# Jev (AI model)

> Source: https://aiwiki.ai/wiki/jev
> Updated: 2026-09-16
> Fact-checked: 2026-09-16
> Categories: AI Inference, AI Models, Machine Learning
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Jev (AI model)." aiwiki.ai, 16 Sept 2026. https://aiwiki.ai/wiki/jev
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Jev is an AI model released in early access on 15 September 2026 by [TypeSafe AI](https://aiwiki.ai/wiki/typesafe_ai), a San Francisco startup founded in 2024 by former [OpenAI](https://aiwiki.ai/wiki/openai) researcher [Diogo Almeida](https://aiwiki.ai/wiki/diogo_almeida) with Erik Gafni and Sasha Sheng. The company describes Jev as the first of a new class it calls "System One Models": instead of generating text one token at a time, the model takes an unstructured input (the "state") together with a set of typed questions, and returns, in a single parallel pass, a typed answer for every question with a probability (or probability distribution) and, for Choice and Score questions, a confidence score.[1][4] TypeSafe says Jev was trained with a method it names Reinforcement Learning for Calibrated Decisions (RLCD), positions it as a "frontier-intelligence function call" for automation rather than chat, and prices it at $0.042 per million input tokens with output tokens free.[1] The model was announced the same day the company emerged from stealth with roughly $40 million in seed funding led by DCVC.[2]

All speed, cost and accuracy figures in this article are TypeSafe's own measurements unless another source is named. The company itself says its published evaluations were "generally run from our laptops on the West Coast", that it cannot prove its pricing is not subsidized, and that the workflows it used were written by its own model capabilities team.[1]

## Announcement

TypeSafe published the launch post "Introducing System One Models & Jev", written by Almeida, on 15 September 2026, and a Business Wire press release the same day announcing that the company had emerged from stealth with $40 million in seed funding led by DCVC, with DCVC general partner James Hardiman quoted.[1][2] Almeida's own post on X that day read, in part: "After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I've spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev", followed by the bullets "20-200x faster" and "40-400x cheaper (w/ output tokens free)".[19]

The "co-inventor of ChatGPT" and "co-inventor of RLHF" descriptions are the company's and Almeida's own. What is independently verifiable is that Almeida is the fourth listed author of the 2022 [InstructGPT](https://aiwiki.ai/wiki/instructgpt) paper, "Training language models to follow instructions with human feedback" (Ouyang et al.), which OpenAI's ChatGPT announcement described as a "sibling model" to [ChatGPT](https://aiwiki.ai/wiki/chatgpt), trained with the same method; the [RLHF](https://aiwiki.ai/wiki/rlhf) technique itself predates that paper.[23] TypeSafe's team page also says Almeida was previously at Google Brain.[27]

Jev is not generally available. The press release describes it as "currently available in early access for select developers" with a waitlist at typesafe.ai, and the launch post says the company is "bringing developers off the waitlist as quickly as we can".[1][2] The public documentation, a web playground (console.typesafe.ai) and client SDKs went live around the launch: the JavaScript and TypeScript SDK's changelog records an initial public release, v0.5.7, on 11 September 2026, and the Python SDK's records v0.5.7 on 14 September, with both moving to v0.6.0 on 15 September.[13][14]

## System One Models

TypeSafe uses "System One Model" for the model class and "Jev" for its first public instance. The class name is borrowed from Daniel Kahneman's book Thinking, Fast and Slow, which distinguishes fast, intuitive "System 1" thinking from slow, deliberate "System 2" reasoning; the company says the emphasis is on "fast, focused judgments", and acknowledges that "System 1 thinking" has also implied error-prone, adding that "for reasons we will get into in the future, we believe System One Models can be made more reliable than its alternatives".[1][5]

The launch post frames the class against existing [large language models](https://aiwiki.ai/wiki/large_language_model) along several axes. The comparison below is the company's own framing, condensed from that post.[1]

| Axis | LLMs, as characterized by TypeSafe | System One Models (Jev), as claimed by TypeSafe |
|---|---|---|
| Training objective | [RLHF](https://aiwiki.ai/wiki/rlhf) (human preference) and [RLVR](https://aiwiki.ai/wiki/rlvr) (verifiable rewards) | RLCD: "answers with epistemically honest probabilities on System One tasks" |
| Input emphasis | Sequential messages | Structured program state |
| Output | Strings, which must be parsed and validated before software can use them | Type-safe structured values whose possible outputs are defined in advance, with calibrated probabilities and confidence scores |
| Sampling | Sequential, one token at a time | Parallel: all outputs in a single query |
| Cost | $0.20 to $10 per million input tokens; output about 5x input | $0.042 per million input tokens; output free |
| Speed | End-to-end response of 3 to 329 seconds for frontier models | 70 ms to 500 ms end to end, "40x-200x faster for the same levels of frontier intelligence for System One shaped queries" |
| Confidence | Prompted confidence estimates "tend to be overconfident and inconsistent" | Confidence and uncertainty returned with every output |
| Intended uses | Human-in-the-loop tasks, verifiable problems, demos | "Smart if-statements", map-reduce over large data, real-time applications, scoring and verifying LLM outputs |

The "3 to 329 seconds" figure for LLM latency links, in the launch post, to a third-party leaderboard mirror of [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) data rather than to a TypeSafe measurement.[1]

The company's documentation is explicit that a System One model "does not write replies, produce code, or generate explanations of their reasoning" and that decomposition is the intended way to handle anything that needs extended reasoning: ask each factor as a separate question and combine the results in code.[4][5] A FAQ answer in the launch post states that "Jev is neither small nor an LLM".[1] The launch post, documentation and press release give no parameter count, architecture details or training compute, and the only size-related figure is the per-request token budget described below.

## How the API works

A request to Jev has three parts: a `state`, a `model` identifier (the documentation examples and SDK default use `jev-latest`), and a dictionary of `questions`. The endpoint is `POST https://api.typesafe.ai/v1/systemone`.[5][12] The state can be a plain string ("My card was charged twice."), a JSON object with named fields, or an array of messages or records; the documentation recommends an object for most requests so each part of the context has a descriptive name.[28]

Each question has an ID, a `type` and `instructions`, and there are three question types, which TypeSafe calls primitives.[6][7][8][9]

| Primitive | Question it answers | Returns | Documented limits |
|---|---|---|---|
| Choice | Which one of these options? | `choice`, `probabilities` over every option, `confidence` | Up to 255 options; each option "costs a few tokens"; the docs suggest adding an `other` option when the list may not cover every input |
| Score | Which level on an ordered scale? | `score` (a position that can fall between levels), `legend`, `probabilities` over levels, `confidence` | At least two and up to 10 levels; the score is the probability-weighted mean of the level numbers |
| Noul | Is this statement true? | `noul`, a probability from 0 to 1 that the answer is yes | No separate confidence value; optional `criteria` describing what yes and no mean |

Every question in a request sees the same state and is evaluated independently and in parallel; the documentation says adding questions "barely changes the response time" and that the number of questions is limited only by a shared request budget of "around 32,000 tokens, roughly 150,000 characters of English text".[4][6] Because questions are isolated, the company argues, adding more of them does not cause the context degradation it calls "context-rot".[4]

The `confidence` field on Choice and Score answers is not a separate model output. The documentation describes it as a statistic computed from the shape of the returned probability distribution: a distribution concentrated on one option gives high confidence, a flat one gives low confidence, and developers are free to compute their own measure from the raw probabilities instead.[10] The documentation's worked examples show how this plays out: a support ticket that clearly names a wrong size returns `wrong_size` at confidence 1.0, while a "requested_resolution" question about a customer who never said what they wanted returns confidence 0.16 because probability is spread across the options.[7] For Score questions, the docs caution that different distributions can produce the same score (a score of 1.0 can mean all probability on level 1, or half on each of levels 0 and 2) and that a confidence of 1.0 "describes the model's answer, not a guarantee that the answer is correct".[8]

TypeSafe's suggested architecture is to gate actions on confidence: act automatically above a high threshold, ask for confirmation in a middle band, and route to a person or a reasoning model below a floor, with thresholds set higher for destructive operations than for read-only ones.[10] The documentation lists named patterns built on this, including "speculative fan-out" (send many questions, including speculative ones, in one call and let code decide what matters), "confidence-gated routing", "composite scoring" and "intent routing".[4]

The company also published an MIT-licensed Python package, `system-one-adapter`, that exposes the same `system_one` call backed by OpenAI-compatible or Anthropic APIs instead of Jev, using native [structured output](https://aiwiki.ai/wiki/structured_output) modes or prompted JSON with client-side validation and corrective retries. Its README says it is "useful for comparing TypeSafe against an LLM on cost/speed/intelligence", and it is the wrapper TypeSafe used for the LLM side of its workflow evaluations; the repository was created on 8 August 2026.[17][1]

## Training: RLCD

TypeSafe says Jev was built on "a new model architecture, parallel sampler for maximum efficiency, and training method we call Reinforcement Learning for Calibrated Decisions (RLCD)".[1] The documentation presents RLCD as a third post-training path alongside RLHF, which "turned pretrained models into chatbots", and RLVR, which "created reasoning models that are strong at tasks such as mathematics, but slower and more expensive". Under RLCD, as the company describes it, the model does not generate text, returns decisions and probabilities, and is trained so that "higher probability should correspond to a greater chance that the answer is correct".[11]

The company's stated rationale is that RLHF optimizes for text a human rater prefers, which it calls the wrong task for automation, and that RLVR "is great for tasks with simple programmatic verification, but most real-world judgement tasks don't fit into that shape", producing what it calls "spikey / non-robust intelligence".[1] Its AI primer adds that preference optimization causes "mode dropping", a narrowing of the output distribution that it likens to a milder form of mode collapse in generative adversarial networks, and that RLHF "can also reward sycophancy and confident-sounding hallucinations".[11] A companion post, "The Bitterest Lesson" (10 September 2026), argues that "doing the right task > data > compute > algorithms", citing the InstructGPT result that instruction-tuned models more than 100 times smaller than GPT-3 were preferred over it.[20]

[Calibration](https://aiwiki.ai/wiki/calibration) is defined in the documentation in the usual statistical sense: across many predictions, outcomes assigned probability 0.2 should occur about 20 percent of the time and those assigned 0.8 about 80 percent. The docs stress that "these rates describe groups of predictions, not a guarantee about any single answer".[11]

On training data, the launch FAQ says "TypeSafe is primarily a data research lab", that "we make all the data ourselves", and that the company would not train on customer data.[1] TypeSafe's launch post, documentation and manifesto describe RLCD only at the level of its objective; none of them links a technical report or preprint. An arXiv search for the phrase "Reinforcement Learning for Calibrated Decisions" on 16 September 2026 returned no results, so the method's details, and the calibration measurements behind the "calibrated" claim, had not been published in a form that outsiders could check as of that date.

## Pricing and speed claims

The following table collects the headline numbers the company and its founder have given, with the source and, where the company states it, the comparison point. None had been reproduced by an independent party as of 16 September 2026.

| Claim | Figure | Who states it, and against what |
|---|---|---|
| Input price | $0.042 per million tokens ("$42 per billion tokens") | TypeSafe launch post and homepage[1][3] |
| Output price | Free ("too cheap to meter") | TypeSafe launch post[1] |
| End-to-end latency | 70 ms to 500 ms | TypeSafe launch post; The Register notes the comparison LLM figures are the company's[1][18] |
| Latency | "less than 100 milliseconds" | TypeSafe press release[2] |
| Speed vs LLMs | 40x to 200x faster "for the same levels of frontier intelligence for System One shaped queries" | TypeSafe launch post[1] |
| Speed and cost vs LLMs | "20-200x faster", "40-400x cheaper" | Almeida's launch tweet[19] |
| Speed and cost vs LLMs | "up to 100 times faster and less expensive than other frontier models" | TypeSafe press release[2] |
| Workflow evals headline | 193.6x faster, 444.6x cheaper | TypeSafe homepage, "based on workflows for System One tasks"; the launch post says these come from the workflow evals and "are on the higher end of real world gains"[3][1] |
| Input price vs Claude Fable 5.1 | 238x lower | TypeSafe homepage[3] |
| Side-by-side demo | Jev 0.114 s and $0.000081 vs "LLMs" 8.566 s and $0.013880 | TypeSafe homepage; the launch post identifies the LLM as [GPT-5.6](https://aiwiki.ai/wiki/gpt_5_6) Terra at default reasoning[3][1] |

The 238x figure is consistent with the published list prices: [Claude Fable 5.1](https://aiwiki.ai/wiki/claude_fable_5_1) is billed at $10 per million input tokens, and $10 divided by $0.042 is about 238. The Register's write-up put the GPT-5.6 Terra comparison at $2.00 per million input tokens and $12 per million output tokens, which matches OpenAI's post-July 2026 Terra pricing.[18] Comparing only input prices flatters Jev less than a blended comparison would, since Jev charges nothing for output while the LLMs it is compared against bill output tokens at a multiple of input.

On sustainability, the launch post says: "We can't prove it isn't subsidized; we'll need the long-term to prove the sustainability of our pricing (which we expect to go down, not up)."[1]

## Evaluations

TypeSafe has said it will not report results on public benchmarks. Its launch FAQ says the company "deliberately chose not to publish performance against public benchmarks", plans "only one-off evals when we make product updates", and asks users to build their own evaluations for their own use cases.[1] A post published four days before launch, "Lies, Damned Lies, and Benchmarks" (11 September 2026), sets out the reasoning: the company argues public evals get "benchmaxxed" through selection even without direct training on them, promises "no standard benchmark table in our model releases", and says new evals will be "dated snapshots and immediately retired once posted rather than hill-climbed".[21] The launch materials therefore rest on two company-run evaluations and several demos.

### Side-by-side demo

The homepage video shows Jev answering a short customer-analysis query with all probabilities emitted at once in 0.114 seconds at a cost of $0.000081, next to GPT-5.6 Terra generating its answer token by token in 8.566 seconds at $0.013880.[3] The launch post's own caveats are that the query was "highly simplified", that the state was a short, dense paragraph chosen to emphasize the sampling difference ("the relatively shorter input paints our model in an advantageous light"), and that the only disagreement between the two models in the recorded run was on "Churn likelihood level", which the company calls "genuinely ambiguous". TypeSafe says it chose Terra with default reasoning because it has found that model "the most comparable at intelligence to Jev on average".[1]

### Workflow evals

The company's main evidence is a set of four "workflow evals" published at evals.typesafe.ai. The premise is that a real automation task is decomposed into a fixed compute graph, a "workflow" in code, in which some steps are programmatic rules and others are narrow Noul, Choice or Score questions put to a model; every model runs the same workflow, and the harness is not tuned per model. Because there is no ground truth, the reference labels are "generated via an average of the responses of GPT-6 Astra and Claude Fable 5.1, both at high thinking, answering every question in the harness", and every other model is run at its provider's default reasoning setting. Each model is also run in a "prompt" configuration in which the whole policy is handed over as one prompt.[15] The four workflows are: security incidents (close, escalate to an analyst, or contain a machine, given an alert and its history), agent trace observability (whether a finished support-agent run needs human review and how soon), invoice processing (pay, hold or return a vendor bill given the order and delivery records) and customer service (what a support assistant should say and do next, given the thread, account and any pending proposal).[15][16]

The overview chart averages accuracy against the consensus labels, cost per case and seconds per case across the four workflows with equal weight. The workflow-mode results as published are:[15]

| Model (provider, per TypeSafe's grouping) | Mean accuracy | Cost per case | Time per case |
|---|---|---|---|
| Jev (TypeSafe) | 67.8% | $0.0004 | 0.4 s |
| GPT-5.6 Sol ([OpenAI](https://aiwiki.ai/wiki/openai)) | 74.1% | $0.0836 | 23.3 s |
| [Claude Opus 5](https://aiwiki.ai/wiki/claude_opus_5) ([Anthropic](https://aiwiki.ai/wiki/anthropic)) | 73.1% | $0.1761 | 37.8 s |
| GPT-5.6 Terra (OpenAI) | 67.9% | $0.0304 | 10.1 s |
| [Claude Sonnet 5](https://aiwiki.ai/wiki/claude_sonnet_5) (Anthropic) | 67.8% | $0.1174 | 78.1 s |
| GPT-5.6 Luna (OpenAI) | 66.8% | $0.0033 | 12.9 s |
| [DeepSeek V4-Pro](https://aiwiki.ai/wiki/deepseek_v4_pro) (via [Fireworks AI](https://aiwiki.ai/wiki/fireworks_ai)) | 65.5% | $0.0413 | 86.5 s |
| [DeepSeek V4-Flash](https://aiwiki.ai/wiki/deepseek_v4_flash) (via Fireworks AI) | 64.4% | $0.0059 | 51.9 s |
| [Claude Haiku 4.5](https://aiwiki.ai/wiki/claude_haiku_4_5) (Anthropic) | 53.6% | $0.0195 | 12.5 s |

Two readings of this table matter. First, Jev is not the most accurate model in TypeSafe's own evaluation: GPT-5.6 Sol and Claude Opus 5 score higher in workflow mode, and Jev's mean accuracy is essentially tied with GPT-5.6 Terra and Claude Sonnet 5. The company's claim is about the cost and latency frontier at a given accuracy level ("owning the Pareto frontier for almost 2 orders of magnitude"), not about topping the accuracy column.[1][15] Second, the per-workflow results vary: Jev scored 76.0% on customer service (behind GPT-5.6 Sol at 78.3%, DeepSeek V4-Flash at 76.8% and DeepSeek V4-Pro at 76.1%), 71.6% on agent trace observability, 61.7% on security incidents, and 61.8% on invoice processing, where Sol reached 79.1% and Opus 5 78.4%.[15][16] The evals site's own statement is that, "averaged across the four example tasks, every model is more accurate, cheaper and faster in the workflow than it is with the same policy as a prompt"; the published per-workflow numbers show a few exceptions (DeepSeek V4-Flash is less accurate in workflow mode than as a prompt on security incidents, and DeepSeek V4-Pro on agent-trace observability), but the averaged claim holds for every model.[15]

The homepage's 193.6x and 444.6x figures do not appear on the evals site itself; the launch post says they come from these workflows and that "we expect that these are on the higher end of real world gains". For reference, the overview table lists Claude Sonnet 5 in workflow mode at the same 67.8% mean accuracy as Jev at 78.1 seconds and $0.1174 per case, against Jev's 0.4 seconds and $0.0004.[1][15]

TypeSafe lists its own caveats: the workflows "were made by individuals on our model capabilities team, so some bias could exist"; using the average of [GPT-6 Astra](https://aiwiki.ai/wiki/gpt_6_astra) and Fable 5.1 as the reference "biases answers towards OpenAI and Anthropic's models" and likely underestimates Jev and the DeepSeek models; and the LLMs were driven through the company's System One adapter, which it says is the most accurate way to get decisions from LLMs but "tends to be slower and more expensive than giving decisions without probabilities".[1] Two further limits are worth stating: the two reference models come from the same two vendors as six of the eight compared LLMs, and the accuracy metric is agreement with those references rather than with human judgment.

## The "can't hallucinate" claim

The launch post says Jev "can't hallucinate" and the homepage advertises "Zero Hallucinations".[1][3] What the company means is narrow, and it says so. Because the answer space for every question is fixed in advance, the model returns a value from that space with a probability distribution over it; it cannot emit a malformed tool call, an option that was not offered, or free text. The launch post puts the "no type errors" claim in the falsifiable category ("this would be an easy thing to falsify with just a single counter-example, but it is mathematically impossible") and, in a chart comparing tool-call [hallucination](https://aiwiki.ai/wiki/hallucination) and type-error rates, states plainly that "our number is not empirical. Schema matching is guaranteed, thus we can confidently add 0% into the plots". The LLM rates in that chart come from [OpenRouter](https://aiwiki.ai/wiki/openrouter), and the company flags that source as biased because "more complex queries might be routed to better models".[1]

Type conformance is not correctness. A Choice answer can put its probability on the wrong option, and TypeSafe's documentation repeatedly says that calibration "does not guarantee that an individual answer is correct" and that a confidence of 1.0 describes the model's distribution rather than the truth.[5][8] The Register's Thomas Claburn made the same point in the first press write-up: "TypeSafe claims that Jev is hallucination-free, which really isn't a fair comparison as its output is not natural language. Jev instead returns structured responses with probabilities, and that does not preclude the possibility of being incorrect."[18] In the company's own evaluation, Jev's answers agreed with the reference labels 67.8% of the time on average, so roughly a third of its decisions differed from the consensus of the two reference models.[15]

The design choice that makes this coherent is the confidence field: the company's argument is not that Jev is always right but that it "always communicates confidence and uncertainty with every output", so software can act on high-confidence answers and escalate the rest.[1][10] Whether the probabilities are in fact well calibrated on customers' data is the claim that outside evaluation would need to test.

## Demos

TypeSafe released two demonstrations alongside the launch. In the first, Jev plays Doom from a structured text representation of the game state, not from images; the company says the engineer behind it "was worried about making 10 queries a second (which ends up costing ~$7/hour)", and concedes that "a non-AI doom bot could play better" and that the point was a bot "reactive to different representations of game state".[1] The Register led its coverage with this demo.[18]

The second is Wikiracing: starting on one Wikipedia page and reaching a target page using only links, where "each step can mean choosing between hundreds to thousands of links". TypeSafe presents it as a test of high-cardinality choice without hallucinating a link that does not exist. The company notes that the speedups here "tend to be a lot less than in previous demos" because the comparison LLMs were run in non-reasoning modes (Astra at its lowest reasoning setting) to keep the videos watchable, and that Jev "tended to finish in fewer steps". It also explains a mechanical detail: Jev supports a cardinality of up to 255 options per Choice, and for the higher-cardinality choices the system runs "a 2 stage-system of scoring independently then making an explicit choice, hence the occasional slowdown".[1]

## Intended uses

The company's stated use cases are "AI-powered workflows / smart if-statements" (classify, route, score, extract or branch inside ordinary software), map-reduce over large corpora, real-time applications where roughly 100 ms latency is acceptable, and verification: scoring, judging, guardrailing and jailbreak detection over LLM prompts, reasoning traces and outputs.[1] The documentation's example use-case map and cookbooks cover extraction, classification, self-consistency checks and batching, and a smart-home assistant demo.[4] Every's head of evals, Mike Taylor, who tested the model at launch, framed it as "a smart if-then statement that determines what happens next when you're automating a workflow", noting that he had previously used DSPy to coax LLMs into typed outputs, and concluded that "just how well it gets the job done is still an open question".[25]

The manifesto that accompanies the launch calls the goal "composable AI" and "machine-native" intelligence: a primitive "any programmer can invoke for semantic judgement and decisions, while still using code for what it's best at: exact computation", summarized by the company as "smart if-statements" and, in its slogan, "Build Prod, Not God".[22]

## Reception

Coverage on the first day was largely descriptive. The Register (Thomas Claburn, 16 September 2026) explained the primitives and pricing, questioned the hallucination framing as noted above, and observed that the Jevons-paradox naming "reflects the AI industry's bet that greater token efficiency will increase token consumption even as token prices decline", a wager it called unsettled.[18] GIGAZINE (16 September) summarized the launch and read the company's cost chart as Jev matching GPT-5.6 Terra's task quality at about one hundredth of the cost.[24] The TestingCatalog account on X described Jev as "a small and customizable System One decision model" and reported "approximately 150ms per call"; the "small" characterization conflicts with the company's FAQ statement that Jev "is neither small nor an LLM", and the 150 ms figure does not appear in the launch post, which gives a 70 ms to 500 ms range, but it matches the "real-time speeds (150ms)" line in the company's use-case documentation.[26][1][4]

## Name

Jev is named for William Stanley Jevons, the 19th-century English economist behind the Jevons paradox, the observation that more efficient steam engines increased rather than reduced coal consumption. TypeSafe says it "expect[s] machine intelligence to follow a similar path to coal" and that "every order of magnitude drop in the cost of intelligence unlocks orders of magnitude more use cases".[1][2] The press release calls the name "a nod to Jevons Paradox".[2]

## References

1. Diogo Almeida, "Introducing System One Models & Jev", TypeSafe AI Blog, 15 September 2026. https://typesafe.ai/blog/introducing-system-one-models-and-jev
2. "TypeSafe AI Emerges From Stealth With $40M in Funding With New Model for Composable AI", Business Wire, 15 September 2026. https://www.businesswire.com/news/home/20260915525333/en/
3. TypeSafe AI homepage, accessed 16 September 2026. https://typesafe.ai/
4. "Introduction", TypeSafe AI documentation, accessed 16 September 2026. https://docs.typesafe.ai/introduction
5. "System One", TypeSafe AI documentation, accessed 16 September 2026. https://docs.typesafe.ai/concepts/system-one
6. "Primitives (Questions)", TypeSafe AI documentation, accessed 16 September 2026. https://docs.typesafe.ai/primitives
7. "Choice", TypeSafe AI documentation, accessed 16 September 2026. https://docs.typesafe.ai/primitives/choice
8. "Score", TypeSafe AI documentation, accessed 16 September 2026. https://docs.typesafe.ai/primitives/score
9. "Noul", TypeSafe AI documentation, accessed 16 September 2026. https://docs.typesafe.ai/primitives/noul
10. "Confidence", TypeSafe AI documentation, accessed 16 September 2026. https://docs.typesafe.ai/confidence
11. "AI primer", TypeSafe AI documentation, accessed 16 September 2026. https://docs.typesafe.ai/introduction/machine-learning-primer
12. "Quick start", TypeSafe AI documentation, accessed 16 September 2026. https://docs.typesafe.ai/introduction/quickstart
13. "Changelog", TypeSafe Python SDK documentation, accessed 16 September 2026. https://docs.typesafe.ai/sdk/python/changelog
14. "Changelog", TypeSafe JavaScript SDK documentation, accessed 16 September 2026. https://docs.typesafe.ai/sdk/javascript/changelog
15. "Workflow evals: Structured workflows for automation tasks", TypeSafe AI, accessed 16 September 2026. https://evals.typesafe.ai/
16. "Customer Service", TypeSafe AI workflow evals, accessed 16 September 2026. https://evals.typesafe.ai/customer_service.html
17. typesafe-ai/system-one-adapter-python, GitHub repository (MIT license), accessed 16 September 2026. https://github.com/typesafe-ai/system-one-adapter-python
18. Thomas Claburn, "TypeSafe AI debuts model for machines that plays Doom", The Register, 16 September 2026. https://www.theregister.com/ai-and-ml/2026/09/16/typesafe-ai-debuts-model-for-machines-that-plays-doom/5296711
19. Diogo Almeida (@CompleteSkeptic), post on X, 15 September 2026. https://x.com/CompleteSkeptic/status/2099925682726002904
20. "The Bitterest Lesson", TypeSafe AI Blog, 10 September 2026. https://typesafe.ai/blog/bitterest-lesson
21. "Lies, Damned Lies, and Benchmarks", TypeSafe AI Blog, 11 September 2026. https://typesafe.ai/blog/antibenchmaxxing
22. "Composable AI: Build Prod, Not God", TypeSafe AI manifesto, accessed 16 September 2026. https://typesafe.ai/manifesto
23. Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida et al., "Training language models to follow instructions with human feedback", arXiv:2203.02155, 4 March 2022. https://arxiv.org/abs/2203.02155
24. "ChatGPT co-developers have created 'Jev,' an AI model that processes tasks using a different method than LLM", GIGAZINE, 16 September 2026. https://gigazine.net/gsc_news/en/20260916-system-one-jev/
25. Mike Taylor, "Mini-Vibe Check: TypeSafe's Jev Judged Everything I've Written in 0.7 Seconds", Every, 15 September 2026. https://every.to/also-true-for-humans/mini-vibe-check-typesafe-s-jev-judged-everything-i-ve-written-in-0-7-seconds
26. TestingCatalog (@testingcatalog), post on X, 15 September 2026. https://x.com/testingcatalog/status/2099968075861008781
27. "Team", TypeSafe AI, accessed 16 September 2026. https://typesafe.ai/team
28. "State", TypeSafe AI documentation, accessed 16 September 2026. https://docs.typesafe.ai/concepts/state