JevBench v1.4.1 — frozen results

• † JevK5 v0.2.0: Author says no JevBench items or outputs were used for training, tuning, or selection; public results are reported. Scan found only one generic instruction shared by 8 public hard items; unreleased teacher/replay corpora were unavailable.
• † Hopper: Author discloses heavy public-benchmark-directed development (26 model/prompt configs, 20+ calibration-map variants observed against the public half); scan of 17 released files vs 231 public tasks found 0 state/instruction matches; training corpus not released so overlap not independently verifiable.
• † Winnow-12B Q8: The submitted Q8_0 GGUF ran through the pinned author's TypeSafe-compatible /v1/systemone server with 8,192 context, four resident decision branches, Q8 KV, and full GPU offload. The private training corpus was not released. The author's checksum-based audit reports zero exact public-item overlap, but that claim cannot be independently reproduced; our scan found no exact public state or instruction text in the released artifacts. Cost uses the $0.05/M-input hosted Gemma 3 12B reference, not free/100.
• † reflex 4B (kshetrajna12): The author's reflex-serve: Qwen3.5-4B with the published LoRA and its per-primitive calibration file; the state is encoded once and each question read from the label logits. Run serially on our GPU; the author discloses that the 231 public items were used four times as a development gate.
• † djev (Maisa, diffusion-gemma): The measured endpoint was Maisa's hosted API in free preview; the cost uses its announced price ($0.035 per million input tokens, output free), and nothing was charged. The self-hostable djev-dev runtime is Apache-2.0 and applies a structured one-step inference method to Google's Apache-2.0 diffusiongemma-26B-A4B-it checkpoint; it adds no separately trained djev weights. Probabilities are djev's own (its docs call them experimental and uncalibrated). v1.4: hosted api.djev.dev was paused by its operator ("Serving is paused by the administrator"); sealed tier measured on the public djev runtime (Davipar/djev-dev 3ce907e, same weights, default mode) self-hosted on an H100; Speed/Cost kept from v1.3.
• † Jev-Omni (akhilaaa3, Gemma-4-12B merged): Akhilaaa3/Jev-Omni revision c050d51354147985d13286cf4acf90f562f2c631, the author's own load_model.py (merged text decision model + 256-way head) and his own predict(), transformers 5.17.0 / torch 2.8.0 from the pod image; built on the CPU and moved to CUDA with every nn.Linear weight cast to bfloat16 first - the same cast his reference loader jev_omni.py applies - because our 46 GB GPU cannot hold his fp32 copy; on our RunPod L40 in Czechia Offline local open-weight inference on the frozen 308-item v1.4 set in a network-disabled, read-only container; no operator endpoint received sealed text.
• † metask-jev-4b: Model card discloses 44.8k+16.1k+390 training rows incl. synthetic families intentionally mirroring JevBench hard families, plus repeated evaluation on all 231 public items (benchmark-directed development disclosed). Scan of 61 files: 0 exact matches.
• † Jobe Qwen3.5-4B (frozen): No trained weights/LoRA/calibration fit; release explicitly rejects fitted temperature/order averaging. Scan of 37 files vs 231 public tasks: 0 matches.
• † local-jev Qwen3.5-4B: Scan of 72 files: 0 matches. Author discloses choosing JSON layout after observing results on the 231 public items (benchmark-directed choice, disclosed).
• † spark-s1-4b-v6 (Open Spark Jev, abhishek085): Abhishek085/spark-s1-4b-v6 revision 93d49ddbfb29212e3296635a75a3e80cf69da027, code github.com/abhishek085/open-spark-jev 30ac6d89b7fa36c644cf86aac68f35c1d276a919, the author's own MenuScorer.decide with his fitted calibration.json temperature, bf16, base Qwen/Qwen3.5-4B, transformers 5.17.0 / torch 2.8.0 from the pod image, flash-linear-attention 0.5.2 installed, causal_conv1d not installable here (no wheel builds against this toolchain), on our RunPod L40 in Czechia Offline local open-weight inference on the frozen 308-item v1.4 set in a network-disabled, read-only container; no operator endpoint received sealed text.
• † jqv (Qwen3-32B zero-shot): A stock Qwen3-32B with no decision training: the state is prefilled once, each question is an isolated branch and the answer is read from the option-letter logits, with one fitted temperature (3.02, 400 MMLU validation items). Re-run in v1.2.8 on our own GPU from the now-public serving code (Octalab-Inc/jqv 0189b67), so all 534 decisions including the held-out hard items were asked; this full run replaces the v1.2.7 partial row, which had been measured on the submitter's machine. Cost is the base model's public per-token tariff, not free.
• † Qwen3-Reranker-4B: Neutral documented reranker adapter; instruction and no-instruction public calibration were run, then frozen before one held-out pass.
• † decider-35b-a3b (Mapika): The author's TypeSafe-compatible server and published FP8 weights, run serially on our H100 NVL. The exhaustive startup batch warmup was skipped; each required serial shape captured lazily before its measured request. Self-host latency receives the standard ×2 + 0.15 s adjustment. Cost uses the closest hosted 35B-A3B input tariff and is not the temporary rental charge.
• † Raw Qwen3 4B Instruct 2507 direct logits: Neutral raw-logit control.
• † OpenSourceJev (Qwen3.5-4B Q4_K_M, native llama.cpp): DM submission, measure-only (JevBench publishing HOLD in force). Same author as the existing simplejev-qwen3.5-0.8b row (sabeel111/Featherless AI). Round-4 audit of an earlier commit could not be measured (no public GGUF, Windows-only DLL loader, unpinned llama.cpp build); this round the author published the exact unsloth Q4_K_M GGUF (hash/size independently verified) and we built llama.cpp CUDA from current upstream master on Linux ourselves -- its ABI matched the ctypes bindings exactly, so only a loader file-naming fix was needed (documented diff), no code/scoring/calibration change. Our public-231 subset exactly reproduced the author-reported table: easy 48/48, standard 67/72, hard 66/111, schema 231/231. Calibration (noul temperature) fit only on Google BoolQ, not JevBench. Repo docs name 3 public task IDs while describing benchmark-directed algorithm fixes on the public half (disclosed); 0 exact state/instruction text matches in a released-file scan.
• † ZeroEntropy zerank-2: Neutral documented reranker adapter; instruction and no-instruction public calibration were run, then frozen before one held-out pass.
• † decision-machine-1 (milliseconds.ai): A closed-weights decision model behind a production API that serves TypeSafe's wire format, so the unchanged typesafe adapter ran it. Run on a free test key (30 requests a minute, 2.2 s between requests); the provider states the inference infrastructure is the same as for paid keys. Cost is the public paid tariff, $0.04 per million input tokens (output free), times the input tokens the API reported.
• † Raw Phi-4 mini direct logits: Neutral raw-logit control, not JevBench-directed.
• † JEV Qwen3.5-9B Base NVFP4: Byte-identical to upstream March-2026 NVFP4 checkpoint, predates JevBench v1.2, no task-specific training added. Had 3 preliminary scope/policy scoring errors corrected during review (pooled-ECE, full-run latency, cost estimate); score above is final corrected value.
• † kev 4B (research preview): Self-hosted from the author's repository at commit 20fa626 through its native TypeSafe-compatible `/v1/systemone` server, BF16 on an RTX 3090; measured serially from Sandy over the internet. The author labels this checkpoint a research preview. 306/308 sealed items answered validly (failures count as wrong)
• † Decision 2B (FlyMy.AI, v59): Flymy-ai/decision-2b-preview revision df57b75db927acc9ad91ec8115508c1e487086eb (checkpoint minicpm5_reduced_v16_4k_v59), base openbmb/MiniCPM5-2B revision 12a3808a956f869c767195e9266b59c4d21d92e2, the submitter's own FlyMyJevPackageAdapter and frozen calibrator, bf16, unmerged adapter, 4096-token packing, transformers 4.57.6 / peft 0.15.2 as pinned, torch 2.8.0 from the pod image, on our RunPod L40 in Czechia Offline local open-weight inference on the frozen 308-item v1.4 set in a network-disabled, read-only container; no operator endpoint received sealed text.
• † Qwen3.5-9B Jev-like data-mix v2: The author disclosed development on the public JevBench set and public-result comparisons; released-data overlap scan found no matches, but 764 gap and 382 replay training rows are unreleased.
• † GPT-6 Luna (low reasoning effort): OpenAI direct API baseline; reasoning effort low; strict JSON-schema probability response; temperature unset; max_completion_tokens=4096; price cost from returned usage at official standard list rates.
• † SimpleJev Qwen3.8-27B: Author's no-login shared demo, model id recorded verbatim, one request at a time at or below its 2 RPS limit. SimpleJev reads answer-token logits and returns the complete distribution; it does not generate an answer. Speed uses the public-demo x2 load adjustment; cost uses a hosted size-class input price and is not free/100.
• † NInfer Qwen3.8-Flash-Next mixed: Engine scan (1,758 files) vs 231 public tasks: 0 exact matches. Submitter discloses no training for Flash-Next but repeated consultation of public items and a public-hard temperature sweep.
• † GPT-6 Luna (default medium reasoning effort): OpenAI direct API baseline; reasoning effort default medium; strict JSON-schema probability response; temperature unset; max_completion_tokens=4096; price cost from returned usage at official standard list rates.
• † open-alternative-jev (Qwen3.5-4B, IkerMoel): With the options in reverse order (A. no, B. yes) the same model scored 21 % instead of 72 % on yes/no answer-judging items — small models are very sensitive to option order.
• † jev-local (Qwen3.5-9B): The author's local Jev-compatible server in its default full configuration: a frozen Qwen3.5-9B scores each option by its mean log-probability (one forward pass per option, no generation, no decision training). Run serially on our GPU. It re-reads the state once per option; if its reported token count covers one pass only, a per-token hosted price would be higher than this estimate.
• † Decision Fast (FlyMy.AI, v53a): Flymy-ai/decision-fast-preview revision 4225d41c66119fe28e95a2631bb0103decae6d56 (checkpoint qwen3_06b_headfirst_ep2a_v53), base Qwen/Qwen3-0.6B-Base revision da87bfb608c14b7cf20ba1ce41287e8de496c0cd, the submitter's own FlyMyJevPackageAdapter and frozen calibrator, bf16, unmerged adapter, 4096-token packing, transformers 4.57.6 / peft 0.15.2 as pinned, torch 2.8.0 from the pod image, on our RunPod L40 in Czechia Offline local open-weight inference on the frozen 308-item v1.4 set in a network-disabled, read-only container; no operator endpoint received sealed text.
• † decider-2b (Mapika): The author's TypeSafe-compatible server and published weights (Qwen3.5-2B-Base with a trained one-pass decision readout), run serially on our GPU. Self-host latency gets the standard ×2 + 0.15 s adjustment.
• † jeff (Logan Markewich, GLiFormer 400M): Self-hosted from its GitHub repo with server defaults, on our CPU (the author recommends a GPU, e.g. an L4), through the same TypeSafe-compatible API as Jev.
• † Laya (Convai Innovations, ModernBERT-large 421M): The English checkpoint (repo root), run on our CPU through its own `laya` package. Its budget is 512 tokens per question, so long hard-tier states are cut by the package itself.
• † lev-350m (Franck Verrot, LFM2.5-350M): Weights franckverrot/lev-350m revision ab08ad8b8f346994d983152917e114224f6adac7, code github.com/franckverrot/lev c48a945dbf629998d7458dcc5c16f58df964db94, the author's own lev.serve /v1/systemone endpoint with its shipped calibration temperature, base LiquidAI/LFM2.5-350M, on our RunPod L40 in Czechia Offline local open-weight inference on the frozen 308-item v1.4 set in a network-disabled, read-only container; no operator endpoint received sealed text.
• † Von (wfzyx, Option-Marker 395M): The author disclosed that its temperature calibration map used the 231 public JevBench items; monotonic scaling does not change accuracy. Estimated cost is USD 0.00551 per 1,000 decisions.
• † NInfer Qwen3.8-27B NVFP4 (T=1.5): T=1.5 is an offline recomputation from raw logits of the same run, not a second execution. T=1.5 was chosen by sweeping public hard items (development-set calibrated, disclosed).
• † NInfer Qwen3.8-27B NVFP4: Raw T=1.0 row. Author discloses repeated public-item consultation and public-hard tuning (applies to both NInfer 27B rows).
• † kev 8B (research preview): Self-hosted from the author's repository at commit 20fa626 through its native TypeSafe-compatible `/v1/systemone` server, BF16 on an RTX 3090; measured serially from Sandy over the internet. The author labels this checkpoint a research preview.
• † JevOne: Scan vs 231 public tasks: 0 matches. Training corpus/provenance not disclosed — overlap unknown.
• † SimpleJev Qwen3.6-35B-A3B: Author's no-login shared demo, model id recorded verbatim, one request at a time at or below its 2 RPS limit. SimpleJev reads answer-token logits and returns the complete distribution; it does not generate an answer. Speed uses the public-demo x2 load adjustment; cost uses a hosted size-class input price and is not free/100.
• † kev 0.6B (research preview): Self-hosted from the author's repository at commit 20fa626 through its native TypeSafe-compatible `/v1/systemone` server, BF16 on an RTX 3090; measured serially from Sandy over the internet. The author labels this checkpoint a research preview. 307/308 sealed items answered validly (failures count as wrong)
• † Raw Qwen3 8B direct logits: Neutral raw-logit control.
• † OpenDecision (ModernBERT-large zero-shot): A zero-shot NLI classifier behind a TypeSafe-compatible server, not a trained decision model: it scores each option as an entailment hypothesis with ModernBERT-large-zeroshot-v2.0. Its choice path runs several NLI passes over the same state, which the reported token count does not include, so a per-token hosted price would be higher than the estimate here. Pre-registered for our CPU in v1.2.7, run on our GPU because the CPU was far too slow.
• † LitJev (Qwen3.8-27B): The author's reproduction of Jev's decision layer on an off-the-shelf model, in its default configuration: Qwen3.8-27B, scores read from the output head, no training and no calibration file (its README says probabilities are not calibrated by default). Run serially on our GPU through an SSH tunnel, because its server binds to localhost; the request still crosses the internet and gets the ×2 + 0.15 s adjustment.
• † openJev Verdict 1.4: Same public weights as the earlier Verdict row, run through the author's fixed v1.4 engine. That engine auto-loads the calibrator for every option count, frames candidate labels as NLI sentences and uses a 512-token context budget. Run locally on our CPU, serially.
• † kev 0.5B: Self-hosted from the author's repository at commit 20fa626 through its native TypeSafe-compatible `/v1/systemone` server, BF16 on an RTX 3090; measured serially from Sandy over the internet. This is the v0.1 release. 307/308 sealed items answered validly (failures count as wrong)
• † Bespoke Nimble 9B (Bespoke Labs): Re-run in v1.2.8 at Bespoke Labs' request after they raised the serving prompt limit from 2,048 to 8,192 tokens (bespokelabsai/nimble PR #4). Same recipe as the v1.1.3 run — the published LoRA merged into Qwen3.5-9B with the author's PEFT safe-merge, served with SGLang and the author's Jev-compatible API — now from current nimble main; the adapter weights are unchanged. Hard-tier accuracy rose from 43.6 % to 65.5 %, yet the score fell: the long hard items that used to fail at once are now answered and priced (so Cost fell), and this pod was in Canada while the v1.1.3 run's was in Sweden, so part of the lower Speed is network distance from our server in Germany. This complete run replaces the earlier row; its old score is kept in the artifact under superseded_rows.
• † openJev Verdict (heman10x, ModernBERT-base 151M): The openJev-verdict-2.0 Hugging Face repo ships no weights; its config is byte-identical to heman10x/rlcd-modernbert-151m, whose published weights we ran with the author's engine. The 'verdict2-base' checkpoint behind the README's numbers is not downloadable yet (Git LFS 404); we will run it once it is.
• † Raw Qwen3 1.7B direct logits: Neutral raw-logit control.
• † reflex-27b (Qwen3.8-27B): The frozen public Qwen3.8-27B checkpoint through reflex at the requested pinned commit, with two option orders averaged and temperature 1. No adapter or fitted calibration file. Run serially on our H100 NVL. Self-host latency receives the standard ×2 + 0.15 s adjustment; cost uses the exact base model's public hosted input tariff.
• † djev (thinking): Experimental full-generation path over the same DiffusionGemma checkpoint as djev-dev: thinking was enabled and the model could generate up to 8,192 tokens before returning its distribution. Current djev-dev itself hard-codes enable_thinking=false, diffusion_max_steps=1 and read_only=true, so this is not a switch in its published typed API. It is substantially slower/costlier, and 72/534 requests exhausted the output budget without a parseable distribution; those are failures. Cost uses measured tokens and a same-size hosted reference, not the H200 rental bill. 220/308 sealed items answered validly (failures count as wrong)
• † GLiNER2 large (Fastino): The large checkpoint of Fastino's earlier GLiNER2 family, same documented mapping as the GLiNER2 row: the question goes in front of the text and the probabilities are the model's own single-label softmax over the labels, read out in full. A general schema classifier, not a Jev rebuild.
• † OpenJev (thinking, BF16): OpenJev's real typed-API thinking switch at think=512, using its own /v1/systemone server over BF16 DiffusionGemma. The thought is generated first, then native probability reads are taken after it. All 534 requests returned valid distributions. Cost counts the server's billed input and thought output tokens.
• † Qwen3.5-0.8B Decision Model (Mourad Ghafiri): JevLite SystemOne on CPU; bundled per-question calibration; no operator endpoint or network access. Local CPU measurement on the existing 534-decision v1.3 set plus 308 sealed v1.4 decisions; no operator endpoint received sealed text.
• † Gemini 3.1 Flash-Lite: 307/308 sealed items answered validly (failures count as wrong)
• † open-jev-deberta-v3-large (local CPU): 297/308 sealed items answered validly (failures count as wrong)
• † smalljev semantic-v9: The public semantic-v9 LoRA and native heads over MiniCPM5-2B-Base, through the mapping frozen before the run. It has a typed Python contract but no TypeSafe-compatible HTTP route. The released training recipe explicitly hill-climbed against JevBench's public shape and source families; this allowed public benchmark-directed development is disclosed. Cost is $0.04/M measured input tokens, not free/100.
• † GLiNER2 (Fastino, gliner2.5-base): A general schema classifier, not a Jev rebuild. The question goes in front of the text; the probabilities are GLiNER2's own single-label softmax over the labels, read out in full (mapping fixed before the run).
• † Open-Jev 9B (Zefan Cai): The author's pinned LoRA adapter, trained scalar decision head and calibration temperature, served by the author's Open-Jev server with prefix caching off, batch size 1 and 4,096-token limit. Serial requests were measured from Sandy over an SSH tunnel to the H100. Self-host latency receives the standing x2 + 0.15 s adjustment. Cost uses the exact Qwen3.5-9B hosted input tariff for 9B and the same conservative same-family proxy for the unlisted 2B; neither receives an automatic 100. Exact normalized comparison found no JevBench public task state or instruction in the 79,116-row public training projection.
• † Open-Jev 2B (Zefan Cai): The author's pinned LoRA adapter, trained scalar decision head and calibration temperature, served by the author's Open-Jev server with prefix caching off, batch size 1 and 4,096-token limit. Serial requests were measured from Sandy over an SSH tunnel to the H100. Self-host latency receives the standing x2 + 0.15 s adjustment. Cost uses the exact Qwen3.5-9B hosted input tariff for 9B and the same conservative same-family proxy for the unlisted 2B; neither receives an automatic 100. Exact normalized comparison found no JevBench public task state or instruction in the 79,116-row public training projection.
• † GLiNER2.5 multi (Fastino, 287M): The multilingual GLiNER2.5 checkpoint (287M), same family and same documented mapping as the GLiNER2 row. JevBench items are English only, so its multilingual training is not exercised here.
• † SimpleJev (Qwen3.5-0.8B, CPU): 143 pinned files scanned: 0 exact matches. No JevBench-specific fine-tuning.
• † GLiNER2.5 small (Fastino, 74M): The small GLiNER2.5 checkpoint (74M), same family and same documented mapping as the GLiNER2 row: the question goes in front of the text and the probabilities are the model's own single-label softmax over the labels, read out in full. A general schema classifier, not a Jev rebuild.
• † Raw Qwen3 0.6B direct logits: Neutral raw-logit control.
• † DeepSeek V4.1 Flash (thinking default): 298/308 sealed items answered validly (failures count as wrong)
• † Mirror: 171 intended HTTP 422 context rejections (over 512-token limit) counted once each as misses; only 363/534 valid distributions returned. Found via Gmail submission (Lewis). 102/308 sealed items answered validly (failures count as wrong)
• † Mixedbread mxbai-rerank-base-v2: Neutral documented reranker adapter; instruction and no-instruction public calibration were run, then frozen before one held-out pass.
• † BAAI bge-reranker-v2-m3: Neutral documented reranker adapter; instruction and no-instruction public calibration were run, then frozen before one held-out pass.
• † Alibaba GTE Reranker ModernBERT-base: Neutral documented reranker adapter; instruction and no-instruction public calibration were run, then frozen before one held-out pass.
• † Certo v1 (AltSlate Labs): The public Certo v1 checkpoint through the author's DecisionModel, serially on our rented GPU. The question instruction is prepended to the state because Certo exposes state + runtime options but no separate question field; the published 64-token state and 48-token option limits are unchanged. The model card says v1 does not yet transfer to arbitrary natural-language prose. Cost is an estimate from same-size hosted encoders times the checkpoint's retained input tokens, not free/100.
• † Open Jev JSON Canvas (JoshuaSP): Returns only a final label, not a probability distribution — calibration counts as 0 in the composite. Scan of 39 files: 0 matches.
• † classifier.dev (fast tier): Its own benchmark page says the fast tier is Jev. Free for us; the price is its published Pro plan ($20/month for 200,000 fast classifications a day) at full use, $0.0033 per 1,000 decisions.
• † Qwen3.8 27B (Chutes TEE): 69/308 sealed items answered validly (failures count as wrong); partial: Chutes rate limit stopped the run after 81/308 items; unranked as in v1.3
• † swanOne: Draft vocabulary derived via AGPL-3.0 generator (provenance recorded separately). Submitter consulted all 231 public tasks and swept temperature over 111 public hard tasks. Score corrected during review from pooled-534 ECE/latency to hard-tier/242-item block. v1.4: Not re-measured: needs its own H100 NVL pod (99 GiB weights); RunPod balance ran low during this job. Ready-to-run recipe kept.
• † Needle 3, options as tools (post-hoc adapter mode): V1.4: Not re-measured: same as needle-3.
• † Needle 3 (Cactus, 2-bit, local CPU): V1.4: Not re-measured: ~100-250 s per item on a rented CPU pod (19 s on Sandy); needs a dedicated CPU host.

Rows without a † have no note beyond the shared provenance: every row was measured or re-run with its recorded recipe, and deviations are in its run manifest.