Use cases hubProject paths checked · agent-guard row · benchmark links · support-ticket row · routing and moderation rows · email row · trading, fraud and judge rows · browser row · reranking, SQL and document rows · real-time control row and failure counts · cookbook figures read
What can you build with Jev (TypeSafe AI)? Use cases by evidence level
Jev is useful for one bounded decision inside a larger workflow: route this ticket, hold this comment, pick this page element. This page lists the patterns found so far, with the evidence behind each and what goes wrong when Jev is wrong or unavailable.
Short answer
For 15 patterns the pool read the code that calls Jev: email and lead sorting (five projects read on 3 Oct), comment moderation (six projects read on 2 Oct), trade entry gates (four bots read on 4 Oct), fraud and payment checks (three gates read on 4 Oct), Jev as an eval judge (four tools read on 4 Oct), browser and computer use (seven acting agents read on 5 Oct), UI composition, an agent completion check, agent tool-call guards, support-ticket routing (five projects read on 1 Oct) model routing (six routers read on 2 Oct), reranking and passage filtering for RAG (five implementations read by 8 Oct), SQL row judgments and document splitting (five projects read on 8 Oct), and real-time control demos (three game and robot-simulation demos read on 12 Oct). Vendor documentation adds marketplace moderation, citation checks and fixed-set function selection. None of these patterns has an independently verified production deployment, and the projects read handle Jev failures in different ways. Across 80 implementations in 12 use cases, on a Jev error 40 stop or hold the action, 22 fall back, 10 let it through unchecked, 5 return no decision and 3 are advisory (when Jev fails: all use cases compared; eval judges counted from 7 Oct, rerankers and document pipelines from 8 Oct, two coding-agent rows from 10 Oct, three more from 11 Oct, three real-time control demos from 12 Oct and two Claude Code guards from 18 Oct).
Does your task fit?
- You can list the answers in advance (queues, labels, yes/no, a 2–10 level rubric).
- Your code, not Jev, performs the action and checks permissions.
- You have a path for low-confidence answers and API errors: a rule, another model or a person.
If any answer is no, read Limitations and fit check first; it has an interactive version of this test and the 17 weaknesses TypeSafe documents for Jev 1.13 (arithmetic, dates, long irrelevant context and others). To see these patterns running, use the demo videos, grouped by the same use cases.
Pattern picker
Patterns on this page by type of evidence
| Use case and question type | Input → Jev output → code action | Evidence and source | What can go wrong | More |
|---|---|---|---|---|
| Email sorting (Gmail) Choice, Score, boolean | Sender, subject, time, snippet → tray Choice, urgency Score, personal-mail question → classification stored with correction data | Documented Project code read: Jevmail classify.ts, schema; checked 28 Sep 2026, re-pinned at commit f6f20af on 3 Oct 2026 (unchanged). One of five email and lead sorters read in source on 3 Oct, chosen by GitHub stars (a reading order, not a sample). | Jevmail asks only for read-only Gmail access, so nothing in the mailbox changes. On a Jev error the email stays unclassified and is retried on the next run; repeated 429s pause the run and a 403 stops it (no-decision). Low confidence is only a flag. No-key /preview is a simulator, not a Jev call. | email and lead sorting · failure matrix · videos |
| Comment moderation (WordPress) Noul, Score | Held comment → spam/scam Noul and toxicity Score → site thresholds set spam, hold or WordPress's own decision | Documented Project code read: Jev Comment Triage plugin at commit d848a5b; checked 28 Sep 2026, re-read at the pinned commit 2 Oct 2026. One of six moderation projects read in source on 2 Oct: four leave the post up when Jev fails, this one holds it (content moderation) | Provider errors keep the comment held, then human review. With no provider configured, WordPress's original decision stands. Default request includes author name, URL and email. | six projects compared · failure matrix · videos |
| Trade entry gate Choice, Noul, Score | Order candidate, market data or open position → entry, side or regime judgments → the bot places, skips or sizes the order | Documented Four trading bots read at pinned commits, 4 Oct 2026, including QuantDinger ai_decision_filter.py at a5a9f4c (first audited 28 Sep from moving main links, which no longer point at the same code). All four default to mock, paper, testnet or a virtual account (trading and fraud gates). | On a Jev error, 2 of 4 can still place a new order (QuantDinger after an LLM fallback; btc-hft-jev through a rules-only path); all 4 have a path that trades without Jev. In two bots "mock" means no Jev, not no trading: the mock model still sends real orders when a wallet key is set. | seven projects compared · failure matrix · videos |
| Fraud, payment and expense checks Noul, Score, Choice | Transaction, payment or receipt fields → risk judgments → approve, step up, review or decline (returned to the caller) | Documented Three gates with a real Jev call read at pinned commits, 4 Oct 2026 (two demos on synthetic or fictional payments, one expense-policy gate) | None of the 3 has a payment call site, so no failure moves money. On a Jev error all 3 review, step up or abort; in one demo a non-numeric answer approves. Cardholder, payer and receipt text is sent to Jev. | three gates compared · data privacy |
| LLM-as-a-judge and eval scoring Noul, Score, Choice | Model output, reference or rubric fields → pass/fail, level or category judgments → eval score, ranking or gate | Documented Four judge and eval tools read at pinned commits, 4 Oct 2026 (DeepEval's JevEval and openlayer jevals as library features, two products). Published judge studies: B02, B06, B08 Reported | On a Jev error DeepEval stops the run by default; the other 3 drop the item from the score, and openlayer jevals still reports it as passed. 1 of 4 codes an escalation for unsure verdicts (opt-in). Counted on when Jev fails from 7 Oct. | four tools compared · judge domain · confidence thresholds |
| Browser and computer use Choice, Noul | Indexed page elements or screen state, goal, recent actions → operation and target Choices → click, type, tap or submit; a separate LLM types free text in some agents | Documented Seven acting agents and one recommendation-only extension read at pinned commits, 5 Oct 2026 (was one project, 25 Sep) | On a Jev error none of the 7 clicks or types, but 5 act on low-confidence answers (no threshold) and only 1 asks a person before send, buy or delete. 5 of 7 run in the user's own browser, phone or desktop session by default. | seven agents compared · videos |
| UI composition Choice | App-owned component candidates → membership and layout Choices → code validates the assembled tree | Documented Project code read: vercel-labs/json-render compose.ts; checked 25 Sep 2026 | Experimental playground with synthetic data; the candidate library limits what UI can be built. | repository guide |
| Agent completion check Choice | Edit/check ledger and a completion claim → advisory judgment → deterministic ledger still blocks unsupported "done" claims | Documented Project code read: Canny src/jev.ts; checked 25 Sep 2026 | Jev does not authorise completion. Keyed mode sends a clipped diff and final message. | products ledger · videos |
| Agent tool-call guardrail (Claude Code, Codex, OpenCode, Cursor and Hermes Agent hooks, MCP, LangChain agents) Choice, Noul, Score | Proposed tool call or shell command, plus goal, recent actions or messages depending on the guard → risk judgments → the hook denies, asks a person, or leaves Claude Code's normal permission rules in charge | Documented Project code read: 7 Claude Code and MCP guards and the official plugin at pinned commits, listed on Claude Code and MCP (29 Sep 2026; two rows added 6 Oct); 22 coding-agent integrations on coding agents (6 Oct 2026; 4 added 10 Oct) | Defaults differ: of the first five Claude Code guards, two ship in shadow or log-only mode, one is advisory, one fails open, one fails closed. A timeout shorter than the round trip turns the guard off. One sends the last 30 messages in full. Of 6 coding-agent integrations that act on Jev's answer, 4 let the tool call run when Jev errors. | failure matrix · coding agents · build it · videos |
| Coding-context selection Choice (keep/drop) | Tool history and task → keep/drop judgments → selected context kept verbatim | Reported Project README: fast-jev-compaction; checked 25 Sep 2026 | A wrong drop can hide needed evidence (see its issue #52). Project evaluations are not independent. | issue #52 · videos |
| Code-review triage Choice, Score | Diff and files → staged risk and severity judgments → local dashboard and reviewer priorities | Reported Project README: jev-review; checked 25 Sep 2026 | Not a secure automated review; no production deployment claimed. | |
| Shell and CI filter Noul, Choice, Score | UTF-8 or JSONL input → typed judgment → exit code, JSON output or escalation | Reported Project README: SemDecide; checked 25 Sep 2026 | Text is sent to TypeSafe. Provider failure has its own exit path, distinct from "false". | |
| SQL row judgments Noul, Score, Choice | Selected rows and a predicate → yes/rank → SQL filter or order | Documented Code read at pinned commits: pg-jev, sqlite3-jev, sqlite-jev; 8 Oct 2026 | A Jev error aborts the statement in all 3 extensions; each sends row text out (pg-jev sends the whole row as JSON). Set a spend guard. Never use it as an access-control WHERE clause. | five SQL and document projects compared |
| Document splitting Choice, Noul | Extracted pages → category and boundary Choices → code builds page groups | Documented Code read at a pinned commit: DocJev; 8 Oct 2026 | On a Jev error DocJev stops; a page whose boundary answer is below 0.5 joins the current document; page text is sent out. The README says it fails on unreadable OCR and that optional cloud OCR changes where data goes (Reported). | DocJev row |
| Real-time control demos one decision per tick or step | Game or simulation state → next action → the game or robot loop moves, stops or keeps its last action | Documented Real-time control demos: 3 game and robot-simulation demos read at pins; what each loop does when Jev fails. Code read at pinned commits: doom-jev bd4e98f, jev-robotics-demo 531de61, JevPlaysSnakeGame b9237bf; 12 Oct 2026. Runnable demos, not products; nothing was run. | On a Jev error, of 3 demos, 2 keep their last action (doom-jev, JevPlaysSnakeGame) and 1 stops (jev-robotics-demo). doom-jev's 2.0 s request timeout is about 17 of its 0.12 s ticks; the game keeps its last action meanwhile and drops stale answers. | five harnesses compared · when Jev fails · videos |
| Support-ticket routing Choice, Noul, Score | Ticket subject and body → queue Choice and optional urgency → code routes, or sends unsure tickets to a person | Documented Five projects read at pinned commits, 1 Oct 2026: four send unsure tickets to a person; one files tickets in an archive lane on any API error (support-ticket triage). Vendor examples: TypeSafe API, Netlify; checked 25 Sep 2026 | Tickets with no or several intents need a review option; never trigger refunds or account changes from the answer alone. | failure policies and safe design · evaluation protocol · videos |
| Model or tool tier routing Choice | Request or task features and permitted models → tier or model Choice → code calls that model, or falls back | Documented Six routers read in source at pinned commits, 2 Oct 2026 (LiteLLM and five open-source routers): on a Jev failure one sends the request to its frontier model; five keep the current model, pick locally or choose no model (model routing). OpenRouter's Jev Router does not document its failure behaviour. Was a pool proposal until 2 Oct. | A failure branch can bill your top-tier model on every outage; long prompts sent in full can fail at Jev. Measure the whole pipeline: cost, errors, fallback and review load, not one call. | six routers compared · cost per decision · videos |
| Refund, support and account-risk checks Choice, Noul, Score | Declared case state → typed judgment → application-owned routing or review | Documented Vendor example: Cloudflare Workers AI examples; checked 25 Sep 2026 | API-shaped examples, not deployed customer workloads. | |
| Marketplace moderation Noul, Choice | Listing fields → policy and contradiction judgments → code publishes, holds or rejects | Documented Vendor example: OpenRouter tutorial; checked 25 Sep 2026 | Evaluate on consented, labelled cases; keep appeal and human review. What six moderation projects read in source do on failure, doubt and appeal: content moderation (6 projects, 2 Oct). | content moderation · videos |
| Retrieval passage selection and reranking Noul, Choice, Score | Shortlisted passages and query → relevance → reordered or cut passages passed to the answering model | Documented Code read at pinned commits: 5 reranking and passage-filter implementations, 5–8 Oct 2026; vendor example: TypeSafe cookbook | On a Jev error, of 5 implementations, 2 stop, 2 fall back and 1 keeps unscreened passages. A high-ranked passage may not answer the question; keep no-answer and citation checks. | studies by type and five implementations compared |
| Semantic find Choice, Noul | Tagged lines and query → line Choice plus presence Noul → line or "absent" | Documented Vendor example: TypeSafe cookbook; checked 25 Sep 2026 | Choice alone always picks a line. The recipe uses Jev 1.12 and cached responses. | |
| Citation check Choice | Claim, quote and context → support Choice → weak citations sent to review | Documented Vendor example: TypeSafe cookbook; checked 25 Sep 2026 | The judgment is not proof; inspect the exact source. | |
| Value extraction from located spans Choice | Pre-located spans → Choice → copy and validate the exact observed value | Documented Vendor example: TypeSafe cookbook; checked 25 Sep 2026 | Cannot recover a value that is not in the text; parse and compare dates in code. | |
| Taxonomy classification with backoff Choice | Known taxonomy → branch and leaf probabilities → narrow label, or broad label / review | Documented Vendor example: TypeSafe cookbook; checked 25 Sep 2026 | Thresholds are specific to each workload. | confidence thresholds |
| Function selection from a fixed set Choice | Request → permitted function and argument Choices → code validates and gates | Documented Vendor example: TypeSafe cookbook; checked 25 Sep 2026 | Ambiguous requests should abstain; not a licence for unattended orders. | videos |
| Mobile GUI action loop Choice | VLM plan and accessibility tree → Jev action selection → guarded mobile action | Reported Author-reported study: arXiv 2609.30186 (24 Sep); checked 25 Sep 2026 | The reported AndroidWorld result was not independently replicated; no deployed app verified. | |
| Document field selection Choice | OCR-extracted candidate values → Choice among observed values plus "missing" → code validates provenance and schema | Unverified Proposed pattern: Pool proposal; checked 25 Sep 2026 | Jev cannot supply a value that is absent; OCR and storage are separate steps. |
"Project code read" means the pool read the file that calls Jev; it does not show the project is running. "Project README" means the project's own description, not checked against its code in the stated run. Rows keep their original check date; they were not all rechecked on 28 September.
Which of these tasks has published accuracy tests?
Some patterns above have been measured by third parties. The results are on the graded benchmarks ledger, each with operator, task, sample size, baseline and grade. All are Reported; the pool ran no benchmark.
- Jev as an LLM judge or evaluator (hallucination checks, preference pairs, summary quality): B02 Arize and B08 CMU; what judge tools do when Jev fails is on Jev as a judge. B02 finds Jev level with Claude Opus 5 once cutoffs are tuned on half the data; B08 finds it "within three points of GPT-6 wherever a verdict can be read off the text" and behind on math, code and logic. B06 Jevals adds yes/no, intent and score tasks graded against human labels.
- Model or tool tier routing: B05 LiteLLM (80 authored prompts; grade C, the operator ships Jev as a router option). What routers do when Jev fails is on model routing.
- Content moderation: no published accuracy test was found (content moderation).
- Retrieval passage selection and reranking: B14 Elastic (e-commerce search; reranking-pipeline study, grade B, not rerunnable at 250 queries; grade A provisional until 7 Oct). Every reranking study by type (relevance-judgment or reranking-pipeline), with grade and conflict-of-interest flags: Jev reranker and Jev for RAG.
- Email sorting: B12 Bryo (German and English business email, grade C).
For choosing between Jev, an LLM and a trained classifier for your task, see Jev vs an LLM vs a classifier.
TypeSafe's own patterns and cookbooks
TypeSafe's documentation index lists 4 patterns, 18 cookbooks and 1 demo (checked 28 Sep 2026). Six of the cookbooks appear with analysis in the pattern picker above. Figures quoted in cookbook descriptions are TypeSafe's own results, not independent measurements. Documented titles and links from docs.typesafe.ai/llms.txt; the pages were not opened one by one
All 23 official patterns, cookbooks and demos
| Kind | Title | Vendor-reported figure in the description |
|---|---|---|
| Pattern | Speculative fan-out | — |
| Pattern | Confidence-gated routing | — |
| Pattern | Composite scoring | — |
| Pattern | Intent routing (on 7 Nov 2026 this address redirects to TypeSafe's "Routing by probability" pattern, patterns/probability-routing; R27-S17) | — |
| Cookbook | Self-consistency: nouls | — |
| Cookbook | Self-consistency: choices | — |
| Cookbook | Parallel questions | "12.2x cheaper" and "11.8x faster" (one GDPR article, 13 questions; TypeSafe's cookbook, read 7 Nov 2026; earlier 10.0x faster; R27-S16) |
| Cookbook | Re-ranking | top-1 accuracy 5% to 30%, top-10 38% to 65% (40 CLERC queries; TypeSafe's own figures, read 7 Nov 2026; earlier TypeSafe's docs gave 5% to 18% and 38% to 62%; R27-S15) |
| Cookbook | Line-by-line search | 218 line IDs in one request |
| Cookbook | Structure recovery | — |
| Cookbook | Function calling | — |
| Cookbook | Skill suggestion | 182 skills (Hermes catalogue) |
| Cookbook | Knowledge-graph entity alignment | 450 candidate pairs |
| Cookbook | Classifying RAG passages | — |
| Cookbook | Double-checking citations | — |
| Cookbook | Guardrails for LLMs | — |
| Cookbook | SDE cascade | — |
| Cookbook | Date extraction | — |
| Cookbook | Pre-parsed value extraction | — |
| Cookbook | Hierarchical classification | — |
| Cookbook | Autoresearch feature discovery | — |
| Cookbook | Classification using confidence | 75 SEC industry groups |
| Demo | Smart home assistant | — |
What happens when Jev fails: three audited projects
Compare all use cases: when Jev fails counts what 80 implementations in 12 use cases do on a Jev error, a malformed answer, an unsure answer, a missing key and a bypass (recomputed 19 Oct 2026; ohmyjev and two other Claude Code guards were added on 18 and 19 Oct, three real-time control demos as a new use case on 12 Oct, three coding-agent rows were added on 11 Oct and two on 10 Oct; rerankers, a prompt filter and two document pipelines were added on 8 Oct). The table below is the original 28 Sep audit.
The 28 September audit read the code paths of three projects that call Jev. Each handles an API error, a missing key, a malformed answer and a low-confidence answer differently. Before copying a project, decide which of these behaviours your own task can accept. Documented source code read 28 Sep 2026
What three audited projects do when Jev fails
| Project | Jev API error | No key or provider not set up | Malformed or unexpected answer | Low confidence |
|---|---|---|---|---|
| Jevmail (Gmail sorting) | no-decisionRetries 429/5xx once; the email stays unclassified and is retried on the next run; repeated 429s pause the run, a 403 stops it. Read-only Gmail access, so the mailbox does not change (corrected 3 Oct) | fail-closedNo classifier fallback; /preview is a separate simulator | fallback-promotionalUnexpected category becomes 'promotional' | acts-anywayTop-two margin under 0.15 is flagged, not held; the lane is still assigned |
| WordPress Comment Triage | fail-closedStays held; after 3 failures, human review | fail-openWordPress's original decision stands; may publish | fail-openMissing fields read as 0.0; may restore approval | heldSite thresholds hold or mark spam |
QuantDinger (trade entry gate)re-read at a5a9f4c, 4 Oct: unchanged | fallback-LLMLLM fallback if set, then entry allowed (fail-open) | fallback-LLMSame: fallback, then entry allowed (fail-open) | fallback-LLMSame: fallback, then entry allowed (fail-open) | fallback-LLMSame: fallback, then entry allowed (acts-anyway) |
- fail-closed or held: no action, or a person decides
- fail-open or acts-anyway: the action goes ahead
- fallback-<what>: a named substitute
- no-decision: nothing returned; left for a later run
f6f20af (email and lead sorting). Terms and colours follow the shared legend on when Jev fails, which counts all three projects (changed 6 Oct 2026; until then this table coloured "Stops" red and "Proceeds anyway" green). Rate limits and retry guidance are on Errors and rate limits. Six Claude Code, MCP and agent-guard integrations are compared the same way on Claude Code and MCP (29 Sep), five support-ticket triage projects on support-ticket triage (1 Oct), six model routers on model routing and six moderation projects on content moderation (2 Oct).What was not verified
- No pattern on this page was run by the pool, and no accuracy figure is established for any of them.
- No audited project was observed running in production. Creator timing and cost figures are their own reports.
- Project README rows were not checked against source code in the run where they were recorded.
- Main-branch links in the 28 Sep rows can change after the check date; that audit did not pin commits. The WordPress row is now pinned at
d848a5b, Jevmail atf6f20af(3 Oct), QuantDinger ata5a9f4c(4 Oct; its 28 Sep line links had gone stale), and the support, email, routing, moderation, trading and judge pages use pinned commits.