Decision model glossary
These terms are English in the source documentation. This page fixes one translation for each and keeps the English original alongside.
- System OneSystem One model
A class of models trained to make fast, structured decisions that software can consume directly. The key difference from generative LLMs is that they produce no text — so there is nothing to parse and nothing to hallucinate. The output is typed values and probability distributions your code can branch, sort and route on.
- state
The material a System One model evaluates: an email, a ticket, a JSON document, a passage of contract text. Every question in a request is evaluated against the same state, independently. A practical consequence: adding questions barely changes latency, and more questions do not degrade the context.
- questionquestion (primitive)
A typed question asked of the state. There are exactly three: choice, score, and noul. They are called primitives by analogy with software primitives — modular, composable, structured. You can mix all three and ask many at once in a single request.
- choice
Pick one option from a defined set. Jev’s Choice accepts up to 255 options. Alongside the selected item you get a probability for each option and an overall confidence. Good for classification, routing, and picking one candidate out of many.
- score
Place the state on an ordered, described scale. The answer is a score, a probability for each level, and confidence. The difference from choice is that the levels are ordered. A common mistake is asking one composite question — decompose into atomic scores and combine them with weights you control in code.
- noul
Ask a yes/no question and get back the probability that the answer is yes (0–1). Not the same thing as confidence: a noul is a probability about the world, confidence is how reliable the model considers its own judgement.
- confidence
How certain the model is about its own judgement. This is the most useful output: the answer tells you what, confidence tells you whether to act on it. The canonical use is gating — auto-apply above a threshold, route to a human below it. Note that not every question type returns it: choice and score do; noul does not, being a probability itself.
- calibration
Whether the probabilities mean what they say: among samples the model calls 70%, about 70% should actually be positive. Both models train for this with reinforcement learning against strictly proper scoring rules (Jev calls it RLCD), and it is the precondition for using confidence as a gate. Uncalibrated confidence is just a number.
⚠️ Training against a calibration objective is not the same as shipping calibrated probabilities. There is public contrary evidence: an independent experiment asked Jev to guess a fair die 400 times and it picked the same face every time at an average self-reported 82.9%, against an actual hit rate of 19.0%; a preregistered out-of-distribution test found opposite bias directions per primitive. In production, treat confidence as a prior to be calibrated and measure the trustworthy band on your own data — see the measurements and threshold selection in the guide.
- escalation
Handing a decision to a human when confidence lands in the middle band. It is a designed tier, not a failure branch: auto-apply above, drop below, escalate the band in between. Two things are easy to miss: the band is a human cost (wider band, more human work), and escalation needs a fail-safe — without one, an unclaimed item sits there forever while the dashboard looks exactly like "all clear".
fast-jev-compaction6,653★ · foreman540★ →
- cascade
Let the cheap tier decide first and send only what it is unsure about to the more expensive model. The saving equals the share of traffic inside the high-confidence region — one measurement settled 37% of cases at ≥90% confidence and removed 37% of the large-model calls — not some fixed percentage. It assumes the smaller model errs as random noise; if it is systematically overconfident on one class of input, that class is never escalated.
jev-review (devagrawal09)586★ · jev-search446★ →
- fan-out
Ask many questions in one request and let your code keep the useful answers. What you save is mostly input: the same state goes out once, so the more questions, the thinner the fixed per-request overhead spreads (one measurement: 76–86% median input-token saving with 8 questions in parallel). ⚠️ It suits independent judgements — classification, scoring, filtering — and not ranking, which requires items to be compared with each other: a benchmark found the batched path failing its ranking-quality gate.
jev-review (devagrawal09)586★ · jev-search446★ →
- guardrail
One structured judgement, applied before content enters or leaves a system, that decides pass / review / block / reroute. Its key trade-off is not the threshold but which way it falls on error: risk guardrails (permissions, secrets, dangerous actions) should fail closed, throughput guardrails (completion checks, format checks) should fail open. A second recurring practice is running deterministic rules first and giving the model only the grey area the rules cannot decide.
fast-jev-compaction6,653★ · foreman540★ →
- abstention
The mechanism a model uses to flag answers it is not sure about, without discarding them. Give a request a
min_confidencethreshold and any answer below it comes back markedlow_confidence— the answer itself is still returned, and the caller decides what to do. The gate reports one of three states:passed,abstained, orunevaluated(the answer carried no usable confidence, so the gate could not decide). With no threshold set, none of the keys appear at all: absence is the report, not a fourth state.- local deployment
The practice of running open decision models on your own machine — Ollaya is one ready-made way to do it. The hallmarks: model weights live on local disk, the service listens on a loopback address such as
127.0.0.1:11435, nothing is billed per call, and your data stays inside your own environment. Deployment and model are two layers: one deployment can serve many models, and the accuracy you get is the model’s, not the deployment’s.