Run 8 report: browser and computer use, framework integrations, Jev on Bedrock or Azure (TypeSafe AI, 5 Oct 2026)

This dated report records the evidence behind the release of 5 October 2026 (regular-2026-10-05-browser-frameworks). It covers 7 Jev agents that click, type or tap; 6 Jev framework packages; whether Jev is listed in the Amazon Bedrock and Azure AI Foundry catalogs; and the failure checklist, which grows from 39 to 51 rows. The title is an internal label for the run; the page titles it feeds use phrases that search suggestions offered on 5 Oct 2026, 05:14–05:15 UTC (a demand signal, not a volume). The pool ran nothing: no browser or computer-use agent, no framework code, and no Jev call.

Run: run:b26864e3-2a62-4080-b445-28fe3dbadc7e (scheduled regular research run 8) · Pool: pool_jev_catalog · Runner: one runner plus two helper agents that read source in parallel. The runner spot-checked 15 + 26 helper lines at the same commits (1 line-number correction) · Checks: 5 Oct 2026; runner started 05:11:41 UTC and returned at about 05:36 UTC (about 25 minutes).

The run's research notes and its source ledger (88 records, IDs R8-S01 to R8-S133: 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 browser or computer-use agent was run; no cloud login, no key, no install and no Jev API call. Failure cells come only from source at pinned commits. Analytics were unavailable for this run. Nothing in this run is Tested. Every finding below is Documented, Reported (a named third party's own claim) or Unverified (a lead only).

Previous release: regular-2026-10-04-trading-failures-judge; report: trading and fraud gates, the Jev failure checklist and Jev as a judge. All dated reports: Research.

1. Summary answer, as of 5 October 2026

1. Access is unchanged since 4 Oct. State open_with_conditions (no new-user credit) since 27 Sep, 22:33:25 UTC. API operational since 29 Sep, 22:05 UTC. Limits 100K tokens per second and 80 requests per second; price $0.042 per million input tokens. All five notices kept (four shown); nothing retired; no new notice. Documented
2. Browser and computer use: of 7 Jev agents that click, type or tap, none (0/7) acts when Jev errors. 5/7 act on low-confidence answers (no threshold), and only 1/7 asks a person before send, buy or delete. 5/7 run in the user's own browser, phone or desktop session by default. Documented
3. Frameworks: of 6 framework packages, 4/6 are pre-1.0 and none (0/6) is from TypeSafe. 4/6 raise Jev errors to the caller. The community Spring AI passage filter and the LlamaIndex reranker fail open. The LangChain middlewares fail closed. Documented
4. Bedrock and Azure: Jev is not listed in the Amazon Bedrock model cards at 5 Oct 2026, 05:18 UTC (19 providers). Jev is not listed in the Azure AI Foundry model catalog at 5 Oct 2026, 05:19 UTC (0 models for "typesafe" and for "jev"). TypeSafe's docs never mention Bedrock, Azure or Foundry. Teams call TypeSafe's API or a gateway, and the data leaves the cloud's model boundary. Documented
5. when-jev-fails grows from 39 to 51 rows in 9 domains. 30/51 stop or hold the action on a Jev error; 6/51 let it through unchecked. Documented

2. Publication

Release regular-2026-10-05-browser-frameworks shipped all three planned page items and the follow-on edits:

• Browser use and computer use (new page).
• LangChain and other framework integrations (new page).
• Is Jev on Amazon Bedrock or Azure AI Foundry? (new section on the where-to-use hub).
• When Jev fails: 12 new checklist rows (39 → 51), the recount, and the copy block.
• Products refresh: 72 rows (A 55, B 16, C 1).
• Small additions: EU hosting on data privacy; Strands Decider 2B on alternatives; the Zapier app on n8n; links to the Langfuse, LangSmith Evals and DeepEval documentation on Jev as a judge; rate-limit check dates moved to 5 Oct on errors and rate limits and limitations; the status refresh on access status and Home.

The publisher's tool test result for this release is recorded on the Research page.

3. Corrections to plan facts and pool pages

1. langchain-typesafe 0.0.1a3 was uploaded 20 Sep 2026 (18:54:39Z), not 4 Oct (PyPI). Documented
2. @effect/ai-typesafe is now 4.0.1 (4 Oct, 21:51Z), not 4.0.0. Documented
3. The Spring AI integration is community (org.springaicommunity:typesafe-spring-ai 0.4.0), not part of spring-projects/spring-ai. Documented
4. AutoModeMiddleware is inside langchain-typesafe (experimental module). A JS version ships in @langchain/typesafe 0.0.2. Documented
5. LlamaIndex has a community integration (llama-index-postprocessor-jev / -selectors-jev 0.1.1). Nothing official exists, within the boundary stated in the research notes. Documented
6. The "Fast browser agent (using Jev)" lead is browser-use/jev-ultrafast. HN item 49849950 links only a tweet, with no source.
7. The AndroidWorld claim comes from arXiv 2609.30186 ("Jev-Mobile …", single author), submitted 24 Sep per arXiv, not 25 Sep. Reported
8. products: Jev for Chrome, computer-use-jev and jev-android were B rows. Their call paths were read in source this run, so they become A.

4. Access status, home notices and timeline

• Rechecked 5 Oct 2026, 05:13:31 to 05:15:03 UTC (R8-S20 to R8-S50). Values unchanged since 4 Oct (see item 1). Documented
• The home status record was written at about 05:16 UTC (minute 4), validated, and sent to the manager. It did not change after minute 15.
• Retire pass: new-user-credit-disabled kept (rank 1); lookalike-resellers kept (rank 2; jevtypesafeai.com is one of the three sites the Eye Security report names, not a new site); rate-limits-changed-2026-10-03 kept (rank 3); subprocessors-added-2026-10-02 kept (rank 4; list unchanged); typesafe-workflow-evals-code kept with shown: false (rank 5; retires at the 7 Oct run). Retired: none. New: none.
• Timeline: the 28 Sep and 29 Sep incidents are already on access status. No additions.

5. Findings per page and ship verdicts

5.1 use-cases/browser-and-computer-use (new): ships

• Page: Jev browser use and computer use (TypeSafe AI): what agents do when Jev fails or is unsure. Not claimed: that Jev browser agents are safe, or any general rate.
• Title query source: "jev browser use" and "jev computer use" are suggested on Google (client=firefox), Bing (osjson) and DuckDuckGo (ac) at 5 Oct 2026, 05:14:58 UTC, confirming the 4 Oct audience row. A suggestion is a demand signal, not a volume.
• Projects and pinned commits: browser-use/jev-ultrafast 1231850, ironbee-ai/ironbee-express 05079b4, forvela/jev-agent-browser eb35251, dougsong/jev-android be8364a, paulsmith/computer-use-jev ff0ad8b, chy4pro/jev-for-chrome d5c24de, razaanstha/ulka bb04433; KaushikSiva/second-look 0c05d16 (recommendation-only; the runner's own read). trycua/cua jev-use ba0033a is listed but not counted (mock provider by default).
• Counts (n = 7 acting agents):
  • on a Jev error, a UI action still happens in 0/7;
  • malformed answer: 0/7 act directly; 1/7 (jev-agent-browser) can click a fallback target picked by word overlap;
  • unsure: 5/7 have no threshold; 2/7 block below 0.65 or 0.5;
  • no key: 6/7 refuse; ironbee-express replays a cached recording unjudged;
  • bypass recorded: 4/7;
  • user's own session by default: 5/7;
  • human gate before irreversible actions: 1/7 (Ulka); Jev confidence gate on mutating actions: 1/7 (computer-use-jev, 0.5); none found: 4/7.
• second-look: on a Jev error or no key, it shows a heuristic "Good deal — buy it" verdict with confidence None.
• Risk box on the page: logged-in sessions; page text and field values leaving the machine (6,000 to 18,000 characters; all 7 exclude passwords); unsure is not the same as failed; irreversible actions; speed claims Reported.
• Ship rule: 3 or more projects with columns 1, 3, 4, 6, 7 and 9. Met by 7.

5.2 build/frameworks (new): ships

• Page: Jev LangChain and other framework integrations (TypeSafe AI): which is mature enough, and what happens when Jev fails? Not claimed: that framework integrations are unsafe, or that LangChain's integration is production-ready.
• Title query source: "jev langchain" is suggested on all three endpoints at 05:14:59 UTC. "jev pydantic", "jev spring ai" and "jev llamaindex" have no suggestions. "jev hermes" appears on all three endpoints; it is material for run 9.
• Pins: langchain 4af7ab8 (0.0.1a3); langchainjs 417ddcf (0.0.2); pydantic-ai 6695132 (v2.54.0); effect 460272d (4.0.1); spring-ai-typesafe 51993bb (v0.4.0); llama-index-jev 421afda.
• Version rechecks: LiteLLM 1.104.0 adds a new lead, a TypeSafe proxy guardrail hook (not audited); Vercel 3.0.12, unchanged; the Pydantic AI Gateway does not route Jev; no Mastra package found.
• Coding-agent pointer (leads only, Unverified, from web-search snippets): the Hermes catalog lists community Jev plugins (pinned in spare time, section 11); Codex, OpenCode and Cursor have only community MCP servers or guides; no vendor plugin was found for any of them.
• Ship rule: 4 or more integrations with columns 1, 2, 4, 7 and 8. Met by 7.

5.3 where-to-use section "Is Jev on Amazon Bedrock or Azure AI Foundry?": ships

• Section: Is Jev on Amazon Bedrock or Azure AI Foundry?
• Checks, all Documented:
  • Jev is not listed in the Amazon Bedrock model cards at 5 Oct 2026, 05:18 UTC (19 providers, no TypeSafe).
  • Jev is not listed in the Bedrock Marketplace model table or on the Bedrock model-choice page at 5 Oct 2026, 05:17 UTC.
  • AWS Marketplace, first page of results only, 05:18–05:19 UTC: "TypeSafe" returns 900 results with no TypeSafe AI listing. "Jev" returns 132 results; the first two are third-party self-hosted look-alike servers, "OpenJev 27B" by Salient Engineering and "openjevx Server" by Deemwar.
  • Jev is not listed in the Microsoft Learn Foundry lists (models sold directly by Azure, models from partners) at 5 Oct 2026, 05:17 UTC.
  • Jev is not listed in the Azure AI Foundry model catalog at 5 Oct 2026, 05:19 UTC: 0 models for "typesafe" and for "jev"; the control search "deepseek" returns 98.
  • Azure Marketplace (05:17 UTC): one third-party Excel add-in, EBITDAI, with an "AskJev" function.
  • TypeSafe llms-full.txt: 0 mentions of Bedrock, Azure, Foundry or AWS Marketplace.
• Secondary (Reported): FactualMinds writes "Jev is not an Amazon Bedrock model". The Microsoft Tech Community post builds a Foundry agent with an OpenAPI tool that calls https://api.typesafe.ai/v1/systemone. jevtypesafeai.com's "Jev on AWS Bedrock" page is a lookalike-reseller fact, not a source.
• Route-matrix row "Jev on Bedrock / Azure AI Foundry", with how-to lines and the data path, is on the where-to-use hub.
• Title query source: "jev on bedrock" on Google; Bing and DuckDuckGo give "jev bedrock" for the same prefix. "jev azure foundry" and "where to use jev" appear on all three endpoints.
• Ship rule: both catalogs read directly at recorded times, plus TypeSafe's absence within a stated boundary. Met.

5.4 build/when-jev-fails: rows and recount

• 12 new rows: 8 browser and 4 frameworks (Spring JevDocumentFilter, LlamaIndex JevRerank, LangChain ModelRouterMiddleware in Python, LangChain autoModeMiddleware in JS). No existing row changes.
• Totals were recomputed by script from the published 39-row tables plus the new rows. Every row sums to n = 51. Below-threshold totals are kept apart from error totals.
• New page meta: "51 Jev implementations in 9 use cases …". The checklist and copy-block counts are on the page (copy block).
• Browser-only figures (n = 47, without the 4 framework rows): 28 stop or hold; 5 let it through.

6. Counting judgements (kept visible)

• second-look is counted as a browser-domain row even though it never acts.
• The JS autoModeMiddleware is counted separately from the Python one already on Claude Code and MCP guardrails.
• Pydantic AI is not counted as a decision row: it raises to the agent the builder writes.
• ironbee-express's no-key cell is counted as fail-open (conditional), following run 7's convention.

7. Scan (S1) highlights

• Window: 4 Oct 2026, 05:14 UTC to 5 Oct 2026, 05:14 UTC. Gap 0, overlap 0.
• Window counts (R8-S01 to R8-S05): Hacker News "jev" 7 stories and 4 comments; 250 GitHub repositories created; 12 npm packages; 18 Hugging Face models. These are search-index counts, not verified integrations or adoption. Reddit and X are unknown, not zero.
• No TypeSafe release, legal, limits, subprocessor or incident change.
• "jev omni": explained as a community open-weights model, akhilaaa3/Jev-Omni (Gemma 4 12B base, Apache-2.0, 20 Sep), plus its copies. It is not TypeSafe's.
• "jev vision": unexplained (candidates listed in the research notes). TypeSafe docs say Jev's input is text only.
• Leads (Unverified): obekt-terminal, an AI agent said to trade crypto with real funds unattended (Reported; not read; no trading or investment advice); DoubtBench (for benchmarks); a WSJ story on copycats (paywalled; not read); local System One runtimes (for alternatives).

8. Drift

• All pinned SDKs, Claude Code and MCP integrations and the five n8n nodes are unchanged. Documented
• LiteLLM 1.104.0 is a minor release of an integration whose routing disposition is published. A targeted diff is due next run.
• @effect/ai-typesafe 4.0.1 is already the pin on the new frameworks page.
• With build/frameworks published, its six pins are due to join the drift list.

9. Products row changes

• Checked against the published page: 67 rows (A 47, B 19, C 1).
• jev-ultrafast stays A (pin refreshed to 1231850).
• B → A: Jev for Chrome, computer-use-jev, jev-android.
• New A: ironbee-express, jev-agent-browser, Ulka, second-look, trycua/cua jev-use (mock by default).
• New total: 72 rows (A 55, B 16, C 1); 0/72 independently verified in production. Arithmetic: 67 + 5 new = 72; A 47 + 3 + 5 = 55; B 19 − 3 = 16.
• Framework packages go under "not counted as products": langchain-typesafe, @langchain/typesafe, Pydantic AI TypeSafeModel, @effect/ai-typesafe, typesafe-spring-ai, llama-index-postprocessor-jev.

10. Small checks (S6)

• Zapier: the "TypeSafe Jev" app has one action ("Ask Questions") and no "built by" field. Failure behaviour is not readable in source. Not counted. Documented
• EU hosting: no TypeSafe EU region found within the stated boundary; all 6 subprocessors are in the USA. Opper is EU-hosted but routes to "TypeSafe AI — United States". The HN request for EU hosting is Reported context.
• Strands Decider 2B: primary source strands-labs/strands-decider 75c9fd3 (Apache-2.0; Qwen3.5-2B base; runs locally behind /v1/systemone). It is a model, interface and runtime substitute. Its JevBench 0.723 (167 of 231) is the authors' own (Reported). The AWS attribution is via The New Stack (Reported).

11. Spare time

1. Hermes catalog: 6 plugin repositories pinned: jev-typesafe b3d29f7, jev cdf59e4, jev-judge 5dddbea, typesafe-skill-router 94fe114, jev-skill-router 76ee09e, hermes-jev-approvals 530fdb0. All 6/6 are listed as Community. Their behaviour was not read.
2. databases-and-documents: 5/5 projects (pg-jev, sqlite3-jev, sqlite-jev, pg-redact, docjev) have a real Jev call path at pinned commits. The count part of the release-4 condition is met; their error behaviour was not read.
3. jev-risk-check-provider fa5e497: no payment call site in the 60 of 361 files scanned. Its credits module bills its own calls. Not a payment-gate candidate on this evidence.

12. Evidence gaps and uncertainties

• No agent or integration was run. Every disposition is read from code; behaviour when run may differ (for example, SDK defaults not read: the @langchain/core backoff, and the typesafe-sdk timeout used by Pydantic AI).
• Test contents were read for only a few projects. jev-agent-browser's npm tarball was not compared to the repository.
• Unknown cells: whether computer-use-jev sends [secure] field values; second-look's fallback card wording; trycua/cua no-key behaviour; the Spring AutoConfig lines (helper-read, not spot-checked).
• Marketplace and catalog boundaries: AWS and Azure Marketplace results were read on the first page only. Private offers and signed-in listings were not checked.
• The coding-agent pointer facts are web-search snippets: Unverified until run 9 reads them.
• The counting judgements in section 6 change totals if read differently.

13. Not done

• LiteLLM targeted diff (next run).
• The LiteLLM guardrail hook.
• The unread same-name browser and computer-use repositories (listed in the research notes).
• The databases-and-documents error behaviour.
• The WSJ story (paywalled).
• obekt-terminal.

Nothing from the planned page items slipped.

14. Timings (UTC, 5 Oct 2026)