Jev (TypeSafe AI) research report, 1 October 2026: Laya vs Kev, data handling by route, and support-ticket triage

Run: run:35d0708f-1681-4f24-9954-581f127a7a34 (scheduled regular research run 4) · Pool: pool_jev_catalog · Runner: single agent · Checks: 2026-10-01, 05:15 to about 05:40 UTC. Nothing in the planned slip order slipped.

The run's full notes and its source ledger (81 sources, IDs R4-S01 to R4-S97: 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. Legal and route texts are quoted, not interpreted ("a reading of the text, not legal advice"). Benchmark results are labelled Reported. Search analytics were unavailable: the run input records the read-only analytics broker as not connected.

Previous release: regular-2026-09-30-evaluation-evidence; report: evaluation evidence.

1. Summary answer, as of 1 October 2026

1. Access is unchanged (checked 05:16 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.
2. The rate-limit change is dated more tightly. Internet Archive captures still show the old limits (250,000 tokens/s, 1,200 requests/min) at 29 Sep, 19:24:32 UTC. The change happened between then and 30 Sep, 20:47 UTC.
3. Laya vs Kev: only one same-items test finds an open model level with hosted Jev: Opper's 362 fresh items put Kev-4B within noise (Opper hosts Kev). JevBench scores Laya and earlier Kev previews well below Jev. Laya's English checkpoint reads about 512 tokens; Kev is English-only.
4. Data handling: on the six routes read, no primary text allows training on your prompts without consent. TypeSafe's MCA grants it a licence "in perpetuity" to use Customer Data to derive telemetry and to monitor for fraud and abuse, and it offers ZDR "for enterprise customers". OpenRouter, Cloudflare and Vercel nonetheless label the TypeSafe provider zero-retention. Vercel also routes Jev through DigitalOcean, which it marks as not ZDR.
5. Ticket triage: four of five projects read in source send unsure tickets to a person. The fifth ignores queue confidence and files the ticket in its archive lane on any Jev API error.

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

No earlier claim was found false apart from the Vercel ZDR reading, which was incomplete rather than wrong. Publisher's note: the comparison page "Jev vs an LLM vs a classifier" said that among gateways "only Cloudflare's listing says 'Yes'" to zero data retention; it now links the data-privacy page instead.

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 1 Oct 2026, 05:16 UTC".
• Notices, in order: new-user-credit-disabled (condition, Documented); lookalike-resellers (condition, Reported); 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: none (no new status-page event since 29 Sep, 22:05 UTC).

alternatives/laya-vs-kev: meets the ship condition

• Licence stack, context and API shape are documented for all three models, or marked unknown with what was checked (Jev's languages, Laya's base-model metadata, Kev-9B's base licence).
• Three graded rows compare at least two of the three: B11 (Jev and Kev-4B), B07 (Jev, Laya and Kev previews) and B09 (Jev and Laya).
• B09: yes/no option names flip 76.9% of Laya's decisions and 32.5% of Jev's (Reported).

where-to-use/data-privacy: meets the ship condition

• Training, retention and ZDR quoted, or marked "not stated", for TypeSafe direct and four gateways (OpenRouter, Vercel, Cloudflare, Opper), each with both layers. Netlify was not read.
• TypeSafe's public documents state no retention period; human review is not stated; SOC 2, HIPAA, BAA and EU residency do not appear. The subprocessor list sits behind a script-only trust page and could not be read.
• Vercel's ZDR docs quote a TypeSafe no-retention clause that is not in TypeSafe's public legal texts.

use-cases/support-ticket-triage: meets the ship condition

• Five projects read at pinned commits (condition: three or more). B, C, D and E gate on confidence (0.5 to 0.9) and route to a person. A routes errors and malformed answers to "archive". Only D validates the answer shape and model and returns 503 on error. Two of five set no timeout.
• Evidence box: TypeSafe's WorkflowEvals customer_service workflow, Jev 76.0% agreement with LLM reference answers at $0.0001 per case (Reported, vendor).

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).
• Ollama's official decision models are nimble and tev1; Laya and Kev are not among them. Ollaya lists both.

4. 24-hour discovery scan in brief

Window: 30 Sep 2026, 05:15 UTC to 1 Oct 2026, 05:15 UTC. 15.5 hours of it overlapped run 3's window and were not re-recorded unless something changed; the gap left by run 3 (29 Sep, 07:05–20:45 UTC) was backfilled for GitHub repository creation.

• Hacker News: 29 Jev-titled stories (4 after run 3's window ended); 13 new comments read, none on access.
• GitHub: 354 repositories created matching "jev" (92 after 20:45 UTC); 10 match "jev triage", 77 "laya".
• npm: 28 "jev" packages with a last publish in the window, including @jkudish/jev-mcp 0.12.0 and ai 7.0.126. DEV.to: 5 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

• Whether DigitalOcean on Vercel forwards to TypeSafe or serves Jev itself, and why Vercel's endpoints API returned only DigitalOcean at 05:30 UTC.
• Whether OpenRouter's ZDR classification of the TypeSafe endpoint rests on a special agreement; whether the clause Vercel quotes is in contracts standard accounts get.
• What trust.typesafe.ai shows (subprocessors, SOC 2): it needs a JavaScript-capable reader.
• Which Kev checkpoints JevBench ran; the JevAny/AnyJev identity; who operates he-jev.
• TypeSafe direct: a retention period, human access to inputs and a Telemetry opt-out are all not stated.
• Search analytics remain unavailable; X, Reddit and TikTok remain unreadable.

6. Not done

• Netlify data terms; the parsecailabs, binubabu and astgl posts; watching the "jev vs laya" videos.
• Re-auditing @jkudish/jev-mcp 0.12.0 and ai 7.0.126.
• Hacker News leads (Jevotron, jevshield, a calibration post, Bongard-mini, AT0M, Lichen) and new Laya/Kev comparison repositories.
• GitHub code search and trending; backlog items carried from run 3.

7. Key sources