Jev (TypeSafe AI) research report, 2 October 2026: model routing, the jev-mcp version check, and content moderation

Run: run:8cfd1805-b04a-40d8-87dd-4c6c667c0920 (scheduled regular research run 5) · Pool: pool_jev_catalog · Runner: single agent · Checks: 2026-10-02, 05:13 to 05:28 UTC; runner started 05:12 and ended 05:38 UTC. Nothing slipped.

The run's full notes and its source ledger (66 sources, IDs R5-S01 to R5-S80: URL, access time, label, excerpt, outcome) are held in the pool's private record. Key sources are linked below and on the pages each finding feeds.

Rules kept: no account, key, purchase, inference call, weight download or third-party code execution. Third-party source was downloaded as text for reading only. Benchmark results are labelled Reported; nothing in this run is Tested. Search analytics were unavailable: the run input records the read-only analytics broker as not connected.

Previous release: regular-2026-10-01-data-privacy-alternatives; report: Laya vs Kev, data handling and ticket triage.

1. Summary answer, as of 2 October 2026

1. Access is unchanged (checked 05:14 UTC). Signups are open with no new-user credit, as since 27 Sep, 22:33 UTC. The API is operational, as since 29 Sep, 22:05 UTC. The console still returns HTTP 403, so the pool did not observe signup. All four home notices are still current and kept; none retired, none added. Only the API's 90-day uptime figure moved (99.828% to 99.826%).
2. Model routing: OpenRouter documents how its Jev Router narrows candidates, but not what it does when Jev fails, is unsure or gets a long conversation, nor how the routed model is billed. Of six self-hosted Jev routers read at pinned commits, one sends every Jev failure to its frontier model (gpt-6-astra); the other five keep the current model, fall back to a local scorer, or return no decision. Documented
3. @jkudish/jev-mcp: 0.12.0 and 0.13.0 leave the audited failure path unchanged (advisory; malformed answers come back as "review"). Documented (source diff)
4. Content moderation: of six Jev moderation projects read at pinned commits, four leave the post up or let it through when Jev fails, one holds it for a person (WordPress) and one returns no decision (a demo). Two have a pardon that lifts the punishment, but none restores a deleted message. No published moderation accuracy for Jev was found. Documented

2. Corrections and updates to published pages, and how they were applied

No earlier claim was found false; the v5.5.0 line was true at its check time. A note on the research plan, not a published page: the plan said Jev's languages are not stated in TypeSafe's docs. The models page does state them qualitatively ("English is the primary training language … Other languages, including CJK scripts, are handled but not equally well"), as the limitations page already showed; the moderation page links that row.

3. Per-page findings and ship verdicts

Home and access status: refreshed

• Headline unchanged: "Available to new users: signups open, but new accounts get no free credit. API operational." Shown as "checked 2 Oct 2026, 05:14 UTC".
• Notices, in order: new-user-credit-disabled (condition, Documented); lookalike-resellers (condition, Reported; the three named sites still answered HTTP 200); rate-limits-changed (event, Documented; retire 7 Oct, 20:47 UTC); typesafe-workflow-evals-code (event, Documented; retire 6 Oct, 16:25 UTC). All four have shown: true.
• Retired notices: none. Timeline additions from the status page: none (no new event since 29 Sep, 22:05 UTC).

use-cases/model-routing: meets the ship condition

• OpenRouter's Jev Router docs read with explicit "not documented" results: on Jev failure, on long conversations and on routed-model billing, across four documents (docs page, model page and FAQ, llms.txt, the 4.16 MB full docs export).
• Six routers read at pinned commits (condition: three or more), with what is sent to Jev, error handling and the failure destination filled from source for all six. Selection: LiteLLM plus the five highest-starred routers in a GitHub search for "jev model router" (313 results).
• Counts: on a Jev error, 1 frontier tier, 1 keep current, 2 local scorer, 2 no decision. 4 of 6 act on confidence. 4 send the full ask with no length cut, 1 clips to 400 characters, 1 sends no prompt text.
• Title query source: Bing and DuckDuckGo autocomplete suggest "jev router" for "jev r", "jev ro" and "jev routing" (05:15 UTC).

build/claude-code-mcp: version check done

• 0.11.0 to 0.12.0: server code moved to a new file with tool logic byte-identical; rate-limit and outage results now come back as tool errors (still not an allow).
• 0.12.0 to 0.13.0: jev_verify adds one Noul per claim and moves claim text into state; a repository example hook (not in the npm package) denies at block ≥ 0.85 and defers to the harness's permission prompt on any failure.

use-cases/content-moderation: meets the ship condition

• Six projects read at pinned commits (condition: three or more), with thresholds, failure handling and reversal filled from source; five code-backed integrations and one runnable demo.
• 3 of 6 delete in a band below certainty; 4 of 6 send author data beyond the text by default.
• Noul-per-harm against Choice compared from TypeSafe's docs. Policy: no clause on automated moderation decisions in the AUP, Terms or MCA. Evidence: TypeSafe's consistency cookbook measures repeatability on one post (90.8% label agreement; 99.2% with 74.2% automatic labels), not accuracy (Reported).
• Title query source: independent publisher titles; no autocomplete suggestion for moderation on any endpoint (05:15 UTC).

Other updates

• No TypeSafe release in the window: no repository push since 29 Sep, 21:11 UTC; no SDK, model or alias change; legal "Last updated" dates unchanged (Terms 19 Sep, AUP 23 Sep, MCA 23 Sep, DPA 24 Apr, Privacy Policy 19 Nov 2025).
• Route data labels unchanged: TypeSafe's endpoint is on OpenRouter's ZDR list; Vercel's endpoints list only DigitalOcean (has_zdr: false); Cloudflare shows "Zero data retention Yes".

4. 24-hour discovery scan in brief

Window: 1 Oct 2026, 05:13 UTC to 2 Oct 2026, 05:13 UTC. Two minutes overlap the 1 October window; no item in that run's ledger fell in them, and there is no gap.

• Hacker News: 19 Jev-titled stories and 80 comments matching "jev"; none about signup, credit or an outage.
• GitHub: 330 repositories created matching "jev"; 20 match "jev router", 18 "jev routing", 1 "jev moderation", 2 "jev spam". The "jev mcp" query hit the unauthenticated search limit: unknown.
• npm: 13 "jev" packages with a last publish in the window, including @jkudish/jev-mcp 0.13.0 and ai 7.0.127; oh-my-claudecode v5.6.0 on GitHub. DEV.to: 2 articles tagged jev.
• Reddit (HTTP 403), X (login redirect) and TikTok could not be read: unknown, not zero.

These are search-index counts, not verified integrations or adoption.

5. Evidence gaps

• What OpenRouter's Jev Router does on a Jev error, timeout, malformed answer or long conversation, and how it bills the routed model. Only OpenRouter can answer.
• Whether any router or moderation project behaves live as its code reads (no runs allowed).
• Moderation accuracy for Jev: none published.
• oh-my-claudecode v5.6.0 failure path (due in the next run).
• Jevdit and the project behind Twittesia issue #261 not located; kerpopule/hermes-jev-skills not read.
• Carried: trust.typesafe.ai (needs a browser), Netlify data terms, LanceDB reranking materials, a calibration lead; whether jev.bevel.software resells TypeSafe's Jev or serves another model.
• Search analytics remain unavailable; X, Reddit and TikTok remain unreadable.

6. Not done

• Spare-time items 2–5 (trust page, Netlify, LanceDB, calibration lead).
• Earlier routing leads and the nine routing repositories created in the window; moderation leads (gg-friggin-ez, jev-sap-commerce, jevmod, verdict-lab, jev-bouncer, shugo, jevchik).
• The oh-my-claudecode v5.6.0 diff; the videos linked on the new pages were not watched.
• GitHub code search and trending.

7. Key sources