JevBench · second fresh regular draw, one row added
Seven official rows from v1.6.13 retain their full 1,500-item results, ranks and categories on draw v1.6-regular-20261012. Microsoft-Decision-1 (Azure Foundry) adds a separate, unranked A2 re-run: 300 supplementary sealed plus 300 public decisions, with an equating pool of four of six self-hosted systems. It does not replace the official Microsoft measurement or enter any ranking. Historical scores keep their original dates.
Aggregate results JSON · SHA-256 e7d11ae08d35e5931000ab324c14009b968f2f70cece53c7b1ef171213dab5e4 · Previous v1.6.13 (seven rows) · Full historical board
JevBench v1.6.14 · headline
JevBench Capability Score
Capability ranking of Jev-class systems
Capability Score averages Intelligence and Calibration. metask-jev-rain-31B-trained leads the Jev-class systems with 78.9.
Jev-class means at most 2× the cost and median latency of Jev (v1.5 reference). How we choose ↘
Adjust cost / latency caps · 2× official
ScoreCCost
ScoreCap.Capability
Score$/1k$/1k decisions
- 1metask-jev-rain-31B-trained70.749.278.9$0.049*
- 2metask-jev-rain-12B-trained67.851.577.9$0.041*
- –Microsoft-Decision-1API59.262.272.2$0.018PRELIMINARY — API supplement A2 (300 sealed + 300 public), not the S1200 full set; equating pool 4 of 6
- 3attest-31B-v1.0API65.250.969.5$0.043*
- 4Torchcast Decision 4B54.261.268.7$0.020*
- 5Uniphore-4B-Decider34.264.757.1$0.015*
- 6Standard One 3B SH32.848.855.6$0.051*
- 7Krino-2B22.054.542.4$0.033*
Wide coloured bar = Capability Score (0–100). Thin bars: cost above, median latency below; shared log ratio scale 0.25× … 64× Jev (v1.5 reference), ticks at 1× and 2× (cap). Shorter is cheaper or faster. * = est. (estimated cost). # counts ranked Jev-class systems; “–” marks unranked or outside systems. Tap ⓘ or hover a row or thin bar for details.
Cost and latency are shown separately because they are nearly independent across systems (Spearman ρ = 0.02, n = 8).
- Closed API (weights not public) (1)
- Open weights · LLM decoder (7)
- green: ≤ reference
- amber: ≤ cap (2× reference)
- red: > cap
Show general-purpose LLMs and other systems outside the limits
Sorted by Capability Score, not numbered. Each row says which limit it misses, measured against Jev ($0.032 per 1,000 decisions, median 0.62 s).
Jev-class = cost per decision at most 2× Jev's (≤ $0.065 per 1,000 decisions) and median latency at most 2× Jev's (≤ 1.23 s), the adjusted p50 — the same median the speed chart plots, not the four-axis Speed score. 8 of 8 systems qualify; the other 0, including the general-purpose LLMs, are listed below the divider in the ranking. The charts below show speed and cost beside Capability Score; the official JevBench Score weighs all four axes.
Capability against cost and speed
Jev-class systems are shown by default. Bubble size follows the official JevBench Score; official rank stays unchanged. The five most capable Jev-class systems are labelled.
Capability vs cost
Upper right is best: more capable and cheaper. Cost is USD per 1,000 decisions on a log scale. The dashed line is 2× Jev’s cost.
Capability vs speed
Upper right is best: more capable and faster. Speed here is the median-latency speed — the same adjusted median (p50) latency the Jev-class limit uses — a log scale, so each 20 points is 10× faster (median under the numbers). Self-hosted systems are plotted at the est. incl. network latency (2x measured + 150 ms); their measured loopback p50 (on-GPU, no network; not directly comparable with API latency) is in the tooltip. The dashed line is 2× Jev’s median latency.
- Open weights · LLM decoder (7)
- faint = outside Jev-class
Model kind
Jev-class
Release
No family reported in this release.
No system in this release reports an exact parameter count.
No system in this release reports an alternative pricing scenario.
JevBench v1.6.14
JevBench Composite Score: 7 ranked systems
Official· four axes 0–100, equal-weight harmonic mean · Method notes ↓
Adjust weights ↓
The official order weighs Intelligence, Calibration, Speed and Cost equally. Each button re-sorts the same systems by one axis, and the column headings sort too.
Greener = stronger within its column.
8 of 8 systems, sorted by official rank, #1 first.
- Microsoft-Decision-1 (preliminary)APInewPRELIMINARY — API supplement A2 (300 sealed + 300 public), not the S1200 full set; equating pool 4 of 6Closed APIBase model: Qwen3.5-9Bsource · Base model disclosed by Microsoft; the Decision-1 endpoint is evaluated as a hosted API offering.71.5I 59C 85S 89K 62$0.018
- 1Torchcast Decision 4Bnewfine-tune · 4BBase model: undisclosed69.6I 54C 83S 94K 61est.$0.020
- 2attest-31B-v1.0APInewbase · 31BBase model: undisclosed68.0I 65C 74S 97K 51est.$0.043
- 3metask-jev-rain-12B-trainednewadapter · 12B · W4A16 INT4Base model: undisclosed67.6I 68C 88S 73K 51est.$0.041
- 4metask-jev-rain-31B-trainednewadapter · 31B · W4A16 INT4Base model: undisclosed66.9I 71C 87S 83K 49est.$0.049
- 5Uniphore-4B-Decidernewfine-tune · 4BBase model: undisclosed27.2I 34C 80S 88K 65est.$0.015
- 6Standard One 3B SHnewfine-tune · 3BBase model: undisclosed21.7I 33C 78S 87K 49est.$0.051
- 7Krino-2Bnewadapter · 2BBase model: undisclosed8.5I 22C 63S 92K 54est.$0.033
Adjust weights ↓
Weights are relative: each axis counts in proportion to its slider. The score stays a weighted harmonic mean with the low-axis gates; the gates still apply when an axis sits at 0. Only equal weights give the official JevBench Score and rank.
Score = 4 / (1/I + 1/C + 1/S + 1/K) (each 0–100; × (axis / 50)² for Intelligence, Speed or Cost below 50)
- Closed API (weights not public) (1)
- Open weights · LLM decoder (7)
- Shown, not ranked
Compare two systems
Pick any two. Radars for the score axes, use cases, chance-corrected competence per request type on the open and sealed sets, and competence per tier on each set. Further out is better on every spoke; the link keeps the pair.
- A: Torchcast Decision 4B — Open weights · LLM decoder · Score 69.6 (#1)
- B: attest-31B-v1.0 — Open weights · LLM decoder · Score 68.0 (#2)
The four score axes
Use cases (TypeSafe categories)
Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.
Low sample, n < 30 — indicative only
| Category (items) | A: Torchcast Decision 4B | B: attest-31B-v1.0 |
|---|---|---|
| LLM guardrails (27) | 48.3 n=27 | 78.8 n=27 |
| Lead generation (25) | 53.8 n=25 | 72.4 n=25 |
| Search and retrieval (24) | 90.9 n=24 | 100.0 n=24 |
| Semantic code linting (24) | 34.2 n=24 | 69.2 n=24 |
| Scientific discovery (23) | 19.8 n=23 | 59.2 n=23 |
| Graphs and knowledge graphs (22) | 25.2 n=22 | 79.3 n=22 |
| Gaming (22) | 22.2 n=22 | 31.1 n=22 |
| Advertising (22) | 19.3 n=22 | 41.0 n=22 |
| Feature extraction for predictive modeling (19) | 0.0 n=19 | 81.7 n=19 |
| Risk assessment (18) | 23.7 n=18 | 48.0 n=18 |
| Recruiting (18) | 50.4 n=18 | 57.8 n=18 |
| Moderation and trust and safety (18) | 9.8 n=18 | 92.1 n=18 |
| E-commerce marketplaces (17) | 36.9 n=17 | 59.6 n=17 |
| Demand forecasting (17) | 29.7 n=17 | 40.2 n=17 |
| Insurance claims (15) | 25.0 n=15 | 45.8 n=15 |
No published value for either system (under 15 answered items, or no per-category values): Customer support (14), Legal and compliance (10).
Not drawn on the radar: with so few items a single answer moves a category score by several points, so these values are noise-prone. These values become spokes once the item pool reaches 30 items per category.
What each category means · items per category
- Model routing — deciding which LLM or model tier should handle a prompt; classifying a request's intent, domain, difficulty or risk to pick a model. 217 items (75 open / 142 sealed)
- Financial crime — suspicious transactions, KYC, fraud alerts, entity matching for investigations. 30 items (4 open / 26 sealed)
Competence per request type, open / sealed
Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.
Competence per tier — open set
Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.
Competence per tier — sealed set
Linear signed scale: −100 at the centre, 0 on the bold middle ring, 100 at the rim.
How the categories were made
Chance-corrected competence (0 = chance, 100 = perfect) over the category's items the system saw, open and sealed pooled, tiers pooled: per request type the method's competence cell, then the mean over the request types present. Not part of the JevBench Score; never equated for API systems; compare systems within a category, not categories with each other. Values can be negative (below chance).
Family and language are authoring metadata of every item in the frozen v1.6 pool; where a uc1 item was written for a TypeSafe use case, that authoring use case is shown. This draw has no subject-topic labels yet, so the topic radar is omitted rather than filled from another draw. All non-English uc1 items are machine-authored and not native-reviewed.
- Cells with fewer than 15 items are omitted; categories under 30 items are marked low-n.
- Core v1.6 items carry no use case; the use-case view covers the use-case (uc1) items only.
All values as a table
| Spoke | A: Torchcast Decision 4B | B: attest-31B-v1.0 |
|---|---|---|
| The four score axes | ||
| Intelligence | 54.2 | 65.2 |
| Calibration | 83.1 | 73.8 |
| Speed | 93.6 | 97.1 |
| Cost | 61.2 | 50.9 |
| Use cases (TypeSafe categories) | ||
| Model routing | 69.9 | 82.4 |
| Financial crime | 56.3 | 51.1 |
| Competence per request type, open / sealed | ||
| Choice · open | 75.5 | 73.2 |
| Choice · sealed | 62.8 | 74.6 |
| Noul · open | 61.9 | 68.7 |
| Noul · sealed | 50.6 | 68.2 |
| Score · open | 42.0 | 52.9 |
| Score · sealed | 32.5 | 53.8 |
| Competence per tier — open set | ||
| Easy | 89.0 | 87.3 |
| Standard | 65.3 | 55.9 |
| Judge | 53.2 | 63.1 |
| Hard | 59.7 | 67.3 |
| Competence per tier — sealed set | ||
| Easy | 69.7 | 82.2 |
| Standard | 62.1 | 69.7 |
| Judge | 43.4 | 67.4 |
| Hard | 49.8 | 63.9 |
Six self-hosted rows and one vendor-hosted API row on the full S1200+P300 draw; the full-set API row is not equated. Category cells use stored full-draw per-item results. Sealed counts refer to self-hosted S (1,200); API rows use their original measured sealed basis.
Languages
Raw chance-corrected competence per item language (0 = chance, 100 = perfect; can be negative), from each system's own measured items: Self-hosted rows use S 1200 + P 300; API rows use their measured sealed subset plus P. Sage (A3) language cells and A2/A3 topic/use-case cells cover public P300 only. Raw and outside the Composite. Cells under 15 completed supported responses are left empty; recorded input refusals count toward coverage. A dagger (†) marks fewer than 30 completed responses. Competence retains all supported scored observations, including failures and refusals. Per-language coverage grows with the expanded uc1.1 multilingual pool, a candidate for a later release that is not part of v1.6.14. English (1,151 items) is listed first; the other 21 languages and the mixed-language group share 349 items. In 13 further groups no system reaches the 15-item reporting minimum, so they get no column (items in the pool shown): Mixed-language (14), French (14), Danish (13), Turkish (13), Swedish (11), Dutch (10), Korean (10), Chinese (10), Finnish (8), Czech (7), Ukrainian (7), Norwegian (6), Indonesian (4) — 127 items, scored like every other item.
7 systems
| 1151 | 39 | 32 | 31 | 28 | 25 | 21 | 16 | 15 | 15 | |
|---|---|---|---|---|---|---|---|---|---|---|
| Torchcast Decision 4Bself-hosted · S+P | 54 | 72 | 61 | 62 | 65† | 46† | 33† | 0† | 80† | 59† |
| attest-31B-v1.0API · S+P | 67 | 58 | 89 | 67 | 74† | 45† | 82† | 26† | 53† | 67† |
| metask-jev-rain-12B-trainedself-hosted · S+P | 65 | 72 | 85 | 74 | 81† | 45† | 59† | 25† | 49† | 77† |
| metask-jev-rain-31B-trainedself-hosted · S+P | 71 | 63 | 90 | 69 | 68† | 46† | 73† | 36† | 57† | 78† |
| Uniphore-4B-Deciderself-hosted · S+P | 32 | 25 | 70 | 71 | 52† | 29† | 0† | 18† | 74† | 15† |
| Standard One 3B SHself-hosted · S+P | 28 | 20 | 77 | 72 | 70† | 22† | 20† | 0† | 68† | 21† |
| Krino-2Bself-hosted · S+P | 16 | 43 | 45 | 12 | 21† | 22† | 0† | 0† | 19† | 12† |
Intelligence gate and Noul decisiveness
The Composite applies a soft gate below Intelligence 50 (× (I/50)²). Rows under it show their value with a "below gate" label. The decisive rate is the share of valid Noul answers with P(yes) ≥ 0.80 or ≤ 0.20 (for systems that return only a label, every valid label counts as decisive); only decisive answers count as answers. Accuracy among decisive answers is a diagnostic and does not enter any score.
| System | Intelligence | Choice | Noul | Score | Noul decisive rate | Accuracy among decisive |
|---|---|---|---|---|---|---|
| Torchcast Decision 4B | 54.2 | native | native | native | 100 % | 78 % |
| attest-31B-v1.0 | 65.2 | native | native | native | 97 % | 88 % |
| metask-jev-rain-12B-trained | 67.8 | native | native | native | 100 % | 86 % |
| metask-jev-rain-31B-trained | 70.7 | native | native | native | 100 % | 88 % |
| Uniphore-4B-Decider | 34.2below gate | native | native | native | 72 % | 82 % |
| Standard One 3B SH | 32.8below gate | native | native | native | 63 % | 84 % |
| Krino-2B | 22.0below gate | native | native | native | 83 % | 71 % |
| Microsoft-Decision-1 (Azure Foundry) — A2 re-run | 59.2 | native | native | native | not supported | — |
All data (8 systems)
Method notes · fresh native cohort · v1.6.14
Each system answered the same 1,200 freshly drawn sealed decisions and 300 public decisions on an offline GPU pod. Scores use methodology v1.6, O1S, with 1,000 bootstrap samples and seed 16.
Capability is the mean of Intelligence and Calibration within the frozen Jev-class cost and median-latency caps. Composite is secondary. Failed and refused answers remain in the full 1,500-decision denominator.
The public-versus-sealed gap reference is the median of this completed native cohort: 9.8 points. Historical scores are shown separately with their original dates and do not enter this cohort’s median or ranking. No API equating is applied.
Latency. API models are timed end to end from our measuring host, network included. Self-hosted models are timed on our own hardware over loopback (on-GPU, no network), so their measured value is not directly comparable with API latency. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms (est. incl. network (2x + 150 ms)). Tables show the measured value first, tagged “on-GPU, no network”, and the estimate second. A model author can provide a hosted endpoint; we then time it exactly like an API, network included, with no assumption (on request: info@productivity-boost.com, or note it in your submission).
Costs use each row’s documented price reference and measured usage on this draw. Subject-topic and use-case radars come from stored per-item results; only aggregates are published.
Sliders, presets and What-If change your view; they do not change the official result. Wrappers and subsidised systems are listed below the ranking.
Results SHA-256 e7d11ae08d35e5931000ab324c14009b968f2f70cece53c7b1ef171213dab5e4 · categories SHA-256 19742a99ca5e60d769d62791834ac50b9f76c39902948fd8c90e1f98034e45f3 · scoring source SHA-256 7abbe2591f8f039b28b4679ced2669f5572cb00c57bd12638af52bd2a70faba1.
Read these scores against this release only
Each JevBench release since v1.6 scores on its own fresh seeded draw, and the gap penalty is measured against the field median of that draw. This release’s median is 9.839526770355711, so its penalty threshold sits at 17.84; the live v1.6.1 board’s median is 2.573387642438244. That makes these numbers not directly comparable with the headline board or with earlier draws. It was frozen at v1.6.10 over its four members, so it is a small-field figure.
The inherited full-set rows keep O1S, 1,000 bootstrap samples, seed 16 and the frozen v1.6.10 field median. The A2 measurement uses the pinned supplementary scorer with the lead-authorized four-system preliminary equating pool; its broad confidence intervals include offset uncertainty. Seven HTTP 422 responses count as wrong. Pool re-runs used H200 versus H100 parents. API latency was measured from Sandy (Germany); Azure Foundry tariff USD 0.042/M input and zero output, with unknown invoice, warmup and refusal charges. Category comparisons cover the seven official rows only.
Refusals, errors and context-bound failures stay in the full 1,500 denominator. Every row kept its own package, model, runtime and context settings; nothing was re-run for presentation.
NOT DIRECTLY COMPARABLE WITH THE LIVE v1.6.1 BOARD OR WITH THE EARLIER FRESH DRAWS. This release is a different draw with its own small-field median: G_med = 9.839526770355711, frozen from the four v1.6.10 members (attest-31B-v1.0 does not move it), against 2.573387642438244 on the live field. G_med_flag_gt10 is false. The gap penalty is max(0, 1 - max(0, (gap - G_med) - 8)/100), so this field's threshold is G_med + 8 = 17.84. On the merged live board this draw is placed with offset 0 while its anchor runs are pending, exactly like the earlier fresh draws. Measured, not estimated: Applying the same penalty formula to the stored base and gap at the live field's median 2.573387642438244, attest-31B-v1.0 Intelligence 65.238 would be 65.238 (gap -0.628); the original four rows are as stated in the v1.6.10 notes: standard_one_3b_sh Intelligence 32.786 would be 31.294 and metask_jev_rain_12b_trained 67.826 would be 67.587; krino-2b and metask_jev_rain_31b_trained are unchanged.
Cost: no item is excluded from the cost basis on this draw. Every row is priced from its own measured usage block at a clearly-labelled ESTIMATE rate fixed before scoring (r99, r101, r103 and r106 pre-registrations); none is a bill any of these systems issued. Every row is under the published v1.6 cost ceiling of USD 0.0646 per 1,000 decisions. The cost normalizer C0 is USD 0.001 per 1,000 decisions with 30 cost points per decade.
No subject-topic radar in this release: this draw's items have not been topic-labelled yet, and the shared topic labels of earlier draws do not cover its items. Families, use cases and languages are authoring metadata and are shown.
v1.6.13 ADDS ONE ROW, attest-31B-v1.0, to the six of v1.6.12 on the same draw, under the same rule as the earlier same-draw additions: the field median stays the frozen v1.6.10 value, so all six earlier rows keep identical point scores, axes and capability; ranks and bootstrap confidence intervals are recomputed for the seven-row roster, and the existing raw_on_A and API-basis offset diagnostics are unchanged because the new row is full-set API, not a new self-hosted equating-pool member. No API offset enters its score.
krino-2b: All 1,500 items answered HTTP 200 with native probabilities. The 15 longest items of the draw (the same 15 that standard_one_3b_sh refuses) ran with the state truncated to 65,536 tokens, as the author's submitted serve command documents (KEV_TRUNCATE_STATES=1); every other item is at most 1,802 input tokens. Reference kernels were used (KEV_FUSED=0) because the documented pip install has no flash-linear-attention. Kev reports usage.output_tokens as the token count of the serialized answer JSON although nothing is generated (165,276 over 1,500 decisions); it is priced as reported at the frozen DeepInfra Qwen/Qwen3.5-2B reference, USD 0.02/0.10 per 1M. Measured on an evaluator-owned Lium H100 pod in an offline container (network none).
metask_jev_rain_12b_trained: All 1,500 items answered HTTP 200. The author's service runs the authors' own int4 (W4A16) quantization of Google's Gemma 4 12B QAT checkpoint with their own rank-16 LoRA applied unmerged by vLLM; per its model card it takes a second pass on part of the items, so it reports generated output tokens (84,238 over 1,500 decisions) and its latency has a long tail (raw p95 8.1 s against a raw p50 of 0.13 s). Measured on an evaluator-owned Lium H100 80GB pod (IP geolocated to Nizhny Novgorod, Russia) in an offline container (network none). google/gemma-4-12b-it is not listed on OpenRouter, so the cost uses the nearest listed sibling, google/gemma-3-12b-it at USD 0.05/0.15 per 1M, times its own usage. It is a different system from the earlier metask_jev_rain_12b row.
metask_jev_rain_31b_trained: All 1,500 items answered HTTP 200. The author's service runs Google's unchanged W4A16 Gemma 4 31B QAT checkpoint with the authors' own rank-16 LoRA applied unmerged by vLLM, at max_model_len 6,656 (per the measuring operator's handoff): longer inputs are shortened by the service's own documented extraction rule rather than refused, so the longest items were answered on a shortened state. The authors tested it on H200; it was measured on an evaluator-owned Lium H100 80GB pod (IP geolocated to Nizhny Novgorod, Russia) in an offline container (network none) with unchanged settings. The service reports 1 output token per decision; cost is the OpenRouter google/gemma-4-31b-it list price (USD 0.09/0.34 per 1M) times its own usage.
standard_one_3b_sh: 15 of 1,500 items returned HTTP 422: they are the 15 longest items of the draw and exceed the model's documented 32,768-token context, so the author's server refuses them rather than truncating; scored wrong per method. Measured on an evaluator-owned Lium H100 pod in an offline container (network none). Priced at the M2 base-model floor, OpenRouter mistralai/ministral-3b-2512 at USD 0.10/0.10 per 1M, because the author's own list price (USD 0.038 per 1M input) is below that floor; the schema head generates no text, so output tokens are 0.
torchcast-decision-4b-v1.1: evaluator-owned Lium GPU pod (H100), offline read-only container (--network none); author's torchcast_decision server unchanged at tag v1.1.0 with the bundled decision_config.json (bf16, one forward pass reading option letters, calibration temperature 0.9, yes/no commit floor 0.801) on the documented stack torch 2.9.1+cu128 / transformers 5.8.1 / flash-linear-attention 0.5.2 / causal-conv1d 1.7.0 (fast kernels active, CUDA graphs on) Paid fast-lane order 529be5d5 (public). Qwen3.5-4B fine-tune served by the author's own torchcast_decision package. Runtime: official pytorch/pytorch:2.9.1-cuda12.8-cudnn9-devel image (digest-pinned) plus an offline hash-locked overlay of the documented packages (Python 3.11 instead of the documented 3.12). decision_config.json sets a disclosed yes/no commit floor of 0.801: a yes/no probability with max(P, 1-P) below 0.801 is moved to 0.801 on its own side (answer unchanged; same mechanism as h2o-lightning-4b). Prompts over 16,384 tokens keep the head and the tail of the state (documented). Output tokens are 0 (no generation). CC BY-NC 4.0 (weights); base Qwen/Qwen3.5-4B Apache-2.0; serving package Apache-2.0. ESTIMATE, v1.5-M2 base-model reference rule: self-hosted open weights, no bookable tariff. Base model Qwen/Qwen3.5-4B (model card base_model). OpenRouter does not list Qwen3.5-4B (host snapshot source/public-docs/OPENROUTER-MODELS.json, fetched 2026-10-12T08:15:26Z, read 2026-10-12). Binding precedent INTERPRETATION-1 I-2/I-3: DeepInfra Qwen/Qwen3.5-4B, USD 0.03/M input, USD 0.15/M output, frozen 25 Sep 2026 price cut-off snapshot (listing later deprecated; frozen reference kept). Logit readout, no generation: server reports 0 output tokens. Isolation is container-level (--network none); no separate host-level netcut probe was taken. The input tar was on the rented pod host disk during the run, and the pod was destroyed after measurement (CUSTODY-RECEIPT.json).
uniphore-4b-decider: 15 of 1,500 inputs exceed the author's documented 16,384-token wrapper limit; the author's code refuses them (HTTP 422, no truncation) and they are scored as wrong per method, so 1,485 items were answered. The native readout is one forward pass, so output tokens are 0 and the cost is the input tokens at the frozen DeepInfra Qwen/Qwen3.5-4B reference, USD 0.03/0.15 per 1M (that listing is flagged deprecated in the frozen snapshot but remains the reference). Measured on an evaluator-owned Lium H100 pod in an offline container (network none), in the vllm/vllm-openai v0.31.0 image with transformers 5.17 and reference PyTorch kernels in place of the documented flash-linear-attention / causal-conv1d kernels, so its latency was measured on the fallback path. The author discloses that the public JevBench set was inspected repeatedly during development.
attest-31b-v1.0: Paid public fast-lane order 07d13315. Vendor-hosted attest.engineering /v1/systemone API; latency measured from Sandy (Germany), using the official per-item latency subset without API adjustment. Full S1200+P300 draw, not API A300 supplement; no API equating. 15 of 1,500 requests returned HTTP 500 after 8 retry invocations and are scored as wrong in the full denominator. Public model documentation declares frozen google/gemma-4-31B-it @842da3794eaa0b77d5f08bae87a17459d91ff475, unmodified bf16, no adapter or quantization; single forward pass reading option-label logits, no generated tokens, one fitted temperature (value not documented). Hosting implementation and deployed weights were not independently inspected. Hosted API access is not an open-weights release by the vendor; base weights are under the Gemma 4 licence. ESTIMATE, v1.5-M2 base-model reference rule: vendor-hosted API, customer list price USD 0.048/M input and USD 0.00/M output effective 2026-10-11 has been in force fewer than 30 days. Base model google/gemma-4-31b-it reference on OpenRouter USD 0.09/M input / USD 0.34/M output, snapshot fetched 2026-10-12T17:55:05Z (SHA-256 81eae2f3f94bf299c8d4499260f842c357b7a869ce82dfa1741ad6a89cf57014), same base-model estimate precedent as metask_jev_rain_31b_trained. Applied to measured response usage; failed HTTP 500 usage/billing may be unreported. No output generation; reported output tokens 0. Estimated, not a vendor bill.
Microsoft A2 re-run: 300 sealed supplementary A2 + 300 public items, not the full S1200 set. Unranked measurement; the official Microsoft row is retained on the merged boards. Equating pool four of six: Krino, Standard One, Uniphore, MeTask 12B. MeTask 31B and Torchcast were not completed after four unsuccessful allocation attempts (price/creation/controller failures); no inferred scores for missing pool systems. Seven HTTP 422 responses are wrong in the 600-item denominator. Pool re-runs used H200 in Helsinki, versus H100 parents; wide equating confidence intervals are reported. Latency from Sandy (Germany). Cost is the Azure Foundry tariff USD 0.042/M input, USD 0/M output; invoice/warmup/refusal charges unknown. No category cells or ranks are assigned to the A2 measurement.
Historical v1.3.0 board: weightings, per-task grid, topic radars and held-out diagnostics
Open this section to load the earlier public-only board and diagnostics.