Updated
Desk
Patterns
Recurring setups from the digest. Each block is this desk's reading of several stories. Compare notes on the articles go into more detail. Ranking a list the host already filtered is the schema case on decision model hallucination.
Put the image in state once
List each file once in assets and put it in state. Questions then refer to it by path. On October 16, 2026, Date header 00:28:50 GMT, four Nouls and four Boxes on one four-color PNG in state used 4,458 input tokens on jev-1.14.0. Four separate POSTs, one color each, summed to 6,612. Aviz billed 2,266 against 8,635 on a different file. On October 15, Date header 11:04:56 GMT, two Nouls on an 850 by 850 PNG in state used 2,259 input tokens. Two Nouls that each wrote the same reference into instructions used 4,209 tokens. Wes Bos billed $0.000096936 for one 850 by 850 question.
Find it in the picture
A Box question returns four edges as fractions of an image. TypeSafe's docs say one box per question. On October 16, 2026, Date headers 08:19:44 GMT to 08:19:45 GMT, one Box for every vehicle on a three-car PNG returned one rectangle, n = 3, confidence 0.19, 0.17, 0.13. A POST with type array returned HTTP 400. Named Boxes in the same request scored IoU 0.89, 0.89, 0.90. TypeLLM's cookbook at 14.00 s prints six vehicle boxes, including Tesla's flying Roadster. Date headers 00:28:47 GMT to 00:28:48 GMT, three boxes on a green Continue button averaged left 0.55, right 0.91, top 0.71, bottom 0.88, confidence 0.91, mean IoU 0.93. Aviz Maeir scored Jev 1.14 Box at 82% on 180 scenes. The same Box on jev-1.13.0 returned HTTP 400.
Rank a legal set
The host enumerates actions the runtime can actually take. Jev picks among those options. Browser Use and Stagehand do this with page elements. Tetris does it with landings. JevPilot does it with sampled steering paths. SREGym ranks diagnostic tests the agent already proposed. Jev macOS Loop ranks native window controls after local perception; pixels stay on the Mac. Jevis ranks Flutter widget actions the test registered, and typed strings have to be supplied by the author. If an illegal action appears, the host listed it. mizchi's fit page counts 489 MOBA decisions and 37 chess moves with 0 illegal answers, and treats that as what happens when an illegal move cannot be written down. maxli tried the same shape on a cube, 18 legal turns, forty times, and the Kociemba distance stayed at 22. The coach that shipped asks whether to speak. Code still names the next algorithm. jev-code asks Jev for the next grammar production and builds the AST locally. On commit 517d0f0 the three-task ladder still fails, including programs that parse and do the wrong thing. Hello world, a smaller smoke test, finishes in 3 turns and 16 requests.
Judge a pile of documents
One schema, many items, often in parallel. Every ran 21 writing questions over 37 documents. Vogel sorted 1,500 emails. Mudholkar routed 1,565 supplier emails into 10 categories and published a Gemini baseline. Elvis scored a morning news wire for 15 brand desks and stopped Opus 5 when Jev finished. Kumar ran 300 items each on four public sets, then 1,005 production pages. LangChain scored five frozen weather traces 100 times and compared Jev with three LLM judges. JevEval is DeepEval's custom metric of the same shape: the caller writes Noul, Score, and Choice questions, Jev answers them in one request, and DeepEval takes a weighted mean. On October 14, 2026 this desk's jev-1.13.0 calls scored its Tool Faithfulness example at 0.254 when the answer invented a breeze, and 0.901 when it did not. Lemkin ran 600 Connect judgments and posted a 32% false-admission rate. DocJev classified 40 public-sector PDFs and split eight packets on the same LiteParse text, with a Luna baseline. RZ recorded 100 synthetic security spans five times each against Terra and Opus. McLaren re-ranked 41 labelled docs queries after a keyword pass. Time, spend, and (sometimes) a confidence band show up more often than agreement with a human rater. Peter's banking-feed post is 612 private rows: Jev 36.1%, Gemini 2.5 Flash 37.3%, Gemini 3.7 Flash 42.5%, with no category list. vase.park's panel is 28,074 translated sentences, one per Jev call, about $0.54, with no gold label. 码农暖爸's post is 100 English items, Laya 92 and Jev 100, and the author says it is not a full benchmark. JevGate asks one function at a time and composes a finding at 0.80. The README says that probability is a judgment, not a measured accuracy. andre's file search asks Jev to rerank the top 128 local matches. nDCG@10 moves from 0.322 to 0.377 on NFCorpus. TypeSafe's CLERC cookbook, 40 queries, moves top-1 from 5% to 18% when jev-1.12 reranks a BM25 shortlist of 30. Hindsight's listwise Choice, 30 LoCoMo candidates, moves recall@1 from 0.800 to 0.950. aifabrice's best NFCorpus row is nDCG@10 0.451, and the four-question gate on that repo scored below ordinary reranking. 七誌 asks five Score questions about each translation. One pass covers 1,072 language files for $0.3165. Watters asks whether a cited passage supports a sentence. Five runs on 18 hard claims leave a 3.3-point gap and p = 1.000. RLCDAlignBench asks a Noul across 7,193 alignment items. Median AUROC for a generic wording is 0.886. Median ECE is 0.168, against a null of 0.074. Doug Turnbull's WANDS run adds 10 times a sum of Jev Nouls to BM25. One question has mean NDCG@10 0.574262. A GPT-5 list of questions has 0.609814. Both medians are 0.5609447680982702. On eight Stripe passages this desk wrote, a 0-9 Score put Roll secret at 8.86 and the other six at 2.42 or below. Pairwise Choice ranked rotating the API key third, and the majority graph had a cycle.
Override only when confident
A slower system still acts. Jev is allowed to refuse or replace that act when its probability clears a threshold. Akshay's TrueForge agent accepts an open-issue match at 80% confidence and a closed-issue match at 90%, and sends the rest to Opus. On 40 requests that review caught 5 wrong matches; REQ-001 still went through at high confidence. He wrote that 40 is too small to establish how reliably that cut routes errors. jevrail denies a shell command at the sample band of 0.80 on destroys_uncommitted and asks from 0.35. The printed corpus leaves the model off, and 15 of 26 safe commands still raise an ask. LangChain's AutoModeMiddleware can refuse a tool call. SREGym's jev_submit sends the agent back to plan. Malis let Postgres plan first and let Jev override join estimates only when it was sure, which is the version that did not blow up a query. Mudholkar's production sketch is the same split: auto-route the high-confidence band, give the rest to a person. Kumar's reject filter skips the existing classifier under a line set from known positives. Poly's 90% precision gate is the same idea on Banking77. AstroHan's coding loop kept a tool chunk when p > 0.5 and stored the rest. LiteLLM's compaction guardrail drops older tool results below 0.2 and fails open if Jev is down. jev compact uses a 0.5 keep-threshold on a session jsonl and keeps kept lines verbatim. n8n-nodes-jev-classification sends low-confidence classify items to a Needs Review output at a default 0.5. Vercel's form router keeps a Choice when TypeSafe's confidence metadata is at least 0.95 and asks Luna Fast to decide the rest. Spring's JevJudge reports a criterion as inconclusive when confidence is under 0.5, and the guardrail can refuse a request before the chat model is called. jev-rubiks applies its own thresholds before the coach speaks, including a streak before stuck help. swift-jev's sample policy escalates under 0.60 and acts at 0.85, and the README says those cuts still have to be checked on the action. spark-s1's tool gate allows a command only at probability 0.995. rev's above(0.9) returns nothing under the line, and a second reading does not raise confidence. postgres-search sinks a top-10 hit whose Noul is under 0.30 and leaves the SQL order alone when the call fails. That 0.30 cut was not calibrated, and the Jev pass was not measured. awlevin's computer-use loop stops under confidence 0.4 and hands the screen to a writer. Lemkin's Connect trial did not keep that split, and a 32% false-admission rate is why he stopped. TypeSafe's jaggedness page warns not to reuse a Noul threshold on a Choice. Eikos reports the error among decisions taken at confidence 0.90 or higher: 2.4% for both model sizes on 7,140 items, and 5.9% for Jev in that README.
Route the expensive call
A cheap typed question picks which slower system should run. Akshay's TrueForge agent shortlists 10 GitHub issues in a script, asks Jev which issue matches, and leaves new-issue drafts and low-confidence review to Opus. Printed decision costs on 40 requests are $0.0015 with Jev and $0.005 to $0.015 without it. JevScout asks Jev which pieces of an MCP result of at least 100 KB to keep, and the README's check is 38% less agent time and 26% less agent cost, with TypeSafe's bill separate. Polylane moved five fixed-answer on-call steps to jev-1.13.0 and prints p90 latency from 4,752 ms to 508 ms. A Pi extension asks which skill names belong in the prompt, and the 20-skill demo cuts about 1,000 tokens. Glean's rerank on 4,855 queries cost about $0.00044 and took 0.195 s at the median. jevgrep asks Jev which files matter and leaves the patch to Sol. On ten tuned tasks the official solves stay 8 of 10, and the combined bill, Jev's list price included, is 25.8% under a saved baseline. toolgate asks Jev to score a shell call after static rules, and version 0.15 stores what a written script would do before a later Bash call runs it. Christoph Görn's router asks a local Kev-4B first, scores 1 of 7 on eight prompts, and fails open to a hosted arm. Lightpanda's post pairs Jev with its browser and prints 6.7 seconds against 14.1 seconds for gemini-3.8-flash. The linked script measures memory and does not call Jev. Hono's semantic router asks Jev whether a request looks like spam, an agent, or an angry customer, then the Worker branches. semgate does the same on Go net/http and ships a public range; by default it sends every header, including Authorization and Cookie. Jevflake puts the question in Snowflake SQL and stores the answer. duckdb-jev batches the same questions from DuckDB and streams across chunks. MotherDuck's prompt_jev() does that inside a paid warehouse and published an AG News accuracy table. jevsearch lets keyword search answer first, then asks Jev to re-rank the top hits. jsort spends its calls on pairs inside one file and prints a scale. jym asks only when a subcommand is missing from help text, then execs the real CLI. Vercel's form router asks Jev for a destination and, under 0.95 confidence, asks Luna Fast the same question. jev-gateway asks which tool a coding agent should call, and in direct mode writes the call itself when every argument is closed. Claude Code stays on a hint the model can ignore. AgentOS can ask Jev which model tier a turn deserves. The local MiniLM classifier stays the default, and a high_risk Noul at 0.7 floors the turn at the deepest tier. Spring's guardrail can return before the chat model is called. jevelry writes act, mark, or fall_back from thresholds stored in a JEVEL.md, and logs the ask so a later outcome can score it. Mejias's pg-jev asks from a PostgreSQL statement and returns jsonb, one call per row. Zachi's pg-jev batches 20 rows, caches the session, and prints about 3.5 seconds on 2,000 rows. poker-table asks Jev for a read of the opponent and lets code pick fold, call, or raise. postgres-search asks one Noul per candidate in a SQL top 10 and can only sink a hit, which the docs have not scored. ASO Workbench asks Jev whether a term is a brand and whether it sounds like demand, then a local model estimates Apple's popularity. The language model on that table talks and takes notes off the decision path. Jevonian asks once, on jevonian/auto, which model and thinking depth a coding turn should use. A named phase or a pinned model id skips the call. The README prints no scored table. awlevin's Mac loop asks Jev which click comes next and calls a writer only for free text or a stopped screen. On one shared screenshot that split is $0.0002 against $0.032 for Opus 5 reading the image. Sam Saffron's term-llm PR classifies each GPT Live utterance in about 650 ms and fails open to the main model. LiteLLM's typesafe guardrail asks whether older tool results still matter and drops those below 0.2 before the generator runs. n8n-nodes-jev-classification emits one wire per category on a self-hosted canvas. OpenRouter, Cloudflare, Netlify, Vercel, Venice, and LiteLLM now list or proxy the same model, so that question can sit in front of someone else's generator. TechCrunch quotes Armin Ronacher on this use. TypeSafe's jaggedness page still says not to ask Jev to write the reply. CLM-8B is built for a fixed action list: action embeddings cache on their own, and a thread post says about 1,000 candidates is 13 times faster than Jev. TypeSafe's Deel charts are production picks from a known set: 6 of 8 quality checks at parity or better, messy metric questions down 16.6 points, live-shadow escalation 4 times. astra-jev-harness asks Jev which bytes a Codex prompt still needs. On one two-file fixture Astra input falls 27.0%, and checks stay 6/6. An older run cut files from 28 to 4 while Astra input fell 1.2%. HydroJEV accepts a benign Jev verdict on its own only when a hand-written rule tree agrees. That gate settles 69 of 200 and 76 of 200 fresh windows. On the look-alike set, 32 of the 76 accepts were wrong, including 21 attacks. jev-doc-search asks Jev to walk a section tree once a single Choice would pass 32k tokens or 255 options. The printed table is two filings, and both answers are the income-statement page. You.com's risk server asks Jev whether a sweep should escalate, then lets Qwen write the briefing. Three escalated sweeps cost $0.43, and raising the Noul cutoff from 0.5 to 0.7 dropped a Gulf briefing the lower cutoff had marked critical. DSPy's patterns repo puts a Noul inside the shell tool: refuse below 0.2, ask below 0.6, run at 0.6 or above. OpenRouter's typesafe/jev-router is a chat model id that asks Jev which model and effort should answer. Arena's October 7 thread prints +3.6% net and a median $0.155 a task on more than 4,700 sessions, and says a direct DeepSeek V4.1 Flash call is cheaper and faster at similar performance.
Recreate the model locally
Jev's weights are not public. The API board opened October 7, 2026, release v1.6.1 under revision v1.7.12, ranks Jev 1.13.0 third at composite 71.5, with capability 77.1. Opened October 8, 2026, revision v1.7.21, that board ranks 26 hosted systems and Jev 1.13.0 is fourth at the same 71.5. On JevBench v1.5.0, opened September 29, Cygnet leads the official harmonic mean at 73.7 and Winnow-12B Q8 is 73.2. The page calls that pair a statistical tie. Jev 1.13.0 is 72.1, official rank 3, and first on capability among 56 Jev-class systems at 80.0. Imajev-4B, a Qwen3.5-4B LoRA that can read up to two photos, scored 67.37 on the v1.4.2.2 board. On the v1.5 addendum it would place sixth at 70.4, outside the ranked order. The page opened October 4 is v1.5.6, and that 70.4 is official rank 10, still tagged roster addendum. Weights and the image-board row are on the imajev-4b page. Plumb-4B, a JevK5 LoRA, scored 65.84 on v1.4.2.1, and the v1.5 addendum would place it fourth at 71.6. House's public-hard runner is still 89 of 111. decider-4b v2 is official rank 5 on v1.5 at 71.3. On the older v1.4.2 page its Intelligence was 49.4 against Jev at 53.1. Sandipan Kundu's 400M encoder is a different project from Palmer's Kev. On 17,476 of its own rows it scores 79.96% against Jev at 73.54%, and on PhishNChips it scores 50.25% against a published Jev figure of 62.6%. SemIf reads option logits from a frozen Qwen3.5-4B and sits at 74.7 on the older geometric mean against Jev 75.4. djev is a hosted DiffusionGemma API with native image input; JevBench 74.3, and Davipar/djev-dev is the local stack. Laya is a ModernBERT-sized encoder with three checkpoints and a Router; JevBench 70.1, and its own table trails published Jev on Banking77. laya-mlx runs those same checkpoints on Apple Silicon. Its README reports selected-answer agreement with upstream Laya on 63 questions, and it does not add a JevBench row. Bespoke Nimble is a 9B Qwen LoRA; it matched 90.12% of its own 324 synthetic labels against Jev at 93.21%. On 13 public subsets the macro is 74.8% against Jev at 76.0%. Verdict is a 151M encoder that runs in a browser tab and scored 48.1% on 337 TypeSafe public cases where Jev scored 90.8%. jeff is a local /v1/systemone server on GLiFormer. On 1,600 public items it trailed Jev (AG News 75.5% versus 90.5%) and was cheaper to serve on a GPU. Kev is a LoRA plus pointer head on Qwen that serves the same HTTP shape; Palmer's new-source table puts Kev-8B at 79.6% against Jev 85.7%, and JevBench listed kev 0.6B at 66.7. SimpleJev points a stock Hugging Face chat model at next-token logits; JevBench listed Qwen3.8-27B at 67.3. Open-Jev is a Qwen3.5 LoRA plus a decision head. On the 231 public JevBench tasks the released 9B checkpoint is 179/231 against Jev 200/231 and Astra 231/231. 27B v1.1, on Qwen3.8-27B, is 197/231 and 80/111 on that slice, and the card prints no latency. That slice is not the 534-task composite, and this Open-Jev is not SemIf. GitHub Next's LocalJev is another /v1/systemone server; it prompts a local model for JSON probabilities on oMLX and says those numbers are not a logit read. Its bake-off compares local checkpoints with each other, not with hosted Jev. llamacpp-jev also serves that HTTP shape, from an unmodified llama-server, and reads label logprobs. It accepts images when a multimodal projector is loaded. The README says the probabilities are not calibrated. On 64-way questions the 0.8B and 2B notes both landed on option 16. Release v0.6.0, published October 5, serves the same path for Nimble and Clef. The Nimble pull names Bespoke-Nimble-9B-v3, Decision Index 0.2.1 at 56.88, CC BY-NC 4.0, which is not the 324-row holdout. Clef can take an image. The October 2 medians stay 3 ms to 43 ms. meraGPT's Decider 1 speaks the same three question types from a hosted API. Its typed-decisions table puts it at 0.768 against Jev 1.13.0 at 0.727. The labels are an average of teacher samples, and the blog's case count differs from the row count on the dataset viewer. rev reads option logits from a frozen Qwen and serves the same HTTP shape. Its Qwen3.8-27B note is 0.784 on 111 public hard items, and the laptop 2B note is 0.550, against a Jev cell of 0.730 in that README. spark-s1-4b-v6 is a Qwen3.5-4B LoRA with a public-proxy Intelligence of 83.1 and a hard cell of 0.595. The README says that proxy leaves out the private judge tier. hotdog-27B is a binary LoRA on Qwen3.8-27B. It leads a 1,631-decision holdout and trails on 74 public binary items. djev-run puts DiffusionGemma-Jev on Cloud Run behind POST /v1/systemone. The latency note is 35 to 60 ms at samples=1. There is no accuracy table on that page. Rizzo Flow reads Spark-X2.5 logits on llama.cpp. On SemIf's authored144 it is 0.812 against SemIf's published 0.819, at 49 ms on an RTX 5060 Ti, and 6 of 36 missing-evidence rows stay confident. Jevre is an 0.85B checkpoint on the TypeSafe Python SDK. Its test split is 0.9323 and held-out sources are 0.8366, with no Jev column. jev_local adds Sarashina2.2 Vision 3B beside LFM. On JGLUE test the Q4 row is 34% and 86% against LFM at 17% and 69%. Ámbar Reflex plays a run-and-gun from the same HTTP shape. The 4B model reaches 24.9% of the level. The 9B model is 82.7% on 554 offline decisions and reaches 17.5% on the clock. Eikos-27B reads option-letter logits from a fine-tune and prints 82.9 on its public JevBench hard row, against Jev at 73.0 in that README. CLM-8B keeps the Qwen3-8B frozen and trains two 20M heads. Its DeepSWE verifier holdout is 31 of 38. jevos-1b is a 1B GGUF behind the same HTTP shape. The README alt text is 0.815 on 2,000 yes/no questions, against Jev at 0.927 and Laya at 0.489. Choice and Score return 422. DiffusionGemma as Jev is the one-step canvas read of google/diffusiongemma-26B-A4B-it. A hand chart adds to 198 of 201 against Jev at 191 of 201, and a routing file puts the djev-spark joint at 32.2% against 61.4%. Clef and Clef-flash are Apache 2.0 weights on frozen Qwen3.8-27B and Qwen3.5-9B. The Workers AI cards list $0.24 and $0.09 per million input tokens. Their Decision Index table is a per-benchmark run, not the composite on the Decision Index page. Strands Decider 2B is an Apache 2.0 checkpoint with 1.9 billion parameters, posted October 1. v19 scores 167 of 231 on the public tasks, median 115 ms on an RTX 3090. The file dates that 2B rank, 3rd of 33, to the September 25 v1.4.2 board, and says the run leaves out sealed items. Bud Decision Studio loads eleven open checkpoints, including Laya, Kev, CLM, and Jev-Omni, and serves them at 127.0.0.1:8420. Its GB10 training table prints Jev-Omni at 62 percent to 77 percent on business workflows and does not print how many examples that was. NIRNAY 450M is a Laya fine-tune, Apache-2.0. Its phase_b JSON is 0.8792 on 3,080 Banking77 test rows, and the card says the Jev cell was not measured in the repo. WaterSheep fine-tunes ModernBERT-base, serves POST /v1/systemone on port 8766, and adds a multi-label question. The README prints 77.8% in distribution and 61.2% on held-out datasets, with no Jev column. StartLux-Decision is five dense sizes from 0.8B to 27B plus a 35B mixture. Code is Apache-2.0 and the weights are CC BY-NC 4.0. The 27B self-run of Decision Index 0.2.1 is 63.88, against the September 28 board's Jev cell of 57.91, and the README says the run is not on the board. basal-1.5 is an Apache-2.0 family on a Bielik base, 1.5B, 4.5B, and 11B. Werdykt v1 puts the 4.5B at 72.12% and Jev 1.13.0 at 81.64%. GitHub main on October 4 still describes basal-1.0. Decision 2.0 is six Apache-2.0 checkpoints from vLLM Semantic Router. Vega's card prints 56.5 on its own Decision Index 0.2.1 run and does not print a Jev row. fsdecide is a 149M ModernBERT trained on ForkScape's five decisions. On 80 review items it scores 91.3% against jev-latest at 56.3%. On the authors' 1,248-request general set, jev-latest is 90.5% and fsdecide is 60.5%. JEV-27B-VL is an Apache-2.0 adapter on Qwen3.8-27B. The card's October 3 section prints 15 of 20 pick-and-place scenes. Its six-benchmark mean is 84.07 against a hosted Jev cell of 83.85, and that JevBench column is not the official board.
September 30 local checkpoints
Maincode's Matilda card, September 30, describes a 26.1B model and a self-run of Decision Index 0.2.1 at 59.26. The run section still says the weights are private. Opened October 8, 2026, the card prints Decision Index 0.3 at 60.16, and the model API lists private false. PostHog/jeeves is a Qwen3.5-9B LoRA plus a pointer head. With thinking it prints 0.865 on 111 public hard items, against a Kev-published Jev cell of 0.730, and the sealed judge tier is left out. Ollaya serves other authors' weights on port 11435 and prints winnow:e4b at 0.722 against TypeSafe Jev at 0.738. The item count for that pair is not on the page we read. Ollaya is a local process, and this desk does not list it as a public door.
October 9 and 10 checkpoints
Clef-omni, posted October 9, 2026, takes text, images, audio, and video in one call. The blog prices it at $0.15 per million input tokens and scores BANKING77 at macro-F1 94.8. The same post lists hosted Clef-flash at $0.038 and a 24k window. PurpleMIST-Flash-1.0, posted October 10, scores 0.668 accuracy on 15,050 translated African-language decisions and serves POST /v1/systemone from serve.py. vegaml, the same day, is a local 0.8B and 4B package. The README scores 75.4% on 800 phishing emails against Jev 1.13.0 at 61.9%. None of these is a door on the Access page.
d1-ultra
Liquid AI posted d1-ultra on October 15, 2026, 17:00 UTC. Opened October 16, Artificial Analysis Decision Index prints 65 against Jev 1.14 at 57. That day's Artificial Analysis thread prints an 85% non-hallucination rate on AA-Omniscience-Decision against 2% for hosted d1. The blog prices input at $0.042 per million tokens, output free. Model id d1-ultra. POST https://api.liquid.ai/decisions/v1/systemone. Vercel lists a 32,768-token window. Liquid's 112-evaluation average of 76.7 is a different table. This desk did not call the API. It is not a door on the Access page.
Celeris-1 Decision
Celeris posted Celeris-1 Decision on October 8, 2026, 23:34 UTC. The blog prints 80.0% on 3,210 jev-bench items at a 67 ms median, $0.04 per million input tokens, output free. Model id celeris-1-decision. POST https://inference.celeris.ai/celeris-1-decision/v1/systemone. Opened October 16, Artificial Analysis Decision Index lists it at 52. GET /v1/models with no key returned HTTP 401. This desk did not send a billed POST. It is not a door on the Access page.
Span-02
Respan posted Span-02 on October 15, 2026, 17:00 UTC. The blog prints Decision Index 0.3 thinking off at 61.2 against Jev 1.13.0 at 58.0. The Hugging Face card prints text index 63.62 on 140,178 requests. Hosted POST https://api.respan.ai/api/v1/decisions, input_cost 0.01 on the public models list opened October 16. This desk did not run the weights. It is not a door on the Access page.
OpenRouter Armor
OpenRouter posted Armor in beta on October 15, 2026, 21:00 UTC. The plugin scores each tool call with a decision model before the call runs. The 492-call chart prints Jev 1.13 at 90.7%, 446 of 492, $0.0021 a call. On October 16, at 01:18 UTC, this desk's jev-1.13.0 noul was 0.98 on git push origin main and 0.01 on ls ./reports. Those two calls are not the 492. Armor is not a catalog row and not a door on the Access page.
hunch student and teacher
Matt Mastracci's October 27, 2026 posts name hunch as an EmbeddingGemma2 /v1/systemone endpoint on CPU, at 20 ms on text and 100 ms on images. Cloudflare Clef teaches a local student from embedding vectors until the student graduates and checks in only occasionally. Open-source launch was planned for later that week. Opened October 28, public search found no hunch GitHub repository. It is not a catalog row and not a door on the Access page.
October 28 to 30 open checkpoints
alphaXiv posted OR-decision-4B on October 30, 2026. The card prints JevBench 193 of 231, 83.55%. The launch chart prints Choice 79.7 against Jev 79.6, and sarcasm 53.6 against Jev 70.7. Apache-2.0 LoRA on Qwen3.5-4B. OpenBMB posted MiniCPM5-2B-Jev on October 28. The project prints 78.79% on JevBench, first among open models at or under 2B, and decision-v7 86.41% with ECE 0.0307. Apache-2.0 on MiniCPM5-2B. Necati Demir posted BranchOne on October 29. Qwen3.5-0.8B-Base plus LoRA returns Choice probabilities. His A100 Colab median on ten short text questions is 0.305 s. Apache-2.0. BranchOne has no published JevBench score. None of these is a door on the Access page.
SubQ Snap
SubQ Snap 1.1, posted October 22, 2026, prints 65 on Artificial Analysis Decision Index, tied with d1-ultra and Elephant 9B. Docs name subq-snap-1.1 at $0.04 per million input tokens. Date headers 02:23:17 GMT to 02:23:20 GMT October 23, three POSTs from this desk returned HTTP 200. The October 12 blog prints Decision Index 0.2.1 at 61.7. It is not a door on the Access page.
Quyet and Blink
Quyet-1.0-Large, opened October 12, 2026, is Apache-2.0 on Gemma 4 31B. The homepage prints frozen JevBench v1.6.0 at official score 71.39 and Capability 81.7. Blink v0.3, last modified October 6, prints a self-run of Decision Index 0.2.1 at 57.48, against the board's Jev cell of 57.91. The card's license field is Apache-2.0. Neither is a door on the Access page.
October 11 open checkpoints
Decision 3.0, posted October 11, 2026, is six Apache-2.0 cards. Opened October 26, Decision Index 0.4.1 Full score for d3 is 67.15. The October 16 card reading of edition 0.3.2 was 66.3. Sibling d3-turbo prints Full 67.31 and vision full 72.46. d3-ultra prints Full 70.18, second of 229. The free API aliases vllm-sr/d3 to d3-turbo. Mizan 4B averages 78.7 on 3,000 real Persian records, against the card's Jev cell of 83.7. THX-01 averages 98.4 on 2,843 support tickets, against Jev 1.13 at 97.4. None of these is a door on the Access page.
October 11 board and two open checkpoints
Opened October 19, 2026, JevBench JSON revision v1.6.29 ranks 184 open systems and 38 hosted ones. basal-omni-1.0-ultra is first of 184 on Capability at 82.5. kettei-fat stays first of 38 hosted at composite 77.8. Jev 1.14.0 is ninth at 70.2. Opened October 17, the API board named 39 hosted systems at that same 77.8. Opened October 11, that board ranked 27 hosted systems, Sage 1.3.0 first at 74.0. H2O-Lightning-4B v1.1, on the October 7 blog, prints open-weights composite 72.5 against Jev 1.13.0 at 71.5. None of these is a door on the Access page.
Hosted decision APIs beside Jev
Fastino posted GLiDE v1.2 on October 22, 2026. Opened October 26, Decision Index 0.4.1 Full 71.5 rank 1 of 229, docs $0.04. Date headers 08:49:09 GMT to 08:49:11 GMT, GET /v1/models omitted GLiDE, POST /v1/systemone HTTP 401. DailyPapers posted Jeva on October 14. The paper prints 0 flips on 600 Choice items and BANKING77 at K=16 91.62 against Jev 82.50. ABE Masanori, October 24, printed $0.055 on Jev Noul against $1.87 on AI_FILTER for 1,774 posts. Date headers 08:49:11 GMT to 08:49:40 GMT October 26, TypeSafe's jev-latest named jev-1.14.0. Base Labs posted mammoth-1 on October 21, 2026. The card prints Decision Index 0.4 Full 71.24. Opened October 26, live 0.4.1 Full 68.4 rank 3. Date headers 02:23:21 GMT to 02:23:23 GMT, GET /v1/models with no key returned HTTP 401. OpenAI Developers posted Decisions API general availability on October 22. Date headers 02:23:12 GMT and 02:23:13 GMT October 23, two POSTs returned HTTP 200. SubQ Snap 1.1 prints 65 on Artificial Analysis Decision Index at $0.04 per million input tokens. Opened October 20, 2026, Nace's Drex Max product page prints Decision Index 0.3.2 public 67.59 against Jev 1.14 at 60.70, $0.06 per million input tokens. GET /v1/models with no key returned HTTP 401. StartLux-Decision v1.1 27B Instant prints Full score 69.12, first of 174 on Decision Index 0.3.2. Sieve-31B prints Full score 63.8. Clairic, posted October 19, is Cloudflare's extraction model at $0.10 per million input tokens. Datadog prints Jev Noul recall 82.90% to 3.88%. Date headers 08:50:46 GMT and 08:50:47 GMT October 20, this desk's jev-1.13.0 is_risky was 0.94 then 0.13. Opened October 19, 2026, basal-omni-1.0-ultra ranks first of 184 open systems on JevBench Capability at 82.5. ToddlerJev, posted October 18, prints 68.70% on MMAU-mini. Qwen-Audio-Jev prints 78.70% on that name in its paper. HiddenLayer prints Jev at 91% Confirm Price on a checkout ad. kettei-fat, opened October 19, ranks first of 38 hosted systems at composite 77.8. POST https://api.sqwish.ai/v1/decide. Docs price input at $0.06 per million tokens. GET /v1/models with no key returned HTTP 401. Perplexity posted pplx-decider-v1.2 on October 16. The card prints Decision Index 0.3 overall 68.09 on joint requests against v1.1 at 62.30, at $0.02 per million input tokens. Databricks posted the same day that TypeSafe's Jev-1.14 is available through ai_classify and ai_decide. SubQ Snap, posted October 12, 2026, names POST https://api.subq.ai/v1/systemone and model subq-snap-preview. Docs price input at $0.05 per million tokens. The blog prints Decision Index 0.2.1 at 61.7 against a Jev cell of 57.9. On October 13, 2026, at about 03:00 UTC, two POSTs from this desk returned HTTP 200. Liquid posted d1-ultra on October 15. Opened October 16, Artificial Analysis Decision Index prints 65 against Jev 1.14 at 57. The blog sample names POST https://api.liquid.ai/decisions/v1/systemone and model d1-ultra, $0.042 per million input tokens. Celeris posted Celeris-1 Decision on October 8. The blog prints 80.0% on 3,210 jev-bench items at a 67 ms median, $0.04 per million input tokens. POST https://inference.celeris.ai/celeris-1-decision/v1/systemone, model celeris-1-decision. That Artificial Analysis board lists it at 52. GET /v1/models with no key returned HTTP 401. Liquid's d1 docs name POST https://api.liquid.ai/decisions/v1/systemone and model d1:free. Output tokens are always 0. The September 29 chart is an internal Decision Index reproduction, 58.9 against 57.9. Nace hosts Drex 1.5 and posts a price of $0.04 per million input tokens. The chart prints 199 of 231 against Jev at 201 of 231. Inception's Mercury Decide post cites JevBench v1.4 and a rate of 14 decisions a second, and prints no score. Opened October 8, 2026, the free model page returned, priced Free, with up to 14 decisions a second and no JevBench score. The free OpenRouter URL returned Not Found. OpenRouter's Mercury 2.5 page says Mercury Decide uses the same /v1/systemone schema as Jev. None of these hosts is a door on the Access page. Rox's chart is a rerank in front of an agent, mean over 75 queries, not a new model card. Perplexity's October 1 docs name POST https://api.perplexity.ai/v1/decisions, model pplx-decider-v1-27b, at $0.04 per million input tokens. Fastino's docs name POST https://api.fastino.ai/v1/systemone, model fastino/glide, at $0.30 per million input tokens and a 40,000-token window. Opened October 26, 2026, the pricing page lists $0.04 per million input tokens and $0 output. The October 8 reading was $0.15. The model catalog names fastino/GLiDE. Those two are also not doors on the Access page. Databricks ai_decide is a Beta SQL function, posted October 1. REST is POST /api/2.0/ai-functions/ai-decide. The docs do not name the model and do not print a price. It is not a door on the Access page. Together's Tev1 page, opened October 7, 2026, lists $0.04 per million input tokens and free output. That chat call returns one letter. Ollama serves the same name on /v1/systemone. Microsoft-Decision-1, posted October 9, 2026, is $0.042 per million input tokens on the Command Line post, with output free. Foundry's path is /providers/microsoft/v1/systemone. The catalog, opened October 10, says the weights are not distributed. It is not a door on the Access page. TypeLLM's homepage, opened October 10, prices input at $0.05 per million tokens and thinking at $0.50. The call is POST /v1/generate, model typellm-latest. The September 23 file scores Qwen3.8-27B at 228 of 231 public JevBench tasks with thinking. That file is not typellm-latest. It is not a door on the Access page.
October 6 and 7 checkpoints
Perplexity posted pplx-decider-v1.2 on October 16, 2026. The card prints Decision Index 0.3 overall 68.09 on joint requests against v1.1 at 62.30. Docs price v1.2 at $0.02 per million input tokens and flash at $0.01. Perplexity's October 7 quickstart priced the API at $0.02 and named v1.1-27b. Decision Index 0.3, opened October 7, ranked v1.1 first at 62.8, with Jev 1.13.0 as a reference at 60.1. Nace's October 6 post calls the open weights Drex 1.1. The card title is Drex DLM. The call is POST /v1/systemone. The card scores Decision Index 0.2 at 52.31 against the kit's Jev cell of 51.67. PolicyLM-1.7B returns a score from 0 to 1 for each category in a policy, up to 16 in one pass. The card prints 0.842 at a 0.5 cutoff and has no Jev column. HiringCafe's October 7 thread prints Spearman 0.74 and 0.71 for Jev against OpenAI on query-to-job, and 0.71 against 0.67 on resume-to-job. Liquid's October 7 Open d1 cards print d1-3B at 48.57 on a self-run of Decision Index 0.2.1, and d1-omni-600M at 15.95. The GGUF card serves POST /v1/systemone on port 8080. Vela 2.0, posted October 7, returns Choice, Noul, Score, Span, and Set. The 9B card prints 41.63 with Noul calibration on and 41.09 with it off. The blog's example server is POST /v1/systemone.
Grade the step when the answer has no key
Step-Jev trains two Qwen3.5-4B-Base LoRAs on a search agent's own traces. The outcome judge returns P(correct) for the final answer. The step judge scores one search as positive, neutral, or negative, and training uses P(positive) minus P(negative). On one BrowseComp-Plus seed the final answer alone falls to 5.4%, with 10.5 searches a question. Adding the step score reaches 19.6%. A rollout with no tool calls never enters the step loop, so the return is only the outcome probability. Hosted Jev, used zero-shot on that page, scores 0%.
Score adult text, leave the picture on the host
jevmod asks one Noul per category and flags adult text at 0.80. On OpenAI's sexual labels that line is F1 0.866, AUROC 0.984, remeasured September 23. The $0.044 per 1,000 messages pays for the whole request. ZeroNSFW scores avatars and posted media with two local vision models, profile bar 40% and media bar 90%. Jev reads clipped profile and message text when a key is set, and a missing answer counts as unavailable. The repo prints no accuracy on adult images.
Start from a public backbone
An open checkpoint that returns a Choice, a Score, or a Noul keeps a language model whose weights can be downloaded, then trains a head or a short adapter for the typed answer. Jeff v1.3 prints a mean of 30.4% for untrained Qwen3.5-0.8B and 92.8% with an adapter, across 13 of its own tests. Unsloth's October 7 documentation fits the same Clef head to Llama 3.2 3B and Gemma 4 E4B. Laya trains its head on ModernBERT. Liquid's open d1 starts from LFM2.5. Jev's own weights stay with TypeSafe.