--- title: "Framework integrations: Pydantic AI, LangChain, Spring AI, Vercel AI SDK, eve, LiteLLM, BAML, TanStack AI, DSPy, Langfuse and other framework-shipped Jev support" type: community source_tier: community tags: [community, integrations, frameworks] created: 2026-09-28 updated: 2026-09-28 confidence: medium sources: - raw/partner/langfuse__docs__evaluation__evaluation-methods__jev-as-a-judge.md - raw/partner/minimax__agent__tools__typesafe-jev.md - raw/partner/langchain__blog__building-prod-with-jev-and-langgraph.md - raw/partner/pydantic-ai__models__typesafe.md - raw/partner/pydantic-ai__models__decision.md - raw/x/pydantic-2100812323027853812.md - raw/partner/langchain__oss__python__integrations__providers__typesafe.md - raw/partner/langchain__oss__javascript__integrations__providers__typesafe.md - raw/partner/spring-io__blog__spring-ai-typesafe-structured-judgment.txt - raw/partner/ai-sdk-dev__providers__ai-sdk-providers__typesafe-ai.md - raw/partner/eve-dev__docs__guides__evaluate.md - raw/partner/eve-dev__docs__evals__judge.md - raw/partner/eve-dev__docs__human-in-the-loop.md - raw/partner/eve-dev__docs__tools__workflows.md - raw/x-repos/vercel__eve.md - raw/partner/litellm__docs__pass_through__typesafe.txt - raw/partner/litellm__docs__auto_router__setup.txt - raw/partner/litellm__docs__proxy__guardrails__typesafe.txt - raw/community/docs-litellm-ai-blog-jev-auto-router-benchmark.md - raw/x-repos/BerriAI__litellm.md - raw/partner/boundaryml__blog__typesafe-ai-jev.txt - raw/partner/boundaryml__blog__baml-0.20.1.txt - raw/x-repos/BoundaryML__baml.md - raw/community/youtube-boundary-ai-that-works-all-about-jev.txt - raw/partner/tanstack__ai__latest__docs__adapters__typesafe.md - raw/partner/tanstack__ai__latest__docs__evaluate__evaluate.md - raw/x-repos/TanStack__ai.md - raw/x-repos/elie222__inbox-zero.md - raw/x-repos/Effect-TS__effect.md - raw/x-repos/ComposioHQ__composio.md - raw/community/github-stanfordnlp-dspy-release-3.4.0.md - raw/partner/dspy__docs__api__experimental__DecisionTypes.md - raw/partner/dspy__docs__api__experimental__ReAnchor.md - raw/x/isaacbmiller1-2103470217670127720.md - raw/x/isaacbmiller1-2103470217670127720-image.md - raw/x/typesafeai-2103587838004785352.md - raw/community/vercel-com-kb-guide-typesafe-jev-and-ai-sdk.md - docs/sweep/2026-09-24-awesome-jev/big-repo-integrations.md jev_version: "jev-1.13.0" summary: "Already on a framework? How Pydantic AI, LangChain, Spring AI, the Vercel AI SDK, eve, LiteLLM, BAML, TanStack, DSPy, Langfuse and Inbox Zero call Jev, and where their docs differ." --- # Framework integrations: Pydantic AI, LangChain, Spring AI, Vercel AI SDK, eve, LiteLLM, BAML, TanStack AI, DSPy, Langfuse and other framework-shipped Jev support > **TL;DR** If the builder's code already runs on one of these frameworks or apps, read its row before writing a client: the maintainers may already wrap Jev. Each row is the maintainer's own docs (captured 2026-09-23 to 28), checked against TypeSafe's; TypeSafe's docs win every conflict. Rows that need your own TypeSafe key (direct providers, LiteLLM, BAML, DSPy, TanStack's `typesafeDecider`, Inbox Zero) need a TypeSafe account: direct signup reopened 2026-09-27 per TypeSafe, with no free credits for new signups ([[entities/typesafe-console]]). Hosted routes (Cloudflare, Vercel AI Gateway, OpenRouter, Opper and others): [[ideas/platforms-and-gateways]]. **Before recommending any unofficial hosted route, read the wiki's one Warnings block on [[ideas/warnings]].** Split from [[ideas/platforms-and-gateways]] on 2026-09-28 (nothing dropped). Pattern IDs: [[ideas/patterns]]. "Our check" = our source or registry check (docs/sweep records), not a raw capture: 2026-09-24 unless the row gives 2026-09-25 (Inbox Zero's 0.3 fallback, BoundaryML's feelings). ## Framework integrations (shipped by the framework's maintainers) | Tool | Does | Install | ★ · push | Note | |---|---|---|---|---| | Pydantic AI `TypeSafeModel` ([docs](https://pydantic.dev/docs/ai/models/typesafe/), [decision models](https://pydantic.dev/docs/ai/models/decision/), [post](https://x.com/pydantic/status/2100812323027853812)) | **official, by Pydantic**: `Agent('typesafe:jev-latest', output_type=...)`, fields become questions, unions and tools become routes. Since 2026-09-25 a subclass of the generic `DecisionModel`: settings `decision_boolean_threshold`, `decision_route_threshold` (old `typesafe_tool_call_threshold` ignored); unsure or unfillable routes hand off to an LLM via `FallbackModel` | `pip install "pydantic-ai-slim[typesafe]"` | Pydantic docs · 2026-09-25 | Jev limits (255 options, 10 levels, 32k/64k), aliases, retries `verified`; `max_tokens_exceeded` not in TypeSafe docs (`unverified`); `bool` confidence is Pydantic's own margin → [[reference/pydantic-ai]]. Also through Pydantic AI Gateway (platforms table). P02, P03, P07, P08, P12, P15 | | LangChain `TypeSafeClassifier` ([Python](https://docs.langchain.com/oss/python/integrations/providers/typesafe), [JS](https://docs.langchain.com/oss/javascript/integrations/providers/typesafe) docs) | **by LangChain**: a Runnable (invoke, batch, pipe). Python `invoke({"state", "questions"})` with `Noul`/`Choice`/`Score`; answers on `.nouls`/`.choices`/`.scores`, plus `model`, `usage`, `request_id`; state may be messages; experimental `ModelRouterMiddleware`, `AutoModeMiddleware`; LangSmith traces. JS: questions set in the constructor, `invoke(state)`, `timeout` 30000 ms, `maxRetries` 2 | `pip install langchain-typesafe`; `npm install @langchain/typesafe @langchain/core` | LangChain docs · 2026-09-24 | same package as @LangChain's `0.0.1a2` post ([[ideas/measurements]]). Key, base URL, answer fields `verified` ([[concepts/primitives]]). Python moved `questions` to `invoke`; JS `stream` yields the whole result. LangChain blog [Building Prod with Jev and LangGraph](https://www.langchain.com/blog/building-prod-with-jev-and-langgraph) (2026-09-25; **vendor**): a LangGraph discovery-review graph asks Jev three questions per page in one request (responsive? personal data → an LLM redacts; privileged → `attorney_review` pauses for a human); LangChain reports Jev "5–6x faster on the classification step" than Sonnet as judge (no sample size; `unverified`), that LangSmith has "dedicated views for decision models like Jev", and that Jev-as-judge scores "barely moved across 100 repeated runs" (`unverified`, no data in the post). P02, P03, P26 | | Spring AI TypeSafe ([post](https://spring.io/blog/2026/09/21/spring-ai-typesafe-structured-judgment), [repo](https://github.com/spring-ai-community/spring-ai-typesafe)), Christian Tzolov | **Spring AI Community** project: `TypeSafeClient` bean; `JevJudge`, `JevSelfRefineAdvisor`, `JevGuardrailAdvisor`, `JevDocumentFilter`, `JevDocumentReranker`, `JevToolIndex`, `JevEvaluator` on Spring AI SPIs; plain Java: `typesafe-java-sdk` | Maven `org.springaicommunity:spring-ai-starter-typesafe:0.1.0`; `TYPESAFE_API_KEY` or `spring.ai.typesafe.api-key` | 0.1.0 · post 2026-09-21 | author: median 275 ms (1 question), 310 ms (3), laptop. Guard thresholds (block > 0.70, review ≥ 0.35) are Spring's. Bare number or boolean `state` → `422` `verified` ([[concepts/state]]). P15, P16 | | Vercel AI SDK via Gateway ([guide](https://vercel.com/kb/guide/typesafe-jev-and-ai-sdk)) | `experimental_evaluate`, `ai` ≥ 7.0.105, `typesafe-ai/jev` | `npm i ai` | Vercel doc | limits, price `verified`; its `boolean` = API `noul`; per-request ZDR `unverified`. Direct route: next row | | `@ai-sdk/typesafe-ai` ([docs](https://ai-sdk.dev/providers/ai-sdk-providers/typesafe-ai)), Vercel AI SDK direct provider | **by Vercel**: `typeSafeAi.evaluationModel('jev-latest')` in `experimental_evaluate`, straight to TypeSafe (no Gateway); `choice`, `score`, `boolean` (sent as `noul`), all questions in one request; `createTypeSafeAi({apiKey, baseURL, headers, fetch})`; Choice and Score confidence under `result.providerMetadata.typesafe.confidence[questionId]`; failures as `APICallError`, AI SDK core retries `429`/`529` with `maxRetries` 2, no provider-side retry; evaluation models only (language, embedding, image factories throw `NoSuchModelError`) | `pnpm add @ai-sdk/typesafe-ai`; key env **`TYPESAFE_AI_API_KEY`** (not `TYPESAFE_API_KEY`); base `https://api.typesafe.ai/v1` | npm 3.0.6, 2026-09-23 (our check) | 255-option max, 2–10 levels, no confidence on a yes/no, `429`/`529` retryable `verified` ([[reference/http-api]], [[reference/rate-limits-and-errors]]); its "1–255 options": docs state only the maximum. "TypeSafe returns scores and probabilities rounded to two decimals" is not in TypeSafe's docs: `unverified`. eve uses it for direct access. Env names: [[reference/environment-variables]]. P02, P16 | | [eve](https://github.com/vercel/eve), Vercel's agent framework ([evaluate](https://eve.dev/docs/guides/evaluate), [judge](https://eve.dev/docs/evals/judge), [approvals](https://eve.dev/docs/human-in-the-loop), [workflows](https://eve.dev/docs/tools/workflows) docs) | **by Vercel**: `typesafe-ai/jev` is the default evaluator in five opt-in features. `evaluate()` (`eve/ai`): typed questions inside tools. `auto()` (`eve/models`): picks the turn's model from described options, reading up to 8 recent messages within 16,000 characters, reused for the whole turn. Tool approval `auto()` (`eve/tools/approval`): `clear` or `caution`; caution, a failed review or incomplete input goes to a person, and the tool input goes to the provider. `agentRouter()` (`eve/tools/agent-router`): input `{ message: string }` (the optional `outputSchema` in the 2026-09-25 capture is gone from the 2026-09-26 one); picks among described subagents; a sole target skips the call. `t.judge` in evals: a yes/no judge scores its probability (eve: "not guaranteed calibrated"); missing credentials or provider errors fail the gate and never switch model | model strings via Vercel AI Gateway (`AI_GATEWAY_API_KEY` or Vercel OIDC); direct: `@ai-sdk/typesafe-ai` + `typeSafeAi.evaluationModel("jev-latest")` | docs captured 2026-09-25 | Any AI SDK evaluation model can replace Jev. eve warns the AI SDK evaluation spec is experimental and can change in patch releases. The approvals page moved from `/docs/tools/human-in-the-loop` to `/docs/human-in-the-loop` and now documents `auto()`. In beta under Vercel's public-beta terms (README). Choice/Score/yes-no semantics `verified` ([[concepts/primitives]]). P02, P03, P05, P15 | | LiteLLM, BerriAI ([pass-through](https://docs.litellm.ai/docs/pass_through/typesafe), [Auto Router](https://docs.litellm.ai/docs/auto_router/setup), [guardrail](https://docs.litellm.ai/docs/proxy/guardrails/typesafe) docs) | **by LiteLLM**, three opt-in proxy features. (1) Pass-through: any path under `/typesafe/` is forwarded (`POST {proxy}/typesafe/v1/systemone`, `GET .../v1/models`) with a LiteLLM virtual key; the proxy adds the TypeSafe key; the response is TypeSafe's, unchanged; spend priced from `usage` and registry entries `typesafe/jev-1.13.0`, `jev-latest`, `jev-preview`, logged under the versioned model. (2) Auto Router `classifier_type: jev`: one Choice `questions.tier` over the tier descriptions picks the completion model; `jev_classifier_config` defaults `jev-latest`, `timeout_ms` 3000, circuit breaker on (30 s cooldown); a timeout, HTTP failure, bad body, unknown tier or open breaker takes the configured fallback (heuristic by default); up to three prior user turns (8,000 characters) go to TypeSafe; Test Routing can incur a Jev charge; the built-in classifier is in OSS, custom instructions and `tier_definitions` need Enterprise. (3) Guardrail `typesafe` (`mode: pre_call`, `default_on` false): one Noul per completed tool exchange ("still needed?"; the 200 most recent, results ≥ `min_chars_to_evaluate` 200, trimmed to `max_result_chars_in_state` 4000); below `relevance_threshold` (0.2 in the docs example) the result is blanked; `unreachable_fallback` `fail_open` by default (30 s budget) | on the proxy: `TYPESAFE_API_KEY`, optional `TYPESAFE_API_BASE` (default `https://api.typesafe.ai`) | all three in stable v1.101.1 and v1.102.1, 2026-09-23 (our check) | Base URL, `jev-preview` alias, $0.042/Mtok input with output free `verified` ([[reference/models-and-pricing]]). `fail_closed`: docs say HTTP 500, the [code at `3fa02ef`](https://github.com/BerriAI/litellm/blob/3fa02ef9fcc287a8648c6d5bc998d9962915699e/litellm/proxy/guardrails/guardrail_hooks/typesafe/typesafe.py#L203) raises 502 (our check): handle both. **LiteLLM reports** ([benchmark post](https://docs.litellm.ai/blog/jev-auto-router-benchmark), Moe Khalil; vendor measurement): on its own author-labelled prompts the Jev classifier matched the expected tier more often than Claude Haiku, faster and far cheaper; labels not independently reviewed, prompts untuned, answer quality not measured ([[ideas/head-to-head-agents]]). P02, P07 | | BAML, BoundaryML ([announcement](https://boundaryml.com/blog/typesafe-ai-jev), [0.20.1 notes](https://boundaryml.com/blog/baml-0.20.1)) | **by BoundaryML** (BAML v1): `client "typesafeai/jev-latest"`, and the function's return type picks the questions: `bool` = Noul thresholded at 0.5, `float` = the Noul probability, enum or finite literal union = Choice (an optional enum adds a `""` abstain option), class = one question per leaf field in one request (nested classes become dotted IDs). **No Score mapping.** Typed results drop probabilities and confidence (read the response envelope through the event hook). Rejected at runtime before any HTTP call: free-form string/int, arrays, maps, recursive or optional classes, `bool?`/`float?`, media, tools, streaming, empty instructions | `typesafeai.Client.new(model, api_key=env.TYPESAFE_API_KEY, base_url, request_timeout_ms, capture_wire)`; default endpoint `https://api.typesafe.ai/v1/systemone` | 0.20.1, post 2026-09-20 (the post says canary; GitHub release not a prerelease, our check); announced 2026-09-17 by Sam Lijin | 255-option max `verified` ([[reference/http-api]]); BAML also demands ≥ 2 options. Wire capture on by default (auth headers redacted). A `route_confidence: float` field asks a new Noul; it does not read the Choice's confidence. P01 | | TanStack AI `@tanstack/ai-typesafe` ([adapter](https://tanstack.com/ai/latest/docs/adapters/typesafe), [evaluate](https://tanstack.com/ai/latest/docs/evaluate/evaluate) docs) | **by TanStack**: `decide({adapter, state, questions})` with `choice()`, `score()` (`levels`), `boolean()`; answers `.value`, `.probability`, `.confidence` (none on `boolean`), `.probabilities`; Score `.value` = nearest level, `.score` = raw fraction; `boolean` `.value` is true at ≥ 0.5 (TanStack's rule); usage and model on `result.meta`. Plain `fetch`, no TypeSafe SDK, no stream. Four transports, same call: `typesafeDecider('jev-latest')` (`TYPESAFE_API_KEY`), `openRouterDecider('~typesafe/jev-latest')` (`OPENROUTER_API_KEY`), `vercelGatewayDecider('typesafe-ai/jev')` (`AI_GATEWAY_API_KEY`), `cloudflareDecider('typesafe/jev')` | `@tanstack/ai @tanstack/ai-typesafe`; `createTypesafeDecider(model, apiKey, {baseURL, defaultHeaders, fetch, timeout})`, default base `https://api.typesafe.ai` | npm 0.1.0, 2026-09-21 (our check) | Model IDs `jev-latest`, `jev-1.13.0`, base URL and the three gateway IDs `verified` (TypeSafe docs; rows below). P01, P05 | | DSPy `dspy.experimental.TypeSafe` ([3.4.0 notes](https://github.com/stanfordnlp/dspy/releases/tag/3.4.0), [decision types](https://dspy.ai/api/experimental/DecisionTypes/), [ReAnchor](https://dspy.ai/api/experimental/ReAnchor/) docs; [post](https://x.com/isaacbmiller1/status/2103470217670127720), Isaac Miller; the release credits the Jev work to @isaacbmiller and @dbreunig) | **by DSPy**, experimental: `dspy.configure(lm=TypeSafe("jev-latest"))`; `Predict` turns a signature's inputs, demos and field criteria into decision requests. Outputs `Noul`, `Choice[...]`, `Score[...]` (`.value`, `.probability` or `.probabilities`, `.confidence`; Score adds `.level`) or native `bool`/`Literal`; a bare `float` is unsupported. Thresholds, Score cuts and Choice weights sit on the predictor (`fields[name]`) and apply locally, so cached answers are reused. No generative fallback, no decision streaming, nested decision outputs not decoded. **ReAnchor** optimizer: fits only those thresholds, cuts and weights to your metric on labelled data; a change must beat the training score and pass a held-out fold check (up to five folds); never edits instructions or demos; no fixed inference budget | `pip install "dspy[typesafe]==3.4.0"`; `TYPESAFE_API_KEY` | 3.4.0 · 2026-09-25 | Question types, fields `verified` ([[concepts/primitives]]). Its Noul `.confidence` is distance from DSPy's threshold (DSPy says so): an extension, the API's Noul has none ([[concepts/noul]]). The extra pins `typesafe-sdk>=0.6.0,<1.0.0`, a range that admits the pre-0.7 releases ([[reference/python-sdk-changelog]]). ReAnchor's claim to calibrate: `unverified`, no results published; it is cut-off tuning on your labels, not a change to Jev's probabilities ([[concepts/confidence]]). TypeSafe quote-posted it on 2026-09-25 ([post](https://x.com/typesafeai/status/2103587838004785352)). P15, P42 | | MiniMax Code "TypeSafe / Jev" page ([page](https://agent.minimax.io/tools/typesafe-jev)) | **by MiniMax**: a launcher with four starter workflows (routing, ranking, value selection, evidence checks) that passes your task to MiniMax Code with a Jev skill instruction; the page says to confirm the skill or plugin in the main app before live use and that live TypeSafe API calls need separate access | from the page | — · captured 2026-09-27 | A coding-agent starter, not a route to Jev: your own TypeSafe (or gateway) key still does the calls. Its advice matches the wiki's (add an `other` route, keep rules in code, a probability is not permission to act) | | [Effect](https://github.com/Effect-TS/effect), [Composio](https://github.com/ComposioHQ/composio) | `@effect/ai-typesafe`; `@composio/typesafe`, `composio-typesafe` | npm / PyPI | 16184, 30293 · 2026-09-23 | READMEs only list them | ## Apps with a built-in Jev option | App | What Jev does | Turn it on | Status | Note | |---|---|---|---|---| | Langfuse "Jev as a judge" ([docs](https://langfuse.com/docs/evaluation/evaluation-methods/jev-as-a-judge)) | **by Langfuse**, marked experimental ("details of the UI and score format may still" change): a decision-model evaluator; you map observation fields into `state`, add 1 to 50 questions (Langfuse's cap), all answered in one call, each written back as one score (Choice → `CATEGORICAL` option, Score → `NUMERIC` expected level, Yes/no → `NUMERIC` P(true)); probabilities, `confidence` and resolved `model` under `metadata.typesafe`; templates for topic, out-of-scope, frustration and a seven-signal conversation check; runs on live rules, batch evaluation and experiments; execution traces in environment `langfuse-llm-as-a-judge` | Settings → LLM Connections, adapter `typesafe`; upstream TypeSafe, Vercel AI Gateway or OpenRouter, billed by that upstream; models `jev-1.13.0`, `jev-latest` | Langfuse docs · captured 2026-09-28 | Limits (255 options, 2 to 10 levels, ~32k for state plus longest question, input-only pricing) `verified` against [[reference/models-and-pricing]] and [[concepts/score]]. **Stale on access:** its FAQ says access "runs through a waitlist" and links the old `/settings/keys` path; TypeSafe reopened signup on 2026-09-27 and the docs link `/keys` ([[entities/typesafe-console]]). State goes to TypeSafe via the chosen upstream: check its data terms. Build and results: [[ideas/builds-apps]] (Langfuse trace judge) | | [Inbox Zero](https://github.com/elie222/inbox-zero) (elie222; open-source email assistant, 12314★ on 2026-09-23; [decision-model code at `de44248`](https://github.com/elie222/inbox-zero/tree/de44248153af42a224b28e0395ea7454c945d318/apps/web/utils/decision-model)) | optional "specialized decision model": Choice over the user's rules, cold-email yes/no, sender categories, thread status, unsubscribe pages, reply-memory selection; raw fetch to `/v1/systemone`, 30 s timeout | operator sets `DEFAULT_DECISION_MODEL=typesafe:` and `TYPESAFE_API_KEY`; `DEFAULT_DECISION_MODEL_ENABLED` (default false) decides whether users who have not chosen get it; users with their own AI key never get it by default; per-user Settings toggle | merged to main 2026-09-19 and 2026-09-21, **no tagged release** (last web tag 2026-05-18); whether the hosted app runs it is not stated | our code read 2026-09-24, not captured (`unverified` beyond that). Any error, or a rule Choice under 0.3 (`MIN_CHOICE_CONFIDENCE`, compared with the answer's `confidence`, or with the chosen rule's probability when it pools content rules; our code read 2026-09-25), falls back to the normal LLM path. Its pricing table lists `jev-latest` at $42 per billion input tokens, output free (= $0.042/Mtok; our code read 2026-09-24; equals the list price, [[reference/models-and-pricing]]). P22 | - BoundaryML's own **feelings** demo (BoundaryML/feelings, names only, our check 2026-09-25) wraps the BAML mapping above as `.feels()`, `.how()`, `.judge()`, `.fill()`; its `.feels()` on numbers is a poor fit ([[concepts/jaggedness-jev-1-13]]). Vaibhav Gupta (BAML) showed it on AI That Works ([video](https://www.youtube.com/watch?v=35PSMmDDKP8), 2026-09-22 stream): a grep-with-vibes command runs `.feels()` on every line of `git log` in parallel and prints the hits (shown: commits that would be risky to revert), and a loop keeps asking an LLM to rewrite a draft while the draft `.feels()` full of corporate jargon, so Jev is the cheap check and the LLM the writer. Asked on stream whether the BAML client returns every option's probability, a host said yes without showing how; per BoundaryML's docs (BAML row) typed results drop them and the envelope is read through the event hook. ## Related - [[ideas/platforms-and-gateways]] — hosted routes: Cloudflare, Netlify, Vercel AI Gateway, OpenRouter, Pydantic AI Gateway, Opper, Fly.io, OpenCode Zen, Zapier - [[reference/pydantic-ai]] — the Pydantic AI integration in depth - [[ideas/tools-and-integrations]] — community MCP servers, skills, plugins; [[ideas/sdks-and-replicas]] — community SDKs by language - [[ideas/head-to-head-agents]] — LiteLLM's Auto Router benchmark in context; [[ideas/builds-apps]] — the Langfuse trace judge build - [[ideas/warnings]] — the wiki's one Warnings block ## Sources Links inline; raw captures in frontmatter (vendor pages under raw/partner/ captured 2026-09-24 to 28; the DSPy posts under raw/x/, images transcribed 2026-09-26; DSPy's 3.4.0 notes under raw/community/, 2026-09-26; LiteLLM's benchmark post 2026-09-25; Vercel's AI SDK guide 2026-09-23; Langfuse's docs 2026-09-28; AI That Works stream transcript under raw/community/; dates in raw/MANIFEST.json). Inbox Zero, LiteLLM's 502 and the version numbers marked "our check" come from the 2026-09-24 source and registry check (docs/sweep/2026-09-24-awesome-jev/big-repo-integrations.md), not from raw captures; Inbox Zero's 0.3 fallback and feelings from the 2026-09-25 vetting (docs/sweep/2026-09-25-mrjev).