Research reportChecks 16 October 2026, 05:21–05:38 UTC

Bypass cells complete, Jev 1.14 preview and Claude Code guardrails rechecked (16 Oct 2026)

This dated report records the evidence behind the update of 16 October 2026. It settles the last open cells on When Jev fails, records TypeSafe's Jev 1.14 preview and its opt-in experimental features, and rechecks the Claude Code and MCP guardrails. Nothing was run: no account, key, API call or third-party code.

Key findings

When Jev fails: For all 77 counted implementations, every failure cell (no key, error, malformed answer, below threshold and bypass) is recorded from source at pinned commits (16 Oct 2026).

Documentedfrom source at the pins; four bypass cells name a narrower read scope (below); recounted by script, 16 Oct 2026

Bypass: Code paths that never ask Jev are now recorded for all 77 implementations (source read at pinned commits, 16 Oct 2026); for 13 of them the cell is 'none found' in the files read.

Documentedthe map counts the 77 implementations that act on Jev's answer

Jev 1.14: Jev 1.14 (jev-1.14.0) is available as a preview under jev-preview since 14 Oct; jev-latest still points to 1.13, and TypeSafe says it moves "in a week".

DocumentedTypeSafe models page and blog post, read 16 Oct 2026, 05:21–05:26 UTC (R17-S05, R17-S09)

Image input (both versions), a Box question type (1.14 only) and 262k long context (both) are opt-in experiments behind "Danger Zone" headers; they are not production-ready per TypeSafe. The models page no longer publishes rate-limit numbers: "Rate limits are set per organization."

DocumentedTypeSafe docs, 16 Oct 2026, 05:21 UTC (R17-S05, R17-S06, R17-S20)

Claude Code and MCP: no failure-path change in the guardrails on the page; the short answer "Only one of them stops the agent when Jev fails" (1 of 7) still holds. 32 related projects seen in scans are now classified in a lead table.

Documentedrepositories read at pins; oh-my-claudecode 5.7.3 compared with 5.6.2, 16 Oct 2026

When Jev fails, 77 implementations: 14 Oct → 16 Oct

fail-closed / recordedfallbackfail-openno-decisionadvisorynot recorded

Bypass, 14 Oct66 recorded, 11 not recorded (judge tools 4, frameworks 3, browser agents 3, email 1)

Bypass, 16 Oct77 recorded (64 with skip paths, 13 "none found"), 0 not recorded

Malformed, 14 Oct35 fail-closed, 24 fallback, 11 fail-open, 3 no-decision, 3 advisory, 1 not recorded

Malformed, 16 Oct35 fail-closed, 24 fallback, 11 fail-open, 4 no-decision, 3 advisory, 0 not recorded

In the bypass bars, green means a recorded cell (a list of skip paths, or "none found" with the files read), not a safety verdict. No-key, below-threshold and error counts did not change. Counts are Documented from source read at pinned commits and recounted by script from the published cells, 16 Oct 2026. Nothing was run.

What changed on 16 Oct 2026 (published value on 14 Oct → value on 16 Oct)

When Jev fails: bypass recorded~~66 of 77~~→77 of 77
The last 11 cells: judge tools 4, frameworks 3, browser agents 3, email 1 (4 + 3 + 3 + 1 = 11).
When Jev fails: malformed not recorded~~1 of 77~~→0 of 77
competitor-hunter is now no-decision: no-decision 3 → 4.
Failure columns with a not-recorded cell~~2 of 5~~→0 of 5
No key, error and below threshold already had 0 not recorded on 14 Oct.
Jev model under jev-preview~~jev-1.13.0~~→jev-1.14.0
jev-latest unchanged at jev-1.13.0; TypeSafe says it moves "in a week".
Published rate limits~~100K tokens/s, 80 requests/s~~→set per organization
Shown in the developer console; the pool has no account, so the numbers are not known to us.
Home notices shown2→4
New: the Jev 1.14 preview and the experimental features. Series A notice retired (it was hidden).
Claude Code and MCP leads~~19 listed~~→32 classified
19 from earlier lists plus 13 from this update's scan, in one table (16 + 12 + 1 + 3 + 0 = 32).
Price per million input tokensunchanged
$0.042 for both jev-1.14.0 and jev-1.13.0; output free (16 Oct 2026, 05:21 UTC).
Benchmarks ledgerunchanged
41 graded (A 12, B 14, C 8, D 7) + 1 not graded. The Decision Index line sits outside the ledger.
Claude Code short answerunchanged
1 of 7 stops the agent when Jev fails (LangChain). No new rows this update.

A dashed box means the value did not change. n stays 77 implementations in 12 domains.

Checks: 16 Oct 2026, 05:21–05:38 UTC (access 05:21–05:26; package drift 05:28; Decision Index 05:31; last completion read 05:38).

The research notes and the source ledger (R17-S… IDs) are held in the pool's private record. Key sources are linked below and on the pages each finding feeds. Checks on the sources: 235 cited IDs checked, 0 missing; every sentence that splits a total was checked by script so that its parts add up and every category with 3 or more items (or 10% of the total) is named.

Rules kept: nothing was installed or run, so no Jev or other provider call was made, and the pool has no account with any of them. Failure cells come only from source at pinned commits. Package facts come from published package files, opened without installing. Vendor quality claims are Reported. Nothing in this report is Tested.

Previous report: 14 Oct research report. All dated reports: Research.

1. What was read

Jev implementations at pinned commits: the 11 open bypass cells and one malformed cell on build/when-jev-fails, plus a completion read of one cell published earlier (JevPlaysSnakeGame). Each cell rests on files and line ranges at the published pin (R17-S100 to R17-S194, R17-S300 to R17-S313, R17-S73); two cells were spot-checked again (R17-S68, R17-S72).
TypeSafe docs and post: the models page, the images and Box pages, the quickstart, the status page, the blog and the post "Jev: For the builders AND dreamers" dated 14 Oct 2026 (R17-S03 to R17-S09, R17-S20, R17-S21).
Route listings: OpenRouter, Vercel AI Gateway, Netlify AI Gateway docs, Cloudflare, Opper, Amazon Bedrock, Microsoft Foundry and the Workers AI catalog (R17-S22 to R17-S35).
npm and PyPI package files: typesafe-sdk 0.7.5, @typesafe-ai/sdk 0.7.0, @ai-sdk/typesafe-ai 3.0.20 and the drift list, read without installing (R17-S37 to R17-S39, R17-S57); pydantic-ai-slim and DeepEval files at their release tags (R17-S58, R17-S59).
Claude Code repositories: the guardrails on the page at their pins, the oh-my-claudecode 5.7.3 release, and 32 leads from scans, classified from README and manifest (R17-S200 to R17-S261).
Standing checks: access and notices, the comparison card's cells, the discovery scan with the 72-hour decision-API look-back, and the drift list.

2. Access status and home notices

Checked 16 Oct 2026, 05:21 UTC: Documented. The access state is unchanged: signups open with conditions.

Signups open; new accounts get no free credit. API operational on the status page (R17-S04).
Console: console.typesafe.ai now serves a sign-in page ("Continue with Google", HTTP 200) where it returned HTTP 403 on 14 Oct. No form was submitted (R17-S01).
Jev 1.14 preview: released on 14 Oct under jev-preview; jev-latest still points to jev-1.13.0 (R17-S05, R17-S09). Details in section 4.
Rate limits: until at least 14 Oct 2026, 16:46 UTC the models page listed 100K tokens per second and 80 requests per second. On 16 Oct 2026, 05:21 UTC it lists no numbers: "Rate limits are set per organization. Your organization's current limits are in the developer console" (R17-S05). This is treated as an access condition, not a notice (editorial decision, 16 Oct). When the text changed is known only as between those two reads.
Price: $0.042 per million input tokens for both versions; output free; "Long context costs the same per token" (R17-S05).
Latency event: a 45-minute latency event on the us-east-1 regional endpoint was resolved on 14 Oct 2026, 22:52 UTC (status page, R17-S04). It is a timeline entry on access status.
Shown on Home (4): new-user credit disabled; Jev 1.14 preview (new); lookalike resellers; images, Box and long context are experimental (new). Retired (1): TypeSafe's Series A announcement, at the end of its seven-day shelf life (checked 16 Oct 2026, 05:24 UTC); it was hidden and never shown on Home (R17-S19).
Subprocessors were not rechecked (the trust portal needs a browser render); the 11 Oct value stands.

3. When Jev fails: the last bypass cells and the malformed cell

A script first reproduced the published counts from the page's cells (14 Oct values). Then the open cells were settled at the published pins and the page was recounted (R17-S67).

Counts of 77 implementations (14 Oct → 16 Oct)
Condition	14 Oct	16 Oct	Check
No key	50 fail-closed, 17 fallback, 7 fail-open, 3 advisory, 0 not recorded	unchanged	50 + 17 + 7 + 3 = 77
Bypass	66 recorded, 11 not recorded	77 recorded (13 "none found", 64 with skip paths), 0 not recorded	13 + 64 = 77
Below threshold	25 no-threshold, 16 held, 15 acts-anyway, 12 fallback, 5 no-decision, 4 advisory, 0 not recorded	unchanged	25 + 16 + 15 + 12 + 5 + 4 = 77
Jev error	39 fail-closed, 22 fallback, 8 fail-open, 5 no-decision, 3 advisory, 0 not recorded	unchanged	39 + 22 + 8 + 5 + 3 = 77
Malformed answer	35 fail-closed, 24 fallback, 11 fail-open, 3 no-decision, 3 advisory, 1 not recorded	35 fail-closed, 24 fallback, 11 fail-open, 4 no-decision, 3 advisory, 0 not recorded	35 + 24 + 11 + 4 + 3 = 77

The 11 bypass cells settled on 16 Oct

Of the 11, 2 are "none found" (DeepEval JevEval, LangChain ModelRouterMiddleware) and 9 list skip paths (2 + 9 = 11). Across the map, 13 cells are "none found" (11 published before + 2 new) and 64 list skip paths (13 + 64 = 77). The full cells, with every file and line, are on the map and the domain pages.

Bypass cells read at the pinned commits, 16 Oct 2026 (all now recorded)
Implementation and pin	Domain	Cell (shortened)	Key lines
DeepEval JevEval at `a200ece`	LLM as a judge	None found in the `deepeval/metrics/jev_eval/` directory; every scoring call asks Jev.	`jev_eval.py` L125-194 (R17-S68, R17-S100, R17-S101)
openlayer jevals at `0a8f895`	LLM as a judge	Some built-in evals score locally; evals missing required fields are skipped and allowed; mock and other backends do not ask Jev; `jevals bench --ragas` also scores with another model (report only). 10 skip paths in 6 groups, plus 1 report-only.	`_bench.py` L26-28, L52-53, L100-108; `_calibrate.py` L88-97 (R17-S72); others on the map
did-they-answer at `e0ee60c`	LLM as a judge	The daily job judges only new questions from the last 21 days that have a reply, within a per-run budget; stored results are reused; the labelling tool records labels from another model or a person. 7 skip paths.	`daily.py` L35, L44-45, L52-54; `compact.py` L50; `label_server.py` L26 (R17-S139 to R17-S145)
competitor-hunter at `40613ea`	LLM as a judge	Re-runs reuse earlier verdicts; pieces judged "paid" by engagement rules are dropped; Reddit posts without a Jev result are still listed; another model writes breakdowns; the results page ranks by views and engagement without Jev. 7 skip paths.	`web/app.js` L438-485; others on the map (R17-S149 to R17-S159)
Spring AI JevDocumentFilter 0.4.0 at `51993bb`	Frameworks	Only an empty passage list is returned without asking Jev. 1 skip path.	`JevDocumentFilter.java` L147-149 (R17-S68, R17-S124, R17-S125)
LlamaIndex JevRerank at `421afda`	Frameworks	Only an empty node list is returned without asking Jev (sync and async). 1 skip path.	`base.py` L127-129, L173-175 (R17-S129, R17-S130)
LangChain ModelRouterMiddleware 0.0.1a3 at `4af7ab8`	Frameworks	None found in `libs/partners/typesafe/langchain_typesafe/`; each run asks Jev once before the first model call.	`model_router.py` L163-174, L194-197 (R17-S134, R17-S135)
jev-for-chrome at `d5c24de`	Browser and computer use	With OpenRouter or Cloudflare selected, those services decide; new tabs are followed; typed text comes from a separate text model; step, budget, repeat and deadlock limits end the task. 3 acting paths, 5 stop paths.	on the map (R17-S160 to R17-S167)
jev-android at `be8364a`	Browser and computer use	With DeepSeek or the scripted sample policy selected, Jev is not asked; a scenario's first navigation comes before any decision; step and time limits end the task. 3 acting paths, 1 stop path.	`MainActivity.kt` L144-149, L163-165; `BilibiliTriplePolicy.kt` L29-57; `JevAgent.kt` L117-118 (R17-S170 to R17-S173)
second-look at `0c05d16`	Browser and computer use	With no key or on a Jev error a simple "Wait" or "Buy now" rule decides; savings, price alerts, the voice assistant and forwarded order emails act without Jev. 8 skip paths.	`sidepanel.js` L1114-1120; others on the map (R17-S176 to R17-S183)
ilyamk Gmail labeler at `ade42bf`	Email and lead sorting	Already-marked messages are not fetched or are skipped; unreadable messages get the review label without the full-text check; a dry run writes nothing; cost and message limits leave the rest unchanged. 6 skip paths.	`CODE.gs` L1593-1600, L1687-1705, L1759, L2508-2517, L1902-1905 (R17-S184 to R17-S189)

Read scope for four cells (editorial decision, 16 Oct)

Four cells count as recorded with a named, narrower read scope. Each cell on the map states its scope:

openlayer jevals: every entry point in full except the documentation helper integrations/_docs.py, read at its function and return lines; it builds a help text and runs no evaluation.
competitor-hunter: all code and web scripts in full; the two HTML pages' markup checked by script for event handlers (none), their inline scripts read.
ilyamk Gmail labeler: entry functions found by a scripted scan of triggers, menus and web-callable handlers (eleven web-callable functions; no timed or menu-driven jobs); helper functions on the path to the Jev call not read in full.
LlamaIndex JevRerank: the postprocessor in full; the Cloudflare and Vercel provider clients at their branch, return and raise lines.

Malformed cell, one completed cell and corrected texts

competitor-hunter, malformed answer: not recorded → no-decision. An unreadable reply is handled like a Jev error and the item is left out of the ranking; a reply missing its answers stops the ranking stage (fail-closed) (jev.py L149, hunt.py L565-566, L778, L796; R17-S155 to R17-S159).
JevPlaysSnakeGame at b9237bf, bypass, completed: the published "none found" cell covered engine.py only. The whole repository is now read (main.py, src/runner.py, src/game/engine.py, src/game/snake.py, src/agents/jev_agent.py). Still none found: every tick asks Jev (engine.py L125-128); a direct reversal it answers is ignored (snake.py L35-37) (R17-S73). Count unchanged.
Eight cell texts on seven implementations corrected; term and colour unchanged: openlayer jevals (malformed: for single-question evals a missing answer is recorded as an error); did-they-answer (error: retries only for HTTP 429, 5xx and 529, inside the 21-day window); competitor-hunter (error: errored Reddit posts are still listed); Spring AI JevDocumentFilter (below threshold: other filter rules can still drop a passage; no key: the startup exception is for a blank key); ilyamk (error: non-retryable errors end the job with status "error"); DeepEval JevEval (error: the 180 s value is outside the library scope read); second-look (malformed: a verdict outside the four known values shows no verdict card on the web route).

4. Jev 1.14 preview

Read 16 Oct 2026, 05:21–05:26 UTC: Documented unless marked.

Models and aliases: TypeSafe lists Jev 1.14 (jev-1.14.0, "available from October 14, 2026") and Jev 1.13 (jev-1.13.0). jev-latest → jev-1.13.0; jev-preview → jev-1.14.0 (R17-S05). TypeSafe's post: "jev-latest will be updated to jev-1.14.0 in a week. jev-1.13.0 will be available for a while", with 60 days' warning before 1.13 is deprecated (R17-S09).
Context: default 64k tokens per request and 32k per question on both versions; 262k with the opt-in TypeSafe-Danger-Zone-Long-Context header (R17-S05).
Experimental features, not production-ready per TypeSafe: images (both versions) with the TypeSafe-Danger-Zone-Images header; the Box question type (1.14 only); 262k long context (both). TypeSafe: images and Box "aren't yet up to our standards to call ready for production" (R17-S05, R17-S06, R17-S20).
Vendor claims Reported: 1.14 brings "small, subtle improvements to everything", with "the largest gain in multi-lingual support" (R17-S09). Not tested by the pool.
Image data terms: not stated on the images page. The models page says requests are not used for training. The post describes an opt-in for jaggedness debugging and says "We will still not train on your queries" (R17-S05, R17-S06, R17-S09).

Routes for Jev 1.14, checked 16 Oct 2026, 05:21–05:25 UTC
Route	Jev 1.14	Source
TypeSafe direct	`jev-1.14.0`, or `jev-preview`	R17-S05
OpenRouter	`typesafe/jev-1.14` (262,144 context, text and image). `typesafe/jev-1.13` is now also text and image, context 64,000 → 262,144	R17-S22, R17-S23
Vercel AI Gateway	`typesafe-ai/jev-1.14.0` (and `typesafe-ai/jev-1.13.0`); context field 32,000, not reconciled with TypeSafe's 64k default	R17-S27, R17-S28
Netlify AI Gateway docs	`jev-1.14.0` listed	R17-S31
Cloudflare, Opper, Amazon Bedrock, Microsoft Foundry, Workers AI catalog	not listed at 16 Oct 2026, 05:21–05:25 UTC	R17-S29, R17-S30, R17-S32 to R17-S35

SDK releases (package files read, nothing installed): typesafe-sdk 0.7.5 adds image, asset and Box types, default model still jev-latest, failure path last read at 0.7.4; @typesafe-ai/sdk 0.7.0 adds experimental image input, default model still jev-latest, its error path not yet re-read; @ai-sdk/typesafe-ai 3.0.20 adds jev-1.14.0 and sends the image and long-context opt-in headers when used (R17-S37 to R17-S39).
Pages updated: versions and aliases, Where to use, limitations, data privacy, Python, TypeScript, and the comparison card (the Jev "Images" and "Documented errors and limits" cells; prices and other columns unchanged).

5. Claude Code and MCP guardrails

Versions checked 16 Oct 2026: official TypeSafe plugin v0.5.7 (makes no Jev call); danna-zhou/jev-mcp 436469a; jev-claude 7d7e876; jev-mcp 0.14.3 on npm (cells read at 0.11.0); langchain-typesafe 0.0.1a3 (4af7ab8); oh-my-claudecode 5.7.3 (Jev hook files compared with 5.6.2: unchanged apart from one comment); jev-permission-gate v0.2.0 (7349bc9); jev-llm-guard 7c7b1b5 (R17-S200 to R17-S215). One unreleased jev-mcp patch (0.14.4, commits only) noted, not re-audited.
Short answer unchanged: of the 7 guardrails on the page, only LangChain's stops the agent when Jev fails, as installed (1 of 7).
Model ID: oh-my-claudecode and jev-mcp send jev-latest by default, so they will move to jev-1.14.0 when TypeSafe repoints the alias (R17-S213, R17-S246 to R17-S248).
Lead table: 32 Claude Code and MCP projects seen in scans from 6 to 16 Oct 2026, classified from their README and manifest (not counted): 16 hooks or plugins that allow, block, hold or reroute; 12 that return Jev's answer to the agent; 1 skill only; 3 not Jev (16 + 12 + 1 + 3 = 32). 19 come from earlier lists and 13 from this update's scan. Three gates were read at pinned commits; they join the page in a later update.
New title (URL unchanged): Jev Claude Code and MCP guardrails (TypeSafe AI): which hook, plugin or server fails open? Query source: Google autocomplete "jev cl" → "jev claude code"; Bing and DuckDuckGo "jev cl" → "jev claude code" (16 Oct 2026; R17-S53). A suggestion is a demand signal, not a search volume.

6. Small items, scan and drift

Decision Index Reported: Artificial Analysis Decision Index v1.0 (13 Oct 2026; third-party index; not graded: one of four evaluations uses 500 private questions; the model ID or route behind "Jev 1.13" is not stated): Jev 1.13 58, GLiDE 54, Microsoft-Decision-1 54. Read 16 Oct 2026, 05:31 UTC (R17-S60, R17-S61). Added to benchmarks outside the graded ledger; no grade or count change.
pydantic-ai-slim 2.56.0 and 2.57.0: TypeSafe provider unchanged; the model adds one check, a request with images raises UserError before any call; the failure path is otherwise unchanged (R17-S58).
DeepEval 4.3.0: JevEval files unchanged (R17-S59).
Drift notes, not re-audited: patches @ai-sdk/typesafe-ai 3.0.19 and 3.0.20, ai 7.0.142, @effect/ai-typesafe 4.0.5, @openclaw/typesafe 2026.10.4, @jkudish/jev-browser 0.8.8; n8n 2.44.2 published (Jev path last read at 2.44.1) (R17-S57).
Scan window (method): this update's scan covered 14 Oct 16:45 UTC to 16 Oct 05:19 UTC. It starts at the previous scan's cutoff, so there is no gap and no overlap (R17-S50 to R17-S56).
72-hour look-back from 13 Oct 2026, 05:19 UTC: Eevo (Upright; launched 14 Oct 2026 as a "consequential decision model"; hosted) is new, seen, not checked (R17-S62); Jev 1.14 is TypeSafe's own release; SubQ Snap, OpenAI's Decisions API (public beta, price unchanged), Microsoft-Decision-1 (Preview, version 1) and Cloudflare Clef: no change found; Cribl decision 1.0 open weights recorded once (R17-S64); a "Pollinations decisions API" named in third-party repositories was not found on the Pollinations API docs page fetched (unverified; R17-S65).

7. Evidence gaps

Subprocessors were not rechecked (the page needs a browser).
Reddit returned HTTP 403; X search was unavailable; YouTube, TikTok, GitHub trending and GitHub releases were not covered. Those surfaces are unknown, not empty.
@typesafe-ai/sdk 0.7.0: the error path was not re-read.
Three new Claude Code gates (read at pins) and mcowdery/claude-model-guard are not yet on the map; 13 of the 16 gate-class leads were read from README only.
Eevo: terms and pricing not read. Pollinations decisions API: unverified.
Vercel AI Gateway's context field (32,000) is not reconciled with TypeSafe's 64k default.
When the models page dropped the rate-limit numbers is known only as between 14 Oct 2026, 16:46 UTC and 16 Oct 2026, 05:21 UTC.

Method notes and what was not verified

Nothing was run: no account, key, API call, package install, image request or third-party code.
Bypass cells rest on reading the files that start work at each pinned commit; for one large Apps Script file the entry functions were found by a scripted scan, and a few helper functions on the way to the Jev call were not read in full.
Vendor claims are Reported and attributed. Third-party source files were read from raw GitHub at the pins; their URLs and line ranges are in the source ledger.

Key sources and check times (16 Oct 2026, UTC)

Sources by section (R17 IDs)

Access, 05:21–05:26: console (R17-S01); status.typesafe.ai (R17-S04); models page (R17-S05); images page (R17-S06); TypeSafe blog (R17-S08); Jev 1.14 post (R17-S09); Series A post (R17-S19); Box page (R17-S20).
Routes, 05:21–05:25: OpenRouter (R17-S22, R17-S23, R17-S26); Vercel AI Gateway (R17-S27, R17-S28); Cloudflare (R17-S29); Opper (R17-S30); Netlify (R17-S31); Foundry (R17-S32, R17-S33); Bedrock (R17-S34); Workers AI (R17-S35).
When Jev fails: published cells (R17-S67); reads at the pins (R17-S100 to R17-S194, R17-S300 to R17-S313); JevPlaysSnakeGame tree (R17-S73); spot-checks (R17-S68, R17-S72).
Packages: typesafe-sdk, @typesafe-ai/sdk, @ai-sdk/typesafe-ai (R17-S37 to R17-S39); drift list, 05:28 (R17-S57); pydantic-ai (R17-S58); DeepEval (R17-S59).
Claude Code and MCP: repositories, releases and leads (R17-S200 to R17-S261); autocomplete, 05:28 (R17-S53).
Decision Index, 05:31: article (R17-S60); methodology (R17-S61).
Scan, window 14 Oct 16:45 UTC to 16 Oct 05:19 UTC: HN Algolia (R17-S50); GitHub (R17-S51); npm (R17-S52); Hugging Face (R17-S54); DEV.to (R17-S55); Reddit, HTTP 403, and Ollama (R17-S56); Eevo (R17-S62); Cribl (R17-S64); Pollinations (R17-S65).