shvydko

CONTEXT → DECISION

Shvydko — AI for decisions

The decision point
in your code.

Shvydko Decision Engine turns your context and criteria into a choice, score or probability. JSON output, ready for the next step.

Open Playground Documentation
  • ↳Choicecategory
  • ≡Scorescore
  • ◌Noulprobability
UA
SHVYDKO DECISION ENGINE

Fewer words.
More logic.

Our decision pipeline combines context, your criteria, and the specified response type. Shvydko Decision Engine returns a value your code can use to determine the next action.

How we think about AI
SIMPLE RULES → COMPLEX BEHAVIORGEN 000
Game of Life · not model inference
01

Generate text

An open-ended response for an explanation, conversation, or new material.

text
02

Define the structure

A JSON schema defines the fields and types of the result.

{ schema }
03

Make a decision

Your categories and criteria define the possible Shvydko responses.

choice · score · noul
01 /THE LANGUAGE OF YOUR DECISIONS

One context.
Three response formats.

Give the model text and questions. Get a category, score or probability in a format your code can use.

01 /Choice
CHOOSE ONE

The right route.

Define the options. Get the model's choice and a probability distribution.

EngineeringA
FinanceB
LogisticsC
02 /Score
RATE ON A SCALE

Your own quality scale.

Describe the levels in ascending order. The model returns a score on your scale.

0123
Lower impactHigher impact
03 /Noul
CHECK A STATEMENT

A signal for the next step.

Write a statement. Get its probability in the context of your request.

Human review is needed?
01

The illustrations explain the formats. Model output appears after you run a request.

FROM CONTEXT TO ACTION / 01

A decision has
a route.

The model makes a semantic choice.
Your code determines what happens next.

SHVYDKO / DECISION CIRCUIT01—03
state.txtINPUT DATA
An online store customer reports: “For the last 15 minutes, no customer has been able to pay for an order by card. The page shows payment unavailable. Payment on delivery works. We checked on three different devices. No money is being charged to the cards.” Routing policy: website failures go to engineering; questions about charges and invoices go to finance; delivery times go to logistics. A payment-method outage affecting many customers requires urgent human review.
RECORDED EXAMPLEANIMATION OFF

Give the model a situation.

A support message, document, or state of your system. Keep the relevant facts together in one context.

This diagram does not call the model. It shows a saved response to this fictional scenario. Animation speed does not represent processing speed.

Replay this request
ONE REQUEST / REAL RESPONSE

From example to verification.

Saved recording · shvydko-latest
Model time
652.5 ms
Reported by the service, excluding the browser network round trip
Input tokens
208
Context and all three questions
Charged for this request
$0 · preview
Free mode at the time of recording

This is one synthetic scenario run through the public API, not a quality evaluation or benchmark. Viewing this page does not execute the request. Playground opens the same context and questions; the response and timing may differ when you run it again.

Measurement details and full JSON
request_id
req_uGar8EBEH-ufcmkx9dEtosrx
API → model → API
784.29 ms
Total HTTPS request time
1029.885 ms
Output tokens
208

Total time was measured by the client that saved this recording, not in your browser. Token counts and the zero charge were checked against API usage history. One recording is not enough to draw conclusions about speed, accuracy, or probability calibration.

Download the request and response
02 /ADD IT TO YOUR STACK

Text becomes
part of
your logic.

Provide context and your own criteria. Use the response to route a request, check a condition or trigger the next step.

Read the API reference
INPUT / STATE + QUESTIONSOUTPUT / STRUCTURED JSON
POST /v1/systemone
request.json
{
  "state": "Card payments are not working. Payment on delivery is available.",
  "questions": {
    "team": {
      "type": "choice",
      "instructions": "Which team should handle this issue?",
      "criteria": {
        "engineering": "Website and integration failures",
        "logistics": "Delivery questions"
      }
    }
  }
}
Your categories. A familiar format.
03 /WHERE IT FITS

Find your
next decision.

Explore all use cases
Support01

Where should this request go?

Identify the team, the issue's impact and whether urgent review is needed, all from one message.

Choice Score Noul
Operations02

Does this return meet the policy?

Check the facts against the store's internal policy and distinguish a standard request from an exception.

Choice Score Noul
Sales03

What is the next step for this lead?

Identify the intent in a company's inquiry and prepare an appropriate handoff to sales.

Choice Score Noul
AI quality04

Does the source support the answer?

Review an assistant's draft response before showing it to a customer.

Choice Score Noul
DEPLOYED VERSION DIAGNOSTICS / 16.10.2026

Compare.
With all the
details.

Shvydko v2 and Jev 1.14 on the same questions. Deployed version evaluation from 16.10.2026: language slices, methodology, and limits of the findings.

Read the full study 4,119 questions out of 13,219 in the panel · 1,119 requests.
A limited diagnostic development dataset. The overall superiority criterion across all three languages was not met.
ACCURACY / FAMILY-MACRO1,662 QUESTIONS
90.50%
Shvydko v2 deployed version
87.57%
Jev 1.14 reference
Difference +2.93 ppInterval 98.33% +0.77 … +5.10 pp

Three equally weighted task families. The lower bound of the interval is above zero on this limited slice.

Paired cluster bootstrap: 20,000 resamples with Bonferroni correction for three languages. Jev: typesafe/jev-1.14-20261014.

Speed, economics, and English-language quality were not evaluated here. This result does not establish overall superiority over Jev. Numbers, versions, and methodology as JSON

FROM SIGNAL TO RULE / 02

Automate
or review?

Move the threshold. See which of the eight recorded Choice responses proceed automatically and which are held for human review.

99.5%
90%100%

Real API responses to eight fictional English learning scenarios, recorded on October 17, 2026. This panel shows only Choice answers; Workflow Lab also shows Score and Noul, including their mismatches with the authored expectations. This is not an assessment of calibration or overall quality. The initial threshold is illustrative; choose your own using a separate validation dataset.

Contexts, expectations, and full responses
CONFIDENCE / ROUTINGRECORDED API / 17.10.2026
8 / 8automatic
0 for review
Automatic For review
EXAMPLE 010.9998 ≥ 0.9950Automatic routingengineering · expected: engineering
04 /TRANSPARENT PRICING

Your volume.
Your estimate.

Pricing is based on input tokens. Output tokens carry no extra charge. Choose your volume and check the estimate.

100,000
1K1M
800
10010 000

Include the context, questions and criteria descriptions. The API returns the actual token count. Each request's charge is rounded up to the nearest microdollar.

MONTHLY ESTIMATE
Loading the rate…
Your estimated volume80 million tokens
Open the console A usage estimate, not an invoice.
NOTES FOR BUILDERS

Ideas. Practice. Shvydko.

Browse the journal
05 /BEFORE YOUR FIRST REQUEST

Questions?

All documentation
Who is Shvydko for?

Product developers who need to turn text into a defined decision: a category, score or probability. Examples include routing support requests, checking internal rules and evaluating AI responses.

Which model can I call?

The current model name and context limits are available through GET /v1/models with an API key, and on the limits page in the console. Shvydko is a decision API; this interface is not a chat or a general-purpose text generator.

How do I get an API key?

Sign in to the console, select a workspace and open API keys. An owner or administrator can create a key. The full value appears only once: store it on your server in a secrets manager.

Can my team use it together?

Yes. A workspace brings together members, API keys, usage and balance. Owners and administrators manage access. Create an invitation link in the console and share it with your colleague.

Where can I discuss an integration or higher usage?

The Enterprise page explains how to prepare a pilot. Use Contact to send the team your task and requirements. SSO, SLAs and private deployments currently require a separate agreement.

What does Shvydko return?

A structured result for each question. Choice selects a key from your options. Score returns the expected value on an ordered scale. Noul estimates the probability of a statement from 0 to 1. You supply the context and questions as JSON.

Can I ask several questions about the same text?

Yes. Add named questions to the questions object. They share the same state. Each question has its own type, instructions and, where needed, criteria. Request limits depend on the service configuration.

Are the examples actual model responses?

The library contains original learning scenarios with fictional companies. They do not contain preset results. Open an example in the Playground and run it to get a response from the configured model.

How should I use probabilities in a business workflow?

Choose a threshold on a separate sample of your own labeled requests and review the errors. Add human review for uncertain or consequential cases. The probability of an individual response does not guarantee that it is correct.

Has Shvydko been shown to outperform other models?

We publish a limited diagnostic sample with methodology and intervals. It does not establish overall superiority. Speed, economics, English-language quality and a final evaluation were not assessed in this sample. See the research page for details.