Typed Decisions, Not Chat

Paper type: For noting.
As of: 16 September 2026.
This paper does not recommend a purchase, a build, or a pause. It states what is on the public record one day after TypeSafe left stealth.

Jev is TypeSafe AI’s first public System One model: a hosted function that reads unstructured state (a ticket, a JSON object, a slice of application memory) and returns typed decisions — a choice from a list you supplied, a score on a rubric, or a yes/no probability — rather than generating a paragraph.

Early access opened 15 September 2026. The HTTP call is POST https://api.typesafe.ai/v1/systemone with model alias jev-latest. List price on the launch post is $0.042 per million input tokens and $0 output, which TypeSafe calls “too cheap to meter.”

The founder, Diogo Almeida, is a starred primary author of the 2022 InstructGPT paper (the OpenAI work that trained GPT-3 to follow instructions with human feedback). He is not the sole inventor of ChatGPT. Cofounders named on the team page are Erik Gafni (CTO) and Sasha Sheng (COO). DCVC led a $40 million seed announced the same day as the model. A Forbes article the same day put post-money at $200 million according to a person familiar with the deal. That valuation is one anonymous source.

What the launch claimed, and what the same blog already hedged

TypeSafe’s launch post says “Extraordinary claims require extraordinary evidence.” The X thread led with 20–200× faster and 40–400× cheaper. The blog under the thread is narrower.

Speed. Vendor end-to-end band is 70–500 milliseconds against 3–329 seconds for “frontier” calls, with evals generally run from West Coast laptops because that is where the service sits. The homepage pair 193.6× faster / 444.6× cheaper is taken from TypeSafe’s own workflow evals; the blog says those figures are “on the higher end of real world gains,” and that the workflows were built by the company’s model-capabilities team, so “some bias could exist.”

Price. The $0.042 / MTok input number is on the price sheet. TypeSafe writes, of that sheet: “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).”

Hallucination. Homepage copy says “Zero Hallucinations.” The launch post’s nuance box says the 0% bar “is not empirical. Schema matching is guaranteed, thus we can confidently add 0% into the plots.” That is a closed output set: Jev cannot emit a tool name or a class that was not in the question. It can still pick the wrong class. TypeSafe’s own docs say calibration does not guarantee that an individual answer is correct.

Intelligence. The pitch is “similar levels of intelligence on System One tasks” compared with existing LLMs. The evaluation TypeSafe published does not use human ground truth. It uses the average of GPT-6 Astra and Claude Fable 5.1 as the reference, on four internal workflows (security incidents, agent traces, invoices, customer service). Secondary reading of that dashboard (OrcaRouter, 16 September 2026) puts Jev around 67.8% mean agreement against 74.1% for the best comparator: cheaper and faster, not more accurate. TypeSafe also notes that forcing LLMs through its structured wrapper is slower and more expensive than asking them for a decision without probabilities, which paints Jev in a favorable cost/latency light.

Architecture. The company says it rebuilt the model, the sampler, and the trainer (RLCD, Reinforcement Learning for Calibrated Decisions). There is no paper, no parameter count, and no public weights. Sean Goedecke’s 16 September 2026 essay argues that prefilling a normal language model and emitting one constrained token already buys most of the speed, and that “can’t hallucinate” is a semantic dodge. That replica is the live technical objection. It does not make Jev fake. It makes “new architecture family” unproven.

Demos. Doom at about ten calls per second and about $7/hour is, in TypeSafe’s own nuance box, a bot over structured text state, “not on images (yet…),” and “a non-AI doom bot could play better.” Wikiracing uses a two-stage score-then-choose pattern once cardinality exceeds 255 options.

Independent measurement in the first day

Mike Taylor at Every (15 September 2026) ran Jev over 27 of his articles plus 10 AI-styled counterparts, 21 questions each: 777 judgments in under 0.7 seconds, about a quarter of a cent. A second small test against planted writing defects was about 25× faster and about 1/580th the cost of Claude Fable 5.1 at high effort, and missed one of seven defects that Fable caught. That is the entire independent accuracy record as of this paper. It is enough to treat latency and list price as real. It is not enough to treat accuracy as frontier-equal.

X and Community Notes

The launch went out from Almeida’s personal account @CompleteSkeptic. A tracker (Doomers) recorded very high view count on that thread. No X Community Note was found on the main launch posts as of 16 September 2026. Notes often lag a 24-hour-old viral post; absence is not clearance.

What the first-day crowd actually argued, in substance rather than in likes: Jev is not a general chat model; “no hallucination” is being redefined as “no illegal string”; the price may be seed-subsidized; a prefill replica may clone the speed without TypeSafe’s unpublished net. A Hacker News comment from an account identifying as a TypeSafe employee said Jev is not a reasoning model and that reasoning should live in code. That employee frame matches the public docs.

All-In and Moonshots episodes available through 11 September 2026 predate the 15 September launch. There is no podcast confirmation of this SKU.

How the object sits in software

Jev is a model. From a harness’s point of view it is also a tool (ask a typed question, get a number, branch). It is not a harness: it does not run a loop, a sandbox, or other tools. TypeSafe’s own “how to build” page wants code to own control flow and side effects, with System One inserted only at narrow semantic judgments. Community glue such as pi-jev uses it as a gate on bash/write/edit. Confidence bands (act / confirm / escalate) live in the caller’s code. Noul answers do not even carry a separate confidence field; the probability is the signal.

State you send is the document that leaves your boundary. Public docs reviewed for this paper do not include a residency table, a zero-data-retention row, or a DPA. The service is described as US West Coast. jev-latest is the documented alias; a silent alias move would move every threshold you tuned.

Automated decisions and an explainable audit trail

Jev can sit in an automated decision path. It does not, by itself, produce a natural-language explanation. That split is in the product contract, not an oversight.

Automation. The documented job is software that acts: send state plus typed questions; get a Choice, a Score, or a Noul plus probabilities; branch, score, or route in code. TypeSafe’s pattern is three bands — act, confirm, escalate — with the cutoffs in the caller’s program. Noul has no separate confidence field; the 0–1 probability is the signal. That is automated decision-making with a human-in-the-loop gate, not a chatbot that narrates a plan.

The remaining error class is wrong-but-typed. Calibration is defined across a group of predictions; it does not guarantee a single answer. Unattended high-stakes use is only as honest as those probabilities on the caller’s own labeled set.

What “explainable” Jev actually returns. TypeSafe’s docs state that System One models do not write replies, produce code, or generate explanations of their reasoning. A Hacker News comment from an account identifying as a TypeSafe employee said the same thing: put reasoning in code. Jev will not emit “because clause 4.2 and the receipt don’t match.”

What a caller can persist as a trail, because it is already structured:

Field logged	What a later reviewer can see
state	The exact document judged
Question instructions plus criteria / options	The policy that was asked, in the caller’s words
choice / score / noul	The decision
Full probabilities	Alternatives the model did not pick
confidence (Choice and Score only)	How peaked that distribution was
model (jev-latest unless pinned)	Which SKU answered
The program’s action and threshold	Why the system paid, held, or paged a human

That record is inspectable: what was asked, over what options, what numbers came back, what the program did. It is not a rationale. Confidence 1.0 means the probability mass sat on one option. TypeSafe says that describes the answer, not that the answer is correct.

How a fuller trail is assembled. Split one fat judgment into atomic questions and combine them in code (composite scoring, confidence-gated routing). The composition is then the audit story: for example, Noul “duplicate charge?” = 0.97; Choice “team” = returns at 0.60 and billing at 0.38, confidence 0.39 → hold and review. TypeSafe’s citation-check cookbook uses Jev the other way: to judge an LLM’s citations (verified / unsupported / contradicted / fabricated), not to author the explanation.

What is not in the product. No chain-of-thought, no “why” string, no feature-attribution trail, and no published logging, retention, or DPA so the vendor side of an audit is still the caller’s to design. jev-latest can move under a tuned threshold unless the alias is pinned.

Jev can therefore give a replayable, typed ledger on an automated path. If an auditor needs “why” in sentences, that text has to come from the questions, the rules, and a human — not from Jev.

Facts / Trends / Bets / Warnings

Facts

Fact	Print	What it is not
Jev early access 15 Sep 2026	TypeSafe launch blog; DCVC same day	General availability
Three primitives only	Docs: Choice, Score, Noul	A JSON-document generator
$0.042 / MTok input, $0 output on the launch sheet	Launch blog	Proof the meter is unsubsidized
$40M seed led by DCVC	DCVC; FinSMEs	The $200M Forbes figure
Almeida starred on InstructGPT	arXiv:2203.02155	Sole inventor of ChatGPT
No Community Note found 16 Sep 2026	X search this pass	A Note cannot appear later
System One does not generate explanations of reasoning	Docs: System One	A natural-language rationale / chain-of-thought
Choice/Score return full probabilities plus confidence; Noul returns only noul	API reference; Confidence page	A vendor-hosted audit log

Trends

Trend	Metric	Classification
Closed-set inference getting cheap	Jev list price vs chat/reasoning output tokens	Stepwise product on a longer cheap-classifier curve, not a published doubling-time exponential
Judges inside agent loops	pi-jev, Every’s “check the agent as it goes”	Measured in demos, not yet a market series
Public benches losing trust among some labs	Almeida anti-bench essay, 11 Sep 2026	Take, not a measured decline in bench use

Bets

Bet	Holder	Falsifier
Cheaper intelligence increases total use (Jevons); TFP +3% in five years	TypeSafe manifesto	TFP does not move; usage stays in copilots
RLCD makes probabilities honest enough to automate	TypeSafe / Almeida	Buyer reliability diagrams sit off the diagonal
A lab ships Terra-System-One and eats the interface	Goedecke (wish, not a timed bet)	No lab SKU, or TypeSafe’s unpublished net still wins the eval

Warnings

Item	Kind
Homepage 193.6× / 444.6× without the blog’s “high end / our team built the workflows” box	Marketing vs footnote
“Zero hallucinations” heard as “never wrong”	Error-class swap
jev-latest plus tuned thresholds	Model-continuity
US-hosted closed API on a live customer decision	Data-boundary / registry row for the use, not for the vendor slogan
Seed-era prices	Vendor says it cannot prove they are not subsidized

How to see this in your own books

This is observational, not a vendor pick. If invoices already show spend on chat completions used only to emit a label (route this ticket, is this fraud, did the agent’s patch look safe), those lines are the comparison class Jev is priced against — not the lines that pay for code generation or long reasoning. If a named group already runs a classifier or grammar-constrained decode for the same labels, that is the replica Goedecke described, and it is the control any later Jev trial would have to beat on your labeled set, not on TypeSafe’s Astra/Fable agreement plot.

Watch list

Date / gap	What would change the record
Architecture paper / params / weights	Moves “new architecture” out of slogan
Public reliability diagram on a non-vendor set	Moves RLCD out of product-name
Twelve-month price sheet	Tests the subsidy hedge
A Community Note on the launch thread	Tests whether X’s crowd-correction layer ever attached
All-In / Moonshots after 15 Sep 2026	First podcast-signal pass that can actually mention the SKU
Second independent accuracy study larger than Every’s 7-defect sample	The accuracy claim is still almost entirely vendor-side
Residency, DPA, jev-latest changelog	Enterprise control-plane gaps
Vendor-side audit log or “why” field	Not on the public API; trail is what the caller persists

Sources

Research appendix (this site)
- Research appendix home
- Research index
- What Jev is
- Architecture and RLCD
- Claims audit
- Company and founder
- Evals and accuracy
- API and audit trail
- Contrarian
- Taxonomy
- Vendor risk
- Question ledger
- External primary sources

Primary sources
- Introducing System One Models & Jev
- TypeSafe docs
- DCVC seed note
- InstructGPT