Compare all local models with matched queue-excluded latency and korjev naming

README.md

@@ -30,51 +30,31 @@ configs:
 **25,418 Korean text-only evaluation items from 9 public datasets and 22 subsets**, normalized to finite-choice decisions such as `A`, `B`, or `C`.
 
 <!-- trained-comparison:start -->
-## Matched trained-model comparison · 8 October 2026
+## Matched model comparison · 8 October 2026
 
-Backbone, v4 and v7 KL epoch1.0 were evaluated on the same single RTX PRO 6000 Blackwell Server96GB, using matched FP8 exports, original LM-head logits and candidate-only masked softmax. The v7 checkpoint was nominated from private validation before public testing.
+All nine local configurations use the same GPU type, semantic evaluation sets, warmup and load-search protocol. **korjev-26b4a-v1** is the previously selected KL epoch1.0 adapter, with unchanged weights. Gemma 4 26B-A4B has 3.8B active parameters; the name uses the rounded 4a label.
 
-| Model | Korean25418 | Δ vs backbone | MMLU-Pro12032 | Δ vs backbone |
-| --- | ---: | ---: | ---: | ---: |
-| Gemma 4 26B-A4B · matched FP8 | 90.44% | — | 45.50% | — |
-| Korean JEV v4 | 92.12% | +1.68pp | 65.08% | +19.58pp |
-| Korean JEV v7 · KL 1.0 | 92.12% | +1.68pp | 65.67% | +20.17pp |
+| Model | Precision | C* | Decisions/s ↑ | Processing p50 ms ↓ | Korean ↑ | English ↑ | Order agreement ↑ |
+|---|---|---:|---:|---:|---:|---:|---:|
+| Gemma 4 31B | FP8 | 64 | 45.2 | 528.7 | 94.40% | 57.64% | 76.89% |
+| Qwen3.8 27B | FP8 | 64 | 52.9 | 384.8 | 89.02% | 60.40% | 74.11% |
+| Gemma 4 26B-A4B | FP8 | 128 | 200.4 | 114.6 | 90.37% | 45.36% | 71.09% |
+| Gemma 4 12B | FP8 | 64 | 114.4 | 208.1 | 89.58% | 39.69% | 70.77% |
+| Cygnet · Gemma 12B | BF16 | 64 | 49.9 | 305.3 | 89.18% | 54.84% | 73.47% |
+| Mica v0.1 4B | BF16 | 64 | 152.0 | 106.8 | 81.47% | 52.96% | 77.68% |
+| K-Decision 4B | BF16 | 256 | 102.7 | 596.7 | 85.41% | 48.94% | 73.47% |
+| Qwen3.5 4B | BF16 | 64 | 185.6 | 108.0 | 78.77% | 46.24% | 63.62% |
+| korjev-26b4a-v1 | FP8 | 64 | 203.6 | 112.7 | 92.16% | 65.64% | 89.36% |
 
-![Five-axis higher-is-better model overview](results/20261008_v7/model_dashboard.png)
+![Nine-model higher-is-better overview](results/20261008_all_models/model_overview.png)
 
-![Updated Korean accuracy, throughput and latency chart](results/figures/pareto_combined.png)
-
-![MMLU-Pro accuracy comparison](results/20261008_v7/english_accuracy.png)
-
-English uses the frozen **zero-shot direct-choice** protocol, without generated reasoning; it is not the official few-shot CoT leaderboard protocol. The backbone-selected concurrency is reused for all three matched models. Their chart throughput/p50 coordinates are shared arithmetic averages of the three full-pass measurements; individual measurements are preserved in the result JSON. p50 includes server queues. The Korean chart retains the original eight local configurations and hosted JEV reference, and adds the three matched runs. The historical RedHat FP8 baseline and new matched export are labeled separately. No public results were used to create training data or select checkpoints. Evaluation data and labels are unchanged. [Matched results and timing](results/20261008_v7/summary.json).
-
-The model overview shows five higher-is-better axes: throughput index, inverse p50 index, Korean accuracy, English accuracy and semantic answer agreement after reversing choices. The three matched models share **210.8 decisions/s and 1209 ms p50 at concurrency 256** on the Korean workload. Speed indices are therefore identical. Order agreement uses 1,259 private development items and is an aggregate diagnostic, not an independent public test. The historical RedHat point used concurrency 64; differences in loaded p50 across these settings do not isolate model speed.
+One RTX PRO 6000 Blackwell Server 96GB per pod. Each model is warmed up and its throughput-maximizing tested concurrency C* selected independently. **Processing p50 excludes the initial server queue** and is measured at C*. Korean: 25,418 items; English: 12,032 zero-shot direct-choice MMLU-Pro items; order agreement: original vs reverse on 1,259 private development items. Native prompts/readouts are preserved. English scores are not official few-shot CoT leaderboard results. [Measurement details and limitations](detail.md#matched-model-measurement--8-october-2026) · [Full metrics](results/20261008_all_models/summary.json).
 
 <!-- trained-comparison:end -->
 
-## Original baseline results · 4 October 2026
-
-The dashed frontiers cover the eight local configurations. **JEV 1.13 is a hosted API reference**, evaluated through OpenRouter on the same 25,418 items; its provider hardware is undisclosed. p50 is measured at throughput-selected concurrency, and JEV latency includes network and retries.
-
-All nine configurations cover **25,418 items each** (2026-10-04). The eight local configurations use **one RTX PRO 6000 Blackwell 96GB GPU** per run. Throughput and latency are from the full evaluation after separate local/API concurrency searches.
-
-| Model / recipe | Precision / serving | Accuracy ↑ | ECE ↓ | Decisions/s ↑ | p50 ms ↓ |
-| --- | --- | --- | --- | --- | --- |
-| [Gemma 4 31B](https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-dynamic) | FP8 | 94.38% | 5.42% | 47.1 | 1,299 |
-| [Gemma 4 26B-A4B](https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-FP8-dynamic) | FP8 | 90.41% | 8.94% | 215.4 | 289 |
-| [Gemma 4 12B](https://huggingface.co/RedHatAI/gemma-4-12B-it-FP8-Dynamic) | FP8 | 89.63% | 9.14% | 117.2 | 529 |
-| [Cygnet (Gemma 4 12B recipe)](https://huggingface.co/google/gemma-4-12B-it) | BF16 | 89.15% | 2.03% | 49.6 | 1,258 |
-| [Qwen3.8 27B](https://huggingface.co/unsloth/Qwen3.8-27B-FP8) | FP8 | 88.98% | 3.31% | 52.1 | 1,187 |
-| [JEV 1.13](results/README.md#jev-113-hosted-api-reference) | Hosted API / undisclosed | 88.97% | 1.19% | 69.2 | 428 |
-| [K-Decision 4B*](https://huggingface.co/ThakiCloud/kd-4b-ko-v0) | BF16 / FP32 head | 85.39% | 9.52% | 92.7 | 2,733 |
-| [Mica v0.1 4B](https://huggingface.co/sky7350/Mica-v0.1-4B) | BF16 | 81.47% | 2.55% | 145.0 | 455 |
-| [Qwen3.5 4B](https://huggingface.co/Qwen/Qwen3.5-4B) | BF16 | 78.74% | 7.96% | 157.8 | 191 |
-
-These compare the listed serving configurations: precision, prompts and probability readouts differ. Seven systems use SGLang; K-Decision uses native Kev serving, and JEV uses OpenRouter's native Choice API. ECE uses ten confidence bins. Local models retain published temperatures; JEV uses its returned probabilities. No model, prompt or temperature was fitted on this benchmark. A matched BF16-versus-FP8 ablation was not run.
-
-*K-Decision uses the recommended adapter with the public Qwen3.5-4B-Base; identity with its private training-base revision is unverified. Its disclosed KLUE/KoBEST training overlap is examined in the results details.
+## Earlier results
 
-**[Evaluation methods, source-level results, Cygnet ablation and API cost](results/README.md)** · [All nine results (JSON)](results/comparison.json)
+[4 October local-model and hosted JEV API results](results/README.md#original-baseline-results--4-october-2026) use queue-inclusive timing and are archived separately from the processing p50 comparison above.
 
 ## Load and evaluate
 

detail.md

@@ -267,3 +267,17 @@ The transformations are source selection, reviewed exclusion and deduplication,
   note = {Version 1; curated Korean finite-choice evaluation data with LLM-assisted audit}
 }
 ```
+
+<!-- matched-model-measurement:start -->
+## Matched model measurement · 8 October 2026
+
+Nine local configurations use one RTX PRO 6000 Blackwell Server Edition 96GB GPU per pod and the same semantic evaluation items. For each model, concurrency 1, 8, 32, 64, 128, 256 and 512 is tested on the same 2,048 Korean performance items. Each setting receives 256 warmup requests; the prefix KV cache is cleared before the timed pass, with normal within-pass prefix reuse. The fastest two coarse settings are repeated. C* is the best observed setting by median decisions/s across its trials, selected without using gold labels. The table reports throughput and processing p50 from those same trials; raw trial counts are retained. This is a bounded concurrency search, not proof of a global optimum.
+
+For SGLang 0.5.21, request processing latency is the interval from `forward_entry_time` to `prefill_finished_time`. Initial scheduler queue, HTTP and tokenization are excluded. Requests spanning several prefill chunks retain the wall-clock interval through the final chunk. These are serving-model intervals, not pure CUDA kernel times. Existing scheduler metadata is exposed through `/v1/score` without changing weights or readout. K-Decision uses its official runtime’s CUDA-synchronized model batch interval, excluding queue and without dividing by batch size. Its pointer head requires a different native runtime. HTTP round-trip and available scheduler queue timings are recorded separately.
+
+All 25,418 Korean and 12,032 English MMLU-Pro test items are evaluated at the selected Korean performance concurrency. English uses zero-shot direct-choice logits without generated reasoning, so it is not comparable to official few-shot CoT leaderboard scores. Order agreement is measured on the same 1,259 private development items in original and exact reverse order, with semantic answer alignment and concurrency 16 for every model. Agreement includes consistent wrong answers; accuracy, JS and both-order-wrong counts are retained in the private report. This development set has been used adaptively and is not an independent held-out stability test. Public scores were not used for training or checkpoint selection.
+
+Native specialized prompts and readouts are retained. Cygnet uses its fixed upstream temperature, top-k and letter-alias readout; Mica uses its fixed upstream temperature and previously validated text-only RoPE configuration. Mica also needs its released `model.language_model.` tensor-name prefix normalized before the SGLang text-body loader; tensor values are unchanged. Unused Gemma 12B vision/audio configurations are disabled for text-only serving. K-Decision uses BF16 backbone / FP32 pointer head. Its reported training sources include KLUE and KoBEST, overlapping benchmark sources; the original private training base revision cannot be independently identified, so the declared public base is pinned. No temperature is fitted on these evaluation sets. Hosted JEV API results remain separate because its GPU configuration and queue-excluded timing cannot be verified with the same method.
+
+**korjev-26b4a-v1** identifies the already selected KL epoch 1.0 adapter. Its weights are unchanged. The [Google backbone card](https://huggingface.co/google/gemma-4-26B-A4B-it) reports 3.8B active parameters, rounded to 4a in the name. Backbone and korjev are exported from the same pristine BF16 revision with the same FP8_DYNAMIC W8A8 recipe; LM head, embedding, router and normalization remain protected. Pinned revisions, concurrency trials and dataset hashes are in [the result summary](results/20261008_all_models/summary.json). Dataset items and labels are unchanged.
+<!-- matched-model-measurement:end -->

publication_manifest.json

@@ -14,8 +14,8 @@
       "bytes": 1274
     },
     "README.md": {
-      "sha256": "0ec8d21fb589090ac754e6c798627378b65a00e68f9f7bef9a8ca276d1e51261",
-      "bytes": 7152
+      "sha256": "a9eacc4e036fa776f6c56200a0ab51b9c7b09f964f4069a1ade5425dddba5ad9",
+      "bytes": 4453
     },
     "audit/adapter_pilot/development_references.jsonl": {
       "sha256": "c02af78b627d4b14135d9fbcbe244584f58a0a2b0c85df944e9c5c332ad3cce4",
@@ -138,12 +138,12 @@
       "bytes": 3199295
     },
     "detail.md": {
-      "sha256": "2a8896fc3c8a01d6a21976e21eb9c692ccf854af1ee38d9f14b97036fabd2188",
-      "bytes": 25391
+      "sha256": "9d9e2107a2de9a3e523351b60aa0b3ada8627af70d1551f1f676c982d756e95b",
+      "bytes": 29123
     },
     "results/README.md": {
-      "sha256": "9f21b44c5c7f1bd0779270cd08b533dcbb68a1d34c0199d9ef3251d48e4947d9",
-      "bytes": 14917
+      "sha256": "0c97c76930cebc24a3a7b26f84b5ab1356685b7932d4312212c0c96e64ae9e00",
+      "bytes": 17518
     },
     "results/summary.json": {
       "sha256": "06a0551df0f60801a62d6e8f31e8d08003487e3c38abb1aec4360eb1fb77e875",
@@ -228,6 +228,26 @@
     "results/20261008_v7/model_dashboard_scores.json": {
       "sha256": "a20239bd71bfb533dea802541f64d3ef17829b6c923448339ee4de42b84a6c87",
       "bytes": 1487
+    },
+    "results/20261008_all_models/summary.json": {
+      "sha256": "497f92476b08342544045efad2c5a2cc24558837a5571030cfdc25d593a9baa3",
+      "bytes": 111611
+    },
+    "results/20261008_all_models/plotted_metrics.json": {
+      "sha256": "64c88c26381dbdddfc887d9a2bf620a75a17c692289dd28b456ec68bb19626a3",
+      "bytes": 5000
+    },
+    "results/20261008_all_models/model_overview.svg": {
+      "sha256": "d227028c9dee4bf2e60223f713b8e7700a9007c137a8155abb1b92cbf7049515",
+      "bytes": 179166
+    },
+    "results/20261008_all_models/model_overview.png": {
+      "sha256": "f37f2de2388368e1387034b73dd57d13dddbff4b1d22122c60f472fbd58470a4",
+      "bytes": 1099127
+    },
+    "results/20261008_all_models/model_overview.pdf": {
+      "sha256": "f2c4e22a10fb62f6137c4727b070d4f602776a96919dfa6376569fa2c2c22b5c",
+      "bytes": 68790
     }
   },
   "documentation_update": {
@@ -242,5 +262,5 @@
     "local_model_count": 8,
     "hosted_reference_count": 1
   },
-  "latest_chart_comparison": "results/20261008_v7/summary.json"
+  "latest_chart_comparison": "results/20261008_all_models/summary.json"
 }

results/20261008_all_models/plotted_metrics.json

@@ -0,0 +1,173 @@
+{
+  "models": {
+    "gemma4_31b": {
+      "scores": [
+        22.20343554943966,
+        20.208082827247623,
+        94.39767094185223,
+        57.63796542553191,
+        76.88641779189834
+      ],
+      "operating": {
+        "concurrency": 64,
+        "runs": 2,
+        "median_decisions_per_second": 45.198894817718156,
+        "median_service_p50_ms": 528.666079044342
+      },
+      "korean_accuracy": 0.9439767094185223,
+      "english_accuracy": 0.5763796542553191,
+      "order_agreement": 0.7688641779189833
+    },
+    "qwen38_27b": {
+      "scores": [
+        26.009901118690326,
+        27.765823761341856,
+        89.01565819497993,
+        60.397273936170215,
+        74.10643367752185
+      ],
+      "operating": {
+        "concurrency": 64,
+        "runs": 2,
+        "median_decisions_per_second": 52.947607241465946,
+        "median_service_p50_ms": 384.7653865814209
+      },
+      "korean_accuracy": 0.8901565819497994,
+      "english_accuracy": 0.6039727393617021,
+      "order_agreement": 0.7410643367752184
+    },
+    "gemma4_26b_a4b": {
+      "scores": [
+        98.45684565568783,
+        93.1964338824841,
+        90.37296404123062,
+        45.3623670212766,
+        71.0881652104845
+      ],
+      "operating": {
+        "concurrency": 128,
+        "runs": 2,
+        "median_decisions_per_second": 200.42576748840344,
+        "median_service_p50_ms": 114.63236808776855
+      },
+      "korean_accuracy": 0.9037296404123062,
+      "english_accuracy": 0.45362367021276595,
+      "order_agreement": 0.7108816521048451
+    },
+    "gemma4_12b": {
+      "scores": [
+        56.191598121756584,
+        51.340922527658876,
+        89.57825163270124,
+        39.694148936170215,
+        70.77045274027004
+      ],
+      "operating": {
+        "concurrency": 64,
+        "runs": 2,
+        "median_decisions_per_second": 114.38761931636569,
+        "median_service_p50_ms": 208.0860137939453
+      },
+      "korean_accuracy": 0.8957825163270123,
+      "english_accuracy": 0.39694148936170215,
+      "order_agreement": 0.7077045274027005
+    },
+    "cygnet": {
+      "scores": [
+        24.522519045649084,
+        34.98817242376884,
+        89.18089542843654,
+        54.837101063829785,
+        73.47100873709293
+      ],
+      "operating": {
+        "concurrency": 64,
+        "runs": 1,
+        "median_decisions_per_second": 49.919786356564764,
+        "median_service_p50_ms": 305.34112453460693
+      },
+      "korean_accuracy": 0.8918089542843655,
+      "english_accuracy": 0.5483710106382979,
+      "order_agreement": 0.7347100873709294
+    },
+    "mica": {
+      "scores": [
+        74.65427823335727,
+        100.0,
+        81.47375875363916,
+        52.95877659574468,
+        77.68069896743447
+      ],
+      "operating": {
+        "concurrency": 64,
+        "runs": 2,
+        "median_decisions_per_second": 151.97156593396363,
+        "median_service_p50_ms": 106.83327913284302
+      },
+      "korean_accuracy": 0.8147375875363916,
+      "english_accuracy": 0.5295877659574468,
+      "order_agreement": 0.7768069896743447
+    },
+    "kdecision": {
+      "scores": [
+        50.43527381976249,
+        17.904018624575667,
+        85.41191281768825,
+        48.93617021276596,
+        73.47100873709293
+      ],
+      "operating": {
+        "concurrency": 256,
+        "runs": 2,
+        "median_decisions_per_second": 102.66963557987714,
+        "median_service_p50_ms": 596.7
+      },
+      "korean_accuracy": 0.8541191281768825,
+      "english_accuracy": 0.48936170212765956,
+      "order_agreement": 0.7347100873709294
+    },
+    "qwen35_4b": {
+      "scores": [
+        91.1740369423918,
+        98.96243119729543,
+        78.7748839405146,
+        46.243351063829785,
+        63.6219221604448
+      ],
+      "operating": {
+        "concurrency": 64,
+        "runs": 2,
+        "median_decisions_per_second": 185.6003633622327,
+        "median_service_p50_ms": 107.95336961746216
+      },
+      "korean_accuracy": 0.787748839405146,
+      "english_accuracy": 0.46243351063829785,
+      "order_agreement": 0.636219221604448
+    },
+    "korjev-26b4a-v1": {
+      "scores": [
+        100.0,
+        94.83467893199378,
+        92.16303407034385,
+        65.64162234042553,
+        89.35663224781572
+      ],
+      "operating": {
+        "concurrency": 64,
+        "runs": 2,
+        "median_decisions_per_second": 203.56712238101738,
+        "median_service_p50_ms": 112.65212297439575
+      },
+      "korean_accuracy": 0.9216303407034385,
+      "english_accuracy": 0.6564162234042553,
+      "order_agreement": 0.8935663224781573
+    }
+  },
+  "definitions": {
+    "throughput": "100 × selected peak requests/s / maximum across nine configurations",
+    "latency": "100 × minimum processing p50 / processing p50 at selected throughput concurrency",
+    "order_stability": "100 × semantic answer agreement, original vs reverse on existing private1259",
+    "no_composite_score": true,
+    "english": "Zero-shot direct MMLU-Pro12032, not official few-shot CoT."
+  }
+}

results/README.md

@@ -148,3 +148,29 @@ Recomputed from the same saved BF16 Cygnet outputs, holding prompts, token aggre
 The hosted snapshot is `typesafe/jev-1.13-20260917`; this is a response-reported serving version, not a public checkpoint revision. Its weights and training procedure are not available for inspection.
 
 Full numeric metrics, source/subset breakdowns and request/prediction hashes are available in [comparison.json](comparison.json). [summary.json](summary.json) retains the eight local runs' shared hardware and runtime metadata; [jev_1_13.json](jev_1_13.json) records the hosted API run and costs. Results reflect these particular adapters, checkpoints and service conditions. The gold labels are inherited and machine-assisted reviewed; public-source exposure and selection effects remain relevant as described in [dataset limitations](../detail.md#limitations).
+
+<!-- baseline-20261004:start -->
+## Original baseline results · 4 October 2026
+
+The dashed frontiers cover the eight local configurations. **JEV 1.13 is a hosted API reference**, evaluated through OpenRouter on the same 25,418 items; its provider hardware is undisclosed. p50 is measured at throughput-selected concurrency, and JEV latency includes network and retries.
+
+All nine configurations cover **25,418 items each** (2026-10-04). The eight local configurations use **one RTX PRO 6000 Blackwell 96GB GPU** per run. Throughput and latency are from the full evaluation after separate local/API concurrency searches.
+
+| Model / recipe | Precision / serving | Accuracy ↑ | ECE ↓ | Decisions/s ↑ | p50 ms ↓ |
+| --- | --- | --- | --- | --- | --- |
+| [Gemma 4 31B](https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-dynamic) | FP8 | 94.38% | 5.42% | 47.1 | 1,299 |
+| [Gemma 4 26B-A4B](https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-FP8-dynamic) | FP8 | 90.41% | 8.94% | 215.4 | 289 |
+| [Gemma 4 12B](https://huggingface.co/RedHatAI/gemma-4-12B-it-FP8-Dynamic) | FP8 | 89.63% | 9.14% | 117.2 | 529 |
+| [Cygnet (Gemma 4 12B recipe)](https://huggingface.co/google/gemma-4-12B-it) | BF16 | 89.15% | 2.03% | 49.6 | 1,258 |
+| [Qwen3.8 27B](https://huggingface.co/unsloth/Qwen3.8-27B-FP8) | FP8 | 88.98% | 3.31% | 52.1 | 1,187 |
+| [JEV 1.13](results/README.md#jev-113-hosted-api-reference) | Hosted API / undisclosed | 88.97% | 1.19% | 69.2 | 428 |
+| [K-Decision 4B*](https://huggingface.co/ThakiCloud/kd-4b-ko-v0) | BF16 / FP32 head | 85.39% | 9.52% | 92.7 | 2,733 |
+| [Mica v0.1 4B](https://huggingface.co/sky7350/Mica-v0.1-4B) | BF16 | 81.47% | 2.55% | 145.0 | 455 |
+| [Qwen3.5 4B](https://huggingface.co/Qwen/Qwen3.5-4B) | BF16 | 78.74% | 7.96% | 157.8 | 191 |
+
+These compare the listed serving configurations: precision, prompts and probability readouts differ. Seven systems use SGLang; K-Decision uses native Kev serving, and JEV uses OpenRouter's native Choice API. ECE uses ten confidence bins. Local models retain published temperatures; JEV uses its returned probabilities. No model, prompt or temperature was fitted on this benchmark. A matched BF16-versus-FP8 ablation was not run.
+
+*K-Decision uses the recommended adapter with the public Qwen3.5-4B-Base; identity with its private training-base revision is unverified. Its disclosed KLUE/KoBEST training overlap is examined in the results details.
+
+**[Evaluation methods, source-level results, Cygnet ablation and API cost](results/README.md)** · [All nine results (JSON)](results/comparison.json)
+<!-- baseline-20261004:end -->