Updated

Desk

Compare

Similar tests already in the stories, placed next to each other. The sets, labels, and comparators differ, so the rows are not a ranking. We did not rerun any of them. One-row write-ups live on evals. Failure modes live on limits.

Email routing

Three inboxes, three label sets. Vogel posted a sample size and no error rate. Mudholkar posted a named-company table with a Gemini baseline. Kumar posted agreement with whether a human kept the task.

WhoSetJevAgainstCaveat
Vogel1,500 personal mailsn = 1,500none postedNo accuracy, no comparator.
Mudholkar / Bryo1,565 supplier mails, 10 categories96.4%Gemini 3.5 Flash-Lite 97.5%; Gemini 3.8 Flash 98.5%364 synthetic mails. Attachments excluded.
Kumar800 inbox mails87.4% vs human keepnone postedNo lossless drop line. 96% on the confident band (69% of the set).

1,500 emails Bryo Filter

Banking77

Same 77-way intent set, different comparators. Kumar ran 300 items against small generative models. Poly ran the full test against open zero-shot encoders and published ECE.

WhoSliceJevAgainstCaveat
Kumar300 items, accuracy76.0%gpt-5.4-mini 78.7%; gpt-5.6-luna 81.7%Jev leads his three short-text sets and loses this one.
PolyHeld-out test, macro-F10.782 ± 0.009ModernBERT-large-zeroshot 0.712 ± 0.020ECE favors ModernBERT (0.081 vs 0.105). Thread only, no repo.

Filter Encoders

Local stand-ins

Jev's weights are not public. Each project scores itself on its own set, except JevBench, which puts several of them on one rubric. LocalJev does not put hosted Jev on the same 120 labels. Laya's Banking77 figure is the author's, against a published Jev number the README says was not measured in that repo. Palmer's 79.6% is his new-source split, not JevBench.

ProjectWhatTheir scoreJev on that testCaveat
Nimble9B Qwen LoRA, 324 synthetic labels90.12%93.21% (jev-1.13.0)Labels unreviewed by people.
jeffGLiFormer, 1,600 public itemsAG News 75.5%90.5%Copies POST /v1/systemone. Isolation of each noul is the default.
SemIfQwen3.5-4B logit readout, JevBench74.7 composite · 59.5% hardJev 75.4 · 74.1% hardClosest open row on that suite. Author 3090 note is a different fixture.
LocalJevPrompted JSON on oMLX, 120 gold labelsQwen3.6-35B-A3B 76.7% short-inputnot comparedAuthors say the probabilities are not a logit read.
llamacpp-jevUnmodified llama-server, Qwen3.5-2B, four questions on a 448 image526 ms median, 32/32 syntheticnot on this fixtureRaw label softmax. 64-way answers collapsed to option 16. Idle L4 was 236 ms.
Verdict151M encoder, 337 TypeSafe public cases48.1%90.8%Cited next to Nimble, not as its own story.
LayaModernBERT-sized encoder, JevBench plus author tableJevBench 70.1; Banking77 0.425Jev 75.3; published Jev Banking77 0.870 (72 labels)Author Jev column was not measured in that repo. 77 vs 72 labels.
KevQwen LoRA plus pointer head, author new-source plus JevBenchKev-8B 79.6% new-source; kev 0.6B JevBench 66.7Jev 85.7% on Palmer's split; Jev 75.4 on JevBenchAuthor suite is decision-v7 / transfer-v4. JevBench rows are Qwen3 checkpoints.
djevHosted DiffusionGemma API, JevBench74.3 composite · 69.5% hardJev 75.4 · 74.1% hardAnnounced $0.035/M, free preview. Native images. Calibration 65.4.
SimpleJev Qwen3.8-27BNext-token logit server, JevBench public demo67.3 composite · 75.0% hardJev 75.4 · 74.1% hardDemo x2 latency adjustment. Cost estimated. Text HF server; site has vision demos.
Open-Jev 9BQwen3.5-9B LoRA plus decision head, 231 public JevBench tasks179/231 (77.49%) · hard 66/111Jev 200/231 (86.58%) · hard 81/111Public slice, not the 534-task composite. SemIf used the name OpenJev earlier. Astra is 231/231 on this slice.

Nimble jeff LocalJev SemIf Laya Kev djev SimpleJev Open-Jev JevBench llamacpp-jev

JevBench public subset

Zefan Cai scored five systems on the 231 public JevBench tasks (72 original, 48 easy, 111 hard). The other 303 tasks are unavailable, so none of these rows is the 534-task composite. Open-Jev ran on one H100 over loopback HTTP. Hosted models used HTTPS. GPT probabilities are verbalized. Candidate order differs on 119 of 139 Choice tasks. One observation per task.

ModelCorrect / 231HardP50Caveat
Open-Jev 2B150 (64.94%)46/111138 msReleased Qwen3.5-2B LoRA. Prefix cache off.
Open-Jev 9B179 (77.49%)66/111189 msReleased Qwen3.5-9B LoRA. P95 839 ms.
Jev 1.13.0200 (86.58%)81/111291 msOne vector renormalized. About $0.038 per 1,000, estimated.
GPT-5.6 Luna206 (89.18%)89/111954 msReasoning none. About $0.179 per 1,000, estimated.
GPT-6 Astra231 (100%)111/1112,206 msReasoning low. About $8.82 per 1,000, estimated.

Open-Jev JevBench

Document classify and split

DocJev ran hosted Jev 1.13.0 and GPT-5.6 Luna on the same LiteParse page text. Forty authentic public-sector PDFs for classify, eight constructed packets for split. Decision times exclude OCR. Human annotation review was not performed.

TaskSetJevLunaCaveat
Classify40 originals, eight per category40/40 · 138.6 ms median40/40 · 794.3 ms median5.73× median-time ratio. Convenience sample.
Split8 five-document packets, 116 pages7/8 exact · 209.6 ms median8/8 exact · 1,352.3 ms medianJev extra cut on a Fed attachment. 6.45×.

DocJev

SQL row judgments

Four listings call a provider from SQL. Jevflake stores answers in Snowflake through dbt. duckdb-jev batches inside DuckDB and published a live Choice throughput table. MotherDuck's prompt_jev() runs inside the warehouse on paid plans and published topic accuracy. pg-jev is a PostgreSQL extension and prints no scored set. All four send row text out of the database.

WhoHostPublished figureCaveat
Kranz / JevflakeSnowflake SQL via dbtv0.1, one live account, ten sample ticketsPacking several rows is experimental. Extra columns changed probabilities.
Prasanth / duckdb-jevNative DuckDB C++2,049 rows in 0.887 s; cache 28.6 msTwelve-template corpus. Throughput, not accuracy. colliber/duckdb-jev is a separate MIT extension.
MotherDuck / prompt_jevHosted SQL, paid plans100k AG News train rows, 89%, 40s, $0.50Training split, seed 43. Terra on the same sample is 88%, 31m 59s, $37.58. Nulls unscored.
Mejias / pg-jevPostgreSQL extension, plpython3uno scored set in the READMEMIT. Returns jsonb. One call per row. Rollback does not undo a request already sent.

Jevflake duckdb-jev prompt_jev pg-jev

Coding-agent tool routing

jev-gateway's README cites 120 sessions on a chess rules engine: six models, fix-bugs and add-a-feature, five runs with routing and five without, no MCP servers or plugins. Cells are medians versus the same model with routing off: output tokens, input tokens, then time. Claude Code rows are hint mode. Five runs per cell is the sample the README flags as small.

ModelClientBug-fixAdd a featureCaveat
GPT-6 AstraCodex-57% / -7% / -39%0% / +2% / +8%The tweet's 57% is this bug-fix output cell, shared with Sol.
GPT-5.6 SolCodex-57% / -40% / -36%-9% / -39% / -16%Both tasks used fewer tokens with routing on.
GPT-5.6 LunaCodex-12% / -10% / +10%-14% / -51% / -14%Feature task solved 3 of 5 with routing, 5 of 5 without.
Fable 5.1Claude Code-13% / -19% / +6%-24% / -27% / -26%Hint mode. The model can ignore the suggestion.
Opus 5Claude Code-7% / -22% / +2%+22% / +61% / +83%Feature task used more tokens and more time.
Sonnet 5Claude Code-41% / -48% / -25%+9% / +16% / +37%Bug-fix got cheaper. The feature task did not.

jev-gateway

Site search ranking

McLaren indexed TypeSafe's documentation and ran the same labelled queries through jevsearch and several keyword libraries. Jev re-ranks a 20-hit keyword pool. A page that never reaches the pool cannot win.

LibraryHit@1Hit@3MedianCaveat
jev-search83%83%278 ms uncached41 labelled queries. About $0.26 per thousand searches.
Keyword pass / Lunr41%59%8.7 ms / 0.7 msThe keyword pass tied Lunr. Jev converts that 20-hit recall.
Pagefind20%29%2.6 msReturned nothing for 12 of 41 queries.

jevsearch

Pairwise ranking

jsort fits a scale from pairwise Noul judgments. The CommonLit check uses teacher Bradley-Terry scores as the reference. The Fed check uses the funds-rate move, not a human ranking. Joe Weisenthal's 4,005-speech post is a separate count from the 95-statement script.

CheckSetjsortOther measureCaveat
CommonLit easiness300 excerpts, -k 10Pearson r 0.824 · $0.046One Noul r 0.754 · SMOG 0.661Teacher reliability about 0.78, ceiling near 0.88. Run 2026-09-19.
FOMC hawkishness95 opening statementsSame-day rate move +0.46Next 180 days +0.37Top six are the 2022 50- and 75-point hikes. 4,005 is Joe's later post.

jsort

Judging other output

Same job, different oracles and sample sizes. LangChain froze five weather traces and used a human label. Lemkin compared Jev with Sonnet on Connect pairs and posted a false-admission count. Every asked Jev 21 writing questions and did not put an LLM judge on the same 777 cells. JevBench's hard tier is a different rubric, written by Opus 5 and GPT-5.6 Sol. RZ recorded 100 synthetic security spans five times each.

WhoSetJevAgainstCaveat
LangChain5 traces × 100, does_pass500/500 vs humanTerra 99.8%; Luna 96.4%; Sonnet 4.6 80.0%One reviewer. Quality variance is the other claim.
Lemkin / SaaStr600 judgments70.5% vs third-model refereeSonnet 77.5%; 32% vs 13.5% false admissionsThread only. No prompts, no repo.
Every37 docs × 21 questions777 answers in <0.7 slater 12-passage check vs Fable 5.1, 6 of 7 defectsNo LLM judge on the 777-cell grid.
JevBench hard tier220 frozen items74.1%Luna 94.5%; SemIf 59.5%Composite still ranks Jev above Luna on cost.
RZ / security triage100 synthetic cases × 565.3% balanced accuracy · 0.57 sTerra 75.1% · 10.23 s; Opus 91.2% · 13.83 sLabels by one author. Safe-class recall 45.5%. CLI vs API latency.

Judge SaaStr Every JevBench Security triage

Legal-set loops

The host enumerates actions the runtime can take. Jev picks among those options. Illegal moves stay out because they were never listed. These are demos, not a shared computer-use eval.

DemoSetFigureCaveat
Browser UseZurich to London Google Flights7.1 s · $0.0039Repo notes; clip at 1x. We did not replay it.
Stagehand / jev-browserComputer use from an a11y tree$0.001, near-instant (Jeong)Author post, not a Browserbase lab table.
TetrisSame 200-piece sequence, legal landings9,200 pts · ~300 ms · 0 illegalOne run. Haiku 8,900 / 1.52 s; Gemini Flash-Lite 9,000 / 1.13 s.
JevPilotSampled steering paths in Three.jsup to 4 requests/s; $1.24 / 15,000A game. Geometry stays in the browser.
macOS LoopSix native GUI tasks6/6; Finder 9 files in 7.39 sSmall functional sample. An earlier Calculator run failed.
JevisFlutter integration_testno published pass rateTyped strings come from the test author.

Flights Stagehand Tetris JevPilot macOS Loop Jevis

Context filters

Tran argued for scoring tool output and dropping it. AstroHan kept the agent and the model fixed and reported whether the task still passed. One run per arm.

WhoSetWith JevWithoutCaveat
Tran / fast-jev-compactionKeep-or-drop tool callsplugin and librarysummarization (the default it replaces)No published pass-rate table in that repo.
AstroHanFrontierHarness 30, deepseek-flash25/30 · meriyah 49/4922/30 · meriyah 0/49Two of 30 the filter lost. 126m vs 123m input tokens. Thread only.

Compaction 30 tasks

Router tier on a labeled set

AgentOS compared an opt-in Jev strategy with its default on-device MiniLM classifier. Same 121 turns, gold labels R0 to R3, classifier only, two repeat runs identical. Blank cells are blank on the chart. A same-day follow-up says it sharpens the R2 and R3 wording. This is not a JevBench score.

MetricLocal MiniLMJevCaveat
Accuracy0.3550.901121 turns. Vietnamese, Chinese, English.
Under-routing0.4130.066Gold is the project's own R0 to R3 labels.
Over-routing0.2310.033Jev is opt-in. Turn text leaves the machine.
English / Vietnamese / Chinese0.513 / 0.312 / 0.2500.946 / 0.938 / 0.806Chinese is the low Jev slice on this chart.
Cost and time$0, on deviceabout $0.005, about 0.9 s per call121 cases. Macro-F1 0.894 and ECE 0.062 are Jev-only on the chart.

Pilot Router

Laya, two machines

Same weight families, different runtimes. The T4 times are from the Laya README cited in the earlier story. The M3 Max times are from the laya-mlx README. Neither row is a score against hosted Jev. laya-mlx also says the selected answer matched upstream on 63 of 63 validation questions.

CheckpointAuthor T4laya-mlx M3 MaxCaveat
English, 421M39.5 ms, one question13.42 ms P50, 13.92 ms P95Different machines. MLX port, FP16, load excluded.
Multilingual, 322M32.8 ms, one question7.39 ms P50, 7.79 ms P95Batched multilingual on the T4 note was 7.2 ms per question at 10 questions.

Laya laya-mlx

Picking a cube turn

maxli asked Jev for the best of 18 face turns, forty times, then removed the experiment. Kociemba in code is the solver that ships. A random walk is the other number in the same README note.

Who picksDistanceTimeCaveat
Kociemba, in codeat most 22 movesabout 20 mscubejs wrapper. This is the solve button.
Jev, best of 1822 to 22forty callsExperiment removed. The coach asks whether to speak instead.
Random walk22 to 21same noteOne comparison the README prints beside the Jev run.

jev-rubiks

Same Flights task

Zurich to London, one way, one adult, economy, stop when results are visible. The Jev column is jev-ultrafast's published measurement. dejevu did not rerun it. dejevu's own runs are September 22 on an M3 Pro, and the Sunday it searches is October 18 2026 because September 20 has passed.

AgentVerifiedTimeCalls and tokensCaveat
jev-ultrafast, Jev 1.13 + Mercury 2.53 of 3, published7.09 s17 calls, 84,650 in, about $0.0036Other machine. The earlier clip on this desk said about 7 s and $0.0039.
dejevu, llama-3.3-70b on Groq3 of 35.63 s10 calls, 15,071 in, $0.0092Faster here, and about 2.5 times Jev's list price on this route.
dejevu, same Llama on DeepInfra2 of 237.5 s$0.0020Cheaper than the published Jev cell, and much slower the day they measured.

dejevu Flights

Typed decisions

meraGPT's September 22 blog, one run, zero-shot for Decider 1 and for the Jev row. The fitted encoders were trained on each workflow's labels. Reference labels are the average of three teacher samples. The blog says 400 cases and 2,000 decisions. The Hugging Face viewer lists 1,600 rows across four subsets. We did not recompute the table.

ModelAccuracyKLBrierCaveat
Prior, ignores the input0.4700.3470.189Floor on the blog's table.
ModernBERT-base, fitted0.6460.2230.119Trained per question. Decider 1 did not get that.
Jev 1.13.00.7271.4420.148Choice gap vs Decider 1 is 0.013. The KL gap is the wide one.
Decider 1 (sd-1)0.7680.0960.052Hosted. 4,096 token cap. At most ten choice labels. $0.03 / 1M input.

Decider 1

Banking-feed categories

Peter's September 22 post. 612 sanitized rows, described as a harder subset of about 33,000. This is not Banking77. Gemini costs are estimates and exclude thinking tokens. No category list was printed.

ModelAccuracyTimeCostCaveat
Jev 1.1336.1%6 s$0.023Closest to Gemini 2.5 on accuracy. Fastest of the three.
Gemini 2.5 Flash37.3%95 sabout $0.088Cost is an estimate from the text-token budget.
Gemini 3.7 Flash42.5%78 sabout $0.180Highest accuracy on this private slice.

Banking feed

Hold'em clocks

Two published timings for Jinzhengxu/poker-table, plus the README's language-model baselines on the same 46 spots. The live pill and the spot bank are not the same measurement. The bb/100 table elsewhere in that README is a rule-policy ablation.

ClockTimeWhat was measuredCaveat
Live pill, September 22 post545 ms average, $0.002 a handThe table the post linksWe did not open poker.ccswitch.online.
Spot bank, Jevabout 1 s median46 spots, 3 repeats, OpenRouterNo accuracy percentage. Three passes under a cent.
Spot bank, single-shot LLM6 sSame spots, reasoning modelREADME says Jev's range read matched it.
Spot bank, tool-loop agent35 sSame spotsThe path Jev replaced in front of the action clock.

Poker