Updated
Glossary
Terms
171 definitions from the Terms block on each article. A word can appear in more than one story. The wording here follows the explainer when we have one, otherwise the longer note.
- __none__
- A required option in jym's Choice, alongside the real subcommands. If Jev picks it, the wrapper does not substitute a documented command.
- 2026-09-19-public-v5
- The frozen recorded run on RZ's dashboard. One hundred synthetic cases, five passes, 1,500 decisions. Earlier v2 through v4 directories are rejected by corpus hash.
- Abort-on-finish
- The Opus 5 column is stopped when Jev completes the wire, so the slower model is not billed for the remaining headlines.
- Accessibility tree
- The structural snapshot of a webpage exposing element roles, labels, and states. These workflows supply this tree to Jev as state rather than transmitting screenshot images.
- Accurate when confident
- In this write-up, accuracy where a pick-one confidence is 0.9 or higher, or a yes/no probability is at or below 0.1 or at or above 0.9, plus the share of items in that band.
- Action budget
- The attempts argument on test(). It is a cap on UI actions executed, not on scenario restarts. Goal evaluation also runs before the first action and after the last allowed action.
- Action space
- The set of valid moves available for the active piece, recomputed on each turn similarly to how browser agents rebuild interactive DOM elements.
- Action table
- The compact state Jev sees: eligible paths, road boundaries, nearby traffic, signals, stop memory, and destination guidance. Shared values are sent once.
- AG News training sample
- MotherDuck's published prompt_jev table uses 100,000 rows drawn with a reservoir sample, seed 43, from the AG News training parquet. It is not the dataset's held-out test split.
- AI Gateway
- A cloud provider's managed proxy for AI APIs handling credentials, billing, and logs, such as those provided by Vercel, Cloudflare, and Netlify.
- anchored scale
- A jsort JSON file of anchor texts and their scores. Later texts are compared only with those anchors, so a new line can be placed without sorting the original file again.
- AutoModeMiddleware
- An experimental wrap_tool_call hook where Jev evaluates tool call safety, returning an error ToolMessage when an action exceeds risk thresholds.
- Awesome list
- A curated catalog of projects. Two independent lists appeared during launch week; project counts on each reflect point-in-time snapshots.
- Balanced accuracy
- Unweighted mean of per-class recall over vulnerable, safe, and insufficient context. The published v5 run puts Jev at 65.3%, Terra at 75.1%, and Opus at 91.2%.
- banking feed
- In this post, rows from a product that categorizes bank transactions. The 612 rows are described as a sanitized, harder subset of an evaluation set of about 33,000. The categories are not listed.
- Borderline-weighted pairs
- Lemkin's 200-pair slice that over-represents close cases. He reports 59.5% agreement with Sonnet there, against 67% on 200 random pairs.
- Calibrated decision
- TypeSafe's phrase, used in the TechCrunch piece, for a typed answer with a probability rather than generated text.
- Calibration
- The degree to which a model's reported probability corresponds to empirical label accuracy, which TypeSafe targets via RLCD.
- Choice
- Selecting one option from a user-defined set (cardinality up to 255), returning a probability distribution and confidence score.
- classifier.dev fast tier
- A hosted classification API whose own page, and JevBench's note, say the fast tier is Jev. It matches Jev on Intelligence and leads the ranking on a cheaper published plan and a faster measured p50 from the bench's server.
- coach facts
- In jev-rubiks, the plain-English description code sends after a hand move: beginner stage, pieces placed, whether the move undid progress, and how long the coach has been quiet. Jev answers whether to speak. It does not return the next algorithm.
- Compaction
- Reducing an agent's context history so subsequent prompts fit model limits. Traditional approaches rely on summarization prompts; this method scores entries and removes irrelevant ones.
- Confidence band
- A slice of answers above a probability cutoff. On this set he reports zero errors in Jev's most confident 85.5%, and that nearly half of answers below 70% were wrong.
- confidence statistic
- TypeSafe metadata on a Jev Choice, keyed by question id. The form router reads it from providerMetadata.typesafe.confidence. It runs from 0 when probability is spread across destinations to 1 when it sits on one destination, and it is not the probability of the selected option.
- Context filter
- AstroHan's loop: Jev answers one boolean per ~2k-character tool chunk (state is task, intent, action, output) and the harness keeps the chunk when p > 0.5.
- Contrastive data curation
- Write two examples that differ in one fact so the correct answer flips. The rest of the context and the question stay the same.
- Control plane
- In this PR, a classifier lane that labels each GPT Live utterance before the main model runs. Config values are off, agent, or classify.
- Decider 1
- meraGPT's hosted model sd-1. The September 22 blog says it answers noul, choice, and score questions in one call, and that a typesafe-sdk client can point TYPESAFE_BASE_URL at https://meragpt.com.
- Decision rate
- How often the sim asks Jev for a new path. The README caps this at 4 Hz in traffic and about 1.5 Hz on an empty road. Single-answer questions are resolved locally.
- Decisions API
- Venice's endpoint POST /api/v1/decisions, which accepts state and typed questions rather than a chat transcript. OpenRouter uses the same name for POST /api/alpha/decisions.
- Direct logit readout
- One forward pass over declared option tokens, with no sampled answer. SemIf's 3090 note: 1.023 s for 21 questions versus 5.332 s for a compact JSON array from the same weights.
- direct mode
- A jev-gateway outcome. Jev is confident about the tool, and every argument is an enum, a boolean, or a constant, so the gateway writes the tool call and does not call the language model.
- djev
- Maisa's hosted decision API on DiffusionGemma, and the name Matt Mastracci used for DiffusionGemma-as-Jev. JevBench v1.2.8 lists the hosted row at 74.3.
- DocJev
- Jerry Liu's Apache 2.0 library for classifying and splitting PDF, DOCX, and PPTX files. LiteParse reads pages locally; hosted Jev returns the category or the page-range boundaries.
- does_pass
- The binary evaluator in jev-as-a-judge. Jev matched the human oracle on all 500 repeated decisions in this run.
- Dry run
- Operating against live market data without private transaction keys, simulating order fills when trade prints cross the resting quote.
- duckdb-jev
- Prasanth J's Apache 2.0 native C++ DuckDB extension. SQL functions call TypeSafe's Jev over HTTP, with batching, streaming, and an optional connection cache.
- Early warning
- Taylor's description of using Jev as a first pass that is cheap enough to run even when a later human or larger model still reviews the result.
- evaluate
- Vercel AI SDK's experimental method for Jev, passing state and questions to receive typed answers rather than text completions.
- Expected Calibration Error
- A score for how far stated probabilities sit from observed accuracy. Lower is better. Poly reports 0.081 for ModernBERT and 0.105 for Jev on this Banking77 run.
- Fail open
- Every error path returns the original input as steer, so a classifier outage does not drop the utterance.
- fail_open
- LiteLLM's default when Jev is unavailable: send the original request to the generator without compaction.
- fallbackReason
- The form router's label for why Luna decided. null when Jev cleared 0.95. low-confidence, missing-confidence, or jev-error otherwise. A failed Luna call does not return a reason or a destination.
- False admission
- A negative case the model treats as acceptable. Lemkin reports 47 of 148 for Jev (32%) and 20 for Sonnet (13.5%) on this SaaStr Connect sample.
- fx auto mode
- A Vercel safety classifier Sharma says previously ran on gpt-5.6-luna. The September 16 post is the source for the 5 to 18 times figure.
- Geometric mean
- The four axis scores are multiplied and the fourth root is taken. A weak axis pulls the total down; Luna's cost score of 28.2 is why a 96.8 Intelligence row ranks 12th at 66.0.
- GLiFormer
- Knowledgator's encoder with task heads at inference time. jeff loads knowledgator/gliformer-large-v1.
- head_max_len
- Laya's option-prompt token budget. At default 256 on the multilingual checkpoint, 77 Banking77 labels get about 3 to 4 tokens each, which the README says is why accuracy falls to 0.425.
- Header denylist
- semgate.WithHeaderDenylist. Without it, Authorization and Cookie are sent to TypeSafe with the rest of the request.
- Helium
- A browser reported to cause account registration failures during the initial onboarding period.
- Hierarchical context
- A proposed structure organizing execution history into a nested task tree to allow logarithmic search instead of sequentially reading complete logs.
- high_risk floor
- In the AgentOS Jev strategy, a Noul at or above 0.7 forces the turn to tier c3 in code, so a short request to delete production data does not depend on the route Choice reading the words literally.
- hint mode
- The jev-gateway path for Claude Code. A short suggestion is added to the request because a forced tool_choice would break extended thinking or the prompt cache. The model can ignore it.
- Hit@1
- Share of labelled queries whose correct document is ranked first. On McLaren's 41-query TypeSafe-docs set, jevsearch is 83% and the keyword pass alone is 41%.
- Hybrid override
- Postgres plans first. If Jev's estimate is confident, that plan can be replaced. The thread reports this avoided the 10x regression from blind overrides.
- Independent verifier
- A check of the app's actual state (file locations and hashes, Calculator display) that the decision policy cannot read.
- Injection Range
- semgate-example, a Cloud Run service with six simulated injection sinks. With a TypeSafe key the guard answers 403; without a key every payload reaches the handler. No real side effects.
- input battery
- The JevGuardrailAdvisor check on the user's request. A block returns before the chat model is called. The output battery checks the model's answer and does not retry.
- Intelligence-per-cost
- Almeida's term for the trade-off between computational cost and decision quality, prioritized over raw demonstration latency.
- Jaggedness
- TypeSafe's word for uneven failure modes on jev-1.13. The docs page lists nine of them and says many will be fixed in later versions.
- jeff
- A self-hosted server that speaks TypeSafe's System One API and answers with Knowledgator's GLiFormer instead of Jev.
- jev_ask
- A Snowflake function that sends several named questions about one row in a single TypeSafe call, so the row is billed once.
- jev_plan
- A tool the agent calls with three to five hypotheses and a read-only test for each. Jev ranks those tests against a fresh namespace snapshot and does not run them.
- jev_stream
- A DuckDB table function in duckdb-jev. It takes a subquery of row id, evidence, and questions, packs HTTP batches across chunks, and returns answers keyed by the supplied ids.
- jev_submit
- A review gate before diagnosis or mitigation is sent to the grader. Every required question had to reach probability 0.70 or the agent had to gather new evidence.
- jev-compaction
- LiteLLM guardrail that asks Jev whether each older tool result still matters for the latest user question, then replaces results below relevance_threshold (default 0.2) with a short notice.
- jev-latest
- TypeSafe's alias for the current stable model release, currently routing to jev-1.13.0.
- JevBench public subset
- The 231 public tasks in JevBench, out of 534. Original 72, easy 48, hard 111. The other 303 tasks are private or judge-only, so a score on this slice is not the full benchmark composite.
- JevBench Score
- Benchmark Heaven's composite for typed-decision systems: Intelligence, Calibration, Speed, and Cost at 25% each, combined with a geometric mean. v1.2.2 reports Jev 1.13.0 at 75.3.
- jevctl
- npm package for Nasrallah-AL/jev-cli. The command is jev. It wraps TypeSafe's Jev as verify, screen, classify, extract, find, rerank, match, route, ask, compact, and batch.
- jevel
- A folder with one JEVEL.md. Frontmatter names the model, the required state fields, and the questions. jevelry reads that file. The body is for people.
- Jevflake
- Luke Kranz's dbt package (and Terraform module) that creates Snowflake SQL functions for Jev's noul, choice, and score questions. Default model jev-1.13.0.
- Jevis
- A Dart package that runs on Flutter integration_test, sends the current screen and a registered action list to Jev, and executes the chosen widget action.
- JevJudge
- A Spring AI TypeSafe builder. Each criterion is a Noul or a Score plus the threshold it must clear. Confidence under the default floor of 0.5 is reported as INCONCLUSIVE rather than FAILED.
- Jevons paradox
- The economic observation that increases in resource efficiency can expand total demand. The product is named around this economic hypothesis.
- jevsearch
- Kyle McLaren's MIT shadcn registry block for site search. A lexical pass answers first; Jev re-ranks the top candidates in one request.
- Join order benchmark
- JOB, a standard set of join-heavy queries over IMDB-style data, used here as the reported test. The thread does not list which query subset or Postgres version.
- jsort
- Khaled Eltokhy's MIT command, installed as jev-sort. It asks Jev which of two texts ranks higher on a plain-English comparative, fits a Bradley-Terry scale, and prints the file in that order with a standard error.
- Judgment
- One typed answer to one question about one document. 37 documents times 21 questions produced the 777 count.
- Judgments model
- Jevflake's recommended dbt table: one stored answer per key and question, resent only when the row, the questions, or jevflake_model change.
- jym
- The command from syumai/jevyoumean. It wraps a CLI, and on an unknown subcommand asks Jev which documented command the token resembles, then can run that correction.
- keep-threshold
- jev compact's default 0.5. A tool call or its result is kept when Jev's probability is at least this. LiteLLM's similar guardrail defaults to 0.2.
- Kev
- Jared Palmer's Apache 2.0 family of Jev-like decision models. Rank-16 LoRA plus a pointer head on Qwen bases. Serves POST /v1/systemone.
- Late
- A status indicating the model response missed the block deadline, resulting in a hold action without updating quotes.
- Laya
- Convai Innovations' Apache 2.0 typed-decision engine. Three checkpoints plus a Router. JevBench v1.2.2 reports 70.1 for the Laya row against Jev 75.3.
- Legal landing
- A valid board placement calculated by the game engine. The model chooses exclusively among these options.
- LiteParse
- The default local parser in DocJev. It extracts complete page text on the machine. Decision timing in the published pilots excludes this step.
- Literal reading
- Answering the words in the instruction rather than the intent behind them. The docs say to put boundary cases in the criteria, or split the judgment into two questions.
- llamajev
- The MIT command in NakliTechie/llamacpp-jev. It launches or attaches to an unmodified llama-server and serves POST /v1/systemone by reading single-token label logprobs.
- LLM fallback
- A small language model used exclusively when the chosen operation requires text input.
- LLM-as-a-judge
- A generative model scores another system's trace. LangChain compared that pattern with Jev on five frozen weather-agent runs.
- Local perception
- ScreenCaptureKit, OmniParser CoreML, Apple Vision OCR, and accessibility labels run on the Mac. Pixels and coordinates are not sent to Jev.
- LocalJev
- GitHub Next's local HTTP server that accepts TypeSafe's System One request shape and answers with a model running behind an OpenAI-compatible Chat Completions API.
- Macro accuracy
- Accuracy with each of the 10 categories weighted equally. He reports 92.0% for Jev, 94.6% for Flash-Lite, and 96.9% for Flash.
- MCP
- Model Context Protocol. typesafe-mcp and jev-mcp expose Choice, Score, and Noul as tools callable by Claude Code or similar clients.
- meriyah
- A DeepSWE task whose unfiltered tool output was 1.46 million characters. Unfiltered 0/49 tests; filtered 49/49.
- Mock model
- The repository's default heuristic that simulates trading signals locally without requiring an API key.
- Model completion
- A Noul on the current screen clearing goalThreshold. Reports mark this completionBasis as model unless a verify callback also passes, in which case it is model+assertion.
- ModelRouterMiddleware
- An experimental middleware hook where Jev analyzes the latest user input to select which chat model processes the turn.
- n8n-nodes-jev-classification
- MIT n8n community node from khmuhtadin. Classify, Score, Check, and Ask Questions against TypeSafe's Jev. Fixed classify mode emits one output per category plus Needs Review.
- Native computer use
- Clicking and dragging controls in a real macOS window rather than driving a browser DOM.
- Needs Review
- The extra classify output when confidence is below the node's threshold, default 0.5. Items still carry jev.needsReview when the node is told to send them to the best category anyway.
- New sources
- Palmer's out-of-domain split: datasets and policy rule types Kev was not trained on. The September 20 post puts Kev-8B at 79.6% there against Jev 85.7%.
- Next-token logit read
- Score the allowed answer tokens from one prefill, with no decode loop. SimpleJev reuses the shared prompt's KV cache across questions in the same request.
- No state retention
- Venice's claim that it keeps no copy of the state you send or the answers it returns, and that it does not train on that traffic.
- Noul
- TypeSafe's term for a boolean question, returning a probability score between 0 and 1 without a separate confidence field.
- Open-Jev
- Zefan Cai's Qwen3.5 LoRA adapters plus a trained decision head. The 2B and 9B checkpoints return Choice, Noul, and Score probabilities. A different project from SemIf, which previously used the name OpenJev.
- opponent read
- The Jev request in poker-table: a choice over range buckets, a noul for folding to each candidate bet size, and a noul for continue strength. The README says hole cards and chat are not in that request.
- Paid team share
- Vercel's figure for the fraction of paying AI Gateway teams that called a model in a launch window. The Jev post uses nearly 13% at 24 hours.
- Parallel constrained decoding
- Read the prompt once, then score the allowed answer tokens instead of generating a string. Nimble's Python code builds the typed result from those scores.
- Parallel questions
- Submitting multiple questions regarding the same state within one API request, distinct from classifying multiple different emails simultaneously.
- Parallel sampling
- Evaluating multiple declared questions over a single input state simultaneously without sequential token generation.
- Pass rate
- Successful attempts out of five per problem. The authors used this instead of time-to-diagnosis.
- Pass-through
- LiteLLM's /typesafe/ prefix. The proxy forwards TypeSafe paths, including POST /v1/systemone, and returns TypeSafe's response unchanged.
- pass@1
- Whether the task passed on the single attempt that was run. 25/30 with the filter, 22/30 without, on the FrontierHarness 30.
- Pilot Router
- AgentOS's layer that picks a model tier for a turn, from c0 for a short reply to c3 for deep work. The Jev strategy is opt-in. The default remains the local MiniLM classifier.
- Planted defect
- A problem inserted into synthetic text so the tester knows which passages should be flagged.
- Play credit
- The hosted demo signs you in through Standard Agents and gives $0.25 of Jev spend. Local development uses your own TypeSafe key and skips login.
- plpython3u
- The untrusted Python language PostgreSQL needs for this extension. The README says the function performs external HTTPS work inside the statement, and that a rollback does not undo a request already sent.
- Pointer head
- Kev's scoring layer. It compares each option's </opt> hidden state with the question's <decide> hidden state, then softmaxes those scores.
- port fidelity
- In the laya-mlx README, whether the MLX port selects the same answer as upstream Laya on a fixed set of 63 validation questions. The README separates that check from accuracy against a human label or against Jev.
- Post-only
- An order that adds liquidity to the order book rather than filling immediately against existing orders, capturing the spread when matched.
- Precision-gated automation
- Fit a confidence cutoff on validation for a target precision, then apply it to held-out traffic. Poly's 90% precision bar auto-handled 67.5% of items with Jev and 35.7% with ModernBERT.
- Promotional pricing
- The Gateway model card's free input and output for typesafe-ai/jev, listed as ending September 25, 2026.
- prompt_jev
- In mejiasd3v/pg-jev, a PostgreSQL function that sends text and typed questions to Jev and returns jsonb. The README's modes are a noul probability, a choice, a score, or several questions in one request.
- Prompted probability
- A probability vector the generative model writes as JSON, then LocalJev validates and normalizes. The README says this is not mathematically equivalent to reading logits.
- Quality-score variance
- How much a judge's continuous 0–1 quality score moved when the agent output was unchanged. LangChain reports Jev at 0.0000149 mean per-case variance.
- Question isolation
- Running a separate encoder pass per noul so other questions in the same request cannot change that probability. jeff defaults to isolating nouls.
- raw label softmax
- llamajev's probabilities are a softmax over the logprobs of the option labels. The README says they are not TypeSafe's calibrated RLCD probabilities.
- Reference-label accuracy
- Share of emails where the model's top category matches the label Mudholkar assigned. The 96.4% figure is this metric on 1,565 items, including 364 synthetic mails.
- Reference-label agreement
- Share of examples where the model's top answer matches the synthetic label. The 324-example table is this metric, not calibration against human judges.
- Reject filter
- Skip the slower model when Jev's P(yes) sits under a line set from known positives. On Kumar's page gates the line came from half the lowest P(yes) among real positives, checked by rotating production runs.
- RLCD
- Reinforcement Learning for Calibrated Decisions, TypeSafe's training methodology intended to align model confidence scores with true outcome probabilities.
- Router
- Laya's entry point that detects script and language, then dispatches to english, multilingual, or typed-decisions. Lazy load without preload can cost 7 to 10 s on a language switch.
- Runnable
- LangChain's standard unit of composition that can be executed, batched, or streamed across agent pipelines.
- Schema
- The structured Choice definition representing possible piece landings, preventing the model from outputting invalid moves.
- Score
- A numeric evaluation based on a custom rubric, returning an estimated position, distribution, and confidence level.
- Selectivity
- The fraction of rows a filter is expected to keep. Malis asked Jev for this number and let Postgres plan from it.
- Semantic route
- An HTTP route handler matched by a descriptive phrase (such as 'a request from an AI agent') instead of an exact method and URL path.
- semgate
- mizutani's Go net/http middleware. It asks Jev typed questions about an incoming request and hands the answers to your function, which may block, route, or call next.
- SemIf
- Theodore Lee's open project, formerly OpenJev, that reads typed option probabilities from a local model. JevBench v1.2.2 reports 74.6 for the Qwen3.5-4B row.
- Set A
- Per-headline questions: a desk, a story type, and six 0-4 scores (magnitude, velocity, novelty, window, heat, risk) that feed a newsworthiness mark out of 10.
- Set B
- Per-company questions on standing and journalist shape, then a tier. A story can land on several of the 15 illustrative desks, or none.
- settle
- After each action, dejevu waits until the DOM is quiet and the requests that action started have finished, with a cap per action type. The README contrasts that with fixed waits, and says jev-ultrafast discards 4 to 6 of 17 calls because the page changed under them.
- Shadow mode
- Classify and log the label, then act as if control_plane is off. The PR default for live.classify.shadow is true.
- Sieve
- A filter placed before a larger model to discard unnecessary context according to calibrated criteria.
- signup pause
- TypeSafe's September 22 statement that new Jev accounts are temporarily not being created, while accounts that already signed up keep working. The post does not give a reopen time.
- SimpleJev
- Featherless's Apache 2.0 classifier server. It reads next-token logits for allowed labels and returns Choice, Score, or Noul JSON. JevBench v1.2.8 lists Qwen3.8-27B at 67.3.
- Skill file
- A SKILL.md instruction file read by coding agents. TypeSafe's file prompts agents to fetch updated documentation rather than relying on stale parameters.
- spot bank
- 46 handwritten Hold'em spots in the poker-table repo. The Jev mode repeats them three times through OpenRouter on typesafe/jev-1.13. The README gives a median near one second and does not print an accuracy percentage.
- SREGym-Lite
- A smaller SREGym problem set meant to run on modest clusters. This experiment used 10 of those problems, not the full 21-problem lite suite.
- Stagehand
- Browserbase's library for automating remote browser sessions through structured interaction steps.
- Starting credit
- The $5 TypeSafe posts for every new account, described as about 120 million tokens at $0.042 per million input.
- State
- The input data provided to Jev, such as a single email or batch. The calling application retrieves and supplies the data.
- status.typesafe.ai
- TypeSafe's public status page for api.typesafe.ai and console.typesafe.ai. The September 21 incident is titled Console is unavailable.
- Structural invariant
- A relationship you might expect between two questions, such as P(yes) on a Choice matching a Noul, or a Noul and its negation summing to 1. The docs say jev-1.13 does not guarantee those identities.
- Structured read
- OpenJev's path, as LocalJev describes it: a one-step DiffusionGemma pass that needs seeded canvases, read-only denoising, and selected-token logprobs, which stock oMLX does not expose.
- System One
- TypeSafe's term for a model returning structured decisions with probabilities rather than text output, inspired by Daniel Kahneman's distinction between fast and slow cognition.
- Target element
- The specific page element receiving the selected action. The repository states that both heads share a request to avoid mismatched actions and targets.
- teacher average
- The blog's name for the reference on typed-decisions: the mean of three samples from teacher models. It says two single samples agree 73.5% of the time, so a model can match the average more often than one teacher matches another.
- Threshold
- The minimum score required for a match (defaulting to 0.5). The first description whose Noul probability exceeds this value triggers its handler.
- TypeSafeClassifier
- A LangChain Runnable that submits state alongside defined Choice, Score, or Noul queries to Jev, returning structured results rather than text generation.
- Verbatim keep
- Retaining original tool outputs unchanged in context without summarization, preventing drift in variable names or error logs.
- verdict
- act, mark, or fall_back, chosen from thresholds in the jevel. act means the probability clears the bar to use the answer. mark means use it and flag it. fall_back means do what the code did before the call.
- verified run
- In dejevu, a run counts only when code checks the final page. The README says the model announcing that it is done is not the check. Flights must show results. Wikipedia must land on the exact article URL.
- Waitlist
- TypeSafe's early-access queue from launch through September 20, 2026, when the company posted that Jev was open with no waitlist.
- web context
- In the Exa demo, search snippets placed into the state Jev judges. The page says Jev is asked the same yes/no question with that state and without it. No scored pair was printed on the page we loaded.
- Whole-document read
- A question whose answer needs the page as a unit, not a short snippet. Kumar's 135-row extraction gate is the example: Jev matched later outcomes 68.6% and the probability swung by 0.5 between runs on the same page.
- Workers AI
- Cloudflare's serverless inference runtime, executing queries via env.AI.run('typesafe/jev', { state, questions }).
- x402 wallet
- A prepaid USDC path Venice lists next to a Bearer API key for authenticating decision calls.
- Zero-probability miss
- The model assigns probability 0 to the correct label. Poly reports this on 6.6% of Jev's decisions and on none of 1,155 decisions from each of the three open encoders. A zero cannot be recovered with a threshold.