The launch post: state in, typed probabilistic decisions out, RLCD training, 70 to 500 ms, $0.042 per MTok
Jev61The model, its open-source rebuilds, the independent evaluations, and the commentary.
System One & Jev 19
What TypeSafe has published: the launch post, the docs, the SDKs, the essays, and the failure modes it admits to.
The three typed question shapes and the probability-per-option answers they return
TypeSafe's documented failure modes: literal reading, counting, dates, indirection, distractor state, adversarial content
Official drop-in that serves the same typed interface from OpenAI or Anthropic models, the baseline for every comparison
72
1,850 points and 485 comments; the CEO confirms the zero-shot classifier reading and the encoder-with-heads shape
Defines the model class: fast structured decisions for software, contrasted with text-generating LLMs
Confidence as a statistic of the returned distribution, and the act, confirm or escalate thresholds built on it
TypeSafe's own account of RLCD as a third post-training regime beside RLHF and RLVR
Speculative fan-out, confidence-gated routing, composite scoring and intent routing as architectural patterns
The request and response contract for POST /v1/systemone
Reproducible recipes for parallel questions, reranking, guardrails, extraction and hierarchical classification
Official synchronous and asynchronous Python client
Official TypeScript client with answer types inferred from the questions
68
Skill files that teach Claude Code, Codex and similar agents to design System One workflows
The first third-party surface: Jev-latest as an evaluation model behind experimental_evaluate, no waitlist
Machine-native composable AI as reliable primitives for programmers, against the assistant framing
Why TypeSafe publishes dated, retired eval snapshots instead of a benchmark table
Diogo Almeida's thread arguing RLCD decision models reach economic value before chat models do
TypeSafe's own essay ranking task choice above data, compute and algorithms, the argument behind System One Models
Open Source 18
Open weights and code that rebuild the System One shape from encoders, small decoders and constrained decoding.
Can something like Jev run on a 3090 at home; open baseline for direct typed option scoring
From-scratch model with Jev's exact shape: text plus N options in, one probability per option out, option-attention head, Doom and chess demos
608
Qwen2.5-1.5B fine-tune plus parallel constrained decoding; all schema fields scored in one broadcast prefill, 5.6x to 7x faster on Apple Silicon
Decorator that routes DSPy typed Signatures to Jev where the signature is a pure decision
56
LLM gateway that mimics the TypeSafe structured-output contract, used for Qwen-on-Cerebras side-by-sides
25
Jev-compatible API endpoint served from open models with prefill-only inference
16
Unofficial study of Jev-style parallel typed decisions on stock 1.5B to 8B models on Apple Silicon, with benchmarks
One-pass scorer over a variable option list: hashed n-gram encoder, rival-aware attention, gated head, temperature scaling
5
Qwen3.5-2B fine-tune that emits typed decisions with calibrated probabilities in one pass
4
Local bilingual probability decisions from context, questions and candidate answers
3
Jev's decision layer on off-the-shelf Qwen checkpoints: one shared prefill, then option logits read per question, no training
3
Encoder-side reproduction: GLiClass ModernBERT base retrained for calibrated label probabilities
Choice, Score and Noul on Qwen3.5-0.8B, the smallest decoder-based reproduction
1
Live demo of the Qwen-2.5-1B-RLCD approach: KV-cache broadcast, logit slicing per candidate, 100 percent schema validity
RLCD-style fine-tune of Liquid AI's LFM2.5-2.6B, the largest open attempt so far
350M-parameter RLCD-style decision model, the smallest open attempt
Qwen3.5-4B base trained as a Score-style rubric rater
DistilBERT-sized System One shape, a floor for how small the idea can go
Independent Evaluations 17
Every independent test of Jev published so far, with the headline number where the source gives one. TypeSafe's own dashboard is listed and marked official.
Jev versus GLiNER2.5 on 300 BTZSC examples with calibration and selective risk: 0.910 AG News, 0.870 Banking77, worse on emotion
6
Jev as reranker over 8 datasets and 1,617 questions: nDCG@10 0.692 versus Cohere Rerank 4 Pro 0.691, at 422ms
Blind prompt-injection run on 662 deepset messages: 96.5% accuracy, 0.9927 ROC-AUC, ECE 0.0588, p50 325ms
1
TypeSafe's own four-workflow dashboard, Jev at 61.7 to 76.0% accuracy and 0.3 to 0.5s per case against frontier baselines
Jev routes 100 mocked tools behind a confidence gate, measuring steps, tool calls, tokens and cost against the LLM choosing directly
45
Claim-level Brier scoring of federal motion-to-dismiss outcomes, a fixed binary with one probability per unit; no Jev row published yet
4
Jev no better than random picking chess moves from a FEN, but F1 0.96 on NPC addressee detection, 0.2s median
2
Three production SaaS decisions published as schemas with auto-post confidence thresholds; LLM baseline rows filled, Jev row still empty
1
Zero-shot spam Noul on 18,514 emails reaching 0.9833 accuracy, matching a TF-IDF classifier trained on 14,800 labels
100 balanced secret-detection cases as single Noul questions scored by accuracy, AUC and Brier; server p50 75 to 90ms
Reproducible harness scoring Jev ultrafast research-browser runs over 11 baseline cases plus 18 human and quant stress cases with QC grades
Tic-tac-toe and connect four pitting Jev against four frontier models on identical legal-move choice options; no aggregate results published yet
Harness asking Jev and frontier LLMs identical typed questions, scoring accuracy, calibration, latency and schema validity; no measured run published yet
50-case event validation: Jev 96% at 0.59s and $0.043 per 1,000, Mistral small 4 84%, Gemini Flash-Lite 86%
777 judgments over 37 articles in 0.7s for a quarter of a cent; caught six of seven planted defects, Fable seven
Poker probe finding 15 to 30 point swings from relabelling the same hand, and 16 of 16 bets against a made flush
One direct Jev question per row against 12 to 14 Jev-scored dimensions with fitted weights on three tasks: 0.9076 vs 0.8373 on Japanese NLI, but 25x the hard-benign false positives, 37.2% vs 1.5%
Commentary & Analysis 7
Reporting and technical commentary that checks the launch claims against the evidence.
Latent Space roundup of the launch and the HN mapping onto encoders, GLiNER, constrained decoding and DSPy
Claim-by-claim audit separating verified Jev pricing and latency from unproven calibration; puts aggregate accuracy at 67.8% versus Opus 5's 73.1%
Press account of the $40M raise, the Doom demo and the caveat that structured output is a different error type, not correctness
Technical comparison of Jev's single-pass sampler with token-by-token decoding and masked diffusion
Essay on Jev as the model that refuses to write sentences and what that implies for product design
Explainer reconstructing the RLCD objective and the parallel sampler from public statements
Secondary analysis of TypeSafe's dashboard putting Jev at about 67.8% mean agreement against 74.1% for the best comparator
Lineage29Twenty-nine earlier works, each labelled with why it is here: named by TypeSafe or the launch discussion, or the same shape as Jev.
The Shape Before Jev 15
Earlier work with the same input and output shape: a fixed answer set, one probability per option, no generated text. Label-conditioned encoders, scalar reward heads, reinforcement learning for calibrated confidence, and the single-pass inference TypeSafe's own forks point at.
Benchmarking Zero-shot Text Classification: Datasets, Evaluation and Entailment Approach
same shape Label set given at inference, an entailment model returns one probability per label, no text generated
Training language models to follow instructions with human feedback
same shape A Bradley-Terry head emits one scalar per response in a single pass, no text, co-authored by Jev's founder
named, same shape 184M DeBERTa returning a binary injection probability, the BERT-style encoder guardrail HN engineers mapped Jev onto
Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty
same shape Adds a Brier-score reward to RLVR so the model emits calibrated confidence, the closest published relative of TypeSafe's RLCD
Large Language Diffusion Models
primary The typesafe-ai GitHub org forked this masked diffusion LM, the strongest public hint at how Jev fills every answer slot in one pass
Generalist Model for Named Entity Recognition using Bidirectional Transformer
named The span-and-label encoder family HN mapped Jev onto, types supplied at inference and scored in one bidirectional pass
Generalist Lightweight Model for Sequence Classification Tasks
named The open analogue HN pointed at: labels and text in one encoder pass, one probability per label, no decoding
Passage Re-ranking with BERT
same shape Landmark cross-encoder: pair in, one scalar relevance probability out, no generation, the ancestor of Jev's Score primitive
Revisiting Text Classification in the Era of Transformers
same shape Controlled comparison of encoder, autoregressive and diffusion classifiers over fixed label sets on accuracy, calibration and ordinality
LLM-based Input-Output Safeguard for Human-AI Conversations
same shape Fixed safety taxonomy with the verdict read off one safe/unsafe token probability, the guardrail classifier Noul replaces
same shape Reports that RLHF destroys the base model's calibration, the finding RLCD is positioned against
A Reinforcement Learning Approach to Calibrated Confidence Expression of Large Language Models
same shape Trains confidence expression by RL on the logarithmic scoring rule, an independent rediscovery of the proper-scoring-rule reward RLCD uses
Balancing Classification and Calibration Performance in Decision-Making LLMs via Calibration Aware Reinforcement Learning
same shape RL that adjusts decision-token probabilities directly, keeping RLVR accuracy while cutting ECE, the closest public analogue of Jev's typed decision heads
Efficient Memory Management for Large Language Model Serving with PagedAttention
primary The typesafe-ai GitHub org forked this engine; paged KV cache plus prefix caching is what makes extra questions over one shared state nearly free
Ultra-Fast Language Models Based on Diffusion
named HN read Jev as a stripped down text diffusion model, and Mercury is that idea shipped commercially with parallel refinement
What Jev Is Sold Against 9
The tools TypeSafe and the launch discussion named as what Jev replaces: constrained decoding, structured outputs, typed prompt programming, LLM judges, routers and guard classifiers.
Efficient Guided Generation for Large Language Models
named The finite-state-machine guided decoding HN named as the incumbent way to get typed values, which Jev claims to replace
Compiling Declarative Language Model Calls into Self-Improving Pipelines
named Typed signatures compiled into prompts, named on the HN thread as the fair comparison for Jev's typed question interface
Learning to Route LLMs with Preference Data
same shape Router scores a fixed two model set and returns win probability per option before any text is generated
named Constrained generation library named on the HN launch thread as what Jev's typed outputs get compared against
21.8k
named The provider-side JSON-schema guarantee the CEO named on HN as what Jev replaces, shape enforced but no probability returned
A Study on the Impact of Format Restrictions on Performance of Large Language Models
named Measures the accuracy format restrictions cost, the study behind the CEO's HN claim that constrained decoding makes models dumber
A Rigorous Benchmark of Structured Outputs for Language Models
primary 10k real schemas scored on validity, coverage and latency, the constrained-decoding route Jev's 0% type errors claim competes against
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
named The landmark LLM-as-a-judge paper, named on the launch thread as the layer Jev's score and Noul primitives replace
Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming
named, same shape Input and output classifiers gating a frontier model on a fixed policy, the deployed slot Noul targets
Where the Name Comes From 5
System 1 in Kahneman's sense, the bitter lesson TypeSafe argues with, and the Jevons paradox the model is named after.
named The source TypeSafe cites for naming Jev after System 1, fast intuitive judgement with no deliberation
named Kahneman's two-system account, intuition returning an answer directly while reasoning deliberates, the split Jev's design copies
named The AI charter for System 1 components that answer from experience without search, what System One Models productizes
named The essay TypeSafe's own Bitterest Lesson argues against, named in TypeSafe's materials as its starting point
named The rebound effect Jev is named for, where cheaper decisions raise total decision volume
Nothing matches that search.