sub10: add Qwen3.8-27B GSQ-RCO IQ2_XS row (asked for by a reader)

README.md

@@ -239,7 +239,7 @@ One-shot investigations that don't fit the leaderboard format — claim verifica
 
 | Report | Finding |
 |---|---|
-| [Qwen3.8-27B under 10 GB, thinking on: Underdog-Saluki-27B against the 3-bit and ternary files](reports/sub10-qwen38-27b.md) · [chart](reports/sub10-qwen38-27b.png) | Underdog-Saluki-27B-1.0 IQ2-mix (7.9 GB, stock llama.cpp) on the second-tier think-on 16k board, 1,583 items: four-task mean 85.68 against Qwen3.8-27B Q6_K 89.67, UD-IQ3_XXS 89.54, Ternary-Bonsai-2-27B PQ2_0 88.69 and PTQ1_0 88.67 (PrismML fork). Saluki keeps 95.6% of the Q6_K mean, the Bonsai files 98.9%, and trails each of them on all four tasks. Paired on the same items: Q6_K right-only 91 vs 25 (p < 0.0001), PTQ1_0 (1.95 GB smaller) 92 vs 50 (p = 0.0005). Saluki spends 1.5x to 1.8x the tokens per correct answer and capped on 122 of 1,583 generations (Q6_K 42). Tool calling, the model card's headline, is not tested here. |
+| [Qwen3.8-27B under 10 GB, thinking on: Underdog-Saluki-27B against the 3-bit and ternary files](reports/sub10-qwen38-27b.md) · [chart](reports/sub10-qwen38-27b.png) | Underdog-Saluki-27B-1.0 IQ2-mix (7.9 GB, stock llama.cpp) on the second-tier think-on 16k board, 1,583 items: four-task mean 85.68 against Qwen3.8-27B Q6_K 89.67, UD-IQ3_XXS 89.54, Ternary-Bonsai-2-27B PQ2_0 88.69 and PTQ1_0 88.67 (PrismML fork). Saluki keeps 95.6% of the Q6_K mean, the Bonsai files 98.9%, and trails each of them on all four tasks. Paired on the same items: Q6_K right-only 91 vs 25 (p < 0.0001), PTQ1_0 (1.95 GB smaller) 92 vs 50 (p = 0.0005). Saluki spends 1.5x to 1.8x the tokens per correct answer and capped on 122 of 1,583 generations (Q6_K 42). Tool calling, the model card's headline, is not tested here. Added 2026-10-15, asked for by a reader: ISTA-DASLab Qwen3.8-27B GSQ-RCO IQ2_XS (8.4 GB, the quant Saluki's card names as its base) scores 87.01, 97.0% of the Q6_K mean, 1.33 above Saluki (paired 82 vs 61, p = 0.094, inside the noise); misses that hit the 16k cap: Saluki 104 of 228, GSQ-RCO 68 of 207, Q6_K 36 of 162. |
 | [This year's AIME and HMMT at a 32k thinking budget: nine local models against their MATH-500 order](reports/matharena-contests-32k.md) · [chart](reports/matharena-contests-32k.png) | The nine second-tier think-on rows on AIME 2026 (30) + HMMT Feb 2026 (33), MathArena sets, `max_tokens 32768`, `ctx 40960`, greedy, strict sympy grader plus an exact-number veto (0 fired; veto takes agreement with MathArena's labels on HMMT from 98.3% to 99.6%). Gemma 4 31B QAT Q4_0 51/63 (third on MATH-500 at 94.0), Qwen3.8-27B UD-IQ3_XXS 48, Q6_K 47, Ornith 1.5 35B 43, Qwen3.6-27B 37, Qwen3.6-35B-A3B 35, Nemotron 3.5 Lightning 34, Ornith 1.5 9B 29, Qwable-27B 29. Gemma's median completion is 8,259.5 / 8,963 tokens and none of its 63 hit the cap; the other eight capped on 6 to 22 problems per contest and are right on 95.5 to 100% of the answers they finish. 13 of 207 capped answers credited. Gemma vs the Qwen3.8 rows is not significant on 63 problems (paired p = 0.45 / 0.34). |
 | [GPQA Diamond at a 32k thinking budget: the three rows that capped at 16k](reports/gpqa-thinkon-32k.md) | Same 198 questions, only `max_tokens` moved (16,384 to 32,768). Qwable-27B 78.28 to 85.35, Nemotron 3.5 Lightning 30B-A3B 60.61 to 67.17, Qwen3.6-27B 76.77 to 79.80. Every answer that finished inside 16k got the same verdict and completion length at 32k (444/444); every gain is a question that had hit the cap. Nemotron still caps on 19.7% at 32k (a floor). |
 | [Speculative decoding and prefix caching on one RTX 5090, Qwen3.8-27B Q6_K: the two free speedups on the model this box now serves](reports/spec-cache-study-qwen38.md) · [chart](reports/spec-cache-study-qwen38.png) | Rerun of the spec-cache study on Qwen3.8-27B Q6_K (the served model since 2026-09-13, own MTP head in the unsloth GGUF), same 304-request design, same flag set (llama.cpp b9653, ctx 64k, parallel 1, KV q8_0, batch 1). MTP n=2 lifts decode from 61 to 109-144 tok/s (1.8x prose and chat, 2.3x code and repetitive) for +9 ms TTFT; n=4 reaches 168-184 on code/repetitive and loses 5-8% on prose/chat; n-gram lookup is 1.00x. Prefix cache: TTFT 1.004 s to 0.069 s on a 2,954-token system prompt (15x), 3.814 s to 0.086 s on 11,570 tokens (44x), decode unchanged. Stacked, a short answer to the 11.5k context goes 4.25 s to 0.30 s. Losslessness: MTP byte-identical to plain greedy decode on 59% of answers; all 13 divergences are the target's own second choice at a near-tie (top-2 gap 0.004 to 0.133 nats, median 0.030). Every cell within noise of the Qwen3.6-27B run a week earlier: the speedups belong to the shape and the serving stack, not the generation. |

dataset/README.md

@@ -239,7 +239,7 @@ One-shot investigations that don't fit the leaderboard format — claim verifica
 
 | Report | Finding |
 |---|---|
-| [Qwen3.8-27B under 10 GB, thinking on: Underdog-Saluki-27B against the 3-bit and ternary files](reports/sub10-qwen38-27b.md) · [chart](reports/sub10-qwen38-27b.png) | Underdog-Saluki-27B-1.0 IQ2-mix (7.9 GB, stock llama.cpp) on the second-tier think-on 16k board, 1,583 items: four-task mean 85.68 against Qwen3.8-27B Q6_K 89.67, UD-IQ3_XXS 89.54, Ternary-Bonsai-2-27B PQ2_0 88.69 and PTQ1_0 88.67 (PrismML fork). Saluki keeps 95.6% of the Q6_K mean, the Bonsai files 98.9%, and trails each of them on all four tasks. Paired on the same items: Q6_K right-only 91 vs 25 (p < 0.0001), PTQ1_0 (1.95 GB smaller) 92 vs 50 (p = 0.0005). Saluki spends 1.5x to 1.8x the tokens per correct answer and capped on 122 of 1,583 generations (Q6_K 42). Tool calling, the model card's headline, is not tested here. |
+| [Qwen3.8-27B under 10 GB, thinking on: Underdog-Saluki-27B against the 3-bit and ternary files](reports/sub10-qwen38-27b.md) · [chart](reports/sub10-qwen38-27b.png) | Underdog-Saluki-27B-1.0 IQ2-mix (7.9 GB, stock llama.cpp) on the second-tier think-on 16k board, 1,583 items: four-task mean 85.68 against Qwen3.8-27B Q6_K 89.67, UD-IQ3_XXS 89.54, Ternary-Bonsai-2-27B PQ2_0 88.69 and PTQ1_0 88.67 (PrismML fork). Saluki keeps 95.6% of the Q6_K mean, the Bonsai files 98.9%, and trails each of them on all four tasks. Paired on the same items: Q6_K right-only 91 vs 25 (p < 0.0001), PTQ1_0 (1.95 GB smaller) 92 vs 50 (p = 0.0005). Saluki spends 1.5x to 1.8x the tokens per correct answer and capped on 122 of 1,583 generations (Q6_K 42). Tool calling, the model card's headline, is not tested here. Added 2026-10-15, asked for by a reader: ISTA-DASLab Qwen3.8-27B GSQ-RCO IQ2_XS (8.4 GB, the quant Saluki's card names as its base) scores 87.01, 97.0% of the Q6_K mean, 1.33 above Saluki (paired 82 vs 61, p = 0.094, inside the noise); misses that hit the 16k cap: Saluki 104 of 228, GSQ-RCO 68 of 207, Q6_K 36 of 162. |
 | [This year's AIME and HMMT at a 32k thinking budget: nine local models against their MATH-500 order](reports/matharena-contests-32k.md) · [chart](reports/matharena-contests-32k.png) | The nine second-tier think-on rows on AIME 2026 (30) + HMMT Feb 2026 (33), MathArena sets, `max_tokens 32768`, `ctx 40960`, greedy, strict sympy grader plus an exact-number veto (0 fired; veto takes agreement with MathArena's labels on HMMT from 98.3% to 99.6%). Gemma 4 31B QAT Q4_0 51/63 (third on MATH-500 at 94.0), Qwen3.8-27B UD-IQ3_XXS 48, Q6_K 47, Ornith 1.5 35B 43, Qwen3.6-27B 37, Qwen3.6-35B-A3B 35, Nemotron 3.5 Lightning 34, Ornith 1.5 9B 29, Qwable-27B 29. Gemma's median completion is 8,259.5 / 8,963 tokens and none of its 63 hit the cap; the other eight capped on 6 to 22 problems per contest and are right on 95.5 to 100% of the answers they finish. 13 of 207 capped answers credited. Gemma vs the Qwen3.8 rows is not significant on 63 problems (paired p = 0.45 / 0.34). |
 | [GPQA Diamond at a 32k thinking budget: the three rows that capped at 16k](reports/gpqa-thinkon-32k.md) | Same 198 questions, only `max_tokens` moved (16,384 to 32,768). Qwable-27B 78.28 to 85.35, Nemotron 3.5 Lightning 30B-A3B 60.61 to 67.17, Qwen3.6-27B 76.77 to 79.80. Every answer that finished inside 16k got the same verdict and completion length at 32k (444/444); every gain is a question that had hit the cap. Nemotron still caps on 19.7% at 32k (a floor). |
 | [Speculative decoding and prefix caching on one RTX 5090, Qwen3.8-27B Q6_K: the two free speedups on the model this box now serves](reports/spec-cache-study-qwen38.md) · [chart](reports/spec-cache-study-qwen38.png) | Rerun of the spec-cache study on Qwen3.8-27B Q6_K (the served model since 2026-09-13, own MTP head in the unsloth GGUF), same 304-request design, same flag set (llama.cpp b9653, ctx 64k, parallel 1, KV q8_0, batch 1). MTP n=2 lifts decode from 61 to 109-144 tok/s (1.8x prose and chat, 2.3x code and repetitive) for +9 ms TTFT; n=4 reaches 168-184 on code/repetitive and loses 5-8% on prose/chat; n-gram lookup is 1.00x. Prefix cache: TTFT 1.004 s to 0.069 s on a 2,954-token system prompt (15x), 3.814 s to 0.086 s on 11,570 tokens (44x), decode unchanged. Stacked, a short answer to the 11.5k context goes 4.25 s to 0.30 s. Losslessness: MTP byte-identical to plain greedy decode on 59% of answers; all 13 divergences are the target's own second choice at a near-tie (top-2 gap 0.004 to 0.133 nats, median 0.030). Every cell within noise of the Qwen3.6-27B run a week earlier: the speedups belong to the shape and the serving stack, not the generation. |

dataset/second_tier.csv

@@ -51,3 +51,7 @@ underdog-saluki-27b-iq2-mix,Underdog-Saluki-27B-1.0,IQ2-mix,7898369152,ifeval,pr
 underdog-saluki-27b-iq2-mix,Underdog-Saluki-27B-1.0,IQ2-mix,7898369152,math500,acc,91.6,458,500,27,0.054,18,3022.8,704.5,"think-on, max_tokens 16384, ctx 24576, zero-shot, greedy",
 underdog-saluki-27b-iq2-mix,Underdog-Saluki-27B-1.0,IQ2-mix,7898369152,humaneval_plus,pass@1,89.02,146,164,12,0.0732,,2748.2,598.0,"think-on, max_tokens 16384, ctx 24576, zero-shot, greedy",
 underdog-saluki-27b-iq2-mix,Underdog-Saluki-27B-1.0,IQ2-mix,7898369152,mbpp_plus,pass@1,77.25,292,378,29,0.0767,,3145.4,426.5,"think-on, max_tokens 16384, ctx 24576, zero-shot, greedy",
+qwen3-8-27b-gsq-rco-iq2-xs,Qwen3.8-27B,GSQ-RCO IQ2_XS,8422841472,ifeval,prompt_strict_acc,86.14,466,541,31,0.0573,0,3007.5,885,"think-on, max_tokens 16384, ctx 24576, zero-shot, greedy",
+qwen3-8-27b-gsq-rco-iq2-xs,Qwen3.8-27B,GSQ-RCO IQ2_XS,8422841472,math500,acc,92.8,464,500,18,0.036,11,2501.7,678.0,"think-on, max_tokens 16384, ctx 24576, zero-shot, greedy",
+qwen3-8-27b-gsq-rco-iq2-xs,Qwen3.8-27B,GSQ-RCO IQ2_XS,8422841472,humaneval_plus,pass@1,90.24,148,164,10,0.061,,2399.6,550.0,"think-on, max_tokens 16384, ctx 24576, zero-shot, greedy",
+qwen3-8-27b-gsq-rco-iq2-xs,Qwen3.8-27B,GSQ-RCO IQ2_XS,8422841472,mbpp_plus,pass@1,78.84,298,378,22,0.0582,,2623.5,367.5,"think-on, max_tokens 16384, ctx 24576, zero-shot, greedy",

reports/sub10-qwen38-27b.md

@@ -2,13 +2,15 @@
 
 **The one line:** Underdog-Saluki-27B-1.0 (7.9 GB) keeps 95.6% of the Qwen3.8-27B Q6_K four-task mean at a 16k thinking budget; the two Ternary-Bonsai-2-27B files (5.9 and 7.2 GB) keep 98.9% and the 11.9 GB UD-IQ3_XXS 99.9%, and Saluki trails each of them on all four tasks.
 
-**Rig:** RTX 5090 32GB (capsule), llama-server, batch 1 · **Window:** Saluki 2026-10-11 to 2026-10-13, inside the box's cheap-electricity hours (idle-queue item `65s-sub10-core`) · **Harness:** [llm-bench-rig](https://github.com/notwitcheer/llm-bench-rig) second-tier lane, artefacts in `results/<slug>-thinkon/` · **Data:** [`dataset/second_tier.csv`](../dataset/second_tier.csv) · **Chart:** [`sub10-qwen38-27b.png`](sub10-qwen38-27b.png) · **Parent board:** [thinking on at a 16k budget](second-tier-thinkon-16k.md)
+**Addendum 2026-10-15:** a reader asked for the GSQ-RCO IQ2_XS file, the quant Saluki's model card names as its base. It scores 87.01 on the same four tasks (8.4 GB, 97.0% of the Q6_K mean), 1.33 above Saluki, which is inside the noise on a paired test. Details, a cap-hit breakdown for every row and a six-row chart are in the [Addendum](#addendum-2026-10-15-gsq-rco-iq2_xs-the-file-saluki-is-built-on) at the end.
+
+**Rig:** RTX 5090 32GB (capsule), llama-server, batch 1 · **Window:** Saluki 2026-10-11 to 2026-10-13, GSQ-RCO 2026-10-13 to 2026-10-15, inside the box's cheap-electricity hours (idle-queue item `65s-sub10-core`) · **Harness:** [llm-bench-rig](https://github.com/notwitcheer/llm-bench-rig) second-tier lane, artefacts in `results/<slug>-thinkon/` · **Data:** [`dataset/second_tier.csv`](../dataset/second_tier.csv) · **Chart:** [`sub10-qwen38-27b.png`](sub10-qwen38-27b.png) (five rows, as posted), [`sub10-qwen38-27b-gsq.png`](sub10-qwen38-27b-gsq.png) (six rows, 2026-10-15) · **Parent board:** [thinking on at a 16k budget](second-tier-thinkon-16k.md)
 
 ![sub-10 GB Qwen3.8-27B card](sub10-qwen38-27b.png)
 
 ## TL;DR
 
-1. **Saluki lands 4.0 points under Q6_K on the four-task mean (85.68 vs 89.67), the only gap on this page wider than the interval.** The propagated Wilson 95% half-width on a four-task mean is 1.6 to 1.9 points per row here, so the Q6_K, UD-IQ3_XXS and both Bonsai rows (88.67 to 89.67) are one band and Saluki sits below it. On the same 1,583 items, Q6_K is right where Saluki is wrong 91 times and the reverse happens 25 times (exact two-sided sign test on the discordant items, p < 0.0001).
+1. **Saluki lands 4.0 points under Q6_K on the four-task mean (85.68 vs 89.67), the only gap among the five rows published 2026-10-13 wider than the interval.** (The GSQ-RCO row added 2026-10-15 also sits below that band; see the Addendum.) The propagated Wilson 95% half-width on a four-task mean is 1.6 to 1.9 points per row here, so the Q6_K, UD-IQ3_XXS and both Bonsai rows (88.67 to 89.67) are one band and Saluki sits below it. On the same 1,583 items, Q6_K is right where Saluki is wrong 91 times and the reverse happens 25 times (exact two-sided sign test on the discordant items, p < 0.0001).
 2. **The smaller ternary file is ahead too.** Ternary-Bonsai-2-27B PTQ1_0 is 1.95 GB smaller and scores 88.67; paired against Saluki it is right-only on 92 items and wrong-only on 50 (p = 0.0005). The Bonsai files need the PrismML llama.cpp fork; Saluki runs on stock llama.cpp, which is the real difference for anyone on an app built on mainline.
 3. **Saluki thinks longer, and the cap catches more of it.** 1.5x to 1.8x the completion tokens per correct answer of Q6_K on every task, 122 of its 1,583 generations reached the 16,384-token cap against 42 for Q6_K, and 104 of its 228 wrong answers were capped. IFEval capped at 9.98%, on the edge of this board's 10% budget-limited flag. The model card says the same in its own words: "with thinking on, it often reasons at length before answering".
 
@@ -31,9 +33,10 @@ Scores are percentages. `cap` is the share of items whose completion reached the
 | Qwen3.8-27B | UD-IQ3_XXS | 11.9 GB | 90.2 | 94.6 | 92.68 | 80.69 | **89.54** | -0.12 | 3.3% | 1,891 |
 | Ternary-Bonsai-2-27B † | PQ2_0 | 7.2 GB | 87.25 | 95.0 | 92.07 | 80.42 | **88.69** | -0.98 | 7.4% | 2,075 |
 | Ternary-Bonsai-2-27B † | PTQ1_0 | 5.9 GB | 87.62 | 93.8 | 93.9 | 79.37 | **88.67** | -1.00 | 7.2% | 2,127 |
+| Qwen3.8-27B ‡ | GSQ-RCO IQ2_XS | 8.4 GB | 86.14 | 92.8 | 90.24 | 78.84 | **87.01** | -2.66 | 5.73% | 2,502 |
 | Underdog-Saluki-27B-1.0 | IQ2-mix | 7.9 GB | 84.84 | 91.6 | 89.02 | 77.25 | **85.68** | -3.99 | 9.98% | 3,023 |
 
-† PrismML llama.cpp fork `4fbda12`; every other row mainline b10371.
+† PrismML llama.cpp fork `4fbda12`; every other row mainline b10371. ‡ added 2026-10-15, measured 2026-10-13 to 2026-10-15 on the same binary and regime as Saluki (see the Addendum; the four-task mean is exactly 87.005).
 
 Saluki per task, with Wilson 95% half-widths and the cost columns (all from `dataset/second_tier.csv`):
 
@@ -82,12 +85,69 @@ Paired against each row on identical items (A-only = the comparison row right, S
 - The comparison rows were measured 2026-09 (Qwen3.8) and 2026-10-08 to 10 (Bonsai), Saluki 2026-10-11 to 13; build, regime and harness held constant.
 - No speed or memory measurement here; the file size is the GGUF on disk, not peak VRAM.
 - Tool calling, the card's headline, is not measured on this rig.
-- ISTA-DASLab GSQ-RCO IQ2_XS (8.42 GB, the base Saluki builds on) is in flight on the same regime; its row is not on this page yet.
+- ISTA-DASLab GSQ-RCO IQ2_XS (8.42 GB, the base Saluki builds on) was added 2026-10-15 after the post went out; the TL;DR, Reads and the first chart describe the five rows published 2026-10-13, the Addendum covers the sixth.
 
 ## Owed
 
-- ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF IQ2_XS on the same four tasks (running in the cheap-electricity blocks).
+- ~~ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF IQ2_XS on the same four tasks~~ done 2026-10-15, see the Addendum.
 - GPQA Diamond (think-on 16k) and AIME 2026 + HMMT Feb 2026 (think-on 32k) for both sub-10 GB rows (idle-queue item `67-sub10-qwen38-27b`).
 - A 32k pass on Saluki's four legs, to measure how much of the gap is the 16k budget.
 
-*Sources: `results/underdog-saluki-27b-iq2-mix-thinkon/{ifeval,math500,humaneval_plus,mbpp_plus}_detail.json` and `*_progress.json` on the rig, read 2026-10-13; `dataset/second_tier.csv` in this commit; model card [ConwayResearch/Underdog-Saluki-27B-1.0](https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0), read 2026-10-13; chart `scripts/chart_sub10_qwen38.py`.*
+## Addendum 2026-10-15: GSQ-RCO IQ2_XS, the file Saluki is built on
+
+A reader asked for this row. [ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF](https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF) is the quant Saluki's model card names as its base: the `base_model` field of [ConwayResearch/Underdog-Saluki-27B-1.0](https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0) lists `ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF` (read 2026-10-15). So this row is the base file before Saluki's tuning, on the same board.
+
+- **File:** `Qwen3.8-27B-GSQ-RCO-IQ2_XS.gguf`, 8,422,841,472 bytes (`stat -L`), Apache-2.0, a standard GGUF.
+- **Conditions:** identical to the Saluki row. Thinking on, `max_tokens 16384`, `ctx 24576`, greedy, zero-shot, one pass per item, mainline llama-server b10371 (`5d16e81dd`), same harness, idle-queue item `65s-sub10-core`. Run window 2026-10-13 to 2026-10-15 (last leg, IFEval, finished 2026-10-15 02:03 CEST). Zero request errors on all four tasks. Grader import check: `math_equal(r'\frac{1}{2}', '0.5')` from the vendored Qwen-Math grader printed `True` in the run venv on 2026-10-15.
+- **Chart:** the posted card above stays as it was. The six-row version, same script with `--with-gsq`:
+
+![sub-10 GB Qwen3.8-27B card with the GSQ-RCO row](sub10-qwen38-27b-gsq.png)
+
+GSQ-RCO IQ2_XS per task (all from `dataset/second_tier.csv`):
+
+| Task | Score | Correct / total | ± (95%) | Capped | Tokens / correct | Median completion | vs Saluki | vs Q6_K |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|
+| IFEval prompt-strict | 86.14 | 466 / 541 | 2.9 | 31 (5.73%) | 3,008 | 885 | +1.30 | -3.51 |
+| MATH-500 | 92.8 | 464 / 500 | 2.3 | 18 (3.6%), 11 unboxed | 2,502 | 678 | +1.2 | -2.4 |
+| HumanEval+ | 90.24 | 148 / 164 | 4.6 | 10 (6.1%) | 2,400 | 550 | +1.22 | -1.83 |
+| MBPP+ | 78.84 | 298 / 378 | 4.1 | 22 (5.82%) | 2,624 | 367.5 | +1.59 | -2.91 |
+
+IFEval inst-strict is 92.15 (740 of 803 instructions).
+
+Paired against each row on identical items (A-only = the comparison row right, GSQ-RCO wrong; G-only = the reverse):
+
+| vs | IFEval A-only / G-only | MATH-500 | HumanEval+ | MBPP+ | All four | p (all four) |
+|---|---:|---:|---:|---:|---:|---:|
+| Qwen3.8-27B Q6_K | 31 / 12 | 17 / 5 | 7 / 4 | 24 / 13 | 79 / 34 | < 0.0001 |
+| Qwen3.8-27B UD-IQ3_XXS | 35 / 13 | 17 / 8 | 5 / 1 | 24 / 17 | 81 / 39 | 0.0002 |
+| Ternary-Bonsai-2-27B PQ2_0 | 29 / 23 | 20 / 9 | 8 / 5 | 21 / 15 | 78 / 52 | 0.028 |
+| Ternary-Bonsai-2-27B PTQ1_0 | 32 / 24 | 16 / 11 | 8 / 2 | 19 / 17 | 75 / 54 | 0.078 |
+| Underdog-Saluki-27B-1.0 | 27 / 34 | 14 / 20 | 5 / 7 | 15 / 21 | 61 / 82 | 0.094 |
+
+Misses that hit the 16k cap, every row. A miss is an item scored wrong; capped means the completion reached 16,384 tokens (`capped: true` in the task's `*_progress.json`), summed over the four tasks (1,583 items per row):
+
+| Row | Misses | Capped misses | Share | Finished and wrong |
+|---|---:|---:|---:|---:|
+| Qwen3.8-27B Q6_K | 162 | 36 | 22% | 126 |
+| Qwen3.8-27B UD-IQ3_XXS | 165 | 38 | 23% | 127 |
+| Ternary-Bonsai-2-27B PQ2_0 † | 181 | 59 | 33% | 122 |
+| Ternary-Bonsai-2-27B PTQ1_0 † | 186 | 68 | 37% | 118 |
+| Qwen3.8-27B GSQ-RCO IQ2_XS | 207 | 68 | 33% | 139 |
+| Underdog-Saluki-27B-1.0 IQ2-mix | 228 | 104 | 46% | 124 |
+
+Per task, capped misses / misses: Saluki IFEval 46 / 82, MATH-500 22 / 42, HumanEval+ 9 / 18, MBPP+ 27 / 86; Q6_K 13 / 56, 5 / 24, 5 / 13, 13 / 69; GSQ-RCO 27 / 75, 13 / 36, 7 / 16, 21 / 80.
+
+### Reads (addendum)
+
+1. **GSQ-RCO keeps 97.0% of the Q6_K four-task mean (87.01 against 89.67), 96.1% to 98.0% per task.** It sits between Saluki (85.68) and the two Bonsai files (88.69, 88.67), and like Saluki it is below Q6_K on a paired test (79 against 34 discordant items, p < 0.0001).
+2. **Against Saluki the gap is inside the noise.** GSQ-RCO is ahead on all four tasks by 1.2 to 1.59 points and 1.33 on the mean, but on identical items it is right-only 82 times against 61 for Saluki (p = 0.094), and the two means carry Wilson half-widths of 1.80 and 1.87. On these four tasks with thinking on, the board does not show that Saluki's tuning cost or gained anything measurable over its base file. Tool calling, where the tuning is aimed, is not measured here.
+3. **Against the Bonsai files:** PQ2_0 (1.22 GB smaller) is ahead 78 to 52 (p = 0.028); PTQ1_0 (2.48 GB smaller) is ahead 75 to 54 (p = 0.078), not significant on its own. Both need the PrismML fork; GSQ-RCO ran on the same mainline binary as Saluki.
+4. **Saluki spends more tokens than its base and hits the cap more often.** Saluki's tokens per correct answer are 1.15x to 1.31x GSQ-RCO's on every task, and 122 of its 1,583 generations reached the cap against 81 for GSQ-RCO and 42 for Q6_K. GSQ-RCO itself spends 1.2x to 1.4x the tokens per correct answer of Q6_K.
+5. **About half of Saluki's misses are cap hits, a third of GSQ-RCO's, a fifth of Q6_K's.** Counting only answers that finished and were wrong, the order changes: Saluki 124, Q6_K 126, GSQ-RCO 139. This does not show that a bigger budget would turn the capped misses into right answers (a capped item can be wrong at any budget), so it is a reason for the owed 32k pass, not a result.
+
+### Limits (addendum)
+
+- One greedy pass, one card, the same limits as the rest of this page. The ISTA-DASLab repo also ships an `-mtp` variant of this file; it was not run.
+- Paired and cap-hit numbers: [`scripts/sub10_paired.py`](../scripts/sub10_paired.py) `--subject qwen3-8-27b-gsq-rco-iq2-xs` on the rig against the `*_progress.json` files (the default subject reproduces the Saluki tables above unchanged).
+
+*Sources: `results/underdog-saluki-27b-iq2-mix-thinkon/{ifeval,math500,humaneval_plus,mbpp_plus}_detail.json` and `*_progress.json` on the rig, read 2026-10-13; `results/qwen3-8-27b-gsq-rco-iq2-xs-thinkon/` (same files), read 2026-10-15; `dataset/second_tier.csv` in this commit; model card [ConwayResearch/Underdog-Saluki-27B-1.0](https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0), read 2026-10-13 and 2026-10-15 (`base_model`); charts `scripts/chart_sub10_qwen38.py` (default and `--with-gsq`).*