Krisis : A Decision Model Trained and Served in C++/CUDA
Krisis is the decision model of the I-Machine platform. It answers the small, typed questions that an agent asks at every step, and it is trained and served in C++26/CUDA, without Python. Pragma, the agent platform, calls it before it calls a large language model.
The problem
An agent works in steps. At each step it must choose the next tool to call, and in most agent frameworks a large language model makes that choice. The model reads the full context, reasons, and returns one action.
This is slow and costly, and the cost is the same for an obvious choice as for a hard one. Many steps have one clear next action. The LLM call still happens.
What Krisis does
Krisis is a compact model that answers typed decision questions. The questions follow a protocol we call “System One”. Each question has one of three types :
choice: which option to take, from a given list ;score: which level, on a given scale ;noul: true or false.
Each answer carries calibrated probabilities. Pragma compares the probability of the best answer to a confidence threshold. Above the threshold, Pragma applies the decision and does not call the LLM. Below the threshold, the LLM decides, as before. Krisis never removes the LLM : it removes the calls that the LLM does not need to make.
Architecture
- Encoder. A bidirectional encoder, EuroBERT-610m. The base model is open and published under the Apache 2.0 licence. The model has 641 M parameters.
- Decision head. Two attention layers read the encoded question and give a score to each option.
- Calibration. One temperature per question type (
choice,score,noul) turns the scores into calibrated probabilities. - Runtime. An in-house neural runtime in C++26/CUDA runs the forward pass, a hand-written backward pass and the AdamW optimiser. No Python is involved.
- Service. The model runs as a standalone HTTP service.
Training
The training data comes from real decision traces of agents on the platform.
For each decision, a reference LLM, Claude Opus, lists every action that is reasonable in that situation. A decision often has more than one good answer, so the model does not learn a single label. Each list becomes a soft target : the probability mass is spread over all the reasonable actions.
Training stops early when the score on a validation set stops improving. When new traces arrive, incremental training updates the model on the new data.
Measurement
The metric is the share of choices that a judge, Claude Opus, finds reasonable. It is not the agreement with the trace. The trace records one choice among several possible ones, and that choice is not always good : the original agent itself scores 0.73 with the judge.
The test set has 464 decisions. This is an internal benchmark, not a public one.
| Decision maker | Judged reasonable |
|---|---|
| Original agent | 0.73 |
| Krisis, previous version | 0.847 |
| Krisis v6, current version | 0.922 |
| Target | 0.95 |
Integration
Pragma calls Krisis over HTTP.
A web interface lets an operator :
- follow the state of the service ;
- compare model versions against the judge ;
- try a decision by hand ;
- start and follow training sessions.
The building blocks of the runtime are not specific to Krisis. The same layers, optimiser and training loop can serve other models, for example on audio or sensor data. The modular AI article describes how small models fit into the platform, and the AI Hub page describes the inference infrastructure. Other research notes are listed on the Research page.