Each agent carries an objective in one sentence, a measure, a check window and a guard measure.
Three choices set an outcome check up: the objective, the measure and the window.
An objective states the result the agent is there to produce, in one sentence a planner and a manager both recognise. The PO cancellation agent in the showcase estate carries this one: cancel purchase order lines that stand above the need, inside their cancellation window. The need is eight weeks of demand, so the measure behind the sentence is the projected cover on the lines it acts on.
A measure turns that sentence into a figure. For this agent the measure is the projected weeks of cover on the lines acted on, read 14 days after the write-back. The measure uses the same ontology definitions as ChatLima, so the number on the check matches the number a planner sees when asking.
A guard measure watches the other side of the objective. Cancelling too much would show up as shortage messages, so the guard measure counts shortage messages raised for those materials within 28 days. A target with a guard beside it keeps a single figure from telling half the story.
The check window is the gap between the action and the reading. Short windows give fast feedback; long windows measure what a manager cares about. Juno holds the verdict as still open until the window closes, so a page of verdicts always says which readings are complete.
The same three columns would describe every role built from the blueprint, from one plant to a whole chain.
| Agent and level | What it was asked to achieve | Measure and window | What the check found |
|---|---|---|---|
| A2 PO cancellation | Bring projected cover on the lines it acts on to eight weeks or below, and keep supply inside the lead-time fence. | Weeks of cover at day 14 · shortage messages at day 28 | 29 actions: cover 11.4 to 6.2 weeks; 1 material raised a shortage message and the line was re-ordered. |
| A2 Safety stock | Hold the service level for A-class materials at plant 1210 at 98% and release held stock where cover allows. | Service level and stock value at day 30 | 64 changes: service level 98.3%; stock value down $310,000 at the plant. |
| A1 Lead time drift | Flag suppliers whose measured lead time moves more than three days from the planning value. | Agreement with the buyer’s judgement at day 21 | 18 flags: 15 confirmed by buyers, 3 returned as seasonal and used to raise the threshold. |
Showcase estate. The PO cancellation row is the agent in production; the safety stock and lead time drift rows are designed roles from the agent blueprint, shown as they would report. The observability page shows the decisions behind each row. blueprint
Juno reports verdicts per agent and per 1,000 decisions: met, missed and still open. Managers read the same three figures for every agent, whichever chain it belongs to and whichever level it runs at.
Rules rest on assumptions, and Juno watches the assumptions as closely as the results.
Every published rule carries the assumptions it was written on: a lead time, a price band, a demand pattern, a supplier class. Juno reads those values from live data on a schedule and compares them with the assumption held in the rule version.
A reading outside the agreed range raises drift. The drift finding names the rule, the version, the assumption, the measured value and the decisions made since the readings started to move. The owner of the rule receives it the same day.
Drift is a signal about the rule, so the answer is a rule decision. Your team widens the range, changes the condition or lowers the agent a level while the case is open. Each answer arrives as a new version, and the audit trail records which decisions were made under which assumption.
Gartner, 11 March 2026, predicts that by 2030 half of agent deployment failures will come from weak enforcement of governance at run time. Outcome checks and drift readings are that enforcement, running on the same data the agents act on.
A missed verdict is a route into the rules, and every step of that route is recorded.
Holding actions of the same kind is a choice your team makes per agent: everything above a value, or only the material family in question. The setting sits with the limits, and the audit trail records when it was applied.
Lowering a level takes one step and is immediate. Raising it again takes a passing test on recorded cases, a named owner’s approval and a live escalation route, so a verdict that missed stays visible in the evidence for the next move.
Showcase estate.
On Monday 5 October, 14 days after the write-back, Juno read the projected cover on the 29 lines D. Okafor approved. Cover on those lines fell from 11.4 weeks to 6.2 weeks, inside the objective. The check recorded a met verdict against each of the 29 decisions and against the agent for that week.
The guard measure told the rest of the story. One material, 7741-X at plant 1210, raised a shortage message on day 19. The line had been cancelled inside its window, and demand moved forward the following week. The buyer re-ordered it, and the case became the first candidate for a lead-time fence condition in a later version.
One record per action: signal, rule version, facts, limits checked, approver, write-back, acknowledgement and outcome, replayable at any date.
The difference is between counting actions and measuring results.
Most automation reports count what the machine did: rules fired, messages sent, orders changed. Activity counts answer whether the thing ran. They leave open the question a finance lead asks, which is whether the result improved.
Gartner, 5 August 2026, found that 55% of chief supply chain officers are unclear on the returns from AI, while 67% of their digital investment goes to it. An objective, a measure and a window per agent turn that question into a figure your own team reads each month.
The owner of the agent writes it with the planner or buyer who does the work today. It is reviewed when the agent moves a level, and it is stored with the rule version it belongs to.
The measure states what the agent is accountable for, and the guard measure states what it may harm. Where an outside event explains the reading, the reviewer records the reason, and the verdict keeps that note.
Yes. A new objective can be read against decisions already made, because the facts, the rule version and the outcome are all held in the record. Teams use this to set a fair target before an agent moves up a level.
Gartner’s decision intelligence platform market, published on 26 January 2026, names outcome assessment and decision governance among its mandatory features. This page and the audit trail page cover both.
Each one builds on the one before it.
Reads your rows in place.
Connect and move ↗One record per real thing.
Master and govern ↗Apps for every decision.
Operate and decide ↗Asks and acts, at your level.
Ask and act ↗Bring two systems and one decision your planners make every week. In the first week we register the data, confirm the mapping with your team, answer your questions in ChatLima and run the first agent at the level you choose.
