# PREREG-BANDS · the later the desktop rows' bands and rules, drawn before R-LB-IQ3's reading

**Registered** 2026-10-05T03:01:33Z (`date -u` on the laptop), by the lead, before any the desktop reading exists (the first window opens
2026-10-05 07:30Z). Internal. These bands are predictions from arithmetic and from laptop readings already on file;
no the desktop reading was an input. A row whose band is not here runs as "no band: a control (drawn after R-LB-IQ3)" and
is never banded later.

**Inputs.** The desktop's larger card's published bandwidth 1,792 GB/s; ~5.159 GB per generated token for IQ3_S (PREREG-W0 §4), so the
per-token ceiling is **347 tok/s** (1,792 ÷ 5.159 = 347.4). The laptop's L-IQ3-t8: client 24.4 vs engine 24.4 tok/s
at 4K (the client clock's overhead 0.0; `SUMMARY.md` row 105). The laptop's S-IQ3 (hybrid regime, 8,120 of 24,576
slots, labelled overlapped): `tokens_per_round_derived` 2.69 / 2.56 / 2.25 at 4K / 32K / 128K; the engine log shows
windows "up to 6 tokens".

| Label | Engine, instrument | Predicted (decode at 4K, or tg128@d0) | Outside | Under the floor | VOID |
|---|---|---|---|---|---|
| R-LB-IQ3-n2 | llama.cpp 3cf03257, llama-bench, 300 W (a noise read) | 85-160 tok/s (R-LB-IQ3's band) | 40-85, 160-347 | < 40 | > 347 |
| R-LB-IQ3-420 | the same at 420 W (the posture row) | 85-160 | 40-85, 160-347 | < 40 | > 347 |
| R-LB-IQ3-fa0 | the same with `-fa off` at depth 0 (`-ctv f16` forced) | 85-160 | 40-85, 160-347 | < 40 | > 347 |
| R-LB-IQ3-tg | the same with `-p 0` (only if a1 crashed in pp512@d0) | 85-160 | 40-85, 160-347 | < 40 | > 347 |
| **R-LS-IQ3** | llama-server 3cf03257, the harness client, `--n-cpu-moe 0`, 300 W | **85-160 tok/s** (the same arithmetic; the client clock adds nothing measurable) | 40-85, 160-347 | < 40 | > 347 |
| **R-S139-IQ3** | Strata v0.1.39, the harness client, 300 W | **in rounds:** 50-130 rounds/s [inferred: 14-37 % of the per-round ceiling, under llama.cpp's 24-46 % because a verify round reads at least one token's bytes AND runs the draft and verify compute for up to 6 tokens] × 2.25-2.69 tokens per round (the laptop's values, read on another box in a hybrid regime) = **113-350 tok/s** | 40-113, 350-(tpr × 347) | < 40 | > `tokens_per_round_derived` × 347 for that run (≤ 934 at 2.69); and > 2,084 (6 × 347.4, as PLAN-v2 states it) whatever the rounds read |
| **R-S125-IQ3** | Strata 8fc40dde, the same | the same band as R-S139 (the engine version is the thing under test; one band for both is the honest shape) | the same | < 40 | the same |

The laptop's S-IQ3 4K reading (85.6 tok/s) is NOT a base for any Strata band here: it is a hybrid regime and it is
labelled overlapped. The third party's dual-card figures (D-095) are a claim on another box, not an input.

**The rules every row carries.** (1) The attempt rule of PREREG-W0 §10.3. (2) The out-of-band rule of PREREG-W0 §10.4.
(3) **The cross-night rule:** a difference between rows from different windows is reported as a difference only when
it exceeds the larger of 10 % or 3 × the larger row's spread across its repetitions; otherwise "within unmeasured
night-to-night variation"; same-window pairs are preferred once the row list exists; every compared row prints the
cap read back, `MemoryMax`, nice/CPU properties, `CUDA_VISIBLE_DEVICES`, the stop set, the desktop's smaller card's seat state, and
the desktop's larger card's temperature and SM clock per test. (4) **Residency:** "whole" only on the receipts PLAN-v2 §2.5/§2.6 name;
any miss is "partial residency (n of 24,576 slots; hit rate h)". (5) The label rules of PLAN-v2 §2.4-§2.6 (box, not
card; the overlapped label; no laptop twin for R-S139). (6) The 32K and 128K readings are exploratory, no band.