Jev (TypeSafe AI) release-1 research: access, cost, trust, build, limits, videos and alternatives

Run: run:manual:e3ce5784cf44ff403a32f4cdaa67052b (migration bootstrap) · Pool: pool_jev_catalog · Checks made: 2026-09-28, 19:41–19:57 UTC.

This report combines three research packages run in parallel on 28 September 2026: A (access, cost and trust), B (build, troubleshooting and limits) and C (videos, daily scan and alternatives). Source IDs in brackets (A-S… for package A, S…/T… for package B) refer to the pool's source ledgers, which are not published; the key sources are linked in section 6 and on the pages each finding feeds.

Previous release: the 28 Sep implementation-path audit, published 2026-09-28 05:21 UTC.

1. Summary answer, as of 28 September 2026

What can I build with Jev?

TypeSafe's Jev (jev-1.13.0) answers typed questions about a text state:

• Choice picks one of up to 255 options you define.
• Score rates on 2–10 ordered levels.
• Noul returns a probability of "yes".

What it cannot do, per TypeSafe's documentation: it does not generate text; it accepts text input only; it is weak at arithmetic, counting, date comparison, multi-hop questions and long irrelevant context.

The documented pattern: your code owns the workflow, Jev makes one narrow decision, and low-confidence answers go to a person or another system.

What exists to learn from:

• TypeSafe lists 18 cookbooks and 4 patterns.
• 46 creator videos show Jev running on concrete tasks, across 13 use-case groups: email and inbox; support tickets; moderation; model routing and agent harnesses; Python; TypeScript; n8n; Claude Code; evaluation and finance checks; paper trading; games and real-time control; browser and scraping; other demos.
• No production claim we checked is independently verified.

What alternatives exist?

• Kev (open weights) is unchanged since 28 Sep. Laya (open weights) released a code-only update, v0.3.21, inside the scan window; no new weights.
• New row Ollaya: an Apache-2.0 local runtime that serves several open decision models behind a System One-compatible API.
• New row Jevper: an Apache-2.0 Python client that puts a Jev-shaped interface on any OpenAI-compatible model. There is no parity evidence.
• All parity figures for alternatives are author-run.

Where can I use it, and at what cost?

Direct from TypeSafe: endpoint POST https://api.typesafe.ai/v1/systemone; price $0.042 per million input tokens, output free. New users can sign up again: the CEO posted this on 27 Sep at 22:33 UTC. In the same post he said free credit for new users is temporarily disabled.

Gateways, all at the same listed input price:

Other cost facts:

• Lookalike resellers charge about 3–16× the list price. The security researchers who found them do not call them malicious.
• What a decision really costs depends on the input tokens counted per request. Third parties report a floor of roughly 250–300 input tokens even for very short requests.

What changed since the previous release (28 Sep, 05:21 UTC)

Corrections to published pages

1. Products page: two rows wrongly labelled official. "TypeSafe router" (A/C, "official-repo example") and "TypeSafe tool-router demo" ("Official-repository Jev routing example") both come from github.com/TypeSafeAI, which describes itself as "TypeSafe Community … This is an UNOFFICIAL organization made for community repos" and was created 2026-09-18 (A-S38). TypeSafe's official organisation is github.com/typesafe-ai, linked from its docs, PyPI and npm (A-S37, A-S39, A-S40). New label for both rows: "community-organisation repo, not TypeSafe official". Their evidence grades otherwise stand.
2. Where-to-use page: route rows.
   • Vercel: Vercel's own models API lists input 0.000000042 (= $0.042/M) and output 0 (A-S32). Its pricing docs say "AI Gateway charges no markup" (A-S33). This replaces "exact provider tariff … unverified". New fields: zdr "none" and no_training "all". The HTML page still rounds to "$0.04".
   • OpenRouter: documents POST https://openrouter.ai/api/v1/systemone for the official SDKs, in addition to /api/alpha/decisions; "There's no waitlist or separate TypeSafe account."; each response carries usage.cost; context is 32,000 tokens for "the state you send plus the questions" (A-S26).
   • New row: OpenRouter Jev Router (typesafe/jev-router, released 25 Sep). See §3.4.
   • Opper: the card now shows 64K context (32K for state plus the longest question) and the example ID typesafe/jev-1.13.0 (A-S35).
   • TypeSafe direct: replace "credits unknown" with "signups reopened 27 Sep; new-user free credit temporarily disabled" (A-S07).
   • Independent proxy (jevtypesafeai.com): prices unchanged since 24 Sep ($0.42, $0.35 and $0.25/M). Eye Security (25 Sep) names it as a lookalike reseller storefront whose traffic passes through a Railway app (A-S42). Move this row to the new reseller page.
3. Confidence and errors (Package B).
   • There is no 250 ms TypeSafe default timeout. Python SDK v0.7.2 defaults to 10 s per HTTP operation, 2 retries and a 30 s total budget. The 250 ms figure was DEFAULT_TIMEOUT_MS = 250 in one project's own hook (oh-my-claudecode 5.5.0), raised to 2000 ms on 22 Sep (B: S08, S23, S38, S60).
   • "2·p_top − 1" is only the two-option case. All 7 documented Choice examples fit (n·p_max − 1)/(n − 1) to within ±0.01; that formula equals 2·p_top − 1 only when n = 2 (B: T02).
   • Score confidence fits the same formula for 3 levels, but not for the documented 4- and 5-level examples. The Score formula is undocumented.

Other corrections to earlier notes (audience notes and plan)

• The 27 Sep reopening is confirmed by the CEO's own post, not only by AI FrontPage (A-S07).
• The $5 new-user credit (20 Sep) rests on secondary reports only, and one first-hand report contradicts it (A-S12).
• The primary security source is Eye Security, published 25 Sep. GBHackers (28 Sep) is secondary. Eye's markup range is 3–11.5×, not 6–11.5×.
• VentureBeat does not mention Bryo or Notra.
• Vercel's command-safety "swap" is contradicted by its CEO's own post (§3.5).
• SDK issues #12 and #15 are not about the Score-criteria 422: #12 is empty Choice criteria, #15 is a near-tie 0.01 below the top probability.
• A bare Noul returns 400; a null state returns 422.
• typesafe-ai on PyPI is a third-party redirect shim, not TypeSafe's package (B: S20).
• "Jev in 25 lines of Python" (nobodywho) is a local imitation built on Qwen3-0.6B, not a TypeSafe tutorial (B: S55).

2. Method, time window and evidence labels

Time window: Package A 19:41–19:52 UTC; Package B 19:41–19:49 UTC; Package C fixed scan cutoff 19:41:00 UTC, checks 19:42–19:57 UTC.

How sources were read

• Direct, unauthenticated HTTP GET; the raw HTML or text was read.
• JSON APIs where they exist: GitHub, PyPI, npm, HN Algolia, Hugging Face, DEV.to, OpenRouter, Vercel, YouTube channel RSS.
• X posts only through X's public oEmbed endpoint, with no login.
• Two VentureBeat articles returned HTTP 429 and were read through a summarising fetch tool with a verbatim-quote prompt.

Evidence labels

• Tested: the pool ran it. Every test in this run was offline.
• Documented: official TypeSafe or route-owner documentation, a first-party post or listing, or source code read at a stated path and tag.
• Reported: a named third party's own claim or measurement, including a reseller's own displayed price.
• Unverified: a lead only.

No pool Jev call. No request went to api.typesafe.ai or to any gateway's inference endpoint. No key, account, purchase or prompt submission was used. Every live status code, latency, token count and accuracy figure in this report is either Documented by the vendor or Reported by a named third party.

Offline tests only (Package B)

• T01, Tested (offline, local mock server, no Jev call). A standard-library Python retry client ran against a 127.0.0.1 mock: a 429 with Retry-After: 1, then a 200: succeeded on attempt 2; three 529s in a row: gave up after 3 attempts; 400, 403 and 422: returned at once, no retry.
• T02, Tested (offline arithmetic). The confidence formula was checked against TypeSafe's documented example responses.
• T03, not completed. The offline SDK install failed: /tmp is noexec and there is no pip. All SDK code in this report is therefore Documented, not Tested.

Other derived values: Package A decoded post times (UTC) offline from X post IDs, and computed reseller price multiples by arithmetic, marked "derived".

Analytics were unavailable for this run (the run's analytics status was "unavailable"). No search volumes are stated.

3. Findings per release-1 page

3.1 Access status (where-to-use/access-status): ships

The latest event has a primary, dated source.

Status line (checked 28 Sep, 19:41–19:50 UTC): signups are open again (CEO post, 27 Sep, 22:33 UTC). New users get no free credit, "temporarily", per the same post. We found no later statement restoring the credit. We could not observe the console signup itself: console.typesafe.ai returned HTTP 403 to our client.

"Is it down?" status.typesafe.ai exists on TypeSafe's domain and is hosted by Better Stack. Over the 90 days it displays (from 1 Jul): API uptime 99.818%, console uptime 99.976%. TypeSafe's homepage and the docs pages we checked do not link it. An aggregator (didcodexreset) reports that TypeSafe's incident post on X did link it (Reported). Suggested wording: "TypeSafe runs a status page at status.typesafe.ai; its website and docs did not link it on 28 Sep."

Stale third-party pages on 28 Sep: jev-ai.live still says "grab $5 in free credit"; jevaiplayground.com still says "New signups paused"; jevmodel.org still shows a 21 Sep "no waitlist" banner; flaviocopes still says signups were paused as of 24 Sep.

What to do now: (1) sign up directly, and expect no free credit; (2) or use a gateway — OpenRouter needs no TypeSafe account; (3) do not buy from lookalike storefronts.

3.2 Official vs reseller (where-to-use/official-vs-reseller): ships

Official surfaces, as TypeSafe itself links them (Documented): typesafe.ai, docs.typesafe.ai, console.typesafe.ai (keys and playground), api.typesafe.ai; status.typesafe.ai exists but is not linked from the site; GitHub github.com/typesafe-ai; PyPI typesafe-sdk 0.7.2 (26 Sep); npm @typesafe-ai/sdk 0.6.0 (15 Sep); agent skill typesafe-ai/skills; X @typesafeai; sales and support email addresses on the typesafe.ai domain.

Not official: github.com/TypeSafeAI, a self-described unofficial community organisation (A-S38); PyPI typesafe-ai, a third-party shim (B: S20).

Five-step check for the page

1. The domain is typesafe.ai or one of its subdomains.
2. The packages are typesafe-sdk or @typesafe-ai/sdk, with code at github.com/typesafe-ai.
3. The key comes from console.typesafe.ai or a named gateway. A reseller's key is not a TypeSafe key.
4. The price is $0.042/M input with free output. Anything higher is a markup.
5. The site names its operator and discloses any affiliation, and you know what it asks for (card, key or prompts). Never paste a TypeSafe key into a third-party site; jev.works asks for one.

Reseller prices as displayed, 28 Sep, 19:46–19:47 UTC (multiple = displayed price ÷ $0.042/M; derived)

Other domains: jev-ai.live, jevaiplayground.com, jevplayground.com, typesafe.pro, typesafeai.app and jev.works are information sites, playgrounds or gateways, not resellers of Jev tokens. jev-router.com is independent and "do[es] not provide access to TypeSafe's Jev model". Package C flagged, but did not open, these trust leads: jev-ai-desktop/Jev-AI-Desktop (its description is a keyword list), the Hugging Face Spaces jevaimodel/*, and JEV-27B-GGUF.

Wording rule, taken from the originals: Eye Security calls these sites "fake storefronts" and a "rip-off", but also says: "This isn't illegal as far as we can tell." Cyber Security News: the report "does not describe a malware infection or establish that operators stole prompts." Only Mallory, an aggregator, uses "fraudulent or unaffiliated". No named security source calls any of these domains malicious. Use: "lookalike reseller storefront (Eye Security, 25 Sep)". Also flag jev-ai.pro's "Jev-Omni" (image, audio and video): TypeSafe's Jev 1.13 accepts text only.

3.3 Cost per decision (where-to-use/cost-per-decision, with worksheet): ships

Billing basis (Documented): "Charged per input token. Output tokens are free." $0.042 per million tokens ($42 per billion) (A-S21). Responses still report output_tokens (18–34 in the docs examples); the worksheet should ignore it for TypeSafe direct. usage.input_tokens is the "Number of input tokens used", or None if not reported (A-S62, A-S63). On OpenRouter, each response includes usage.cost (A-S26).

What counts toward input tokens (not fully documented): the state is paid once per request — the parallel-questions cookbook: "N single-question calls pay for it N times … the batched call pays once" (A-S61). Question keys are "not sent to the underlying model". Whether question and criteria text is billed is not stated; arithmetic on the cookbook's outputs suggests about 57 tokens per extra question (derived; assumes a price constant we could not see). The docs' own examples report 296–318 input tokens for a one-sentence state and one short question.

Third-party measurements (all Reported)

Worksheet design: the user enters their own measured input_tokens per request; show the reported overhead (~257 tokens, Opper) as a reference, not a default; fallback share, fallback model price and retry rate are entered by the user; no competitor LLM prices are published.

3.4 Route matrix (where-to-use hub, checked 28 Sep 19:41–19:50 UTC)

Name collision: jev-router.com is unrelated to OpenRouter's typesafe/jev-router.

Python base URLs from the SDK docs (B §1.5): OpenRouter https://openrouter.ai/api, model ~typesafe/jev-latest; Vercel https://ai-gateway.vercel.sh/typesafe, model typesafe-ai/jev; Pydantic AI Gateway https://gateway-us.pydantic.dev/proxy/typesafe.

3.5 Production-claims ledger (products)

None of these claims is independently verified.

Keep the bounded finding "no independently verified production deployment in the audited set", re-dated 28 Sep.

3.6 Python (build/python): ships

Official SDK (Documented): package typesafe-sdk, imported as typesafe_sdk; version 0.7.2 (PyPI upload 2026-09-26T21:20:23Z; GitHub release 21:20:32Z); Python 3.10 or later; MIT; repository github.com/typesafe-ai/typesafe-sdk-python. Install: pip install typesafe-sdk; the [http2] extra is new in 0.7.2. Defaults: base URL https://api.typesafe.ai, model jev-latest, timeout 10 s. Release history: 0.5.7 (11 Sep on PyPI; the vendor gives 12 or 14 Sep), 0.6.0 (15 Sep), 0.7.0 (18 Sep, Pydantic migration), 0.7.1 (21 Sep), 0.7.2 (26 Sep).

Minimal call (Documented, not executed): SDK TypeSafeClient().system_one(state=…, questions={…Choice/Score/Noul…}, model="jev-1.13.0"). Raw HTTP POST https://api.typesafe.ai/v1/systemone with a Bearer key and a body of state, model and questions. Response fields: model; answers.<id> — Choice: choice, probabilities, confidence; Score: score, legend, probabilities, confidence; Noul: noul only; usage.input_tokens and usage.output_tokens; header x-typesafe-request-id.

Raw-HTTP retry client: Block C in Package B is Tested (offline, local mock server, no Jev call) and copies the SDK's defaults.

Gotchas (B §1.4, 12 rows)

Docs inconsistency: the jaggedness page uses model="jev-1.13", which is not listed on the Models page and was not tested. Use jev-1.13.0.

3.7 Errors and rate limits (build/errors-and-rate-limits): ships

Status codes: the API reference documents only 401, 422, 429 and 529. SDK v0.7.2 (src/typesafe_sdk/_core/errors.py, source read) also maps 400, 403 and 404, and sends every 5xx to one server-error class. Live behaviour, reported in named, dated issues: 400 (not 422) for rule violations — a bare Noul, more than 255 options, more than 10 levels, an empty key, or an unknown model; 403 JSON "Must supply an API key!" when the key is missing, while an invalid key returns 401; a Cloudflare HTML 403 when state contains a shell curl … https:// command (check Content-Type before rotating keys); 422 only for schema failures — a null state, a wrong type, or Score criteria sent as a map.

SDK retry defaults (Documented): retries 408, 429 and every 5xx (including 529), plus connection errors and timeouts; 2 retries; backoff starts at 0.5 s and doubles up to 5 s, minus up to 25% jitter; honours Retry-After and retry-after-ms; 10 s timeout per operation; 30 s budget per call.

Limits (Documented): 255 Choice options; Score "at least two" levels, maximum 10; 64k tokens per request, 32k for state plus the longest question; 250,000 tokens/s and 1,200 requests/min, "adjusting dynamically"; text-only input.

Observed behaviour (Reported): two independent testers saw no 429 on the direct route — one sent 145 sequential questions in 81 s, the other 120 requests/min for 1–2 minutes. Whether 529 occurs in practice is unverified.

Route differences (Documented by route owners): OpenRouter adds 402 (credits), 403 (moderation), 408, 502 and 503. Vercel's free tier returns 429 rate_limit_exceeded. Cloudflare uses its own Workers AI codes. Reported: the Bifrost gateway drops the Retry-After and request-id headers.

Fallback procedure: (1) classify the failure; (2) retry only transient failures, within a budget; (3) before shipping, choose per decision whether to fail closed, fail to review, or fail open; (4) test the no-key path and the outage path separately. Worked examples: Jevmail, WordPress Jev Comment Triage and QuantDinger, carried forward from the 28 Sep audit and not re-audited.

3.8 Confidence thresholds (build/confidence-thresholds): ships

What the fields mean: Choice confidence is "derived from" the probabilities; it measures how peaked they are. In TypeSafe's examples it matches (n·p_max − 1)/(n − 1), which is 2·p_top − 1 for two options (T02, offline check on the documented examples). It is not the probability that the answer is correct. Score: the confidence formula is undocumented; the 4- and 5-level examples do not fit. Noul: no confidence field; noul is P(yes). Probabilities are reportedly rounded to 2 decimals, which can flip near-ties (#15).

TypeSafe's own claims (vendor claims): Jev is trained with RLCD to return "calibrated" decisions; no calibration metric is published on the pages we read; thresholds "depend on your domain".

Independent calibration results (Reported): AnthusAI (19 Sep, no licence): 8,801 constructed sentiment examples on jev-1.13.0; raw ECE 0.117, 0.052 after Platt scaling, 0.008 after isotonic recalibration. Adilmp (19 Sep, no licence): 8,000 judgments; raw ECE 0.157–0.209, 0.006–0.023 after recalibration; AUC about 0.91.

Critiques: Alex Molas, "Jev can't be calibrated" (23 Sep; HN 65 points, 61 comments; thread text not read). Prefactor (25 Sep), a vendor blog.

Tools: two unrelated repositories are both named jevcal: abhixhek/jevcal (MIT, uses a simulator) and Adilmp/jevcal (MIT); link each by owner. jevkit-calibrate 0.1.0 (MIT).

Procedure (7 steps): freeze the question and pin the model; label real cases; record the probabilities; check calibration and recalibrate if needed; sweep cutoffs per question, weighing the cost of wrongly accepting against the review load; evaluate once on held-out cases; re-run whenever anything changes. Project defaults may appear only as examples of one project's choice: QuantDinger 0.55; Jevmail 0.15 margin; WordPress 0.75 / 0.65 / 0.45. The site recommends no threshold.

3.9 Limitations and fit check (limitations): ships

Sources: the official jaggedness page ("Applies to jev-1.13. Last reviewed 2026-09-17") and the Models page (checked 28 Sep). 17 limitation rows, each with the documented workaround: (1) literal reading; (2) arithmetic and counting ("Jev is not a calculator"); (3) a Score is not a number; (4) dates are read as text; (5) multi-hop indirection; (6) large, irrelevant state; (7) adversarial content; (8) contradictory instructions; (9) no structural invariants (example: 0.72 + 0.47 = 1.19); (10) no text generation; (11) text-only input; (12) 64k/32k context limits; (13) English is the primary language, CJK and other languages handled "not equally well"; (14) a closed answer set; (15) questions answered independently; (16) rate limits are unstable; (17) no per-customer fine-tuning. These are vendor-documented failure modes, not measured rates.

Fit check: all three existing questions (bounded answer space; code owns the action; a fallback for uncertainty) are directly supported by official lines, for example: "Low confidence: Do not act. Route to a human, request clarification, or fall back to a different system." The tool runs locally and calls no model.

3.10 Use-cases hub: "documented examples" rows

docs.typesafe.ai/llms.txt (28 Sep) lists 4 patterns, 18 cookbooks and 1 demo; titles and URLs are in B §5.

• Patterns (4): speculative fan-out, confidence-gated routing, composite scoring, intent routing.
• Cookbooks (18): self-consistency (nouls; choices), parallel questions, re-ranking, line-by-line search, structure recovery, function calling, skill suggestion, entity alignment, classifying RAG passages, citation check, LLM guardrails, SDE cascade, date extraction, pre-parsed value extraction, hierarchical classification, autoresearch feature discovery, classification using confidence.
• Demo (1): smart-home assistant.

Figures in cookbook descriptions are vendor-reported, for example "12.2x cheaper and 10.0x faster" and top-1 accuracy rising from 5% to 18%. The HTML pages were not opened one by one; the titles come from the index.

3.11 Videos (videos): ships

Counts: 78 distinct videos inspected (76 YouTube, 1 X, 1 Loom). 46 kept (48 table entries) across 13 use-case groups. 13 held (demo start not verifiable, or author or date not established; includes the X launch video and the Loom smart-home demo). 19 excluded: overview, news, review or interview 10; talks about Jev without showing it 4; compilation 2; alternative model only 2; mock data 1. This clears both the goal (at least 12 videos over at least 6 use cases) and the ship rule (at least 6 entries over at least 4 use cases).

Kept demos also include Italian, German, Portuguese, Russian, Chinese and Japanese videos.

How start times were established: every start time is the creator's own chapter marker, which also appears as a timestamp line in the video description (method creator-chapter for 47 entries, description-timestamp for 1). No video was watched. YouTube's player showed a login wall, so no frames and no transcripts were viewed. That Jev is being called on screen at the linked second is the creator's labelling, not the pool's observation.

Required statement for the page: "Each timestamp is the creator's own chapter marker, also listed in the video description. The pool has not watched the frames or read transcripts. Figures quoted in videos are the creators' own and were not reproduced."

Row flags: vendor participant HHUsHkYhkcM (LangChain × TypeSafe), QkPnAoHBXwo (ThursdAI); replayed saved outputs, not live calls cDhFRtHKS7E; paper trading ymgH8jS6Wb8; paper or live not stated, and the creator's page mentions running mocked without a key 8DgRqDukf-U; mocked tools QEcOXG85aIs; sponsored segments bA8WeHYmJko, l87PznsmM-g, 4u6-uiDpJ6o; use only from 12:10 onward QbYBRjOaGOo; conflicting first-answer figures, quote neither pmnq5e5Xp4s.

All kept rows, with timestamped links, are published on the videos page.

3.12 Alternatives delta (alternatives)

Licences are recorded per artifact and never carried over from one artifact to another.

New leads from the window, not verified ("seen, not yet checked"): Jeva.cpp (llama.cpp fork, MIT); Gevva0 (Gemma gateway, Apache-2.0); Rene-1 31B (claims "+9 over Jev"; a benchmark claim only); llama-system1 (npm shim); OllamaMQ; kev-onnx (CPU ONNX server for Kev). AnyJev v0.2.0 was released in the window; whether it is the catalogue's "JevAny" row was not verified; do not merge them.

4. Standing preceding-24-hour discovery scan (Package C)

Cutoff: 2026-09-28 19:41:00 UTC. Window: 2026-09-27 19:41:00 UTC (inclusive) to 2026-09-28 19:41:00 UTC (exclusive). Checks ran about 19:42–19:57 UTC.

Result: a very active window. Items with an exact source timestamp inside the window: 27 Jev-related HN stories; 478 GitHub repositories created that match "jev" (57 with the jev topic); releases Laya v0.3.21, Ollaya v0.7.4 and v0.7.5, jevgrep v0.4.0–0.4.4, AnyJev v0.2.0; at least 23 npm "jev" packages with a last publish inside the window; 7 DEV.to articles tagged jev; 19 YouTube videos with an RSS publish time inside the window; PyPI laya 0.3.21. These are search-index counts, not verified integrations or adoption.

Notable timed items (all in C's candidate timing ledger): Laya v0.3.21 at 19:41:23Z, 23 seconds inside the window (code change only); Ollaya v0.7.4 and v0.7.5; five jevgrep releases; kev-onnx, created 19:44:29Z; HN posts Jeva.cpp, Rene-1, Gevva0, Jevdit, Polylane (17:37:04Z), and the critique "You shouldn't use Jev for coding agents and routing" (19:36:13Z); trust-flagged lookalike repositories and Spaces (not opened).

proposal:jev-x-monitoring-access remains unapproved; the manager carries it forward on the v0.2 harness. Nothing was bought. The 25 Sep 40-minute/500-batch experiment was not repeated.

5. Evidence gaps and open questions

1. Account-level facts. Unknown: whether a new account gets any credit today, live quotas, and an actual bill. The console returned 403, and the rules forbid creating an account. There is no primary TypeSafe source for the "$5 / 120M tokens" amount.
2. No live Jev behaviour was observed by the pool. The 400/403/422 behaviour, the 429/529 headers and the scope of the rate limits (per key or per account) are Reported only. The SDK was not executed (T03 failed). model="jev-1.13" was not tested.
3. Confidence. The Score confidence formula for 4 or more levels is unknown. The calibration audits each cover a single dataset and one model version.
4. Billing. Whether question and criteria text is billed is undocumented; the cookbook arithmetic is only an indication. How Jev Router bills the routed model, and what it does on failure, is undocumented.
5. Data terms, deferred to a later run. OpenRouter's retention and logging for Jev; the text of TypeSafe's MCA, privacy policy and DPA (not re-read); the billing entity for direct accounts.
6. Resellers. jev-agent.com (named by Eye Security) was not checked; jev-ai.pro's monthly prices were not shown in its default view; the body of the cyberpress article was not quoted; the trust-flagged repositories and Spaces were not opened.
7. Production claims. TechCrunch's Vercel article was not checked; there is no primary source for Notra; FundaAI's method is paywalled; both VentureBeat pages were read only through a summarising tool (HTTP 429).
8. Videos. No frames or transcripts were viewed; the route or SDK is not stated for several kept videos; whether 8DgRqDukf-U used a real key is unknown; truncated repository links were not resolved; the "AI Revolution" farm channels were not re-inspected.
9. Alternatives. Ollaya's per-model licences and the origin of its benchmark were not checked; Laya's System One compatibility and the card model's context were not re-checked; Kev's base-model licences were not checked; whether AnyJev is the catalogue's JevAny is unresolved; the six new leads are unverified.
10. Scan coverage. Reddit, X, TikTok, Lobsters and GitHub code search are unknown for the window; DEV.to coverage is complete only for the jev tag; npm and Hugging Face counts are capped.
11. Benchmarking terms. jevcal claims TypeSafe's agreement restricts publishing benchmark results. This was not checked; deferred.

6. Sources

Ledgers. The pool keeps four source ledgers for this run (78 sources for package A, 82 entries for package B including the offline tests T01–T03, the scan and alternatives sources for package C, and one line per inspected video). They are not published; each line records access time, source date, label, an excerpt of at most 25 words and the outcome.

Key primary sources (all accessed 2026-09-28)

TypeSafe documentation, code and packages:

• Launch post (15 Sep): https://typesafe.ai/blog/introducing-system-one-models-and-jev
• Models: https://docs.typesafe.ai/models
• API reference: https://docs.typesafe.ai/api
• Confidence: https://docs.typesafe.ai/confidence
• Jaggedness page (reviewed 17 Sep): https://docs.typesafe.ai/model-jaggedness/jev-1.13
• Docs index: https://docs.typesafe.ai/llms.txt
• Parallel-questions cookbook: https://docs.typesafe.ai/cookbooks/parallel_questions
• SDK v0.7.2 source, errors: https://github.com/typesafe-ai/typesafe-sdk-python/blob/v0.7.2/src/typesafe_sdk/_core/errors.py
• SDK v0.7.2 source, retry: https://github.com/typesafe-ai/typesafe-sdk-python/blob/v0.7.2/src/typesafe_sdk/_core/retry.py
• PyPI: https://pypi.org/project/typesafe-sdk/
• npm: https://www.npmjs.com/package/@typesafe-ai/sdk

TypeSafe status page and posts:

• Status page: https://status.typesafe.ai/
• @typesafeai, 20 Sep: https://x.com/typesafeai/status/2101786156572823624
• @typesafeai, 22 Sep: https://x.com/typesafeai/status/2102281508950307159
• CEO, 27 Sep: https://x.com/CompleteSkeptic/status/2104338649999626397

GitHub organisations: official https://github.com/typesafe-ai; unofficial https://github.com/TypeSafeAI

Route owners:

• OpenRouter: https://openrouter.ai/typesafe/jev-1.13
• OpenRouter Jev docs: https://openrouter.ai/docs/guides/community/jev
• OpenRouter Jev Router: https://openrouter.ai/typesafe/jev-router
• Cloudflare: https://developers.cloudflare.com/ai/models/typesafe/jev/
• Vercel model page: https://vercel.com/ai-gateway/models/jev
• Vercel models API: https://ai-gateway.vercel.sh/v1/models
• Vercel pricing: https://vercel.com/docs/ai-gateway/pricing
• Netlify: https://www.netlify.com/changelog/typesafe-jev-ai-gateway/
• Opper: https://opper.ai/typesafe/jev-1-13-0

Security: Eye Security, "Rise of the Jev-Clones" (25 Sep): https://research.eye.security/rise-of-the-jev-clones/

Cost measurements:

• AY Automate: https://www.ayautomate.com/blog/jev-pricing-cost-per-decision
• Opper: https://opper.ai/blog/jev-vs-kev-open-decision-model
• NavyaAI: https://www.navyaai.com/blog/jev-typesafe-limitations-production

Production claims:

• Brizz: https://www.brizz.ai/blog/replacing-llm-classifier-with-jev
• Polylane: https://polylane.com/blog/we-swapped-our-llms-for-jev/
• Vercel CEO: https://x.com/rauchg/status/2100307962262872105
• Bryo CTO: https://x.com/nikhilmudholkar/status/2100604560335139083

Community issues (Reported):

• https://github.com/typesafe-ai/skills/issues/1
• https://github.com/typesafe-ai/typesafe-sdk-js/issues/6
• https://github.com/typesafe-ai/typesafe-sdk-js/issues/15
• https://github.com/Yeachan-Heo/oh-my-claudecode/issues/4091

Calibration audits (Reported):

• https://github.com/AnthusAI/Jev-Calibration
• https://github.com/Adilmp/does-jev-confidence-mean-anything

Alternatives:

• https://github.com/ollaya-dev/ollaya
• https://github.com/zhulinchng/jevper
• https://github.com/jaredpalmer/kev
• https://github.com/NandhaKishorM/laya