On September 15, 2026, TypeSafe AI came out of stealth with Jev — a model that returns typed decisions instead of generating text. LangChain called it the missing piece for agent loops. The AI world called it a new category: System One models.
We built one 9 months ago. Before anyone had a name for it.
The problem we saw first
Every enterprise AI agent needs to classify what kind of action a command requires before it executes. Not what to do — when and how.
Execute now? Wait for a trigger? Act before something happens? Run forever automatically? Those are four completely different execution types. An LLM making that decision on every query is like using a power station to light a single candle. It works. But it costs 400× too much and takes 200× too long.
We built ConceptNet to solve exactly this — a dedicated classifier for enterprise intent. Four deterministic outputs. No text generation. No hallucination possible. Just a structured decision your agent can act on immediately.
What we built vs what Jev built
TypeSafe AI built a general-purpose System One model — it answers any question you define in advance. Impressive engineering by a world-class team led by Diogo Almeida, ChatGPT's co-creator.
ConceptNet built a dedicated enterprise intent classifier — it answers the specific question every enterprise AI agent asks thousands of times a day: what kind of action does this command require?
The difference matters in production:
| Factor | Jev (TypeSafe AI) | ConceptNet |
|---|---|---|
| Benchmark accuracy | 68% | 100% — peer reviewed |
| Token cost | $0.042/million tokens | Zero — token-free |
| Air-gap deployment | API required | Full local deployment |
| Open source | Closed | MIT licensed |
| Languages | Not specified | 9 live today |
| Government/defence | Not mentioned | DASA submitted |
| Output type | Choice, Score, Yes/No | 4 enterprise intent layers |
What independent verification actually looks like
Before TypeSafe launched. Before LangChain published. An independent ML researcher from the Hugging Face community found ConceptNet and tried to break it.
He used grouped lexical-family holdout splits — specifically designed to catch train/test leakage and memorisation. He confirmed 99.3% on the adversarial holdout. We fixed 4 issues in 24 hours. Retrained to 100%.
That's peer review. Not a benchmark. Not a marketing claim. Someone tried to break it and couldn't.
Why enterprise needs something different
Jev is designed for software that knows its possible answers in advance — support ticket routing, lead scoring, trust and safety checks. Those are important problems.
Enterprise workflow automation has an additional requirement: the classification layer must work in environments where the internet doesn't. Government data centres. Military installations. Hospitals with strict data sovereignty requirements. Companies in markets where US AI infrastructure is restricted or politically unacceptable.
ConceptNet runs entirely locally. No API call. No cloud. No data leaving the device. Sub-5ms on CPU hardware. That's not a feature for enterprise. That's a procurement requirement.
The category Jev validated
We're genuinely pleased TypeSafe AI launched. Not because it validates our business — though it does. Because it proves the category is real.
The next two years will see dedicated classification models replace LLMs for every bounded decision task in enterprise software. Intent classification is the first and largest of those tasks. Every AI agent makes this decision thousands of times a day.
The question isn't whether System One models are real. LangChain just published 1,500 words answering that.
The question is which one you trust with your enterprise workflows, your government data, and your autonomous systems.
Try ConceptNet now — no login, no API key needed:
See it for yourself
Type any enterprise workflow command. See it classified instantly into the right execution type. No login. No API key. No setup.