TypeSafe Jev is a real closed commercial model, not a stand-in invented by open competitors. This survey describes it only from TypeSafe’s public materials and two third-party benchmark projects. Its architecture and training implementation are not public, so those parts remain unknown.
First principles
A typed decision is a closed question. The caller supplies state and a schema, and the model returns one of a fixed set of outputs: a choice from listed options, a score, or a yes-or-no probability. A decoder model writes its answer one token at a time. A typed-decision model has nothing to write, so it can score every option in one parallel pass. That difference is the basis of the speed and cost case for this class of model.
The output is a probability, so accuracy is only half the story. Callers act on the number, not just the top label. Two models with the same accuracy are not equally useful if one of them puts zero probability on the true answer for some inputs. Calibration (whether stated probabilities match how often the model is right) and behavior when no option fits need their own measurements.
A closed model can only be known by its behavior. With no weights, architecture or training paper, any statement about Jev’s internals is a guess. What can be checked is what goes in, what comes out, how fast and at what price, under a stated protocol. Those measurements also inherit each benchmark’s choice of prompts, data and network path.
A shared acronym is not a shared method. RLCD names three unrelated techniques below. Matching names say nothing about matching mechanisms, so each claim has to be traced to its own source.
Public record
TypeSafe AI launched Jev in early access on 2026-09-15. It describes Jev as the first of its System One Models: a caller supplies unstructured state and a schema, and the application programming interface (API) returns typed probabilistic decisions. The public product framing covers choice, score, and boolean-style decisions in a parallel, non-autoregressive pass. The closed API is api.typesafe.ai/v1/systemone; no open weights, model repository, or training paper were located.
TypeSafe’s own pricing is $0.042 / 1M input tokens, with free output tokens. Its marketing cites 70-500ms end-to-end latency, 40x-200x faster and 40x-400x cheaper claims, and peak figures of 193.6x / 444.6x. TypeSafe notes that its internal benchmarks are likely to be at the high end of real-world results. These figures are vendor claims, not independent measurements.
Three different RLCDs
The acronym RLCD names three unrelated methods. They do not share a lineage.
| User | RLCD means | What is known |
|---|---|---|
| TypeSafe, Jev | Reinforcement Learning for Calibrated Decisions | Public marketing says it optimizes probabilities against outcomes rather than human-rater preference. There is no published technical mechanism. |
| Laya | Reinforcement learning against strictly proper scoring rules | Its public code combines log score, 0.5 spherical score, and 1.0 ranked probability score in a REINFORCE or Group Relative Policy Optimization (GRPO) style setup. |
| Academic literature | Reinforcement Learning from Contrastive Distillation | Yang et al., arXiv:2307.12950 creates contrastive preference pairs, trains a reward model, then uses preference-alignment reinforcement learning (RL) for chat models. |
Independent benchmark evidence
Two public third-party projects give the most concrete evidence about Jev’s behavior. They are not vendor benchmarks. They are also new projects with their own choices of prompts, data and environment, so their figures are not a controlled measurement of Jev’s internals.
AbdelStark/jev-benchmarks compares Jev with GLiNER. It reports AG News 0.910 versus 0.700, and Banking77/BTZSC 0.870 versus 0.610, in Jev’s favor. On DAIR Emotion it finds a statistical tie on accuracy but worse Jev calibration: Brier 0.846 versus 0.668, negative log-likelihood (NLL) 5.588 versus 1.381, and zero probability on the true label for 16% of examples. Its latency run reports local GLiNER at around 44ms median (p50) on an Apple M4 Max central processing unit (CPU), and Jev at 236-256ms from France; on the 72-label task Jev is 246ms and GLiNER 296ms.
nibzard/decision-model-benchmark tests Jev against 8 constrained large language models (LLMs), 5 suites and 3 deterministic baselines, with $28.34 in measured provider spend and raw logs. It reports Jev at 76.3% on Banking77 77-way classification, behind gpt-oss-120b at 81.3% and glm-5.3 at 80.4%; Spam is 93.0%, while GLM models lead at 94.9%/91.4%. Jev p50 is 264-276ms across 2-255 options, described as 1.2x faster than gpt-oss-120b on Cerebras and 10-16x faster than thinking-mode models. The same benchmark reports an expected calibration error (ECE) of 0.246, where LLMs range from 0.039-0.122, forced-uncertainty bluffing 50.3% versus LLMs admitting ignorance 97.3-100%, $0.07/1,000 decisions, and a hard 255-choice cap with 400 Too many choices at 256+. It also finds a 13% option-permutation flip rate.
Comparison
| System | Availability | Architecture or objective | Latency | Important caveat |
|---|---|---|---|---|
TypeSafe Jev 1.13.0 |
Closed proprietary API | Architecture undisclosed; TypeSafe’s RLCD is undisclosed | 70-500ms vendor claim; 236-276ms p50 independently |
ECE 0.246 in one benchmark; 255 choice maximum |
| Laya | Apache-2.0 package and checkpoints | ModernBERT-large or mmBERT-base plus a small Transformer head; proper-scoring-rule RL | 32.8-39.5ms p50 on T4, self-reported; 79.6ms mean on Apple Metal Performance Shaders (MPS) for 3 questions in this report |
Near-chance zero-shot; default high-cardinality option limit |
| open-jev-deberta-v3-large | Apache-2.0 open reproduction | DeBERTa-v3-large, 435M; objective UNVERIFIED |
UNVERIFIED | In-domain 0.854, out-of-domain 0.690, self-reported and not affiliated with TypeSafe |
| GLiNER / GLiClass | Open | BERT-like encoder, supervised zero-shot label scoring | GLiNER2.5 about 44ms p50 local CPU |
Narrower named entity recognition (NER) and classification scope, no RL-calibration claim |
| SetFit | Open | Sentence Transformer plus classifier, contrastive few-shot fine-tuning | Very fast, no generation | A few-shot classification toolkit, not a typed-decision API |
| Decoder LLM baselines | Mixed | Autoregressive decoder models | 300ms-5.6s in the benchmark |
Often more accurate and better calibrated, at greater cost or latency |
What follows from the evidence
Jev is a useful commercial reference point. In the available independent measurements it is fast and cheap, but its calibration and forced-uncertainty behavior are not settled in its favor. Laya shows that the same interface can be built in the open, with its architectural tradeoffs visible. GLiNER, GLiClass and SetFit show that direct encoder classification already has a mature neighboring ecosystem.
The most important unknown remains Jev itself. Its weights, architecture, training data, and evaluation protocol are closed. Any claim about those details is UNVERIFIED. The useful comparison is therefore behavioral and conditional. Use public third-party measurements for Jev, use public documentation for open systems, and keep self-reported, independent and unverified claims apart.
Sources
TypeSafe public materials, including its Jev launch post; AbdelStark/jev-benchmarks; nibzard/decision-model-benchmark; RLCD, arXiv:2307.12950; ModernBERT, arXiv:2412.13663; mmBERT, arXiv:2509.06888; GLiNER, arXiv:2311.08526; and GLiClass, arXiv:2508.07662.