Model card: add measured tok/s (Meridian, base, VibeThinker-3B), VibeThinker-3B comparisons in Track A/B tables, tuned VibeThinker summary; keep table edit from main

README.md

@@ -33,9 +33,9 @@ It improves in-domain formal-logic performance by **+28% relative** over its bas
 macro is level with the base (0.7942 → 0.7901) and the 14-dataset macro improves slightly
 (0.7332 → 0.7425). On the same harness and sampled rows it reaches the highest Track A macro
 gate, strict-7 and six-lane average of any arm in the tables below for which each can be
-computed, including Qwen3-8B and gpt-oss-120b (the 120B has no gate or strict-7 value).
+computed, including Qwen3-8B, VibeThinker-3B and gpt-oss-120b (the 120B has no gate or strict-7 value).
 
-![Meridian-smaller formal and general reasoning benchmarks against gpt-oss-120b, Qwen3-8B, LFM2.5-8B-A1B and TwIL-LM3](benchmarks.png)
+![Meridian-smaller formal and general reasoning benchmarks against VibeThinker-3B, Qwen3-8B, gpt-oss-120b, LFM2.5-8B-A1B and TwIL-LM3](benchmarks.png)
 
 ## Highlights
 
@@ -55,14 +55,21 @@ computed, including Qwen3-8B and gpt-oss-120b (the 120B has no gate or strict-7
   Qwen3-8B (0.8493 / 0.7591) and gpt-oss-120b (0.8689 / 0.8086). It gains on BBH-logic
   (0.9013 → 0.9540), MATH-500 (0.6567 → 0.7467) and MuSR (0.5922 → 0.6409), and gives back
   GSM-Symbolic (0.8900 → 0.8267) and ARC (0.8933 → 0.8600).
+* **Against VibeThinker-3B, a reasoning model of similar size.** Meridian-smaller leads it on all
+  six Track A lanes — macro gate 0.5539 against 0.4118, strict-7 0.2879 against 0.2021 — but not
+  on Track B, where VibeThinker-3B is ahead on the 10-dataset macro (0.8097 against 0.7901) and
+  Meridian-smaller is ahead on the 14-dataset macro (0.7425 against 0.7262). That 14-dataset lead
+  comes entirely from BBH-logic (0.9540 against 0.6107); without that row Meridian-smaller trails.
 * **Structured formal output.** Tuned for the objects rather than the prose: FOL translation,
   entailment labels, semantic parses, Lean statements and Lean proof critique.
 * **Runs anywhere.** 3.66B parameters in bf16 (6.82 GiB), with a Q4\_K\_M GGUF at 2.09 GiB for
   CPU or 4 GB of VRAM.
 
-It is **not efficient**: it reasons at length. Track A generations average 1,902 tokens, against
-564 for TwIL-LM3, and 24.2% of them hit the length cap. It is also not a general assistant —
-there is no safety or preference tuning here beyond what Granite 4.2 carries. See
+It is **not the most efficient**: it reasons at length. Track A generations average 1,902 tokens,
+against 564 for TwIL-LM3, and 24.2% of them hit the length cap. Its raw decode rate is high
+(21,169 tok/s †) but the long answers bring it to about 11 completed
+answers per second †, against 28.1 for TwIL-LM3 and 4.5 for Qwen3-8B. It is also not a general
+assistant — there is no safety or preference tuning here beyond what Granite 4.2 carries. See
 [Limitations](#limitations-and-caveats).
 
 ## Model Details
@@ -92,30 +99,31 @@ here.
 
 ### Track A — in-domain formal logic
 
-Meridian-smaller and its base were run through the same harness, prompts, decoding settings and
-sampled rows described under [Evaluation protocol](#evaluation-protocol), and the peer columns
-are the values already published for TwIL-LM3 on its card, produced by that same harness (see
-[Comparability](#limitations-and-caveats)). Throughput rows are reported where a dedicated
-throughput measurement exists: `ans/s` is defined throughout as `tok/s ÷ mean generation length`,
-so it measures completed answers rather than raw decode rate.
-
-| lane / metric | Meridian-smaller | Granite-4.2-3B base | TwIL-LM3 | Llama-3.2-3B | LFM2-2.6B | LFM2.5-8B-A1B | Qwen3-8B | gpt-oss-120b ‡ |
-|---|---:|---:|---:|---:|---:|---:|---:|---:|
-| lean_formalize token_f1 | 0.5092 | 0.2943 | 0.5869 | 0.3690 | 0.1321 | 0.4655 | 0.4022 | **0.6306** |
-| rule_induction derivation | 0.4195 | 0.2267 | 0.3192 | 0.0825 | 0.0615 | 0.1936 | 0.3680 | **0.6518** |
-| entailment_label accuracy | 0.6700 | 0.3000 | 0.5750 | 0.3300 | 0.4700 | 0.5400 | 0.5800 | **0.7750** |
-| mcq_answer accuracy | **0.4100** | 0.1200 | 0.1100 | 0.0000 | 0.0150 | 0.0750 | 0.0000 | 0.0700 |
-| semantic_parse token_f1 | 0.4295 | 0.3910 | **0.4416** | 0.3102 | 0.3665 | 0.3778 | 0.4257 | 0.4331 |
-| lean_critic accuracy | **0.7950** | **0.7950** | 0.6600 | 0.5300 | 0.5900 | 0.5500 | **0.7950** | 0.5550 |
-| lm_corpus perplexity ↓ | **2.3130** | 2.6334 | 2.8972 | 2.8478 | 4.3815 | 4.9472 | 2.5440 | 912.23 § |
-| math_corpus perplexity ↓ | **3.6983** | 4.4864 | 3.8229 | 4.7531 | 6.7472 | 8.3323 | 4.0083 | 1045.63 § |
-| average, 6 lanes | **0.5389** | 0.3545 | 0.4488 | 0.2703 | 0.2725 | 0.3670 | 0.4285 | 0.5192 |
-| **macro gate** | **0.5539** | 0.4313 | 0.4218 | 0.2925 | 0.3473 | 0.3757 | 0.5336 | — |
-| **strict-7** | **0.2879** | 0.1821 | 0.1971 | 0.1229 | 0.1579 | 0.1714 | 0.2093 | — |
-| macro_primary | **0.5875** | 0.4825 | 0.4475 | 0.3450 | 0.4188 | 0.4213 | 0.5750 | — |
-| tok/s | — | — | 15880 | 16160 | **25230** | 22480 | 9420 | 3374 |
-| mean gen length | 1902 | 2951 | **564** | 696 | 2296 | 1830 | 2094 | 1005 |
-| **ans/s** | — | — | **28.1** | 23.2 | 10.9 | 12.0 | 4.5 | 3.4 |
+Meridian-smaller, its base and VibeThinker-3B were run through the same harness, prompts, decoding
+settings and sampled rows described under [Evaluation protocol](#evaluation-protocol), and the
+other peer columns are the values already published for TwIL-LM3 on its card, produced by that
+same harness (see [Comparability](#limitations-and-caveats)). Throughput rows come from a
+dedicated decode-throughput protocol: `ans/s` is defined throughout as
+`tok/s ÷ mean generation length`, so it measures completed answers rather than raw decode rate.
+Cells marked † need the engine note below.
+
+| lane / metric | Meridian-smaller | Granite-4.2-3B base | VibeThinker-3B | TwIL-LM3 | Llama-3.2-3B | LFM2-2.6B | LFM2.5-8B-A1B | Qwen3-8B | gpt-oss-120b ‡ |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| lean_formalize token_f1 | 0.5092 | 0.2943 | 0.2087 | 0.5869 | 0.3690 | 0.1321 | 0.4655 | 0.4022 | **0.6306** |
+| rule_induction derivation | 0.4195 | 0.2267 | 0.2038 | 0.3192 | 0.0825 | 0.0615 | 0.1936 | 0.3680 | **0.6518** |
+| entailment_label accuracy | 0.6700 | 0.3000 | 0.5500 | 0.5750 | 0.3300 | 0.4700 | 0.5400 | 0.5800 | **0.7750** |
+| mcq_answer accuracy | **0.4100** | 0.1200 | 0.1700 | 0.1100 | 0.0000 | 0.0150 | 0.0750 | 0.0000 | 0.0700 |
+| semantic_parse token_f1 | 0.4295 | 0.3910 | 0.2422 | **0.4416** | 0.3102 | 0.3665 | 0.3778 | 0.4257 | 0.4331 |
+| lean_critic accuracy | **0.7950** | **0.7950** | 0.6000 | 0.6600 | 0.5300 | 0.5900 | 0.5500 | **0.7950** | 0.5550 |
+| lm_corpus perplexity ↓ | **2.3130** | 2.6334 | 16.9253 | 2.8972 | 2.8478 | 4.3815 | 4.9472 | 2.5440 | 912.23 § |
+| math_corpus perplexity ↓ | **3.6983** | 4.4864 | 27.0318 | 3.8229 | 4.7531 | 6.7472 | 8.3323 | 4.0083 | 1045.63 § |
+| average, 6 lanes | **0.5389** | 0.3545 | 0.3291 | 0.4488 | 0.2703 | 0.2725 | 0.3670 | 0.4285 | 0.5192 |
+| **macro gate** | **0.5539** | 0.4313 | 0.4118 | 0.4218 | 0.2925 | 0.3473 | 0.3757 | 0.5336 | — |
+| **strict-7** | **0.2879** | 0.1821 | 0.2021 | 0.1971 | 0.1229 | 0.1579 | 0.1714 | 0.2093 | — |
+| macro_primary | **0.5875** | 0.4825 | 0.4637 | 0.4475 | 0.3450 | 0.4188 | 0.4213 | 0.5750 | — |
+| tok/s | 21169 † | 21916 † | 28010 † | 15880 | 16160 | 25230 | 22480 | 9420 | 3374 |
+| mean gen length | 1902 | 2951 | 2688 | **564** | 696 | 2296 | 1830 | 2094 | 1005 |
+| **ans/s** | 11.1 † | 7.4 † | 10.4 † | **28.1** | 23.2 | 10.9 | 12.0 | 4.5 | 3.4 |
 
 ‡ **gpt-oss-120b** runs MXFP4 weights at tensor-parallel 2 — quantized and multi-GPU, so its
 throughput rows are not directly comparable to the single-GPU BF16 arms. Its `procedural` lane
@@ -126,11 +134,17 @@ average cannot be computed for it; that is what the — cells mean, not a zero.
 format and tokenizer make the corpus lanes score a different quantity. The number is reported
 for completeness but is not a comparable measurement, and is excluded from the bolding.
 
-**Throughput cells marked — for Meridian-smaller and its base.** The peer throughput columns come
-from a separate dedicated throughput measurement that was not run for these two models, so
-`tok/s` and `ans/s` are left blank rather than mixed with the in-run decode rate (which is on a
-different basis). `mean gen length` is measured directly under the shared protocol and is
-comparable across every column.
+† **Throughput for Meridian-smaller, its base and VibeThinker-3B** was measured with the same
+dedicated protocol and prompt file as the peer columns (128 prompts × 512 generated tokens, EOS
+ignored, greedy, `gpu_memory_utilization` 0.45, idle GPU), run on a single H200 with vLLM
+0.19.1 in a later session, whereas the other columns are figures recorded earlier on vLLM 0.11.2. Engine effects
+are architecture-dependent: in the same session TwIL-LM3 re-measured at 15,248 tok/s
+against its published 15,880 (-4%), while VibeThinker-3B measured
+28,010 against a published 18,116 (+55%). The †
+cells are therefore comparable with each other, and only approximately with the older columns.
+Meridian-smaller's two runs gave 20,454 and 21,884 tok/s (the first paid for cold
+kernel caches) and the table shows their mean. `mean gen length` is measured directly under the
+shared evaluation protocol and is comparable across every column.
 
 **`average, 6 lanes`** is the plain mean of the six objective rows above it, each at whatever
 scoring that row reports (Lean F1, rule derivation, entailment, MCQ strict, semantic F1, critic).
@@ -165,6 +179,13 @@ Qwen3-8B the gate gap is only 0.020, but strict-7 is 0.2879 against 0.2093 — a
 does not depend on loose-match credit — and Meridian-smaller leads on MCQ (strict 0.4100 against
 0.0000), rule induction (0.4195 against 0.3680) and both perplexity lanes.
 
+VibeThinker-3B — WeiboAI's 3B reasoning model, with 37.1% of its Track A generations truncated
+— trails Meridian-smaller on every objective lane and every summary row: macro gate 0.4118 against
+0.5539, strict-7 0.2021 against 0.2879, six-lane average 0.3291 against 0.5389. The gap is widest
+on Lean formalisation (0.2087 against 0.5092) and narrowest on entailment (0.5500 against 0.6700).
+Its corpus perplexities (16.93 and 27.03) are not comparable with the Granite-tokenizer models',
+because perplexity is per token and its vocabulary is 151,936 against 100,352.
+
 It does not lead every lane. gpt-oss-120b is clearly stronger on `rule_induction` (0.6518),
 entailment (0.7750) and Lean formalisation (0.6306), and TwIL-LM3 remains ahead on Lean
 formalisation (0.5869 against 0.5092) and semantic parsing (0.4416 against 0.4295). The two weak
@@ -173,27 +194,27 @@ spots in absolute terms are `procedural` (strict 0.1200, loose 0.2350) and FOL t
 
 ### Track B — held-out benchmarks
 
-| dataset | Meridian-smaller | Granite-4.2-3B base | TwIL-LM3 | Llama-3.2-3B | LFM2-2.6B | LFM2.5-8B-A1B | Qwen3-8B | gpt-oss-120b ‡ |
-|---|---:|---:|---:|---:|---:|---:|---:|---:|
-| gsm8k | 0.9433 | 0.9533 | 0.8733 | 0.8300 | 0.8767 | 0.9133 | 0.9567 | **0.9767** |
-| svamp | **0.9500** | 0.9200 | 0.8500 | 0.8200 | 0.9000 | 0.9133 | 0.9400 | 0.9400 |
-| gsm_symbolic | 0.8267 | 0.8900 | 0.7567 | 0.8067 | **0.9767** | 0.9267 | 0.8133 | 0.8467 |
-| arc_cot | 0.8600 | 0.8933 | 0.8467 | 0.7967 | 0.8667 | 0.9033 | 0.9633 | **0.9667** |
-| logicbench | 0.7933 | 0.7733 | 0.7167 | 0.5733 | 0.6267 | 0.7200 | **0.8567** | 0.8533 |
-| strategyqa | 0.5967 | 0.6267 | 0.6500 | 0.6533 | 0.6433 | 0.6667 | 0.7400 | **0.7867** |
-| drop | 0.7367 | 0.7467 | 0.7467 | 0.6733 | 0.6900 | 0.6633 | **0.8833** | 0.8500 |
-| csqa | 0.7667 | 0.7733 | 0.7367 | 0.7500 | 0.7433 | 0.7700 | **0.8633** | 0.8367 |
-| musr | 0.6409 | 0.5922 | 0.4957 | 0.4932 | 0.4867 | 0.5703 | 0.6301 | **0.6852** |
-| mmlu_redux | 0.7867 | 0.7733 | 0.6667 | 0.6000 | 0.7133 | 0.8367 | 0.8500 | **0.9467** |
-| ifeval | 0.7500 | 0.7633 | 0.6433 | 0.7167 | 0.7300 | **0.8900** | 0.8400 | 0.7900 |
-| rudas_ood | 0.0437 ¶¶ | 0.0012 | 0.0365 | **0.0733** | 0.0017 | 0.0061 | 0.0468 | 0.0000 ¶ |
-| bbh_logic | 0.9540 | 0.9013 | 0.6633 | 0.5333 | 0.5713 | 0.7700 | 0.6367 | **0.9980** |
-| math500 | 0.7467 | 0.6567 | 0.6900 | 0.4233 | 0.7133 | 0.7800 | 0.6100 | **0.8433** |
-| **macro (10 CoT datasets)** | 0.7901 | 0.7942 | 0.7339 | 0.6997 | 0.7523 | 0.7884 | 0.8493 | **0.8689** |
-| **macro (all 14)** | 0.7425 | 0.7332 | 0.6694 | 0.6245 | 0.6814 | 0.7378 | 0.7591 | **0.8086** |
-| tok/s | — | — | 15880 | 16160 | **25230** | 22480 | 9420 | 3374 |
-| mean gen length | ≈792 | ≈1282 | **482** | 510 | ≈796 | ≈1327 | ≈1931 | 801 |
-| **ans/s** | — | — | **32.9** | 31.7 | ≈31.7 | ≈16.9 | 4.9 | 4.2 |
+| dataset | Meridian-smaller | Granite-4.2-3B base | VibeThinker-3B | TwIL-LM3 | Llama-3.2-3B | LFM2-2.6B | LFM2.5-8B-A1B | Qwen3-8B | gpt-oss-120b ‡ |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| gsm8k | 0.9433 | 0.9533 | 0.9600 | 0.8733 | 0.8300 | 0.8767 | 0.9133 | 0.9567 | **0.9767** |
+| svamp | **0.9500** | 0.9200 | 0.9367 | 0.8500 | 0.8200 | 0.9000 | 0.9133 | 0.9400 | 0.9400 |
+| gsm_symbolic | 0.8267 | 0.8900 | 0.8167 | 0.7567 | 0.8067 | **0.9767** | 0.9267 | 0.8133 | 0.8467 |
+| arc_cot | 0.8600 | 0.8933 | 0.9033 | 0.8467 | 0.7967 | 0.8667 | 0.9033 | 0.9633 | **0.9667** |
+| logicbench | 0.7933 | 0.7733 | **0.8600** | 0.7167 | 0.5733 | 0.6267 | 0.7200 | 0.8567 | 0.8533 |
+| strategyqa | 0.5967 | 0.6267 | 0.6467 | 0.6500 | 0.6533 | 0.6433 | 0.6667 | 0.7400 | **0.7867** |
+| drop | 0.7367 | 0.7467 | 0.8467 | 0.7467 | 0.6733 | 0.6900 | 0.6633 | **0.8833** | 0.8500 |
+| csqa | 0.7667 | 0.7733 | 0.7367 | 0.7367 | 0.7500 | 0.7433 | 0.7700 | **0.8633** | 0.8367 |
+| musr | 0.6409 | 0.5922 | 0.5799 | 0.4957 | 0.4932 | 0.4867 | 0.5703 | 0.6301 | **0.6852** |
+| mmlu_redux | 0.7867 | 0.7733 | 0.8100 | 0.6667 | 0.6000 | 0.7133 | 0.8367 | 0.8500 | **0.9467** |
+| ifeval | 0.7500 | 0.7633 | 0.6633 | 0.6433 | 0.7167 | 0.7300 | **0.8900** | 0.8400 | 0.7900 |
+| rudas_ood | 0.0437 ¶¶ | 0.0012 | 0.0056 | 0.0365 | **0.0733** | 0.0017 | 0.0061 | 0.0468 | 0.0000 ¶ |
+| bbh_logic | 0.9540 | 0.9013 | 0.6107 | 0.6633 | 0.5333 | 0.5713 | 0.7700 | 0.6367 | **0.9980** |
+| math500 | 0.7467 | 0.6567 | 0.7900 | 0.6900 | 0.4233 | 0.7133 | 0.7800 | 0.6100 | **0.8433** |
+| **macro (10 CoT datasets)** | 0.7901 | 0.7942 | 0.8097 | 0.7339 | 0.6997 | 0.7523 | 0.7884 | 0.8493 | **0.8689** |
+| **macro (all 14)** | 0.7425 | 0.7332 | 0.7262 | 0.6694 | 0.6245 | 0.6814 | 0.7378 | 0.7591 | **0.8086** |
+| tok/s | 21169 † | 21916 † | 28010 † | 15880 | 16160 | 25230 | 22480 | 9420 | 3374 |
+| mean gen length | ≈792 | ≈1282 | ≈1789 | **482** | 510 | ≈796 | ≈1327 | ≈1931 | 801 |
+| **ans/s** | 26.7 † | 17.1 † | ≈15.7 † | **32.9** | 31.7 | ≈31.7 | ≈16.9 | 4.9 | 4.2 |
 
 ‡ MXFP4 weights, tensor-parallel 2 — quantized and multi-GPU, so not directly comparable to the
 single-GPU BF16 rows. ¶ 74% of its `rudas_ood` generations hit the length cap, so that cell is a
@@ -202,11 +223,14 @@ truncation artefact rather than a measured score; excluding the row, its 13-data
 cell is likewise a truncation artefact rather than a measurement of the model's ability.
 
 Lengths marked ≈ are derived from stored generations rather than read from the run. For
-Meridian-smaller and its base they were re-tokenized directly with the model's own tokenizer and
+Meridian-smaller, its base and VibeThinker-3B they were re-tokenized directly with each model's
+own tokenizer and
 averaged over the 14 datasets (MuSR counted once); the same method reproduces TwIL-LM3's measured
 482 to within 1%. The peer lengths marked ≈ use each model's characters-per-token ratio and are
-carried over from the TwIL-LM3 card. `tok/s` and `ans/s` are blank for Meridian-smaller and its
-base for the reason given under the Track A table.
+carried over from the TwIL-LM3 card. † `tok/s` is the same throughput figure as in Track A (see
+the engine note under that table) and `ans/s` divides it by the mean generation length shown. The
+VibeThinker-3B Track B rows come from the same external-baseline run and vLLM 0.11.2 engine as
+the other external peers.
 
 The honest summary of this table is that Meridian-smaller does not lead it. Larger models score
 higher, and gpt-oss-120b leads seven of the fourteen dataset rows. Three things are worth
@@ -216,6 +240,14 @@ in-domain, which is the point of the WiSE-FT stage. Second, the 14-dataset macro
 over the base, driven by BBH-logic, MATH-500 and MuSR. Third, it edges LFM2.5-8B-A1B on both
 macros at less than half the parameters and leads the table outright on SVAMP (0.9500).
 
+VibeThinker-3B is the stronger held-out model on the 10-dataset macro (0.8097 against 0.7901). It
+is ahead of Meridian-smaller on seven datasets — GSM8K, ARC, LogicBench, StrategyQA, DROP,
+MMLU-Redux and MATH-500 — and behind on the other seven: SVAMP, GSM-Symbolic, CSQA, MuSR, IFEval,
+BBH-logic and `rudas_ood`. Meridian-smaller's 0.0163 lead on the 14-dataset macro comes entirely
+from BBH-logic (0.9540 against 0.6107): on the other thirteen datasets it averages 0.7263 against
+VibeThinker-3B's 0.7350. VibeThinker-3B also writes much longer answers on Track B (about 1,789
+tokens against 792).
+
 Track B here was run with the chat template's thinking mode **disabled** for Meridian-smaller and
 its base (the prompt ends in an empty `<think></think>`), as it was for TwIL-LM3, whereas Track A
 uses the default thinking mode. The Track B numbers therefore describe non-reasoning behaviour;
@@ -223,18 +255,41 @@ they are not a measure of what a thinking-mode generation would score. Some Trac
 truncation-limited: MATH-500 hits the cap on 14.0% of rows, and SVAMP, GSM-Symbolic and
 MuSR-team sit slightly above the 2% cap-hit threshold (3.0%, 3.0% and 2.8%).
 
+### Against the tuned VibeThinker-3B
+
+The tables above use the public VibeThinker-3B checkpoint. The same post-training pipeline was also
+applied to it, and the best tuned configuration (WiSE-FT, λ = 0.50) is the closest same-scale
+comparison to Meridian-smaller. These values come from the family comparison tables rather than
+from a per-lane raw report, so they are shown as a summary only:
+
+| model | macro gate | macro_primary | B10 | B14 | Track A truncation |
+|---|---:|---:|---:|---:|---:|
+| Meridian-smaller | **0.554** | **0.588** | 0.790 | **0.743** | 24.2% |
+| VibeThinker-3B, WiSE-FT λ = 0.50 | 0.541 | **0.588** | **0.802** | 0.728 ◊ | 14.3% |
+
+◊ There is no B14 row for the λ = 0.50 configuration; the figure is the best tuned VibeThinker-3B
+B14 in the family table (SLERP, d = 0.5, t = 0.5).
+
+The two are effectively tied on Track A (macro gate 0.554 against 0.541, `macro_primary` equal at
+0.588, both within sampling noise at n = 200), and the tuned VibeThinker-3B is ahead on the
+10-dataset macro (0.802 against 0.790) with a lower truncation rate. Meridian-smaller's edge is the
+14-dataset macro (0.743 against 0.728), a gap that cannot be broken down per dataset from the
+summary values. Read gaps to the
+*untuned* VibeThinker-3B within one source: the family table records that checkpoint at a gate of
+0.374, from a separate run of the same weights than the 0.4118 in the lane tables above.
+
 ### Checkpoint selection
 
 MGPO checkpoints were compared on both tracks, and step 2580 was published because it has the
 best (or tied-best) held-out score rather than the best in-domain gate:
 
-| checkpoint | macro gate | macro_primary | B10 | B14 | 
-|---|---:|---:|---:|---:|---:|
+| checkpoint | macro gate | macro_primary | B10 | B14 |
+|---|---:|---:|---:|---:|
 | WiSE-FT α = 0.15 (RL initialiser) | 0.531 | 0.583 | — | — |
-| MGPO step 1200 | **0.569** | **0.598** | 0.7819 | 0.7362 | 
-| MGPO step 2000 | 0.557 | 0.581 | 0.7882 | 0.7425 | 
-| MGPO step 2200 | 0.567 | 0.583 | 0.7865 | 0.7406 | 
-| **MGPO step 2580 (published)** | 0.554 | 0.588 | **0.7901** | **0.7425** | 
+| MGPO step 1200 | **0.569** | **0.598** | 0.7819 | 0.7362 |
+| MGPO step 2000 | 0.557 | 0.581 | 0.7882 | 0.7425 |
+| MGPO step 2200 | 0.567 | 0.583 | 0.7865 | 0.7406 |
+| **MGPO step 2580 (published)** | 0.554 | 0.588 | **0.7901** | **0.7425** |
 
 Track A gate here uses the same definition as the tables above (rule induction included).
 Step 1200 leads the gate by 0.015 and `macro_primary` by 0.010, but step 2580 has the highest
@@ -340,8 +395,8 @@ hit the cap. That is far above the 2% threshold our protocol requires to mark a
 than exact. The base is worse (49.2%), and both are pessimistic because a truncated response
 scores zero regardless of reasoning quality — so the true Track A gap over the base is probably
 narrower than +0.123, and part of the improvement is shorter generations rather than better
-answers. For context, Qwen3-8B truncates 23.9% of rows, LFM2.5-8B-A1B 17.9%, LFM2-2.6B 41.3% and
-Llama-3.2-3B 10.0% under the same budget, TwIL-LM3 4.4%. Some Track B stages are also flagged
+answers. For context, Qwen3-8B truncates 23.9% of rows, VibeThinker-3B 37.1%, LFM2.5-8B-A1B 17.9%,
+LFM2-2.6B 41.3% and Llama-3.2-3B 10.0% under the same budget, TwIL-LM3 4.4%. Some Track B stages are also flagged
 unrankable for the same reason (`rudas_ood` 93.7% cap-hit, `math500` 14.0%, plus marginal excess
 on SVAMP, GSM-Symbolic and MuSR-team); the held-out core, retention, IFEval and log-likelihood
 stages are rankable.
@@ -360,14 +415,16 @@ makes no claim about those. The weak absolute areas inside the specialisation ar
 safety tuning beyond whatever the base model carries, and no instruction-following alignment
 work — IFEval is 0.7500 against the base's 0.7633.
 
-**Comparability.** For Track A, Meridian-smaller, its base, Qwen3-8B, LFM2-2.6B, LFM2.5-8B-A1B
-and Llama-3.2-3B were checked to share the same sampled-row manifest and dataset hash, seed and
-decoding; the TwIL-LM3 and gpt-oss-120b values are carried over from the TwIL-LM3 card, which
-describes the same harness. For Track B, the arms checked share the same sampled rows and
-decoding, but the serving engine differs between arms (vLLM 0.19.1 for Meridian-smaller, its base,
-Qwen3-8B and LFM2.5-8B-A1B; vLLM 0.11.2 for TwIL-LM3 and Llama-3.2-3B), and the engine version is
-part of the protocol hash. With n = 200 per lane on Track A and n = 300 per dataset on Track B,
-differences of two to three points are within sampling noise.
+**Comparability.** For Track A, Meridian-smaller, its base, VibeThinker-3B, Qwen3-8B, LFM2-2.6B,
+LFM2.5-8B-A1B and Llama-3.2-3B were checked to share the same sampled-row manifest and dataset
+hash, seed and decoding; the TwIL-LM3 and gpt-oss-120b values are carried over from the TwIL-LM3
+card, which describes the same harness. For Track B, the arms checked (including VibeThinker-3B,
+on all 18 tasks) share the same sampled rows and decoding, but the serving engine differs between
+arms (vLLM 0.19.1 for Meridian-smaller, its base, Qwen3-8B and LFM2.5-8B-A1B; vLLM 0.11.2 for
+TwIL-LM3, Llama-3.2-3B and VibeThinker-3B), and the engine version is part of the protocol hash.
+Throughput has its own, separate engine caveat (see the † note under the Track A table). With
+n = 200 per lane on Track A and n = 300 per dataset on Track B, differences of two to three points
+are within sampling noise.
 
 ## Evaluation protocol
 
@@ -378,6 +435,10 @@ differences of two to three points are within sampling noise.
   `repetition_penalty = 1.0`, chat template applied with thinking disabled, vLLM backend.
 - Both tracks use the same protocol for the model and its base, in a paired run over identical
   sampled rows.
+- Throughput: 128 prompts drawn from a fixed Track A prompt file, 512 generated tokens each with
+  EOS ignored, greedy, vLLM `gpu_memory_utilization` 0.45, `max_model_len` 4096, on an otherwise
+  idle GPU (a single H200 for the Meridian-smaller, base and VibeThinker-3B runs); the reported rate is generated tokens over decode time, excluding engine start-up and
+  compilation.
 
 `repetition_penalty = 1.0` is load-bearing. A 1.1 penalty produced apparent 20-point swings on
 Track B that were pure decoding artefact; the decoding kwargs are hashed into the protocol