Jev (TypeSafe AI) catalogue research history and evidence method

Every claim on this site comes from a dated research run. This page lists all reports, newest first, explains the evidence labels, and records what changed in each release, including where every earlier URL now points.

Reports by date and topic

1. 7 Oct 2026
   Judge tools counted when Jev fails, coding-agent cells settled, and new Jev benchmarks gradedBuild and integrations

   When Jev fails now counts 64 implementations in 11 use cases, including 4 eval judges and 2 Hermes Agent plugins; on a Jev error 35 stop or hold the action and 8 let it through. Coding agents: all 5 open cells settled; 8 of 18 act on Jev's answer. Benchmarks: 9 new studies graded (33 graded rows); two short-answer clauses changed. Access unchanged (checked 7 Oct 2026, 05:19 UTC). Nothing was run; no Jev call.

2. 6 Oct 2026
   Coding agents beyond Claude Code, and Jev model IDs and aliases by routeBuild and integrations

   Read 18 Jev integrations for coding agents (Codex, OpenCode, Cursor, OpenClaw, Hermes Agent) at pinned commits: 1 is vendor-published (OpenClaw) and none comes from TypeSafe; 6 act on Jev's answer, and 4 of those let the tool call run when Jev errors. Read the model IDs and aliases on 6 routes: 4 accept a versioned ID, Vercel and Cloudflare only an unversioned one; jev-latest points to jev-1.13.0 on TypeSafe's models page. The failure checklist grows to 58 implementations in 10 use cases: 32 stop or hold the action on a Jev error and 7 let it through unchecked. Products: 92 rows (A 75), none independently verified. LiteLLM 1.104.0 failure path unchanged. Vercel AI Gateway now lists TypeSafe as a second provider for Jev. Access unchanged (checked 6 Oct 2026, 05:15 UTC). No agent, plugin or hook installed or run; no Jev call.

3. 5 Oct 2026
   Browser and computer use, framework integrations, and Jev on Bedrock or AzureProjects and use cases

   Read 7 Jev browser and computer-use agents at pinned commits: none clicks or types when Jev errors, but 5 act on low-confidence answers and only 1 asks a person before send, buy or delete; 5 of 7 run in the user's own session by default. Read 6 framework packages: 4 are pre-1.0 and none comes from TypeSafe; 4 raise Jev errors to the caller, while the community Spring AI filter and the LlamaIndex reranker fail open. Jev is not listed in the Amazon Bedrock model cards (5 Oct 2026, 05:18 UTC) or the Azure AI Foundry catalog (05:19 UTC). The failure checklist grows to 51 implementations in 9 use cases: 30 stop or hold the action on a Jev error and 6 let it through unchecked. Products: 72 rows (A 55), none independently verified. Access unchanged (checked 5 Oct 2026, 05:13 UTC). No pool Jev call, agent run or install.

4. 4 Oct 2026
   Trading and fraud gates, the failure checklist across use cases, and Jev as a judgeProjects and use cases

   Read four Jev trading bots and three fraud or payment gates at pinned commits: 2 of the 4 bots can still place a new order when Jev errors, all 4 can trade without Jev, and none of the 3 gates has a payment call site. Recomputed the failure checklist across 39 implementations in 7 use cases: on a Jev error, 21 stop or hold the action and 5 let it through unchecked, with "not recorded" cells counted. Read four Jev judge tools: DeepEval stops the run on a Jev error; the other 3 drop the item from the score, and openlayer jevals still reports it as passed. Products: 67 rows (A 47), none independently verified. TypeSafe's trust center lists 6 subprocessors (Nebius and CoreWeave added 2 Oct). Access unchanged (checked 4 Oct 2026, 05:15 UTC). No pool Jev call, trade, payment or eval.

5. 3 Oct 2026
   Email and lead sorting, the oh-my-claudecode v5.6.0 check, the Products refresh and n8n nodesProjects and use cases

   Read five email and lead sorters at pinned commits: none changes the mailbox when Jev errors; of the 3 that can write to the mailbox, only 1 sends unsure mail to a Review label. oh-my-claudecode v5.6.0 keeps its failure path; its default timeout is now 2,000 ms. Recounted the Products ledger: 61 rows, 39 grade A, none independently verified in production. Identified TypeSafe's own n8n node (the scoped @typesafe-ai package); all 5 nodes read stop the workflow on a Jev error by default. Listed request limit now 80 per second; access unchanged. No pool Jev call.

6. 2 Oct 2026
   Model routing, the jev-mcp version check, and content moderationProjects and use cases

   Read OpenRouter's Jev Router docs (failure, long-context and routed-model billing not documented) and six self-hosted Jev routers at pinned commits: on a Jev failure one sends the request to its frontier model, five keep the current model, pick locally or choose no model. Checked @jkudish/jev-mcp 0.12.0 and 0.13.0: failure path unchanged. Read six moderation projects at pinned commits: four leave the post up when Jev fails, one holds it, one returns no decision; no published moderation accuracy found. Access unchanged. No pool Jev call.

7. 1 Oct 2026
   Laya vs Kev, data handling by route, and support-ticket triageAlternatives, data handling and use cases

   Compared hosted Jev, Laya and Kev on the same rows: only Opper's test (B11) finds an open model, Kev-4B, within noise of Jev. Quoted training, retention and ZDR terms for TypeSafe direct and four gateways with both layers apart: no text allows training without consent; Vercel also routes Jev through DigitalOcean, marked not ZDR. Read five ticket-triage projects in source: four send unsure tickets to a person, one archives on errors. Access unchanged; rate-limit change narrowed to 29 Sep 19:24 – 30 Sep 20:47 UTC. No pool Jev call.

8. 30 Sep 2026
   Evaluation evidence and home status: graded benchmarks, terms and accessAccuracy and evidence

   Graded 14 published Jev studies at their primary source and 10 shorter rows (A–D with flags; all results Reported); found that TypeSafe's MCA barred publishing benchmarks until its 19 Sep version; set out 10 criteria for Jev vs an LLM vs a classifier; built the home status dataset (signups open, no new-user credit, API operational; console HTTP 403). Rate limits changed to 100K tokens/s and 40 requests/s. No pool Jev call.

9. 29 Sep 2026
   Integration trust: Claude Code and MCP guardrails, TypeScript paths and testing without a keyBuild and integrations

   Read the failure paths of five Claude Code, MCP and agent-guard integrations at pinned commits, and the official plugin in full: only LangChain AutoModeMiddleware fails closed. Compared four TypeScript paths, twelve no-key test facilities and ten failure branches; ran one hook and one retry client offline. Corrected the oh-my-claudecode timeout note and the jev-mcp product row. Access and price fields unchanged. No pool Jev call.

10. 28 Sep 2026
    Release 1 question pages: access, cost, trust, build, limits and videosFull scope

    Checked signup and credit status (reopened 27 Sep, no new-user credit), route prices and limits, lookalike resellers (3–16× list price), billing basis and third-party token counts, production claims, the Python SDK, error codes, confidence semantics, 17 documented limitations, 46 demo videos and new alternatives. Corrected two wrongly 'official' project labels. No pool Jev call.

11. 28 Sep 2026
    Jev implementation-path audit and daily discoveryProjects and use cases

    Source-path audit of Jevmail, WordPress comment triage and QuantDinger distinguishes error, no-key and bypass policies from observed production use; no exact-time new item verified in the bounded 27–28 September monitor. Route lists showed no material established change.

12. 27 Sep 2026
    Seven-day Jev access checkpoint and daily discoveryAccess and price

    Vercel now shows rounded $0.04/M input instead of its 26 Sep Free display; exact tariff, Jev free-credit eligibility and bill unverified. Other route lists rechecked; QuantDinger documented as optional fail-open integration, not verified live operation.

13. 26 Sep 2026
    Post-promotion route and comparison checkAccess and price

    Vercel still displayed Free after its stated 25 Sep promotion end; effective Jev rate unknown. Cloudflare listed $0.042/M input and zero output/cached input. Added Opper route and one operator-reported Jev–Kev suite; narrowed Laya and JevAny artifact boundaries.

14. 25 Sep 2026
    Second full-scope Jev source passFull scope

    Expanded projects, use cases and layered alternatives; corrected Vercel Free promotion boundary and monitored exact preceding 24 hours. No authenticated call, independent benchmark or verified production deployment.

15. 25 Sep 2026
    Full Jev catalogue rebuild and 24-hour monitorFull scope

    Direct-source rebuild of use cases, independent Jev-like alternatives, project integrations, route pricing and access; no production product or independent benchmark verified. Monitored the preceding 24 hours with documented platform gaps.

16. 25 Sep 2026
    Support-ticket workload-evaluation protocolMethod

    Prepared an unfilled Choice evaluation method for one support-ticket route. It separates consecutive and challenge cohorts, freezes labels and comparators, includes review and failure accounting, and records that no Jev call or measured result occurred.

17. 24 Sep 2026
    Bounded freshness and change-detection passAccess and price

    Corrected the independent proxy tier record, recorded the MCA displayed Sep 23 update with exact diff unknown, dated the waitlist observation, and preserved older source-check dates and bounded negative findings.

18. 24 Sep 2026
    Builder fit and maintenance synthesisMethod

    The 24 September synthesis added a pilot, route fallback, substitute, wait, and stop branch; preserved bounded no-verified-open-weight-Jev and no-verified-production-product findings; proposed source refresh rules; and recorded that no independent workload measurement was run.

19. 23 Sep 2026
    Bounded adoption and product verification auditProjects and use cases

    Direct-source audit of killmyidea, Jev Showcase, the TypeSafe playground router, jev-usecases, and the Awesome Jev lead index. killmyidea is a direct-source application with public-deployment metadata but unverified public access and adoption; the other leads remain an unofficial integration, an explicitly mocked demo, an MIT reference harness, or an index. No verified production product was found.

20. 22 Sep 2026
    Alternatives substitution matrix refreshAlternatives

    Direct-source audit of exactly six named alternatives, separating model parity, System One-like interface, local/privacy boundary, workflow role, and trade-offs; no verified open-weight Jev model was found.

21. 21 Sep 2026
    Alternatives verification refresh (local-only)Alternatives

    A local-artifact-only refresh of the six named alternatives retained direct URLs and substitution layers, upgraded no row, marked all current project facts unknown, and made the no-verified-open-weight-Jev finding explicit without implying that none exists.

22. 20 Sep 2026
    Access, pricing, and availability refresh for JevAccess and price

    Current official Jev model, endpoint, aliases, price, limits, text-only input, access ambiguity, legal/data terms, OpenRouter and Cloudflare routes, inaccessible Vercel status, and the independent proxy price were checked; remaining account and gateway gaps stay explicit.

23. 20 Sep 2026
    Bootstrap evidence pass for JevFull scope

    First source-led catalogue of Jev use cases, open-source substitutions, directly evidenced projects, and dated access/pricing routes.

Evidence labels

Release log

What changed, by date

This dated list used to be on the home page. The home page now shows only current notices; past changes stay here.

Release notes: 7 October 2026

• Counting rule for judges (decided 7 Oct): a judge or eval tool whose own code turns Jev's answer into the score, pass/fail result, ranking or gate that its run, report or CI uses is an action and counts on when Jev fails. An error that becomes a pass is fail-open; a dropped item is no-decision. Reason: a judge that turns a failed Jev call into a pass lets a broken release through as surely as a hook that lets a tool call run. DeepEval counts although it is a library, because its own test runner aborts the run or fails the test case. The Jevals harness is not counted: its code is not public.
• When Jev fails: 6 new rows (4 judge tools, 2 Hermes Agent plugins) and a new "LLM as a judge" domain; n = 64 in 11 use cases. On a Jev error: fail-closed 35, fail-open 8, advisory 3, fallback 13, no-decision 5. The action goes ahead in 13 of 64. The key finding was edited in publication review to name every category (35 + 13 + 8 + 8 = 64). The copy block was regenerated, and the line saying judges were "not yet in this count" was removed.
• Jev as a judge: three cells that read "conditional" now show the counted term with the condition in the note, so that both pages show the same cell and colour: openlayer below threshold (no-threshold) and no key (not recorded), and did-they-answer malformed (fail-closed). deepeval 4.2.8 and jevals 0.1.4 are still the latest releases.
• Coding agents and Products: jev-approvals and jev-curator are counted; both do nothing until switched on. opencode-tool-gate's "Yes" is confirmed under OpenCode's default rules; rh-guard's default mode is enforce; osuki was checked at 0.2.8 (unchanged). Products moved the two Hermes rows to the counted group (8 act, 10 not counted); row and grade counts are unchanged.
• Benchmarks: rows B28–B36 added and Elastic B14 regraded from A (provisional) to B. The title was changed to the searched phrase "jev benchmark" (Bing and DuckDuckGo completions, and "jev 1.13 benchmarks" on Google, rechecked 7 Oct 2026, 05:18:59 UTC). TypeSafe's legal texts still have no clause about publishing benchmarks. Reranking studies: 2 of 2 needed found beyond Parallel (B04) and the earlier MindStudio lead, so the condition for a reranking and RAG page is met for the next run.
• Drift: ai 7.0.130 and @ai-sdk/typesafe-ai 3.0.15 noted on TypeScript and frameworks; oh-my-claudecode v5.6.2 and @jkudish/jev-mcp 0.14.1 noted on Claude Code and MCP as not yet read; six framework packages and LiteLLM 1.104.0 unchanged.
• Cross-page consistency: the four judge rows carry the same cell, term and colour on Jev as a judge and when Jev fails. The settled coding-agent cells match on coding agents, when Jev fails and Products. osuki's version (0.2.7 read, 0.2.8 checked) is stated the same way on all three. The hubs (Build, Use cases), confidence thresholds and Home carry the new counts.
• Home and access status: rendered from the run's status file: status unchanged, checked 7 Oct 2026, 05:19 UTC; the API's 90-day uptime moved from 99.826% to 99.828% with no new incident listed. Four notices shown; the Vercel provider notice is current but not shown; typesafe-workflow-evals-code retired. The access timeline has a 7 Oct recheck row.
• Tools: both tools were opened in a headless browser at tools/<name>.html, at tools/<name>/ and inside their pages at 1280 and 390 pixels wide. Every button and field was used: all 8 answer combinations, the two-fix case, full and partial reset and the Enter key in the fit check; every field with checked results, seven out-of-range values, reset and typed input in the cost worksheet. 294 of 294 checks passed, with no script error and no failed load of a pool file.
• Evidence: 7 Oct research report.

Release notes: 6 October 2026

• New pages: coding agents (18 integrations at pinned commits: 6 counted rows in the shared failure terms with a "does the tool call still run?" column, 12 not counted, split 8 / 2 / 2; the Hermes catalog's installed commits; jev-judge is shadow by default and its cells describe enforce mode; osuki 0.2.8 not read) and model IDs and aliases by route (6 routes with their own dates, or "undated; read at" where the source shows none). Release 3 is complete (9 of 9); release 4 has 1 of 4.
• Editorial correction to the coding-agents key finding: the research draft said "12 only advise the agent". Of the 12 integrations not counted, 8 return Jev's answer to the agent, 2 give advisory hints and 2 are not settled (jev-approvals hands its verdict to Hermes core, which was not read; jev-curator was only partly read). The page and the report use the corrected sentence.
• Claude Code and MCP: two guard rows (jev-permission-gate counted, fallback to the built-in classifier; jev-llm-guard "returns verdict; not a decision"), the count sentence now reads seven guardrails, a pointer to coding agents for jev-kit, jev-review and mcp_jev, and the oh-my-claudecode v5.6.2 tag note.
• When Jev fails: 7 new rows (1 Claude Code guard, 6 coding agents) and a new "Coding agents" domain; totals recomputed for n = 58 in 10 use cases; the action goes ahead on a Jev error in 12 of 58; the copyable checklist was regenerated. The "without trading" sub-row was removed because it no longer matched the domains.
• Products: 92 rows (A 75, B 16, C 1), 20 rows added; two older B rows with similar names are different projects and stay.
• Where to use and data privacy: Vercel lists two providers for Jev (TypeSafe AI, zero data retention; DigitalOcean, no zero data retention) and picks by uptime and latency unless you set the order; route fields re-dated only where rechecked on 6 Oct. Links to the model-ID page from the route matrix, TypeScript, Python and access status.
• Drift: LiteLLM 1.104.0 failure path unchanged (with the commit-order caveat) on model routing; framework version box rechecked; ai 7.0.128 and @ai-sdk/typesafe-ai 3.0.13 noted on TypeScript.
• Cross-page consistency (new standing check): every failure verdict cell on 13 pages was compared with the row of the same pinned commit on when Jev fails (support-ticket triage, model routing, content moderation, email and lead sorting, trading and fraud gates, Jev as a judge, browser and computer use, n8n, Claude Code and MCP, framework integrations, coding agents, when Jev fails, Products) plus the Use cases hub and the errors page. Fixed: Claude Code and MCP, the Use cases hub table and the errors-page disposition column used the reverse colours (stops red, proceeds green); support-ticket triage, model routing, content moderation and n8n used their own colour schemes and plain words; Products tallies and chips used blue for fail-closed and amber for fallback; browser below-threshold "no-threshold" cells were red; QuantDinger and btc-hft-jev fallbacks were red on trading; fraud-jev's conditional malformed cell was amber; ilyamk's unsure cell was red. All now use the shared legend and lead with the shared term; value columns (timeouts, data sent, appeal) are no longer coloured. LangChain AutoModeMiddleware now links commit 4af7ab8 (tag langchain-typesafe==0.0.1a3) instead of PyPI on when Jev fails, Claude Code and MCP, and Products. Open: oh-my-claudecode is shown as a heuristic fallback on Claude Code and MCP (the path once Jev is asked) but counted as advisory on when Jev fails (its default shadow mode), and Products links its v5.6.0 commit while the other pages link v5.5.0; Jev as a judge's five rows are not on the checklist.
• Home and access status: rendered from the run's status file: status unchanged, checked 6 Oct 2026, 05:15 UTC; the API's 90-day uptime moved from 99.829% to 99.826% with no new incident listed. Four notices shown; the new Vercel provider notice and typesafe-workflow-evals-code are current but not shown. None retired. The access timeline has a 6 Oct recheck row. Home task paths link the two new pages.
• Tools: both tools were opened in a headless browser at tools/<name>.html, at tools/<name>/ and inside their pages at 1280 and 390 pixels wide. Every button and field was used: all 8 answer combinations, the two-fix case, reset and the Enter key in the fit check; every field with checked results, six out-of-range values, reset and typed input in the cost worksheet. 220 of 220 checks passed, with no script or load error and no scroll inside either frame. Only the cost worksheet's price check date changed (6 Oct 2026, 05:15 UTC). No new tool this release.
• Evidence: 6 Oct research report.

Release notes: 5 October 2026

• New pages: browser and computer use (seven acting agents and one recommendation-only extension at pinned commits; a "does a failure still click, type or submit?" column; trycua/cua listed, not counted) and framework integrations (six packages at pinned tags or commits; a maturity matrix; coding agents beyond Claude Code are planned for a later page).
• Where to use: new section on Amazon Bedrock and Azure AI Foundry with a route-matrix row; availability is written as "not listed in the catalog at the check time", with catalog links. The lookalike reseller's Bedrock page is cited only as a reseller fact. The "Access today" line now matches Home (checked 5 Oct 2026, 05:13 UTC).
• When Jev fails: 12 new rows (8 browser, 4 frameworks), totals recomputed for n = 51 in 9 use cases; the four counting judgements stay visible on the page; a copyable checklist in plain text repeats the eight questions and their counts.
• Products refresh: 72 rows (A 55, B 16, C 1); 3 rows upgraded from B to A, 5 added, jev-ultrafast re-pinned; six framework packages listed as not counted.
• Home and access status: rendered from the run's status file: status unchanged, checked 5 Oct 2026, 05:13 UTC; the API's 90-day uptime moved from 99.826% to 99.829%. Four notices shown; typesafe-workflow-evals-code stays current at rank 5, not shown. None retired. The access timeline has a 5 Oct recheck row.
• Small additions and re-dating: Zapier note on the n8n page, EU-hosting line on data privacy, Strands Decider 2B on alternatives, tool-documentation links on Jev as a judge; limits re-dated to 5 Oct on errors and rate limits and limitations; list price re-dated on the cost worksheet. Links to the new pages from the Use cases and Build hubs, Python, TypeScript, Claude Code and MCP, model routing, videos, official vs reseller and data privacy.
• Tools: both tools were opened in a headless browser at tools/<name>.html, at tools/<name>/ and inside their pages at 1280 and 390 pixels wide. Every button and field was used: all 8 answer combinations, reset and the Enter key in the fit check; every field with checked results, six out-of-range values, reset and typed input in the cost worksheet. 250 of 250 checks passed, with no script or load error and no scroll inside either frame. Only the cost worksheet's price check date changed (5 Oct 2026, 05:13 UTC). No new tool this release.
• Evidence: 5 Oct research report.

Release notes: 4 October 2026

• New pages: trading and fraud gates (seven projects at pinned commits; a "does the failure reach the order or payment?" column; QuantDinger's order call sites traced), when Jev fails (counts recomputed from the 7 published domain tables, n = 39, with "not recorded" counts and an 8-question checklist) and Jev as a judge (four tools at pinned commits; benchmark rows B02, B06 and B08 copied unchanged).
• Products refresh: 67 rows (A 47, B 19, C 1); 2 rows upgraded from B to A, 6 added, QuantDinger re-pinned; 2 library features listed as not counted.
• Home and access status: rendered from the run's status file: status unchanged, checked 4 Oct 2026, 05:15 UTC. By the manager's decision, subprocessors-added-2026-10-02 is shown at rank 4 and typesafe-workflow-evals-code stays current at rank 5, not shown (it retires 6 Oct). None retired. The access timeline has a 4 Oct recheck row.
• Corrections and updates: email and lead sorting (two relabels, cost range), Claude Code and MCP (advisory wording for oh-my-claudecode at default settings; v5.6.1 note), testing without a key and the Use cases hub (QuantDinger, jev-trader and Jev Trade re-read at pinned commits; the hub now lists 26 patterns, 11 with project code read), data privacy (subprocessors, SOC 2 listed but not seen, a Netlify docs sentence), re-dated checks on Where to use, errors and rate limits, limitations and cost per decision. Links to the new pages from the Build and Use cases hubs, the six domain pages, confidence thresholds, benchmarks, Jev vs LLM vs classifier, videos and errors and rate limits.
• Tools: both tools were opened in a headless browser at tools/<name>.html, at tools/<name>/ and inside their pages at 1280 and 390 pixels wide. Every button and field was used: all 8 answer combinations and reset in the fit check; every field and reset in the cost worksheet. 184 of 184 checks passed, with no script or load error and no scroll inside either frame. Only the cost worksheet's price check date changed (4 Oct 2026, 05:15 UTC).
• Evidence: 4 Oct research report.

Release notes: 3 October 2026

• New pages: email and lead sorting (five projects read at pinned commits; cost per 1,000 emails by pool arithmetic, shown apart from a Reported demo figure) and the Jev n8n node (official node and four community nodes read at pinned commits; npm maintainer and provenance records). Each opens with the run's key finding, its label and check date, and uses the title proposed from search language.
• Products refresh: 24 new grade-A rows (triage, routing, moderation, email and lead, n8n), claims rechecked, repository status checked, LiteLLM listed under "not counted as products".
• Home and access status: rendered from the run's status file: status unchanged, checked 3 Oct 2026, 05:16 UTC; the notice rate-limits-changed was retired as superseded and rate-limits-changed-2026-10-03 (80 requests per second) added; the other three notices kept. One timeline event added.
• Corrections and updates: Claude Code and MCP (oh-my-claudecode v5.6.0 row, version line and check 6), errors and rate limits (80 requests per second with its dated bracket; the 2,000 ms oh-my-claudecode default released in v5.6.0), limitations (rate-limit row), Where to use (limit and check dates), official vs reseller (n8n npm publisher answered), Use cases hub (Jevmail row). New pages are linked from their hubs and related pages.
• Tools: both tools were opened in a headless browser at tools/<name>.html, at tools/<name>/ and inside their pages, and every button and field was used. No tool content changed.
• Evidence: 3 Oct research report.

Release notes: 2 October 2026

• New pages: model routing (OpenRouter's Jev Router docs read with explicit "not documented" results; six routers read at pinned commits) and content moderation (six projects read at pinned commits; Noul-per-harm and Choice compared from TypeSafe's docs). Each opens with the run's key-finding sentence, its label and check date, and uses the title proposed from search language.
• Home and access status: rendered from the run's status file: status unchanged, checked 2 Oct 2026, 05:14 UTC; all four notices kept (none retired); API 90-day uptime 99.826%.
• Corrections and updates: Claude Code and MCP (version check replaces the drift note; stale oh-my-claudecode line fixed), TypeScript (ai 7.0.127), Where to use (Jev Router row links model routing; check dates), data privacy (check dates), Use cases hub (routing row, moderation rows, WordPress commit), and links from benchmarks, confidence thresholds, limitations, support-ticket triage and videos. The Jev vs Laya vs Kev page's heading now starts with "Jev vs Laya", as people type it; its URL is unchanged.
• Tools: both tools were opened in a headless browser at tools/<name>.html, at tools/<name>/ and inside their pages, and every button and field was used. Only the cost worksheet's price check date changed.
• Evidence: 2 Oct research report.

Release notes: 1 October 2026

• New pages: Jev vs Laya vs Kev (licence stack, context and API shape documented for all three; three graded rows compare at least two), data handling by route (TypeSafe direct and four gateways quoted at both layers; Netlify not read) and support-ticket triage (five projects read at pinned commits). Each opens with the run's key-finding sentence, its label and check date.
• Home and access status: rendered from the run's status file: status unchanged, checked 1 Oct 2026, 05:16 UTC; all four notices kept (none retired); the rate-limit notice now shows the narrowed time range.
• Corrections: where-to-use (Vercel data cell; two caveats replaced with links to the data page; each route's data cell rewritten), alternatives (Kev and Laya rows, he-jev note), rate-limit dating (access status, errors and rate limits, limitations, Home), version-drift notes (TypeScript, Claude Code and MCP), and the residency row on Jev vs an LLM vs a classifier.
• Tools: both tools (fit check and cost worksheet) were opened in a headless browser at tools/<name>.html, at tools/<name>/ and inside their pages, and every button and field was used; the tools themselves were not changed.
• Evidence: 1 Oct research report.

Release notes: 30 September 2026

• Tool fix (owner request of 29 Sep): the host serves tools/fit-check.html at tools/fit-check/, so the relative path ../assets/fit-check.js pointed to a file that does not exist; the same held for the cost worksheet. Both tools now load their own script from /p/jev-catalog/assets/…, which resolves the same way at either address; the shared resize helper is loaded as before. Tested in a headless browser at tools/<name>.html, at tools/<name>/, through an imitation of the host redirect and inside each embedding page: every button and field changed the output and no script failed to load.
• New pages: benchmarks (12 studies met the plan's rule of at least 8 read at the primary source with operator, task, n, baseline and materials; 9 compare Jev with a baseline) and Jev vs an LLM vs a classifier (10 criteria with cited evidence, 13 head-to-head rows).
• Home: rebuilt around the owner's order: access status, notices, short answer, direct paths. The status matches the access status page exactly (same values and check time).
• Corrections: rate limits (where to use, errors and rate limits, limitations); the Molas row on confidence thresholds; the home chart's accuracy basis. No page of this site had presented TypeSafe's 67.8% as accuracy; the benchmarks page now explains it.
• Re-dated only (rechecked, unchanged): TypeSafe price and aliases; OpenRouter, Vercel, Cloudflare and Opper listings; typesafe-sdk 0.7.2 and @typesafe-ai/sdk 0.6.0; the six Claude Code and MCP integrations. Netlify, reseller prices and other fields keep their earlier dates.
• Evidence: 30 Sep evaluation-evidence report.

Old-to-new URL map

No page moved in release 1 or on 29 September, so no redirects are needed. Every earlier address still opens the same topic. Base address: https://shaduf.ai/p/jev-catalog/.

Release notes: 29 September 2026

• New pages: Build hub, TypeScript, Claude Code and MCP guardrails and testing without a key. Each met the research plan's rule for shipping. For example, the Claude Code page needed four or more integrations read at a pinned commit; six were read.
• Corrections: oh-my-claudecode's 2000 ms Jev timeout exists only on its dev branch; v5.5.0 and main still default to 250 ms. The Products row "jev-mcp" is now "jev-mcp (jkudish)", regraded from B to A and marked advisory.
• Unchanged, re-dated: signup status, new-user credit, status page, model aliases, the TypeSafe price, the OpenRouter, Vercel, Cloudflare and Opper listings, and both SDK versions.
• Evidence: 29 Sep integration-trust report.

Release notes: release 1, 28 September 2026

The owner asked for an evidence-backed resource ordered by reader question. Release 1 adds or substantially rebuilds 13 pages and lightly extends two. The reason for each page is in the URL map below. Pages whose evidence was not ready were not published thin: the planned pages on TypeScript, Claude Code and MCP, benchmarks, data handling by route, ticket triage and model routing are scheduled for later releases.

• Corrections in this release: two project rows were labelled as official TypeSafe repositories but come from the unofficial github.com/TypeSafeAI community organisation; Vercel's exact price is now documented ($0.042 per million in its models API); the "250 ms default timeout" is not TypeSafe's; the confidence formula 2·p_top − 1 applies to two-option Choices only; the $5 new-user credit rests on secondary reports with a conflicting first-hand report.
• New tool: the cost per decision worksheet, explained on its page.
• Evidence for this release: release 1 research report.

How pages were chosen: no search data

Search impressions, clicks, query coverage and indexing data, Search Console, keyword volumes and site analytics were not available for release 1 (28 September 2026), and were still unavailable for the 29 and 30 September and 1, 2, 3, 4, 5, 6 and 7 October runs (the 6 October run input again records analytics as unavailable). Topics were chosen from qualitative signals: search autocomplete, titles published by independent sites, Hacker News and GitHub issue threads, and video suggestions. No page on this site claims a search volume. When search data becomes available, it will be used to choose and rank the next pages.

Audience research refreshed 4 Oct 2026: autocomplete on Google, YouTube, Bing and DuckDuckGo plus competitor and forum checks; search analytics still unavailable; Reddit and X not accessible.

Evidence gaps after the 7 October run

1. Nothing was run. No eval, judge, agent, hook, plugin or benchmark was installed or run. The four judge rows were mapped from the 4 Oct reads at the 4 Oct pins, not re-read.
2. Cells not recorded. openlayer jevals' default no-key branch (one file); jev-curator's malformed, below-threshold and no-key cells and its cache lifetime; the step in which Hermes core hands the question to the plugin; the osuki 0.2.8 tarball against its tag.
3. Benchmarks. B28: n per set and code; B35: code; B36: cohort list and judge.py; LanceDB: benchmark script; the Hugging Face and LanceDB post dates.
4. Drift diffs due. oh-my-claudecode v5.6.2 and @jkudish/jev-mcp 0.14.1 (not read; the helper was interrupted).
5. Leads not read. About 15 new coding-agent and MCP leads; 7 more Hermes catalog jev-* plugins; OpenAI's Decisions API (an alternatives lead); self-run eval repositories; real-time-control harness code.
6. Scan coverage. Analytics still unavailable; Reddit (403) and X (not attempted) unknown.

Evidence gaps after the 6 October run

1. Nothing was run. No coding agent, plugin, hook, skill or MCP server was installed or started; two npm packages were downloaded and read, not executed. Every failure cell is read from source at a pinned commit.
2. Host behaviour after a hook decision. Codex, Cursor, Hermes core and OpenClaw core were not read, so what each host does with a hook's answer (or an exit code) is not recorded. OpenCode's default permission configuration, rh-guard's default mode, thresholds and redaction, how Hermes core handles jev-approvals errors, and jev-curator's plan path are not recorded.
3. Versions not read. @osuki-dev/opencode-osuki-agent 0.2.8 (5 Oct) and whether the 0.2.7 tarball matches its tag; test-file contents; commit dates for four MCP repositories.
4. Model IDs. What each route actually serves (no request was sent); whether OpenRouter accepts patch-level IDs; route changelogs (only TypeSafe's was checked, HTTP 404); SDK default models beyond TypeSafe's statement.
5. Leads not read. About 20 coding-agent leads, including 7 more Hermes catalog jev-* plugins and 0xNatoshi/jev-codex-router (a model router); abaljeu/jev-mcp ("for thejevai.com", not a TypeSafe domain); the reranking-rag condition check; database and document error behaviour; the 6-byte change to TypeSafe's legal.md (legal "Last updated" dates unchanged).
6. Scan coverage. Analytics still unavailable; Reddit (403) and X (login) unknown; the Internet Archive returned a connection error.

Evidence gaps after the 5 October run

1. Nothing was run. No browser or computer-use agent, framework package or cloud console was run, installed or logged into; every failure cell is read from source at a pinned commit, and live behaviour may differ (for example the SDK defaults that LangChain and Pydantic AI inherit were not read).
2. Unknown cells. Whether computer-use-jev sends [secure] field values; second-look's fallback card wording; trycua/cua's no-key behaviour; the Spring auto-configuration lines (read by a helper, not spot-checked). jev-agent-browser's npm package was not compared with its repository.
3. Catalog boundaries. AWS Marketplace and Azure Marketplace results were read on the first page only; private offers and signed-in listings were not checked.
4. Leads not read. LiteLLM 1.104.0 (a TypeSafe proxy guardrail hook; targeted diff due), more same-name browser repositories, the error behaviour of five database and document projects, obekt-terminal, the WSJ copycats story (paywalled). Coding-agent facts (Hermes, Codex, OpenCode, Cursor) are web-search snippets: Unverified until read.
5. Analytics and scan coverage. Analytics still unavailable; Reddit (403) and X (login) unknown; TikTok, YouTube search, Lobsters, Product Hunt and Discord not attempted.

Evidence gaps after the 4 October run

1. Not recorded cells. Across the 39 implementations on when Jev fails: no-key behaviour not recorded for 17, bypass for 22 (mostly pages published before the bypass column), below threshold for 5 (n8n nodes).
2. Trading. No live order, fill, return or payment by any project was observed. jev-trader's no-key path runs inside @ai-sdk/typesafe-ai (not read). Whether QuantDinger's worker upgrade of a queued signal order to live skips the filter was not traced end to end; its LLM fallback prompt was not read.
3. Judges. did-they-answer's abort on a malformed answer was traced in source but not run; competitor-hunter loads whatever Jev caller a user has installed (not pinned); openlayer jevals' PyPI version was not checked; the Jevals harness code is still not public.
4. Data handling. TypeSafe's SOC 2 report is behind "Request access" and was not seen; what Slack and Google Workspace process for TypeSafe is not recorded.
5. Leads not read. Three more jev-trader repositories, caiovicentino/jev-risk-check-provider and other fraud classifiers; TradeRank not found; JevTRPG, the "Judge Jev" post, DeepEval's ConversationalJevEval; jev-safety-eval, jev-zh-tw-eval and system-one-security (for a later benchmarks refresh).
6. Carried. The two VentureBeat-relayed Products claims (HTTP 429 again); the two guard rows read in spare time wait for the next Claude Code and MCP edit.
7. Analytics and scan coverage. Analytics still unavailable; Reddit (403) and X (login) unknown; YouTube, TikTok and GitHub code search not attempted.

Evidence gaps after the 3 October run

1. Rate-limit timing. The change from 40 to 80 requests per second is known only to a 24-hour bracket (2 Oct, 05:14 UTC to 3 Oct, 05:16 UTC); there is no archive capture and the docs changelog returns HTTP 404.
2. Email. ilyamk's granted OAuth scope (no manifest in the repository); spam-leadgen's no-key behaviour and model (inside the SDK); three email projects set no HTTP timeout; whether question text is billed and the tokens-per-character ratio are assumptions in the cost table. withoneai/mastra-inbound-router, anshulkhare7/jev-gmail-triage and the n8n Gmail-leads workflow from the Nyarko video were not read.
3. n8n. How n8n core handles the node's error-output setting is known only from the official node's code comment; 10 of the 15 Jev/TypeSafe n8n packages found on npm were not read; the "declined verification" claim is Reported only.
4. Products. The two VentureBeat-relayed claims were not rechecked; grade-A repository status was checked through HTML pages, not the API.
5. Helper reads. Two helper agents read source in parallel; their key lines were spot-checked against the raw files, not re-read line by line. Unchecked cells keep their path and line references.
6. Leads not read. Red Hat benchmark, Synthpop calibration, a Substack calibration post, Vercel's "Jev for Python Engineers", madisonrickert/jev-permission-gate.
7. Carried. TypeSafe's trust page (needs a browser), Netlify data terms. jev.bevel.software ("EU/Jev") now answered from its own pages: a Jev-compatible API serving another model, not a Jev reseller (Reported, the operator's own claims).
8. Analytics and scan coverage. Analytics still unavailable; Reddit (403), X (login), TikTok and YouTube not read in the scan.

Evidence gaps after the 2 October run

1. Jev Router. What OpenRouter's Jev Router does on a Jev error, timeout, malformed answer or long conversation, and how it bills the routed model, is not documented; only OpenRouter can answer.
2. No live behaviour. None of the six routers or six moderation projects was run; moderation accuracy for Jev is not published.
3. Version drift. oh-my-claudecode v5.6.0 (1 Oct) not yet checked; ai 7.0.127 not re-audited.
4. Leads not read. kerpopule/hermes-jev-skills, nine routing repositories created on 1–2 Oct, Jevdit, the project behind Twittesia issue #261, and other moderation repositories.
5. Carried. TypeSafe's trust page (needs a browser), Netlify data terms, whether jev.bevel.software resells TypeSafe's Jev.
6. Analytics and scan coverage. Analytics still unavailable; Reddit, X and TikTok blocked; GitHub code search not run.

Evidence gaps after the 1 October run

1. Analytics still unavailable. No search impressions, queries or site analytics were available to choose or rank pages.
2. Signup not observed. The console still returns HTTP 403; X cannot be searched for newer TypeSafe posts.
3. Data handling. TypeSafe's subprocessor list and any SOC 2 report sit behind a script-only trust page; whether DigitalOcean on Vercel forwards to TypeSafe; whether OpenRouter's ZDR listing rests on a special agreement; a TypeSafe retention period, human access and a Telemetry opt-out are not stated; Netlify not read.
4. Alternatives. Which Kev checkpoints JevBench ran; Kev-9B's base licence; the JevAny/AnyJev identity; who operates he-jev; the "jev vs laya" videos (titles only).
5. Ticket triage. Live operation of the five projects; the JS SDK's default timeout; 70 or more triage repositories not read.
6. Version drift. ai 7.0.126 and @jkudish/jev-mcp 0.12.0 not re-audited.
7. Scan coverage. Reddit, X and TikTok blocked; GitHub code search and trending not run.

Evidence gaps after the 30 September run

1. Signup not observed. The console returned HTTP 403; there is no second first-hand signal after 28 Sep; X cannot be searched.
2. Rate-limit change time unknown. No docs changelog records when the limits changed.
3. Provisional grades. The linked code of B13, B14, B15 and B24, and JevBench's code, were not opened; B08's reproducibility package was not located.
4. Unidentified or unextracted results. MindStudio's primary study (B10); B14's Jina reranker row; LanceDB's results.
5. Missing comparisons. No graded head-to-head with a fine-tuned classifier or zero-shot NLI with published materials; no graded non-English study except Bryo (grade C).
6. Unchecked items. OpenAI's reported decision API at its own source; who operates LocalLLaMA/typed-decisions; jevcal's terms claim; the publisher of the npm package n8n-nodes-typesafe-ai; Vercel's provider-terms pass-through.
7. Scan gap. 13 h 40 min between the 29 Sep window and the 30 Sep window were not scanned on GitHub, npm, Hugging Face or DEV.to; Reddit, X, TikTok, Lobsters and Product Hunt stayed blocked.

Evidence gaps after the 29 September run

1. No live behaviour. No integration was run against the real API. The danna-zhou hook's 401/403 path is read from code; whether TypeSafe rejects model: "kev-latest" is unknown.
2. Unread dependencies and files. jkudish/jev-mcp's TypeSafe-direct transport (@jkudish/jev-agent-tools), jev-claude's rule list and Codex adapter, and oh-my-claudecode's per-point hook events were not read. The danna-zhou MCP server's behaviour on a hanging endpoint was not tested.
3. SDKs not executed. The official JS SDK was not installed (the Node test mirrors its defaults in pool code); the Python SDK install is still not done. The AI SDK's default timeout is undocumented, and its first version with experimental_evaluate conflicts (7.0.103 by package contents, 7.0.105 in Vercel's guide).
4. Access. The console still returns 403; the credit status rests on the 27 Sep CEO post and the absence of later statements.
5. Scan coverage. Reddit, X, TikTok and Lobsters were blocked; YouTube, GitHub trending and GitHub code search were not attempted; new hook and MCP repositories are leads only.

Evidence gaps after release 1

1. Account facts. Whether a new account gets any credit today, live quotas and an actual bill are unknown; the console returned HTTP 403 and the pool does not create accounts. No primary TypeSafe source gives the "$5 / 120M tokens" amount.
2. No live Jev behaviour observed by the pool. Status codes, headers and rate-limit scope are reported by third parties; the SDK was not executed; model="jev-1.13" was not tested.
3. Confidence. The Score confidence formula for 4 or more levels is unknown; the independent calibration tests each cover one dataset and one model version.
4. Billing. Whether question and criteria text and failed requests are billed is undocumented; how OpenRouter's Jev Router bills the routed model, and what it does on failure, is undocumented.
5. Data terms. OpenRouter's retention and logging for Jev, the text of TypeSafe's MCA, privacy policy and DPA, and the billing entity for direct accounts were not reviewed (planned for a later release).
6. Resellers. jev-agent.com was not checked; jev-ai.pro's monthly prices were not visible; flagged lookalike repositories and Spaces were not opened.
7. Production claims. TechCrunch's Vercel article was not checked; there is no primary source for a Notra claim; FundaAI's method is paywalled; VentureBeat was read only through a summarising tool.
8. Videos. No frames or transcripts were viewed; several videos do not state the route or SDK; whether one trading video used a real key is unknown.
9. Alternatives. Ollaya's per-model licences and benchmark origin, Laya's System One compatibility, Kev's base-model licences, and whether AnyJev is the catalogue's JevAny are unchecked; six new leads are unverified.
10. Scan coverage. Reddit, X, TikTok, Lobsters and GitHub code search could not be read; DEV.to coverage is complete only for the jev tag; npm and Hugging Face counts are capped.
11. Benchmarking terms. A tool author's claim that TypeSafe's agreement restricts publishing benchmark results was not checked.

Standing 24-hour discovery scan

Every research run checks what appeared in the preceding 24 hours. For the 7 October run the window was 6 Oct 2026, 05:17 UTC to 7 Oct 2026, 05:17 UTC, plus a backfill of the 3 minutes after the 6 October window (05:14–05:17 UTC), which found nothing. The main checks ran at 05:18–05:21 UTC.

No TypeSafe release, model or alias change, route ID change, legal-date change, limits change, subprocessor change or incident was found.