. So every cell's bytes equal the 3090's for the same tag. `dry/` ran `dry-cb7-anchor-4096` (`--limit 3`) from 18:50:15Z to 18:51:47Z, rc 0. Dry rows are never cells (A-25 (4)), and none is scored here.
- **What the operator set by hand** (`laptop-reading-0.1/OPERATOR-CONDITIONS.md`). an operator's turbo fan was on from about 17:4xZ, in an operator's words in D-20261001-050; the board reports `fan.speed [N/A]`, so nothing else records it. The quiet hold's timers (`scribe-embed.timer`, `s2s-memory-prewarm.timer`) were stopped at 18:50:15Z and restarted at 20:07:07Z.

## §2 · Accuracy under parser v2 (strict is the headline; lenient and v1 beside, never in a ratio)

The source is `series/v2score.score_pass()` on each row file. Strict means class `clean` or `letter+name` with the strict choice equal to the label (PREREG-S0 §3.2). Wilson 95 % is `series/stats.wilson()` (PREREG-S0 §4.1). B-chance is PREREG-S0 §5's test on `c`: strict above 24.0 % and the Wilson lower bound above 16.7 %. Balanced accuracy on `d` is PREREG-S0 §4.3's, with its per-class counts. The last column is `analyse_laptop.py`'s lane parser (`v2-benchbox-1`), the count its J-per-correct figures divide by; it equals v2's on all 13 tasks.

| Cell | Task | v2 strict / n | v2 strict (%) | Wilson 95 % (%) | Rate status | Lenient | v1 | Format failures (classes) | B-chance (`c`) · balanced accuracy (`d`) | Lane parser strict |
|---|---|---:|---:|---|---|---:|---:|---|---|---:|
| `b-gemma2-27b` | `c` | 74 / 108 | 68.5 | 59.3–76.5 | `rate` | 74 | 74 | 0 | Clears | 74 |
| `b-gemma2-27b` | `d` | 39 / 48 | 81.3 | 68.1–89.8 | `rate` | 39 | 39 | 1 (`json_verdict` 1) | 87.5 % (admit 12 of 12, refuse 27 of 36) | 39 |
| `b-llama3.1-8b` | `c` | 43 / 108 | 39.8 | 31.1–49.2 | `rate` | 43 | 43 | 0 | Clears | 43 |
| `b-llama3.1-8b` | `d` | 24 / 48 | 50.0 | 36.4–63.6 | `rate` | 32 | 33 | 21 (`json_verdict` 17, `unparsed` 4) | 33.3 % (admit 0 of 12, refuse 24 of 36) | 24 |
| `b-mistral-7b-v0.3` | `c` | 19 / 108 | 17.6 | 11.6–25.8 | `rate` | 19 | 19 | 0 | Does not clear | 19 |
| `b-mistral-7b-v0.3` | `d` | 0 / 48 | 0.0 | 0.0–7.4 | `rate` | 13 | 12 | 48 (`json_verdict` 46, `unparsed` 2) | 0.0 % (admit 0 of 12, refuse 0 of 36) | 0 |
| `b-mistral-nemo-12b` | `c` | 46 / 108 | 42.6 | 33.7–52.0 | `rate` | 46 | 46 | 0 | Clears | 46 |
| `b-mistral-nemo-12b` | `d` | 34 / 48 | 70.8 | 56.8–81.8 | `rate` | 37 | 34 | 3 (`json_verdict` 3) | 50.0 % (admit 1 of 12, refuse 33 of 36) | 34 |
| `t-deepseek-r1-32b-on` | `c` | 1 / 108 | 0.9 | 0.2–5.1 | `rate` | 86 | 14 | 106 (`budget_exhausted` 3, `late_letter` 103) | Does not clear | 1 |
| `k0a-gemma4-anchor-cd` | `c` | 84 / 108 | 77.8 | 69.1–84.6 | `rate` | 84 | 84 | 0 | Clears | 84 |
| `k0a-gemma4-anchor-cd` | `d` | 36 / 48 | 75.0 | 61.2–85.1 | `rate` | 36 | 36 | 0 | 83.3 % (admit 12 of 12, refuse 24 of 36) | 36 |
| `k1-mistral-v0.1-cd` | `c` | 10 / 108 | 9.3 | 5.1–16.2 | `rate` | 10 | 10 | 0 | Does not clear | 10 |
| `k1-mistral-v0.1-cd` | `d` | 0 / 48 | 0.0 | 0.0–7.4 | `rate` | 21 | 12 | 48 (`json_verdict` 48) | 0.0 % (admit 0 of 12, refuse 0 of 36) | 0 |

Every pass has 108 or 48 rows, 0 missing, 0 stray and 0 duplicate items, and its registered fingerprint. 1,044 rows in all.

**Four doorman rows are format verdicts, not capability readings.** `b-mistral-7b-v0.3` answers 46 of 48 doorman items in the doorman's native JSON (`json_verdict`), and `k1-mistral-v0.1-cd` answers all 48 that way. Both score 0 strict on this letter recast (PREREG-S0 §1.1: only `doorman36-json` scores the native contract, and it was not run), and both read the same on the 3090 (§6). `b-llama3.1-8b` `d` has 17 `json_verdict` and 4 `unparsed` replies, and admits none of the 12 controls. deepseek's `c` is a format verdict too: 103 of its 108 replies state their letter last (§5).

## §3 · The gates and the posture

### 3.1 What the registered gates read

| Gate (its rule) | What it read | Source |
|---|---|---|
| The memory gate (A-L01: swap used < 80 %, MemAvailable ≥ 10 GiB, PSI memory full avg60 < 10 %, `/dev/shm` < 4 GiB) | `ok` on 7 of 7 arms: swap 33.56–47.88 %, MemAvailable 48.7–50.0 GiB, PSI full avg60 0.00–0.18 %, `/dev/shm` 4 KiB | `arm.json` `memory_gate` |
| The posture gate (A-L04, R-20: nvidia-powerd active, `clocks.max.sm` 3090 (lock off), AC online, one board, the enforced limit in (95, 175] W, reported per task as min / median / max) | All 40 posture reads (before and after each arm, start and end of each task): 150.00 W, 3090 MHz, powerd `active`, AC `1`, one board. Every one-second sample in all 13 task windows read 150.00 W (2,846 samples; 2,989 over the whole trace files, idle reads included). All 1,044 rows carry `x_enforced_w` 150.0, `x_max_sm_mhz` 3090, `x_posture` `dynamic-boost`. Battery `Full`, 99 % | `arm.json` `posture_before`, `tasks_run.<t>.posture_start` / `posture_end`, `postflight.posture_after`; `<cell>.<t>.ext.csv` column 9; the rows |
| Disk floor (60 GiB) | 388.9–389.0 GiB free on every arm | `arm.json` `disk_floor` |
| Fit and pin (whole when the board grew ≥ 90 % of the store's bytes and `/api/ps` holds no CPU share) | Whole on 7 of 7, `/api/ps` size = size_vram on each, the digest equal to its pin on each | `arm.json` `pin` |
| The temperature guard (≥ 87 °C on 6 consecutive 5 s reads → STOP-HOT, rc 6) | Never tripped; the guard's maximum 67 °C (deepseek); the 1 Hz trace's maximum 68 °C | `arm.json` `temp_guard`; `ext.csv` |
| The co-tenant (the system daemon on :withheld-2026-10-14, read only) | `/api/ps` empty before and after every arm; 0 requests inside any task window | `arm.json` `seat_before`, `postflight.seat_after`, `tasks_run.<t>.seat_requests_task` |
| The unload (`/api/ps` empty, the board back near baseline) | On every arm: empty, and back to its 15 MiB baseline; `POSTARM` agrees | `postflight.after_unload`; `queue.log` |
| Rule 1.1-a (the set fingerprint) and §6 (failed > 2 %, template-suspect > 2 %) | 13 of 13 fingerprints registered; 0 failed, 0 template-suspect | `series/v2score.score_pass()` |
| `voided_by_gate` | Not a field of this instrument: the laptop has no cap to hold (R-20), and its refusals are the gates above (rc 4 before a task, rc 6 for heat). No `*.VOID.json`, `REFUSED.json`, `FAILED.json` or `STOP-HOT.json` in any cell folder | the cell folders |
| The R-20 receipt (PREREG-S0 §2.2; the R-20 clarification of 2026-09-23T19:35:15Z) | +0 on six arms; **+337 on `t-deepseek-r1-32b-on`** (§3.2) | `arm.json` `serve_cgroup_memory_before` → `postflight.serve_cgroup_memory_after` |

The posture, per task (A-L04's form, read from each task's own 1 Hz trace inside its window). The load average and the CPU package's watts sit beside it, as `CB7-QUIET-HOLD.md` §5 asks:

| Cell | Task | Task window (UTC) | Enforced limit min / median / max (W; samples) | AC, start · end | `clocks.max.sm`, start · end | Card temp max (°C) | Load average, median (max) | Package-0 mean W | System daemon requests |
|---|---|---|---|---|---|---:|---|---:|---:|
| `b-gemma2-27b` | `c` | 19:10:16 → 19:11:05 | 150.00 / 150.00 / 150.00 (50) | 1 · 1 | 3090 · 3090 | 50 | 0.96 (0.97) | 23.0 | 0 |
| `b-gemma2-27b` | `d` | 19:11:18 → 19:11:47 | 150.00 / 150.00 / 150.00 (29) | 1 · 1 | 3090 · 3090 | 53 | 0.94 (1.04) | 22.9 | 0 |
| `b-llama3.1-8b` | `c` | 19:13:16 → 19:13:36 | 150.00 / 150.00 / 150.00 (21) | 1 · 1 | 3090 · 3090 | 47 | 0.68 (0.87) | 23.9 | 0 |
| `b-llama3.1-8b` | `d` | 19:13:49 → 19:14:01 | 150.00 / 150.00 / 150.00 (13) | 1 · 1 | 3090 · 3090 | 48 | 0.78 (0.80) | 22.9 | 0 |
| `b-mistral-7b-v0.3` | `c` | 19:15:11 → 19:15:24 | 150.00 / 150.00 / 150.00 (14) | 1 · 1 | 3090 · 3090 | 49 | 0.70 (0.72) | 23.3 | 0 |
| `b-mistral-7b-v0.3` | `d` | 19:15:38 → 19:15:46 | 150.00 / 150.00 / 150.00 (9) | 1 · 1 | 3090 · 3090 | 51 | 1.08 (1.08) | 23.2 | 0 |
| `b-mistral-nemo-12b` | `c` | 19:16:56 → 19:17:30 | 150.00 / 150.00 / 150.00 (32) | 1 · 1 | 3090 · 3090 | 49 | 1.22 (1.88) | 26.1 | 0 |
| `b-mistral-nemo-12b` | `d` | 19:17:44 → 19:17:59 | 150.00 / 150.00 / 150.00 (16) | 1 · 1 | 3090 · 3090 | 51 | 1.45 (1.53) | 22.9 | 0 |
| `t-deepseek-r1-32b-on` | `c` | 19:19:19 → 20:02:10 | 150.00 / 150.00 / 150.00 (2,569) | 1 · 1 | 3090 · 3090 | 68 | 1.29 (2.91) | 25.1 | 0 |
| `k0a-gemma4-anchor-cd` | `c` | 20:03:05 → 20:03:48 | 150.00 / 150.00 / 150.00 (44) | 1 · 1 | 3090 · 3090 | 53 | 1.47 (1.59) | 25.6 | 0 |
| `k0a-gemma4-anchor-cd` | `d` | 20:04:03 → 20:04:26 | 150.00 / 150.00 / 150.00 (24) | 1 · 1 | 3090 · 3090 | 52 | 1.18 (1.21) | 23.4 | 0 |
| `k1-mistral-v0.1-cd` | `c` | 20:05:38 → 20:05:52 | 150.00 / 150.00 / 150.00 (15) | 1 · 1 | 3090 · 3090 | 52 | 0.67 (0.87) | 24.1 | 0 |
| `k1-mistral-v0.1-cd` | `d` | 20:06:05 → 20:06:14 | 150.00 / 150.00 / 150.00 (10) | 1 · 1 | 3090 · 3090 | 53 | 0.76 (0.78) | 22.4 | 0 |

**The load, flagged descriptively** (`CB7-QUIET-HOLD.md` §5). The load average is `rapl.csv`'s `loadavg1`, polled once a second inside each window. The block's median of the 13 task medians is 0.96, with a range of 0.67 to 1.47. This fold's own flag line is a task median at least twice the block's median and at least 1.0 above it. No task reaches it. The highest single poll, 2.91, came inside deepseek's 43-minute window, where the median was 1.29. The enforced limit did not move off 150.00 W in any sample, so no CPU load moved the GPU's power limit in this block (R-20's mechanism).

### 3.2 The R-20 receipt on `t-deepseek-r1-32b-on`: what the file says, and what it does not decide

**The rule, as registered.** PREREG-S0 §2.2 (laptop row): "Rows of record under R-20 and its clarification (the receipt is mechanical: an arm is superseded when the serve unit's `memory.events max` grew between `serve_cgroup_memory_before` and `postflight.serve_cgroup_memory_after` in its `arm.json`)". A superseded arm's "replies stand for determinism checks, their energy is not a series figure". The 16G `k0a` row that rule superseded "stands as a choice figure only" (PREREG-S0 §8.4 item 5). The clarification (`RULINGS-2026-09-23.md`, 19:35:15Z) was asked about the 16G arms. It states the receipt in general words, though, and adds "Every 32G arm reads +0". A-L03's own reason expected 32G not to bind: "32G is a ceiling that should not bind, and every arm.json carries the cgroup's memory.events before/after to prove it". No amendment scopes the receipt to 16G.

**What the files read.**

| Arm | `memory.events max`, before → after | `memory.pressure` "full" total, growth | The serve unit's MemoryCurrent, before → after (GB, bytes ÷ 10⁹) |
|---|---|---|---|
| `dry-cb7-anchor-4096` (not a cell) | 88,900 → 88,920 (+20) | +0.01 s | 34.35 → 22.87 |
| `b-gemma2-27b` | 88,920 → 88,920 (+0) | +0.94 s | 16.69 → 17.44 |
| `b-llama3.1-8b` | 88,920 → 88,920 (+0) | +0.32 s | 17.44 → 14.53 |
| `b-mistral-7b-v0.3` | 88,920 → 88,920 (+0) | +0.37 s | 14.53 → 16.85 |
| `b-mistral-nemo-12b` | 88,920 → 88,920 (+0) | +0.24 s | 16.85 → 18.40 |
| **`t-deepseek-r1-32b-on`** | **88,920 → 89,257 (+337)** | **+1.13 s** | **18.40 → 25.15** |
| `k0a-gemma4-anchor-cd` | 89,257 → 89,257 (+0) | +0.94 s | 25.15 → 18.93 |
| `k1-mistral-v0.1-cd` | 89,257 → 89,257 (+0) | +0.06 s | 18.93 → 14.87 |

The "after" reads come after each arm's unload (`postflight`), so they show the page cache left behind, not the arm's peak.

- **When the counter grew.** It read 88,900 before the dry run. The serve unit's MemoryCurrent stood at 34.35 GB then, 33.29 GB of it page cache (`stat.file`), left by the seven pulls the prep ran through the same unit. Those 88,900 belong to no arm. Inside the block, only deepseek's arm moved it.
- **What deepseek's rows show, item by item.** Prompt eval ran a median of 174.7 ms and at most 233.7 ms. Load ran a median of 100.9 ms and at most 118.9 ms. Decode ran 27.71 to 31.99 tok/s per item (median 31.71). One runner process served all 108 items. Nothing resembles the 1.4 to 7.2 s prefill stalls that made the 16G arms superseded (A-L03's reason). The memory-pressure stall over the 43 m 43 s arm (1.13 s) is close to gemma2's over its 3 m 05 s arm (0.94 s), and gemma2 read +0.
- **In reading 0 on this card**, all 37 arms at 32G read +0 (each `arm.json` on master that carries the receipt). Among them are the six reasoning-on arms at `num_ctx` 8,192, which ran 4.3 to 82.4 minutes each (`t-gemma4-12b-on` the longest) on stores up to 19.0 GB (`t-glm4.7-flash-on`). deepseek's store is 19.85 GB, and its arm is the first 32G arm on this card whose receipt grew.
- **What the fold does with it: nothing.** By the receipt's letter, `t-deepseek-r1-32b-on` is superseded. Its 108 replies stand as a choice figure and a determinism receipt, and its energy is not a series figure. The row of record would be a re-run's. The letter also cuts against the plan's expectation that 32G never binds. So this is a contract call (s2s-agent-workflow §9b (c)), and the lead rules on it. The two readings:
  - **(i) Apply the receipt as written** (fail closed). The laptop's 2025H1 ★ cell holds replies and counts, no row of record, and no energy figure. A row of record needs a re-run under a cap that does not bind: an amendment to A-L03 before its rows, an operator's fan, about 45 minutes of the laptop GPU (this arm took 43 m 43 s), and the tag still in the store.
  - **(ii) An amendment that scopes the receipt**, for example to arms whose rows show a stall. That would be a rule written after its rows were read, so it must say so (PREREG-S0 §8.3's practice).
- **The fold's rec** (the lead's call): take (i) for the series. The cell's strict reading is a format verdict either way (§5). Its J per correct divides by one correct answer either way (§4). And the 3090's row of the same cell is of record. Re-run it only if the series wants the laptop's think-on J per decode token. Before any later long think-on arm on this card, the instrument needs a cap or receipt amendment (§9).

## §4 · Energy and speed, as `analyse_laptop.py` computes them

`analyse_laptop.py` (sha `24da34be…`, from `483eeafb`) ran at 2026-10-02T15:00:11Z on a scratch copy of `laptop-reading-0.1/rows/`. It ran with `PYTHONDONTWRITEBYTECODE=1` and `TZ=UTC`, and its result is "13 task rows, 6 probe rows". It prints only that line and writes `figures.json`, so the figures below are that file's, at its own rounding. The copy is kept here as `analyse-figures-cb7.scrubbed.json`, with the card's UUID replaced by a name.

**Which energy the laptop's rows carry, and what it can and cannot claim.**
- **The board** is the energy of record for the card: `nvidia-smi power.draw` by UUID at 1 Hz, trapezoid-integrated over each task window (`run.py` `Boards`; `<cell>.<t>.watts.csv`).
- **RAPL package-0** sits beside it: the CPU package's counter differences over the same windows (`/run/rapl/energy`, R-10d; `rapl.csv`). Its `psys` domain is recorded and reads 0.3 to 0.4 W under every load, so it is IMPLAUSIBLE (the R-10d addendum) and never enters a figure here.
- The box has no wall, platform or UPS meter (R-6). So the laptop's energy is board plus package, each labelled.

What that can claim:
- The GPU board's own joules per task, per item and per correct answer: gross, net of the card's resident idle, and compute-phase.
- Per-token board energy from the probes.
- The CPU package's joules over the same windows, as their own figure.

What it cannot claim:
- What the laptop drew from the wall or as a whole. Display, memory, storage, fans, regulators and the charger are in neither instrument.
- A figure comparable with the 3090's wall or the desktop's larger card's box delta. R-6 compares board with board only.
- That the package joules are the bench's alone: package-0 counts every CPU process on the box in the window.
- A fixed power posture. Under Dynamic Boost the GPU limit may float with the CPU's draw; it read 150.00 W throughout here (§3.1).

**Read these with three notes.**
- **"Net of the resident idle"** subtracts the card's draw with the model loaded, read for about 9 s just before each task (`tasks_run.<t>.idle`: 22.5 to 23.9 W across the 13 tasks). That is the kit's form. PREREG §7.1's 60 s co-tenant-quiet form is not what this instrument records.
- **Two cells divide by few or no answers.** `b-mistral-7b-v0.3` `d` and `k1` `d` have no correct answer, so they print no J per correct. deepseek's J per correct divides by one answer, so it is uninformative: it does not clear B-chance (PREREG-S0 §7.2's caution).
- **The task windows are short.** They run 9 to 50 one-second samples, except deepseek's 2,569, so the short windows' means are coarse.

| Cell | Task | Board J (s; mean W) | Resident idle (W) | Board J per correct, gross | Net of the resident idle | Compute-phase | Board J per item | Package-0 J (mean W) | Package-0 J per correct | Median s per item (p95) | Median load (ms) |
|---|---|---|---:|---:|---:|---:|---:|---|---:|---|---:|
| `b-gemma2-27b` | `c` | 5,230.2 (49.01 s; 106.7 W) | 22.5 | 70.68 | 55.77 | 41.86 | 48.43 | 1,091.3 (23.0 W) | 14.75 | 0.4555 (0.5187) | 122.87 |
| `b-gemma2-27b` | `d` | 3,091.7 (28.01 s; 110.4 W) | 22.9 | 79.27 | 62.82 | 49.54 | 64.41 | 614.3 (22.9 W) | 15.75 | 0.5946 (0.625) | 122.05 |
| `b-llama3.1-8b` | `c` | 1,340.6 (20.01 s; 67.0 W) | 22.9 | 31.18 | 20.53 | 9.07 | 12.41 | 443.3 (23.9 W) | 10.31 | 0.1919 (0.2111) | 118.04 |
| `b-llama3.1-8b` | `d` | 977.9 (12.01 s; 81.5 W) | 23.0 | 40.75 | 29.26 | 17.7 | 20.37 | 259.4 (22.9 W) | 10.81 | 0.2551 (0.2951) | 123.11 |
| `b-mistral-7b-v0.3` | `c` | 1,582.5 (13.01 s; 121.7 W) | 22.9 | 83.29 | 67.64 | 61.03 | 14.65 | 264.5 (23.3 W) | 13.92 | 0.1262 (0.1421) | 15.61 |
| `b-mistral-7b-v0.3` | `d` | 927.8 (8.0 s; 115.9 W) | 23.0 | — | — | — | 19.33 | 143.6 (23.2 W) | — | 0.187 (0.1931) | 16.87 |
| `b-mistral-nemo-12b` | `c` | 2,937.4 (33.96 s; 86.5 W) | 22.9 | 63.86 | 46.93 | 27.9 | 27.2 | 862.1 (26.1 W) | 18.74 | 0.28 (0.3065) | 119.91 |
| `b-mistral-nemo-12b` | `d` | 1,449.2 (15.01 s; 96.5 W) | 23.1 | 42.62 | 32.42 | 22.2 | 30.19 | 331.3 (22.9 W) | 9.74 | 0.314 (0.3672) | 122.34 |
| `t-deepseek-r1-32b-on` | `c` | 369,348.5 (2,570.89 s; 143.7 W) | 23.0 | 369,348.48 | 310,299.27 | 365,858.47 | 3,419.89 | 64,535.8 (25.1 W) | 64,535.84 | 20.9476 (42.9396) | 100.87 |
| `k0a-gemma4-anchor-cd` | `c` | 2,462.7 (43.01 s; 57.3 W) | 23.9 | 29.32 | 17.09 | 8.13 | 22.8 | 1,084.4 (25.6 W) | 12.91 | 0.407 (0.4401) | 247.61 |
| `k0a-gemma4-anchor-cd` | `d` | 1,371.9 (23.01 s; 59.6 W) | 23.6 | 38.11 | 23.04 | 10.94 | 28.58 | 507.7 (23.4 W) | 14.1 | 0.4866 (0.5293) | 262.04 |
| `k1-mistral-v0.1-cd` | `c` | 1,481.3 (14.01 s; 105.8 W) | 23.3 | 148.13 | 115.49 | 88.16 | 13.72 | 323.9 (24.1 W) | 32.39 | 0.1327 (0.1458) | 29.85 |
| `k1-mistral-v0.1-cd` | `d` | 1,019.6 (9.01 s; 113.2 W) | 23.4 | — | — | — | 21.24 | 185.1 (22.4 W) | — | 0.1888 (0.1949) | 32.45 |

**The probes** (`probes.json` and `probe-*.board.csv`). Decode is the median of 3 runs of 256 tokens; prefill is the median of the nonce-prefixed cold runs (about 2,200 to 2,450 prompt tokens each); TTFT is the client clock on localhost over the first 12 six-way prompts, streamed. J per token is the probe window's mean board W times the phase seconds, divided by its tokens.

| Cell | Decode (tok/s, median of 3 × 256) | J per 1k decode tokens (board, phase) | Board mean W, decode | Prefill (tok/s, median) | Board J per prompt token (phase) | TTFT median (p95), s | Package-0 mean W, decode |
|---|---:|---:|---:|---:|---:|---|---:|
| `b-gemma2-27b` | 37.16 | 3,565.3 | 132.5 | 1,526.6 | 0.06946 | 0.4234 (0.4839) | 23.0 |
| `b-llama3.1-8b` | 126.33 | 922.5 | 116.5 | 6,092.1 | 0.01679 | 0.1766 (0.1885) | 22.7 |
| `b-mistral-7b-v0.3` | 138.12 | 877.3 | 121.2 | 6,026.1 | 0.02014 | 0.0988 (0.104) | 22.9 |
| `b-mistral-nemo-12b` | 84.68 | 1,472.0 | 124.7 | 4,045.6 | 0.02846 | 0.2151 (0.2474) | 23.0 |
| `t-deepseek-r1-32b-on` | No probes registered (its line says probes `no`). The task's own decode: 31.61 tok/s (Σ eval tokens ÷ Σ eval seconds) | — | — | — | — | — | — |
| `k0a-gemma4-anchor-cd` | 148.6 | 716.0 | 106.3 | 4,379.5 | 0.01872 | 0.4047 (0.431) | 23.4 |
| `k1-mistral-v0.1-cd` | 153.63 | 709.5 | 108.9 | 6,677.9 | 0.01722 | 0.1 (0.1087) | 21.2 |

deepseek's task also reads 4.545 board J per decode token (analyse's task form: mean W times Σ eval seconds, divided by eval tokens). By the R-20 receipt's letter (§3.2), every deepseek energy figure here is printed as the tool computed it and is not a series figure until the lead rules.

**No ratio, no dollar.** No figure here is divided by another card's (the ratio rule; R-6; PREREG-S0 §7.1's 2× rule). THE RATIO and every `$` come from `series.jsonl` only, and these rows are not in it.

## §5 · deepseek-r1 32B, think on: the budget, the classes, and D-20261002-014's parenthesis checked

### 5.1 The cell

The line is benchbox's `instrument/queue.txt` line 148, byte for byte: think on, `num_ctx` 8,192, task `c` only, budget 2,048, no probes (A-11's 2025H1 ★; A-17 (2); A-25 (4)). Every row carries `x_num_predict` 2,048 and `x_think` true. The arm ran 19:18:36Z → 20:02:19Z, and its task window ran 19:19:19Z → 20:02:10Z (2,570.89 s).

- **The budget.** 3 of 108 items reached 2,048 eval tokens with `done_reason` `length` and an empty visible reply, so they are `budget_exhausted`, counted wrong (PREREG-S0 §6). They are kit 2, 20 and 50, the ids `c-darwin-01-dessert-01`, `c-einstein-01-dessert-02` and `c-hypatia-01-starter-01`. 5 items spent over 1,500 tokens. Eval tokens ran 324 to 2,048 (median 652.5; 79,917 in all), with 363,959 characters of thinking.
- **The classes.** `late_letter` 103 (the reply explains, then states its letter last), `budget_exhausted` 3, `clean` 2. Strict is 1 of 108, from the two clean items: kit 69 answered `ibn_sina`, correct, and kit 54 answered `sagan` where the label is `ibn_sina`. Lenient is 86 and v1 14, both beside and never in a ratio.
- **Time.** 10.46 to 65.82 s per item (median 20.95 s).

### 5.2 What stands where (the 2025H1 ★ think-on cell, on three cards)

| Card | Reading | Rows | v2 strict | Classes | Lenient | v1 | Status, from the files |
|---|---|---|---:|---|---:|---:|---|
| The reference RTX 3090 (benchbox), 300 W | 0 | 108 of 108 | 1 | `late_letter` 104, `budget_exhausted` 3, `clean` 1 | 83 | 11 | **Of record:** `series.jsonl` row `withheld-2026-10-14` on master (`rate`, `complete`, ★), its rows `rtx-3090-reading-0/rows/t-deepseek-r1-32b-on/` (sha `1c4e6d61…`) |
| This laptop card (the laptop), Dynamic Boost 150.00 W | 0.1 | 108 of 108 | 1 | `late_letter` 103, `budget_exhausted` 3, `clean` 2 | 86 | 14 | **Whole as replies, not of record:** rc 0 and a v2 `rate`, but only on branch `bench/coi-cb7-laptop-0.1` at `483eeafb` (not on master; not in `series.jsonl`, T-12), and by the R-20 receipt's letter superseded (§3.2) |
| the desktop's larger card (the desktop), 300 W window | 0.1 (CB-5b) | 8 of 108 (kit 0–7) | 0 of 8 | `late_letter` 7, `budget_exhausted` 1 | 6 | 2 | STOPPED by I-1 at 08:19:34Z on 2026-10-02; counts only, energy void (CB5B-FOLD §6) |

**Item by item.** The laptop against the 3090:
- 105 of 108 items share the same strict choice, exactly the S-7 line. All 105 have no strict choice on either card, so the bridge passes on a shared format failure, not on a shared answer.
- The strict discordance is 1 against 1: the laptop alone is right on kit 69, the 3090 alone on kit 75. Exact McNemar p = 1.000, and the paired difference is 0.0 points [−2.7, +2.8].
- The format class matches on 101 of 108 items, the lenient choice on 97, and the reply text byte for byte on 2.
- Both cards exhausted the budget on kit 2 (the laptop also on 20 and 50, the 3090 also on 21 and 37).

Against the desktop's larger card's 8 items, the class matches on 8 of 8, the lenient choice on 8 of 8, and the text on 1. Kit 2 exhausted the budget on all three cards.

**D-20261002-014's parenthesis, checked.** The lead's rec (a) reads "record the desktop's larger card's 2025H1 think-on cell as NOT RUN with this fold as its receipt (the same row stands whole on the reference 3090 and the laptop)". Against the files:
- **It holds as a statement about the replies.** Both cards read the same registered line on the same digest (`edba8017331d`) over all 108 items. They give the same strict count, 1 of 108, and the same verdict: a model that states its letter last, with 3 items at the budget.
- **It does not hold as a statement about rows of record today.** Only the 3090's row is of record (reading 0, in `series.jsonl` at master). The laptop's enters the series only through the data PR and the series PR. By the R-20 receipt's letter it is superseded, so its replies stand as a choice figure.
- **They are also not one row.** They are two cells of one registered line: the 3090 at 300 W in reading 0, and the laptop in reading 0.1 under Dynamic Boost.
- **A wording that matches the files:** "the 2025H1 think-on row stands of record on the reference 3090 (reading 0); the laptop's whole 0.1 replies agree with it, and with the desktop's larger card's 8, on class and count".

**What the laptop's row says for the desktop's larger card's open cell.** A second card now reads the whole think-on row as a format verdict: 103 late letters and 1 strict answer, beside the 3090's 104 and 1. The desktop's larger card's 8 items match both cards' classes and lenient choices. So a whole the desktop's larger card re-read would most likely add the 96 GB card's think-on speed and J per token, not a capability reading. This is a conditional, not a prediction of record. It supports rec (a)'s reasoning. It does not supply rec (a)'s "stands whole on the laptop" as a row of record until the lead rules on §3.2 and both PRs land.

## §6 · The comparisons the series needs (item by item, same items, same parser)

Every comparison joins on `prompt_sha256` (rule 1.1-b) over two complete passes of the same S0 set, each scored by the same `score_pass()`. Every pass has its registered fingerprint, so nothing crosses item sets. the desktop's larger card's stopped deepseek arm is the one exception, joined on its 8 items as counts.

The columns:
- **Same strict choice** is S-7's agreement (PREREG-S0 §2.5): no strict choice on both sides counts as the same. **Carried** means the same digest and agreement of at least 105 of 108 or 47 of 48. Below that line, the two cells stay separate as an instrument property.
- **Exact McNemar p** is on the discordant strict pairs.
- **The paired difference** is PREREG-S0 §4.2's registered form: laptop minus the other, 10,000 resamples, seed 20260923, clusters (16 `dinner_id`s) on `c` and items on `d`.
- **Median load ms** is shown because D-10's 2× rule applies to any later cross-card J ratio.

Every other side's strict count equals its `series.jsonl` row where it has one. That checks this fold's scoring path against the series' own build.

### 6.1 Beside the reference RTX 3090 (benchbox, reading 0, 300 W, `rtx-3090-reading-0/rows/` on master)

| Cell | Task | 3090 strict | Laptop 0.1 strict | Same strict choice (line) | Carried | Both right · both wrong | Laptop only · 3090 only | Exact McNemar p | Paired difference, points [95 %] | Byte-identical replies | Median load ms, laptop · 3090 |
|---|---|---:|---:|---|---|---|---|---:|---|---:|---|
| `b-gemma2-27b` | `c` | 73 | 74 | 105 of 108 (105) | yes | 72 · 33 | 2 · 1 | 1.000 | +0.9 [−1.9, +4.0], clusters | 101 of 108 | 123 · 432 |
| `b-gemma2-27b` | `d` | 40 | 39 | 47 of 48 (47); 1 with no strict choice on either side | yes | 39 · 8 | 0 · 1 | 1.000 | −2.1 [−6.2, 0.0], items | 47 of 48 | 122 · 417 |
| `b-llama3.1-8b` | `c` | 43 | 43 | 104 of 108 (105) | no | 41 · 63 | 2 · 2 | 1.000 | 0.0 [−3.6, +3.8], clusters | 94 of 108 | 118 · 413 |
| `b-llama3.1-8b` | `d` | 23 | 24 | 43 of 48 (47); 19 with no strict choice on either side | no | 21 · 22 | 3 · 2 | 1.000 | +2.1 [−6.2, +10.4], items | 42 of 48 | 123 · 409 |
| `b-mistral-7b-v0.3` | `c` | 20 | 19 | 102 of 108 (105) | no | 18 · 87 | 1 · 2 | 1.000 | −0.9 [−4.0, +1.9], clusters | 93 of 108 | 16 · 66 |
| `b-mistral-7b-v0.3` | `d` | 0 | 0 | 48 of 48 (47); 48 with no strict choice on either side | yes (a shared format failure) | 0 · 48 | 0 · 0 | 1.000 | 0.0 [0.0, 0.0], items | 48 of 48 | 17 · 66 |
| `b-mistral-nemo-12b` | `c` | 47 | 46 | 107 of 108 (105) | yes | 46 · 61 | 0 · 1 | 1.000 | −0.9 [−2.9, 0.0], clusters | 106 of 108 | 120 · 414 |
| `b-mistral-nemo-12b` | `d` | 33 | 34 | 44 of 48 (47); 2 with no strict choice on either side | no | 32 · 13 | 2 · 1 | 1.000 | +2.1 [−4.2, +8.3], items | 38 of 48 | 122 · 411 |
| `t-deepseek-r1-32b-on` | `c` | 1 | 1 | 105 of 108 (105); 105 with no strict choice on either side | yes (a shared format failure) | 0 · 106 | 1 · 1 | 1.000 | 0.0 [−2.7, +2.8], clusters | 2 of 108 | 101 · 363 |
| `k0a-gemma4-anchor-cd` | `c` | 88 | 84 | 102 of 108 (105) | no | 83 · 19 | 1 · 5 | 0.219 | −3.7 [−7.7, 0.0], clusters | 101 of 108 | 248 · 768 |
| `k0a-gemma4-anchor-cd` | `d` | 35 | 36 | 47 of 48 (47) | yes | 35 · 12 | 1 · 0 | 1.000 | +2.1 [0.0, +6.2], items | 41 of 48 | 262 · 763 |
| `k1-mistral-v0.1-cd` | `c` | 10 | 10 | 107 of 108 (105) | yes | 10 · 98 | 0 · 0 | 1.000 | 0.0 [0.0, 0.0], clusters | 101 of 108 | 30 · 125 |
| `k1-mistral-v0.1-cd` | `d` | 0 | 0 | 48 of 48 (47); 48 with no strict choice on either side | yes (a shared format failure) | 0 · 48 | 0 · 0 | 1.000 | 0.0 [0.0, 0.0], items | 47 of 48 | 32 · 125 |

The 3090 rows and their `series.jsonl` rows, each `rate`, `complete`, 300.00 W enforced with persistence Enabled:

| Cell | `c` rows (sha) and series row | `d` rows (sha) and series row |
|---|---|---|
| `b-gemma2-27b` | `69df3183…`, `withheld-2026-10-14` | `dcc2768f…`, `withheld-2026-10-14` |
| `b-llama3.1-8b` | `bbca4b51…`, `withheld-2026-10-14` | `97e15acd…`, `withheld-2026-10-14` |
| `b-mistral-7b-v0.3` | `68f6de69…`, `withheld-2026-10-14` | `0cdad975…`, `withheld-2026-10-14` |
| `b-mistral-nemo-12b` | `17f4f417…`, `withheld-2026-10-14` | `88ba149a…`, `withheld-2026-10-14` |
| `t-deepseek-r1-32b-on` | `1c4e6d61…`, `withheld-2026-10-14` | — |
| `k0a-gemma4-anchor-cd` | `withheld-2026-09-29` (B-seat's reference row, PREREG-S0 §5), `withheld-2026-10-14` | `193261c0…`, `withheld-2026-10-14` |
| `k1-mistral-v0.1-c` / `-d` | `6148bf6d…`, `withheld-2026-10-14` | `fd99ff2b…`, `withheld-2026-10-14` |

**What the table says.** S-7 carries 8 of the 13 tasks: gemma2 `c` and `d`, mistral-7b `d`, nemo `c`, deepseek `c`, `k0a` `d`, and `k1` `c` and `d`. Three of those (mistral-7b `d`, deepseek `c` and `k1` `d`) carry only a shared format failure. Five tasks are not carried: llama3.1 `c` (104) and `d` (43), mistral-7b `c` (102), nemo `d` (44) and `k0a` `c` (102). Those stand side by side as an instrument property, never merged. The `k0a` six-way reads 102 of 108 against the 3090, the same count as in reading 0 (`LAPTOP-READING-0.md` §0). The disagreement is a stable property of these two cards for the anchor, not something new in 0.1.

### 6.2 Beside the desktop's larger card (the desktop's RTX PRO 6000 Blackwell; CB-5 and CB-5b, 300 W windows)

The series does not keep CB-5's or CB-5b's rows on master yet. They ride the series PR into `rtx-pro-6000-reading-0.1/` (A-17 (8), A-24 (5)). So they are read from the bench archive, beside the reading-0 beside rows on master.

| Cell | Task | The desktop's larger card row | The desktop's larger card strict | Laptop 0.1 strict | Same strict choice (line) | Carried | Both right · both wrong | Laptop only · the desktop's larger card only | Exact McNemar p | Paired difference, points [95 %] | Byte-identical replies | Median load ms, laptop · the desktop's larger card |
|---|---|---|---:|---:|---|---|---|---|---:|---|---:|---|
| `b-gemma2-27b` | `c` | CB-5b `st-gemma2-27b-cd` (the row of record, A-24 (2)) | 72 | 74 | 106 of 108 (105) | yes | 72 · 34 | 2 · 0 | 0.500 | +1.9 [0.0, +4.7], clusters | 101 of 108 | 123 · 533 |
| `b-gemma2-27b` | `d` | CB-5b `st-gemma2-27b-cd` | 41 | 39 | 46 of 48 (47); 1 with no strict choice on either side | no | 39 · 7 | 0 · 2 | 0.500 | −4.2 [−10.4, 0.0], items | 46 of 48 | 122 · 528 |
| `t-deepseek-r1-32b-on` | `c` | CB-5b `st-deepseek-r1-32b-on-c` (STOPPED; 8 of 108) | 0 of 8 | 0 of the same 8 | 8 of 8; all 8 with no strict choice on either side | — (counts only) | 0 · 8 | 0 · 0 | 1.000 | counts only | 1 of 8 | 101 · 415 |
| `k0a-gemma4-anchor-cd` | `c` | CB-5 `b-gemma4-anchor-cd` | 86 | 84 | 104 of 108 (105) | no | 83 · 21 | 1 · 3 | 0.625 | −1.9 [−5.6, +1.8], clusters | 104 of 108 | 248 · 1,003 |
| `k0a-gemma4-anchor-cd` | `d` | CB-5 `b-gemma4-anchor-cd` | 35 | 36 | 47 of 48 (47) | yes | 35 · 12 | 1 · 0 | 1.000 | +2.1 [0.0, +6.2], items | 39 of 48 | 262 · 1,002 |
| `k1-mistral-v0.1-cd` | `c` | CB-5 `b-mistral-v0.1-cd` | 10 | 10 | 108 of 108 (105) | yes | 10 · 98 | 0 · 0 | 1.000 | 0.0 [0.0, 0.0], clusters | 104 of 108 | 30 · 178 |
| `k1-mistral-v0.1-cd` | `d` | CB-5 `b-mistral-v0.1-cd` | 0 | 0 | 48 of 48 (47); 48 with no strict choice on either side | yes (a shared format failure) | 0 · 48 | 0 · 0 | 1.000 | 0.0 [0.0, 0.0], items | 47 of 48 | 32 · 179 |

The desktop's larger card rows read:
- **CB-5b's** `rows/st-gemma2-27b-cd/` (`c` `1e8fdd04…`, `d` `3ad768e3…`) and `rows/st-deepseek-r1-32b-on-c/` (`2d9d94f7…`; its digest from `preflight.json`, since a stopped arm writes no `arm.json`), under `~/bench-archive/coi-cb5b-2026-10-02/`.
- **CB-5's** `rows/b-gemma4-anchor-cd/` (`394cfe0f…`, `bf173723…`) and `rows/b-mistral-v0.1-cd/` (`9778e52d…`, `7e58b945…`), under `~/bench-archive/coi-cb5-2026-09-27/`. They are byte-identical to the copies in CB-5b's mirror. Every complete CB-5 and CB-5b task above held its 300 W cap, with `voided_by_gate` null. The stopped deepseek arm's one task never ended, so it has no gate verdict (counts only).
- **Two equalities that need no rows of their own.** CB-5's gemma2 run (the same-posture repeat, `rows-superseded/st-gemma2-27b-cd.20261002T081131Z/`) gives exactly CB-5b's counts, because the two are byte-identical (CB5B-FOLD §3). The desktop's larger card's reading-0 beside rows on master give exactly CB-5's counts for both anchors, because they are byte-identical on all 312 items (CB5-FOLD §3). Those are `b1-gemma4-anchor-cd`, series `withheld-2026-10-14` and `withheld-2026-10-14`, and `b2-mistral-7b-v0.1-q4_0-cd` (Q4_0 = `b17615239298`), series `withheld-2026-10-14` and `withheld-2026-10-14`, at 420 W. This fold re-checked both by reply bytes: 156 of 156 for gemma2 and 312 of 312 for the anchors, and its joins against the laptop give the same counts either way.

**D-10, for the series PR's ratios.** The median per-call `load_ms` differs by more than 2× on every cross-card pair above. The laptop reads 16 to 262 ms, the 3090 66 to 768 ms, and the desktop's larger card 178 to 1,026 ms. So any later cross-card J or `$` ratio must print the compute-phase figure beside it (PREREG-S0 §7.1). The same-card pairs in §7 sit within 2×.

The three other DNS-lost models (llama3.1-8b, mistral-7b-v0.3, mistral-nemo-12b) have no the desktop's larger card row in any arms file, on master or in either archive.

## §7 · The two anchors A-14 needs, and what they anchor

**The rule** (PREREG-S0 A-14): "Completeness: a ratio prints on a set × card only where both local anchors were re-read on that card in 0.1 (c + d48) and ≥ 1 numerator row exists; otherwise 'no ratio on that set'. The anchors are in every card's 0.1 session (CB-3, CB-4, CB-5, CB-7, morning B)."

**The two anchors on this card in 0.1.** `k0a-gemma4-anchor-cd` is `gemma4:26b-a4b-it-q4_K_M` at `5571076f`, the seat's bytes (PREREG-S0 §2.5). `k1-mistral-v0.1-cd` is `mistral:v0.1` at `b17615239298`. Each read `c` and `d48`, rc 0, between 20:02:19Z and 20:06:45Z, with every gate held and the R-20 receipt at +0. So A-14's anchor condition is met for the laptop on both sets. The numerator rows exist: the four DNS-lost cells and gemma2's ★ row, plus deepseek's ★ row pending §3.2. Once the series PR carries these rows, `ratio.py` may print a 0.1 ratio for this card on `c` and `d48`. This fold prints none.

**What they anchor: this card did not move between its two readings.** Joined item by item against the card's own reading 0, on master:

| Anchor | Task | Reading 0 strict | 0.1 strict | Same strict choice | Byte-identical replies | Paired change, points [95 %] | Exact McNemar p |
|---|---|---:|---:|---:|---:|---|---:|
| `k0a` (the 32G re-run of 2026-09-23T22:10:06Z, the row of record; series `withheld-2026-10-14`, `withheld-2026-10-14`) | `c` | 84 | 84 | 108 of 108 | 108 of 108 | 0.0 [0.0, 0.0] | 1.000 |
| `k0a` | `d` | 36 | 36 | 48 of 48 | 48 of 48 | 0.0 [0.0, 0.0] | 1.000 |
| `k1` (2026-09-23T18:38:22Z under MemoryMax 16G, a row of record by its +0 receipt; series `withheld-2026-10-14`, `withheld-2026-10-14`) | `c` | 10 | 10 | 108 of 108 | 108 of 108 | 0.0 [0.0, 0.0] | 1.000 |
| `k1` | `d` | 0 | 0 | 48 of 48 | 48 of 48 | 0.0 [0.0, 0.0] | 1.000 |

- **All 312 of 312 replies are byte-identical.** `k0a`'s also equal reading 0's superseded 16G run (156 of 156). So the change between the readings is 0.0 points on every task: "no detectable change at n = 108 / 48" (PREREG-S0 §4.1).
- **What stayed fixed between the readings:** the digests, the runtime (0.32.13, binary `39c8dea5…`), the KV type (f16), the driver (595.91.07), the kernel (7.0.0-34-generic), the chat templates (`b507b9c2…` and `22e1b2e8…` in both readings), and the posture (150.00 W on every sample in both).
- **What moved:** `k1`'s MemoryMax (16G → 32G) and its `coilaptop.py` (`4d07c3cd…` → `0b5dcb41…`). The box also got quieter. The load average at `k1`'s start read 3.49 then and 0.65 now, and package-0 averaged 30.7 W then and 24.1 W now on `k1` `c`. None of those moved a reply byte.
- **Energy and speed, side by side** (the same `analyse_laptop.py` on both readings; reading 0's anchors ran in their own scratch copy at 15:06:06Z, kept as `analyse-figures-laptop-r0-anchors.scrubbed.json`). No ratio is formed. These are this card's two readings of the same bytes, and PREREG-S0 §4.1's between-reading rule is registered for accuracy, not joules.

| Anchor | Task | Board J per correct, gross (r0 · 0.1) | Net (r0 · 0.1) | Compute-phase (r0 · 0.1) | Board mean W (r0 · 0.1) | Package-0 mean W (r0 · 0.1) | Median load ms (r0 · 0.1) |
|---|---|---|---|---|---|---|---|
| `k0a` | `c` | 28.64 · 29.32 | 17.14 · 17.09 | 7.84 · 8.13 | 55.9 · 57.3 | 27.5 · 25.6 | 241.75 · 247.61 |
| `k0a` | `d` | 35.76 · 38.11 | 21.52 · 23.04 | 10.72 · 10.94 | 58.5 · 59.6 | 25.3 · 23.4 | 244.04 · 262.04 |
| `k1` | `c` | 146.16 · 148.13 | 114.43 · 115.49 | 87.61 · 88.16 | 104.4 · 105.8 | 30.7 · 24.1 | 32.24 · 29.85 |
| `k1` | `d` | no correct answer | — | — | 83.6 · 113.2 (12.01 s · 9.01 s windows) | 40.3 · 22.4 | 32.28 · 32.45 |

| Anchor | Decode tok/s (r0 · 0.1) | J per 1k decode tokens (r0 · 0.1) | Prefill tok/s (r0 · 0.1) | TTFT median, s (r0 · 0.1) |
|---|---|---|---|---|
| `k0a` | 149.74 · 148.6 | 701.1 · 716.0 | 4,433.6 · 4,379.5 | 0.3887 · 0.4047 |
| `k1` | 154.02 · 153.63 | 723.6 · 709.5 | 6,683.9 · 6,677.9 | 0.1013 · 0.1 |

**What they do not anchor.** They are not B-seat's reference: PREREG-S0 §5 names only the 3090's `k0a` row (`withheld-2026-09-29`). They carry no other tag's accuracy across cards, because S-7 is per tag and digest (§6). They say nothing about deepseek's receipt, since both anchors read +0. Their cross-card standing is §6's. `k0a` is not carried on `c` against the 3090 (102) or the desktop's larger card (104) and is carried on `d` against both (47). `k1` is carried on both sets against both cards.

## §8 · What this fold does not claim

- **No ratio.** THE RATIO comes from `series.jsonl` only (PREREG-S0 §4.4). These 0.1 rows are not in it (0 of its 359 rows name `laptop-reading-0.1`), and they stay out until the series PR (T-12). No J or speed figure here is divided by another card's or reading's.
- **No dollar figure.** None is printed. The laptop's cost rows come from the series' own build.
- **No ladder verdict and no bar test.** There is no B-anchor, B-seat or B-frontier test. The B-chance column is PREREG-S0 §5's test, read for context.
- **No ruling on the R-20 receipt.** §3.2 lays out what the files say and the two readings. The lead rules. No file was moved or relabelled.
- **No wall or whole-laptop energy.** Board and package are each labelled. psys is never used, and package-0 is not attributed to the bench alone.
- **For deepseek:** no capability reading beyond the format verdict, and no informative J per correct (1 correct answer). No claim that another budget, cap or shape would change its class mix.
- **"Carried" is not "the same answer".** Where S-7's agreement comes from items with no strict choice on either side (deepseek `c`, mistral-7b `d`, `k1` `d`), the bridge carries a shared format failure, not a capability.
- **Not a determinism result in general.** The anchors' byte-identity is two models on one card across one gap of eight days.
- **Not a claim that the quiet hold was total.** The load was measured (§3.1). The fan rests on an operator's word, since the board reports no fan speed.
- **No row-of-record claim for any 0.1 row.** All seven wait for the data PR's audit and the series PR.

## §9 · What is owed, in order

1. **The lead's ruling on `t-deepseek-r1-32b-on`'s R-20 receipt** (§3.2), before the data PR is declared ready: (i) apply it as written (the fold's rec), or (ii) an amendment that scopes it, stated as written after its rows. Whichever way, the series PR has to see it. `series/lanes.py` skips only directories named `superseded*` and reads no receipt, so a superseded arm left in place would enter the series as a row of record unless the series PR reads the receipt or the lane moves it.
2. **The data PR** for `483eeafb` (`bench/coi-cb7-laptop-0.1`, 172 files: the seven cells, the prep pulls' receipts and the dry run's rows) into master, **with one audit** (R-4; the gaps plan's A-12). The audit should attack, not only read (R-3). Suggested attacks: a planted wrong count in a scratch copy of this fold, a row whose `x_enforced_w` leaves 150.00 W, and a planted `memory.events` growth that the receipt check must name.
3. **The series PR** (the gaps plan's A-4, booked Sunday 10-04) before any 0.1 row enters `series.jsonl` (T-12). It carries CB-5's, CB-5b's and CB-7's rows into the 0.1 directories, with `cells.py` placing a listed cell by its rows' measured time (A-17 (8); D-20260927-016). It needs `lanes.py`'s reader for `laptop-reading-0.1/` and the receipt ruling of item 1. The four DNS-lost cells are reading-0 cells landed after the close, so they are 0.1 by measured time (PREREG-S0 §9.1, A-25 (4)). The ★ and anchor rows are on no reading-0 list.
4. **an operator's word on the desktop's larger card's 2025H1 think-on cell** (D-20261002-014 (a) or (b)), with §5.2's facts. Its parenthesis should name the 3090's row as the one of record.
5. **Kit notes for the next laptop block** (no figure moves):
   - A-L03's 32G bound once, on a 43-minute think-on arm with a 19.85 GB model. Before any later long think-on arm on this card, the cap, or a receipt that reads stalls, needs an amendment, registered before its rows.
   - The R-20 receipt is not among the dry run's D-2 checks (`CB7-DRY-RUN-PLAN.md` §4). The dry run's own +20 would have shown it, so it could be.
6. **The seven tags stay in the bench store** until item 1 is ruled: a re-run needs deepseek's tag. Then the lead removes them with `pull.py --lane … rm` (`CB7-DRY-RUN-PLAN.md` §6 (6)).

## §10 · Provenance: what was read, and how

- **The rows.**
  - **CB-7:** `git archive` of estate `483eeafb`'s `bench/cost-of-intelligence-2026-09-23/`, into this lane's session scratch folder (no worktree; the clean checkout at `~/estate-worktrees/coi-cb7-run` was only read). Its `laptop-reading-0.1/` equals the checkout's file for file (`diff -r`), except the three empty, git-ignored `queue.lock` files.
  - **Master's side:** `git archive` of estate origin/master `99f676ec`'s same folder, for the series code, `KIT.lock.json` (`0bca44b4…`), `PREREG-S0.md` (`withheld-2026-09-29`), `series.jsonl`, the 3090's, this card's and the desktop's larger card's reading-0 rows, and the registering queue files.
  - **The desktop's larger card's 0.1 rows:** `~/bench-archive/coi-cb5-2026-09-27/` and `~/bench-archive/coi-cb5b-2026-10-02/` as their L6 lanes brought them home.
- **The tooling:**
  - `parse_v2.py` `689254c5…` (PREREG-S0 §3.1's pin, checked at import; "parse_v2 selftest: 48 cases OK").
  - `series/v2score.py` `f6fecb59…`, `series/stats.py` `86727c4f…` and `series/kitlock.py` `9edc2f9d…`.
  - `analyse_laptop.py` `24da34be…` (= PREREG-S0 §10's working copy). It imports benchbox's `analyse.py`, which reads `74395ddc…` at both commits. That copy carries `PARSER = "v2-benchbox-1"` and the same `classify()` as PREREG-S0 §10's `08a158f6…`: only fields were added between them.
  - The three checks against independent code: five strict counts re-scored with `parse_v2.classify()` directly, `k0a` `c`'s McNemar recomputed by hand (1 · 5, p 0.219), and gemma2 `c`'s board joules re-integrated from `watts.csv` (5,230.2 J over 49.01 s, 50 samples). Each agrees.
- **This fold's own files** (in `fold/`):
  - `cb7_fold.py`, the script that computed every table from the files.
  - `cb7-fold-tables.json`, its output, with every figure and its source path. Card UUIDs become names, with no address and no clock offset. It carries its own `generated_utc`, so a re-run changes its sha, never its tables.
  - `analyse-figures-cb7.scrubbed.json` and `analyse-figures-laptop-r0-anchors.scrubbed.json`, `analyse_laptop.py`'s two `figures.json` outputs with the card's UUID replaced by a name. The raw outputs' sha256 were `967486d3…` and `cb6906ba…`.
  - Nothing else was written, and no `series.jsonl` was written.
- **What this page leaves out on purpose:** card UUIDs, addresses, account names, and the local clock's offset. Every time here is UTC, from file contents or `date -u`.
- **Shas at the final run:** `cb7_fold.py` `c13be86f…`; its output `cb7-fold-tables.json` `9446528c…` (generated 2026-10-02T15:31:04Z); `analyse-figures-cb7.scrubbed.json` `a9218dc3…`; `analyse-figures-laptop-r0-anchors.scrubbed.json` `93395d5d…`.