Jev: The Decision Layer
Most AI systems are asked to generate an answer. Many business systems need something narrower: a decision that software can inspect and use. Jev returns bounded, probability-backed signals rather than paragraphs, and leaves the action to your application. This ten-part series explains the pattern from first principles: the three question primitives, the control boundary around the signal, why a decision model is not an agent, how BMAD can specify and build the integration, South African use cases, thresholds and evaluation, security and POPIA, cost and operations, and when not to use it at all. One fictional municipality and its service-request channel carry the examples. This is an independent CloudNala analysis, with product details checked on 13 October 2026.
10 articles
- Part 1Jev in plain EnglishMost AI systems are asked to write an answer. Plenty of business software needs something narrower: a decision it can inspect and act on. Part 1 explains what the Jev decision model is, what it accepts and returns, what it deliberately does not do, and why that narrow role is the point.8 min read
- Part 2Choice, Score and NoulJev answers three shapes of question: pick one option, place something on a scale, or estimate the probability of yes. Part 2 explains each primitive, what it returns, how it fails, how several questions share one state, and the six-question test a decision must pass before anyone automates it.7 min read
- Part 3State in, signal out, code actsA decision model is only safe inside a boundary your application controls. Part 3 sets out the five layers of that boundary, from preparing minimised state to recording what actually happened, with a worked server-side example and a failure-handling design that never loses a case.7 min read
- Part 4Jev is not an agentA decision model, an agent, a harness and a policy layer are four different components with four different kinds of authority. Part 4 separates them, shows how a coding agent can ask Jev for a risk signal without handing it permission, and sets out the rule that tells you which layer failed.7 min read
- Part 5Jev and BMAD togetherBMAD organises how a team decides what to build and carries that context into the code. Jev answers typed questions inside the running system. Part 5 shows how the two fit: what a decision feature's specification must contain, a template to start from, the brownfield adapter pattern, and when combining them is unnecessary.7 min read
- Part 6Six South African places a decision layer could helpLocal value comes from operating reality, not from a South African flag in the prompt. Part 6 works through six candidate patterns for typed AI decisions, from municipal service triage to tender classification, and shows where the decision service stops and the organisation's authority begins. It closes on language and code-switching.8 min read
- Part 7Thresholds, evals and human reviewA polished number such as 0.91 can look more certain than the system has earned. Part 7 covers how to build a labelled evaluation set, what to measure for Choice, Score and Noul, why thresholds belong to actions rather than models, and what human review has to answer before it is a control rather than a phrase.8 min read
- Part 8Security, POPIA and procurement for a decision serviceTyped output makes integration cleaner. It does not answer where data goes, who can see it or how long it is kept. Part 8 sets out the POPIA questions to settle before personal information reaches a decision service, a threat model with the layer that carries each control, and the evidence pack to request before sensitive use.8 min read
- Part 9Cost, reliability and operationsA cheap model call can still sit inside an expensive workflow. Part 9 builds a cost model that includes review effort, sets the service-level indicators that matter for a decision service, defines which errors to retry and which to fix, and names the owners a decision workflow needs once it is running.7 min read
- Part 10When to use Jev, and when not toThe final part turns the series into a decision. It sets out when a typed decision service is a good candidate, when deterministic code, a general-purpose model or a person should decide instead, a four-week pilot with a gate at the end of each week, and a production-readiness checklist, with the sources behind the series.8 min read
Working through any of this in your own organisation?
These articles come out of delivery work rather than theory. If the problems in this series look familiar, a short scoping conversation is usually the quickest way to work out what is worth doing first.
Talk to CloudNala