GLOSSARY

The vocabulary of System One models.

19 terms, each defined in a sentence you can quote, then explained with an example. Start with what a System One model is.

System One model: A System One model is a decision model: it reads a state and answers typed questions — choose an option, score on a scale, or give the probability that a condition holds — in one forward pass, with calibrated confidence and no generated text.
System Two model: A System Two model reasons by generating text token by token — in practice, a large language model. The term contrasts it with System One models, which return a typed decision in a single pass.
Decision model: A decision model maps an input to an answer from a known set, with a probability for each answer, instead of producing open-ended text. System One models are decision models that take their questions at run time.
Typed question: A typed question is one question put to a System One model together with the shape of its answer: a choice among listed options, a score on a defined scale, or a noul probability. The type fixes what the model may return.
Choice question: A choice question asks a System One model to pick exactly one option from a list you supply, and returns the chosen option with a probability for every option.
Score question: A score question asks a System One model to place the state on an ordered scale you define — for example 0 to 4, with a meaning for each level — and returns the score with a probability for every level.
Noul: Noul is the System One question type for yes/no conditions: it returns the probability that a condition described in the question holds for the state, as one number between 0 and 1.
State: The state is the input a System One model reads: the text, JSON, log or conversation that its typed questions are about. One state can be asked many questions in the same request.
Calibrated confidence: A model’s confidence is calibrated when its probabilities match how often it is right: of all the answers it gives with 0.9 confidence, about 90% are correct. Calibration is what makes a threshold on the probability mean something.
Calibration error (ECE): Expected calibration error (ECE) measures how far a model’s confidence is from its accuracy: answers are grouped by confidence, and the gap between average confidence and accuracy in each group is averaged, weighted by group size. Lower is better; 0 is perfect.
Decision accuracy: Decision accuracy is the share of questions a System One model answers correctly on an evaluation set, reported as decision_accuracy — a number between 0 and 1 — in systemone.yaml.
Valid action rate: Valid action rate is the share of a model’s answers that are allowed answers — an option that was offered, a level that is on the scale. A model that only scores the allowed answers reaches 100% by construction.
Decision latency: Decision latency is the time a System One model takes to answer a request, usually reported as the median (p50) and the 95th percentile (p95) in milliseconds, on stated hardware.
Confidence gating: Confidence gating acts on a model’s answer only when its probability clears a threshold, and sends everything else to a fallback — a person, a larger model or a safe default. The threshold trades coverage for accuracy.
systemone.yaml: systemone.yaml is the manifest a System One model is published with: what it decides, its architecture and base model, licence, runtime, parameters, context length, evaluation numbers and, for hosted models, the provider and price.
Model tree: A model tree is a model’s lineage: the base model it was fine-tuned, quantized or converted from, that model’s own base, and so on — together with the fine-tunes built on it.
Open weights and hosted API: An open-weights System One model publishes its checkpoint for anyone to download and run; a hosted-API model is available only as a paid endpoint from its maker. Some models are both.
Laya: Laya is an open-weights System One model from Convai Innovations: a fully fine-tuned ModernBERT-large encoder with a decision head that scores one marker per option. It answers choice, score and noul questions in a single pass, under Apache-2.0.
LoRA fine-tuning: LoRA (low-rank adaptation) fine-tunes a model by training small low-rank matrices added to its weights instead of the weights themselves. It makes adapting a System One model to your own decisions fast and cheap enough for a laptop.