AI that knows how sure it is.
Krites: decision models that return a calibrated probability for every question you ask, in one call that writes no text.
Models now act. They can't say when not to.
Every step is a decision
Models approve refunds, route tickets and call tools. A wrong step costs money, and there are thousands of steps a day.
A string has no threshold
A chat model answers in text. The confidence it states is not a probability you can set a policy on or audit.
Review is becoming law
Article 14 of the EU AI Act requires human oversight of high-risk systems. Someone has to decide which cases a person sees.
Krites Decisions. Questions in, probabilities out.
- One call, no text. The model reads the case once and scores every question from that pass.
- Typed answers. Choice, score or yes/no, each a full distribution. No JSON to repair.
- Your questions, at run time. Options and rubrics are plain text, so a new task needs no retraining.
- A cut-off you can price. Automate above it, send the rest to a person.
// a real response from Krites-2-4B, from the technical report // ticket: card charged twice, customer wants a refund today {"team": {"choice": "billing", "confidence": 0.98}, "urgency": {"score": 2.03, "confidence": 0.85}, "refund": {"noul": 0.98, "confidence": 0.96}, "usage": {"input_tokens": 362, "output_tokens": 0}} // probabilities per option are returned too; trimmed here
Level with Jev. Zero-shot.
On the public Typed Decisions benchmark, Krites-2-27B scores 72.2% against 72.7% for TypeSafe's closed Jev 1.13.0, a gap well inside the margin of error.
| Model | Access | Accuracy | Brier |
|---|---|---|---|
| Krites-2-2B | Open weights | 89.2% | 0.142 |
| Krites-2-4B | Open weights | 91.7% | 0.111 |
| Krites-2-9B | API | 92.5% | 0.099 |
| Krites-2-27B | API | 93.9% | 0.085 |
Calibration error (ECE), held-out
Faster: 32 questions in one pass vs one at a time
Generated per answer
Six weeks from first commit.
One founder, self-funded. The corpus, the training recipe, the kernels and the serving stack were built in house.
First commit
A framework that turns any open language model into a typed decision model.
Krites-1
1.08M typed decisions and the first models, up to 27B.
Audit-driven rebuild
Questions that fit their scenario: 77.2% to 93.2%.
Krites-2
1.25M decisions, 131 domains, 17 languages, in 174 single-GPU hours.
Launch
Decisions API beta, open 2B and 4B weights, the report on arXiv.
A category is forming. The incumbents are closed.
The reference product
Closed decision API, $0.042 per million input tokens. Krites matches it on the public benchmark and speaks its wire format.
Validation from the top
A limited preview was announced on 29 Sep 2026. Decision APIs are becoming a standard building block for agents.
Competing on price
Hosts sell open decision models from about $0.02 per million tokens. Price alone is not a moat; calibration and coverage are.
- Our position: Jev-compatible, calibrated and open at the small end, with two products nobody else offers: Map (any taxonomy) and Opinions (how people will respond).
Open at the bottom. Paid at the top.
Open 2B and 4B
Apache-2.0 weights and the framework. Developers run them locally and adopt the wire format. This is the funnel.
Hosted API
Prepaid, priced per input token; generated output is free because there is none. 9B and 27B are API-only.
Dedicated and private
Dedicated deployments, custom calibration on the customer's own labels, and data that stays in region.
Moving from Jev is a change of base URL. Switching cost is low by design; calibration and Map and Opinions keep customers.
Decisions first. Then the same engine, further.
Krites Decisions
Private beta of the hosted API for agent guardrails, ticket triage, moderation and routing.
Krites Map
Walks a customer's own taxonomy with a probability at every level and stops where it is unsure.
Krites Opinions
Estimates how a defined group would answer a question, scored against real survey data.
All three run on one serving stack: a shared base model, a small adapter and a read-out head per product.
Harnick Khera, founder.
Machine-learning researcher and first author at WASPAA 2025. Built Krites end to end: the corpus, the training recipe, the kernels and the serving stack.
- Speed. Four models and a technical report six weeks after the first commit.
- Rigour. Results on a public benchmark, held-out splits by schema, and an audit that rebuilt the corpus.
- Range. Research, data, training and serving, all in one person, before any funding.
We are raising a seed round.
Launch the APIs
Decisions first, with design partners, then Map and Opinions.
Larger and better
Larger models and the compute to train and serve them.
Human-response data
Real outcomes to calibrate against, the hardest asset to copy.