We took the opposite research direction.
Generative models predict text sequentially token by token, accumulating latency, cost, and hallucination. System One uses a bidirectional encoder with direct decision heads to produce calibrated probabilities and structured choices in a single forward pass.
Single-pass forward evaluation on batch workloads without token decoding delay.
Bidirectional mmBERT-base encoder coupled with specialized decision heads.
Native multilingual representation across Latin and non-Latin character sets.
Direct classification heads eliminate generative drifting and schema corruption.
What are System One Models and Jev?
The System One paradigm was introduced by TypeSafe AI (founded by Diogo Almeida, former OpenAI research lead behind ChatGPT instruction-following). Inspired by Daniel Kahneman's distinction in Thinking, Fast and Slow, System One models deliver fast, intuitive, and mathematically calibrated decisions rather than slow, open-ended generative string completions.
System One vs Generative LLMs
Traditional LLMs predict one token at a time in recurrent loops. They are flexible for chat, but introduce high latency, type drift, and hallucination risks into automated code. System One models trade away string generation to execute parallel decisions directly over program state with zero type errors and zero hallucinations.
Jev & The Jevons Paradox
Jev is TypeSafe AI's flagship frontier model. Named after economist William Stanley Jevons, it embodies the thesis that dropping the cost and latency of decision intelligence by two orders of magnitude unlocks massive demand for automated software workflows, real-time routing, and deterministic guardrails.
Drop-in SDK & API Compatibility
Compatible with TypeSafe Jev client libraries and OpenAPI schemas
This server implements the exact TypeSafe Jev API contract. You can point any TypeSafe-compatible SDK or standard HTTP client to this self-hosted endpoint simply by changing the base URL:
import httpx
response = httpx.post(
"http://localhost:8000/v1/systemone",
json={
"state": "Payment failed for customer #4421.",
"model": "laya",
"questions": {
"is_urgent": {"type": "noul", "instructions": "Is this time sensitive?"},
"dept": {"type": "choice", "instructions": "Target team", "criteria": {"billing": "Billing", "tech": "Tech"}},
"severity": {"type": "score", "instructions": "Incident severity", "criteria": ["Low", "High"]}
}
}
)
print(response.json()["answers"])
Interactive Playground
Execute live inference requests directly against the local POST /v1/systemone endpoint.
Stored in session storage. Leave blank for public access (rate limited to 10 requests per minute per IP), or enter your configured API key for unlimited access.
Mandatory for the Compare tab to query typesafe/jev-1.13 on OpenRouter. Handled directly by your browser via HTTPS. Never sent to or accessible by our backend server.
Context string evaluated across all configured questions.
Map of question keys to question objects (types: noul, choice, score).
Ready for evaluation.
Select a preset or enter input state and click Run System 1 Evaluation.
Laya (Local) vs TypeSafe Jev (OpenRouter)
Enter your OpenRouter API key on the left to benchmark both models side by side.
Laya Multilingual Decision Backbone
System One runs on Laya, a non-autoregressive decision model developed by Convai Innovations and Nandha Kishor. Laya replaces generative token prediction with direct probabilistic classification heads.
FastAPI service wrapper, TypeSafe Jev specification compatibility, and developer playground interface.
Laya model architecture design, pretraining, and RLCD alignment implementation.
Research laboratory specializing in efficient decision-first architectures.
Official model checkpoints, tokenizer configurations, and weight releases.
Bidirectional encoder with 22 layers, hidden dimension 768, and 256k vocabulary.
Compact footprint suitable for local edge execution and high-density server nodes.
Up to 1024 tokens context per question with 256 tokens dedicated to instructions.
Reinforcement Learning from Calibrated Decisions across 15,987 updates and 4 epochs.
Evaluated across 51 languages in the MASSIVE benchmark with superior accuracy on non-Latin scripts.
Batched evaluation throughput on standard hardware with ~32.8 ms single forward pass.
Permissive commercial and research license for weights and deployment code.
Frequently Asked Questions
How does System 1 operate without autoregressive decoding?
Standard generative models predict text token by token in a recurrent loop. System One uses a bidirectional encoder backbone coupled with non-autoregressive decision heads. Both the state context and question instructions are processed in a single forward pass, producing deterministic classifications and calibrated probabilities without token decoding delays.
How are the output probabilities calibrated?
Probabilities are calibrated through temperature scaling and Reinforcement Learning from Calibrated Decisions (RLCD). When System One returns a probability of 0.85 for a binary decision or choice category, that classification is empirically accurate approximately 85% of the time across test distributions.
How do I integrate System One into an existing backend?
Send an HTTP POST request to the
/v1/systemone
endpoint with your state context and questions
schema. The endpoint responds with structured
JSON containing typed answers and token usage.
curl -X POST "http://localhost:8000/v1/systemone" \
-H "Content-Type: application/json" \
-d '{
"state": "Customer charged twice for order #8841. Requesting full refund.",
"questions": {
"is_refund": {
"type": "noul",
"instructions": "Is the customer requesting a refund?"
}
}
}'
What question types are supported?
System One supports three structured question types:
- noul: Binary decision returning a calibrated float between 0.0 and 1.0.
- choice: Categorical selection among named criteria options with probability distribution.
- score: Ordinal rating across ranked criteria levels with score index and confidence.