Typed decisions, plus shaped text via generate
Chat LLMs
Free text you parse
System 1 decision models
Typed decisions
Coming soon
“I was charged twice for order #4471.”
Which team owns this?
one glance · no thinking
margin 0.91
reasoning = "auto"
confident (margin ≥ 0.50)✓
stable when options reordered✓
→ answer directly
idle · not needed for this one
{
"team": {
"choice": "billing"
"confidence": 0.94
"thought": false
}
}
generatereply, one line
“Refund for the duplicate charge is on its way.”
requests0
answered directly0
thought first0
thinking rate–
runs oncloud · on-prem
Not Chat
Reinforcement learning from human feedback taught language models to please people. They became brilliant at following instructions, and that is what we now call chat. The same training brings overconfidence and unreliable answers, so every LLM workflow still needs a human watching it.
Not Just System 1
Fast decision models answer every question in a single glance. That is perfect for easy calls and brittle for hard ones. They read only text, and they forget everything between requests.
System 1.5
Wity decides at the speed of instinct and thinks only when a question calls for it. It sees images and hears audio, keeps a working memory across a whole task, and writes the short piece of text an action needs. And it runs wherever your data is allowed to be.
Typed outputs, with a confidence on every answer, that your software can act on directly.
Instant on easy calls, deliberate on hard ones, and the output type never changes.
Our cloud or your premises. Same model, same API.
Five differences
A Wity state isn't limited to text. Put photos, screenshots, scanned documents and audio in the same request as your text, and ask typed questions across all of it at once.
answers = wity.decide(state=[image("dashcam.jpg"),audio("voicemail.wav"),"Policy 4471 · comprehensive cover",],questions={"at_fault": Noul("Was the insured driver at fault?"),"damage": Score("How bad is the damage?",["cosmetic", "repairable", "total loss"]),},)
Every question you ask Wity takes one of four shapes. You pick the shape, Wity fills it, and your code gets back a value it can branch on. Decisions come with a confidence; generated text comes back in the shape you defined.
A yes-or-no call.
Noul(question)“Was the insured driver at fault?”
Where on a scale it sits.
Score(question, levels)“How bad is the damage?”
Pick one of many.
Choice(question, options)“Next action on this page?”
The words an action needs.
generate(instruction, shape)“Origin city, as printed on the ticket”
shape [A-Za-z ]{1,40}✓ fits
Illustrative examples. Values are not benchmark results.
Benchmarks
We ran hosted Jev and Wity on the same questions. On one-glance decisions they are close. On anything that needs thought, Wity pulls away.
90/90
Hard reasoning, Wity with thinking
Jev: 61/90
+32
Points over Jev on hard reasoning
100% vs 67.8%
0
Answers changed by option order or naming
Jev: 5 of 104
95.5%
Accuracy with auto-thinking
thinking on 29% of questions
Accuracy (%) · higher is better
90 questions with exact answers across six families
Accuracy (%) · higher is better
15 questions per family · hover a group for counts
Accuracy (%) on hard reasoning · higher is better
Same 90 questions · Jev has no thinking mode, shown as a line
Accuracy (%) vs. share of questions thought about
110 questions, 50 held out from public benchmarks
Lower is better
Options reversed and option IDs replaced by opaque hashes
Single-pass accuracy (%) on standard decision suites · higher is better
No thinking on either side. Both sit near the ceiling here; the gap opens on questions that need thought.
Built For Automation
Wity returns typed decisions with a confidence on every answer, so your code can reason about uncertainty. Set where it acts on its own and where it hands off for review, and let it think on the calls that need it. Chain decisions inside one session and stay in full control of how the intelligence is used.
Between instinct and deliberation. Fast when it's sure, thinking only when it isn't.
Typed decisions, plus shaped text via generate
Chat LLMs
Free text you parse
System 1 decision models
Typed decisions
On every decision
Chat LLMs
None
System 1 decision models
On every decision
Text, images, screenshots, documents, audio
Chat LLMs
Varies by model
System 1 decision models
Text
Sessions: send only what changed
Chat LLMs
Re-send the whole chat
System 1 decision models
Each request stands alone
Thinks automatically when unsure
Chat LLMs
Always think, or never
System 1 decision models
Split them up yourself
Our cloud or your premises
Chat LLMs
Vendor cloud
System 1 decision models
Hosted API
Deployment
Some decisions can't leave the building. Wity runs the same way in our cloud and on your own hardware.
System 1 thinking is fast and intuitive; System 2 is slow and deliberate. A System 1.5 model moves between the two: it answers most questions in a single fast pass and thinks only when a question needs it, and either way it returns a typed decision. Wity is the first System 1.5 model, built for decisions inside software.
Five ways. It's multimodal, so it reads images, documents and audio as well as text. It's stateful, so a session carries context across a whole task. It adds a generate primitive for the short text an action needs. It thinks automatically when an answer is uncertain. And it can run on your own premises.
Text, images, screenshots, scanned documents and audio, mixed however you like in one request or across a session. Every question is answered against all of it.
You open a session and add to it as your workflow moves: a screenshot, a message, a file. Wity keeps that context, so each step sends only what changed and later decisions can build on earlier ones. An agent loop becomes a sequence of small calls instead of re-sending everything every time.
generate writes short text for an action: a value to fill in, a city to type, a one-line reply. It runs in the same session right after a decision, and its output follows a shape you define, so you can validate it like any other input. Decisions stay typed; generate covers the few places where software genuinely needs words.
With reasoning set to "auto", Wity answers directly when it is clear and thinks first when an answer is uncertain or unstable. You can also turn thinking off for the lowest latency, or always on. The output type is the same in every mode.
Yes. We deploy Wity on your premises, inside your network, with the same API and SDKs as Wity Cloud. Your data never leaves your environment, and we handle installation, tuning and updates with your team.
Structured outputs constrain the format of an LLM's text, but the model is still writing prose underneath with no sense of how sure it is. Wity's answer is a typed decision by construction, with a confidence attached.
Yes. No model is perfect. Use the confidence on each decision to decide when to act and when to escalate, and validate generated text before it touches anything important, exactly as you would with any other input.
Create an account to get an API key, read the docs, or join our Discord. For on-prem deployments and enterprise questions, email info@alphanimble.com.