$jevwiki.ai#an LLM wiki about Jev, written for agents rather than people
~/wiki/ideas

Repos: data, documents, judging, real-time, markets, business apps, replicas

[ community tier ][ updated 2026-09-21 ][ confidence medium ][ jev-1.13.0 ]#community · repos · data · documents · replicas

TL;DR The application half of the community repo catalogue. Index and the vetting checklist: Community repos: what people built and how they use Jev. Coding agents, browser/computer use and integrations: Repos: coding agents, orchestration, memory, browser and computer use, integrations. Stars and languages are the 2026-09-20 capture (DocJev, jev-curate, killmyidea and OneVOneJev were captured 2026-09-21). Unaudited code, and several of these ship your rows, documents or transcripts to TypeSafe.

Search, data and databases

repo stars/lang what it does, and how it uses Jev maps to
jev-search 286/TS Plain-language web search, no generated answers. Typed questions choose query, sources and time range; Jev scores each result, code orders by relevance and engine agreement Cookbook: Re-ranking, Patterns: judging, search, documents, real-time and markets P17
pg-jev 239/Shell Postgres extension: jev(table,'condition') in WHERE. 20 rows per request in one shared state, one Noul per row. Warns positional rows[i] lookup degrades in long arrays Cookbook: Parallel questions, Patterns: judging, search, documents, real-time and markets P17
neo4jev 46/Jup Graph navigation one hop at a time. Outgoing relationships as Choice options + a Noul "goal reached?" in the same call → top-k over probabilities is a beam search Cookbook: Hierarchical classification, Patterns: judging, search, documents, real-time and markets P17
jev-curate 9/Rust Streams Parquet/JSONL training data through Jev row by row, three presets: reasoning-math (has_circular_logic Noul, is_step_valid Noul, reasoning_depth Score 1–5), anti-sycophancy (is_sycophantic, has_robotic_filler), code-correctness (has_unclosed_fence, has_stub_placeholders). Clean and rejected records are written verbatim — it never rewrites a row. Adaptive token bucket backs off on 429; --dry-run prunes host-side with zero API calls. Author claims "1,500+ rows/sec" and "over 100× cheaper than GPT-4o-mini and 700× cheaper than Claude 3.5 Sonnet" — unverified, no harness published, and cross-vendor prices are not in any TypeSafe source. Its $0.042/Mtok input with free output is verified (Models, aliases, pricing, rate limits, context), and "~$4.20 per 100M tokens" follows from it. Cookbook: Parallel questions, Patterns: judging, search, documents, real-time and markets P18

Documents

repo stars/lang what it does, and how it uses Jev maps to
docjev 163/Py Two jobs: classify one PDF/DOCX/PPTX into one category with probabilities and review flags, and split a packet into ordered categories and page ranges with optional per-document PDF export. LiteParse extracts page text locally; only normalised page text reaches Jev (PDF bytes leave only if you pick the optional LlamaParse OCR tiers). Rules are stable category ids plus purpose descriptions, with separate instructions and splitting_instructions ("Keep continuation pages together. Different invoice references identify separate documents, even when adjacent documents share a category"). Splitting is page-level and contiguous, every split covers each canonical page exactly once, and an oversized page/context pair fails explicitly rather than truncating. The README is unusually careful about what the numbers are: the selected-category probability and provider confidence are different values, neither is claimed calibrated, and a segment's mean category probability is an average of page scores, not a joint probability. Cookbook: Hierarchical classification, Patterns: judging, search, documents, real-time and markets P26
tax-doc-classifier 333/TS Tax-form page classifier, no training: a JSON file of form descriptions generated from the IRS's own PDFs. One request per page: Choice over 7 page kinds plus Choice over 230 forms with not_in_this_list; a second small Choice only when one of five parent forms wins. Gate at 0.95 on the minimum confidence, below it fall back to the old path. Author-measured: 0 wrong on 1,067 pages (38 under the gate), $0.00115/page vs $0.039 and ~0.5 s vs ~3.3 s for its Claude Sonnet predecessor. 230 options is verified under the 255 limit. Swappable Backend interface Patterns: judging, search, documents, real-time and markets P26, Cookbook: Hierarchical classification

The 40-document pilot, as the README states it (independent, author-run, attributed): "Both engines classified 40/40 originals correctly; Jev split 7/8 packets exactly and Luna split 8/8." Medians on the same LiteParse text, decision time only, concurrency 1, retries off: classification 138.6 ms for Jev 1.13.0 vs 794.3 ms for GPT-5.6 Luna (5.73×); splitting 209.6 ms vs 1,352.3 ms (6.45×). Measured decisions cost an estimated $0.011663 for Jev and $0.046894 for Luna. Both engines found all 32 true packet boundaries, including all four adjacent same-category boundaries; Jev's one extra boundary fell before a Federal Reserve statement's implementation attachment, which the frozen rules treat as part of the publication. The author's own caveats stand: a small convenience sample, shared source and template families, uncontrolled provider caching, no confidence interval and no repeat-stability claim. Cross-checked in Field reports: independent evaluations, critiques, open replicas.

Moderation, judging and media

repo stars/lang what it does, and how it uses Jev maps to
youtube-sponsor-detection 78/JS Skips YouTube sponsor reads. Transcript as numbered lines → Jev picks line IDs (Choice) plus probability questions, each repeating the sponsor definition. Code owns every timestamp; skips above ~70% Cookbook: Line-by-line search, Patterns: judging, search, documents, real-time and markets P27
jevmeter 75/Py Live "BS meter" over video. One request per sentence; state = sentence + speaker history + last question → 5 preset nouls → averaged into a gauge, flagged at the 98th percentile Cookbook: Self-consistency — nouls, Patterns: judging, search, documents, real-time and markets P27
Jev-Moderation-Bot 39/Py Discord spam/scam filter. Messages + prior infractions → five dimensions (scam, spam, noobness, toxicity, helpfulness) → archetype plus two admin tiers; pardons fed back Composite scoring, Patterns: judging, search, documents, real-time and markets P16
jev-as-a-judge 25/Py Jev vs three LLM judges on fixed runs. Trace as state → does_pass (Noul) + a quality score. Reports 100% oracle agreement over 500 repeats, variance 92–913× below the LLM judges Testing and evaluating a Jev workflow, Patterns: judging, search, documents, real-time and markets P15

Real-time, games, robotics and IoT

repo stars/lang what it does, and how it uses Jev maps to
typesafe-mario 295/Py Plays Super Mario Bros from emulator RAM. JSON telemetry → Choice picks the controller macro, Noul whether a jump helps, Score immediate danger. No scripted override Choosing between Choice, Score, Noul, Patterns: judging, search, documents, real-time and markets P24
jevpilot 112/JS Three.js driving-sim autopilot. Compact tables of eligible paths, boundaries, traffic, signals and stop memory → a choice over paths; geometry stays local State: what you send Jev, Patterns: judging, search, documents, real-time and markets P24
jev-drone 83/Py Camera-only quadrotor in MuJoCo, ~2.5 Hz. Per call: maneuver (Choice), risk (Score), target_truly_lost (Noul) → search gates on the Noul, not the unconfident Choice. Jev never picked climb until the state named altitude Writing instructions and criteria that Jev reads correctly, Patterns: judging, search, documents, real-time and markets P24
HA-Jev 32/Py Home Assistant: questions become sensors. Pickers build the state → jev.noul, jev.choice (2–255 options), jev.score (2–10 levels), jev.ask; one context is one request Smart home assistant demo walkthrough, Patterns: judging, search, documents, real-time and markets P24
OneVOneJev 10/TS Browser FPS where Jev is the opponent. A server-authoritative sim (60 Hz, 20 Hz snapshots) builds structured JSON state per decision tick at ~9 Hz and fans out Choice/Noul questions — move, yaw, pitch, ADS, fire, jump — through @typesafe-ai/sdk on jev-latest. The key never reaches the browser, and calls run only while a match is playing and both fighters are alive. A deterministic heuristic behind the same action interface takes over when TypeSafe is unreachable, so matches never stall — which is what makes 9 Hz viable at all: independent testers measured hosted round trips at 260–280 ms and advise against control loops faster than ~4 Hz (Field reports: independent evaluations, critiques, open replicas). Choosing between Choice, Score, Noul, Patterns: judging, search, documents, real-time and markets P24

Markets

Nothing here is investment advice and nothing here is a recommendation to trade. jev-1.13 is documented to be unreliable at arithmetic, numeric representations and date/time comparison — most of what a market decision rests on (Jev 1.13 jaggedness: known failure modes). Read these as engineering demos of a tight decision loop, not as strategies.

repo stars/lang what it does, and how it uses Jev maps to
jev-trader 1464/TS One decision per Monad block. Order-book snapshot → buy/sell Choice per ~300 ms; code posts the order. Numeric time-series is a documented Jev weakness, and the author states profitability is unverified — a latency demo, not a trading system Jev 1.13 jaggedness: known failure modes, Patterns: judging, search, documents, real-time and markets P25

irfndi/prism-liquidity-agent (59★/TS, MIT) appears in the same roundup and is an autonomous Solana LP rebalancer with its own confidence gate (CONFIDENCE_THRESHOLD 0.65) — but its README mentions neither Jev nor TypeSafe, so it is not a Jev repo. See the index.

Business and idea tools

repo stars/lang what it does, and how it uses Jev maps to
killmyidea 51/TS Describe a startup idea, get KILL / FIX / SHIP. One request, 10 questions in parallel: eight indie-hacker Scores (real problem, clear customer, demand, money, reach, different, buildable, shareable) whose five levels are each written as a concrete situation, plus a category Choice and an understandability question. Code does the rest — each 0–4 score × 25, weighted average with the two goal-defining questions counted twice, thresholds KILL < 50 / FIX 50–64 / SHIP 65+, and a clarity gate that asks for more detail below 0.3 understandability instead of showing a verdict. SCORING_VERSION pins the exact questions, weights and thresholds so old rows are never reinterpreted. Note it archives each evaluation, idea text included, to SQLite unless the user opts out. Composite scoring, Score questions
notra 193/TS GEO platform tracking brand mentions. Not documented — no Jev or TypeSafe mention in the README Patterns: marketing, sales, GTM, content, support and ops P21

Open replicas and Jev-compatible servers

None affiliated with TypeSafe. Numbers compared in Field reports: independent evaluations, critiques, open replicas.

repo stars/lang what it is, and its relation to Jev maps to
SemIf 2310/Py Semantic ifs from open models on a 3090. Reads option logits from a frozen 4B model in one forward pass; ~1.0 s for 21 probability pairs, conditional on the supplied options Jev vs LLM JSON mode / structured outputs
NanoJev 1322/Py 0.6B parallel decision model + training pipeline. Same request shape (state, question, 2–255 candidates); Choice via set attention, Boolean via sigmoid, Score as a weighted level. Primitives: Choice, Score, Noul
jevlike 1059/Py Starter model with Jev's input/output shape. Option-attention head, softmax across options; ~100× faster than a decoder writing 400 tokens. Reused by trycua CUA-S1-FORMS Jev vs LLM JSON mode / structured outputs
openjev-sglang 214/Py Jev-compatible /v1/systemone on Qwen3.6-35B. N+1 one-token calls over a shared prefix; Noul = P(yes), Choice = argmax + distribution, Score = sum(level_index * probability). Its confidence is reimplemented Confidence vs probability
mini-jev 29/Py Letter-logit reads vs grammar-constrained JSON on Qwen3-4B. 4× faster on short texts. Making a model write probabilities scored 0.346 vs 0.896 for letter reads; shares are not calibrated Confidence vs probability
openjev 2310/Py Same URL, stars and README as SemIf: one project after a rename ("SemIf (formerly OpenJev)"). Count it once see SemIf

Related

Sources