The last chapter left one question open: where the language model sits, and what it gets to see.
Two halves, one handoff: the planning tables released as version 3.
Go back to the Friday of chapter 5. At 09:00 L. Tan typed one sentence: “Cancel open purchase order lines where projected excess is above eight weeks of demand, the line is inside its cancellation window, the line value is under $25,000 and the supplier class is standard. For lines of $25,000 and above, draft the cancellation and ask the buyer.” A cloud language model of the company’s choice read it, and five rows came back.
That is design time, the one place a cloud model appears: it reads the intent and drafts rows from a fixed action vocabulary. Token cost applies here: $420 over 62 calls to author the policy, $260 over 38 calls to simulate and refine it.
Now the Monday. At 06:00 the Lima Rules Engine read version 3 from the rule store and evaluated it on the 38 lines. For the basis under each draft it called the small language model inside your estate. The agent’s 40 runs over eight weeks carried zero token cost.
The language model drafts, explains, discovers candidate rules in past planner behaviour and detects drift. The rules engine decides and executes, so every write-back traces to a versioned rule and, at A2, to D. Okafor.
Showcase estate. Bars are shares of the agent’s $2,500 budget; $1,284 over eight weeks, all at design time.
The line is drawn on the model link and checked on every call.
On the Friday the cloud model needed to know what the tables are called and which words the company uses. So the names of the registered tables and columns crossed, with the vocabulary P. Sharma keeps (NCNR, the lead time fence, “strat supplier”) and the sentence L. Tan typed.
On the Monday the 38 lines, their values, the stock and demand figures behind the excess and the suppliers’ names stayed in SAP S/4HANA and Oracle Fusion, read in place. They stayed inside your estate when a basis was worded.
Each check is recorded, and a limit that holds an action turns it into a draft for a person.
| Limit | What it does | On the cancellation agent | On the Monday |
|---|---|---|---|
| Value threshold | Lines under the value become drafts; lines above it go to a person. | $25,000 a line | 29 lines under it became drafts worth $187,300; 9 above it reached D. Okafor as buyer, for his judgement, with the reason for each. |
| Daily action budget | Caps the actions one agent may take in a day. | 60 actions a day | The run used 29 of 60; the counter sits in every record. |
| Confidence floor | Holds a draft under the floor; a person receives the finding. | 0.80 | Line 00020 of order 4500067123 scored 0.91 and passed. |
| Rate limit | Spaces the writes an agent sends to a source system. | 12 writes a minute | The 29 approved writes reached SAP S/4HANA over 9 minutes, each acknowledged. |
| Two-person approval | Calls for a second named approver on sensitive fields. | Bank details · payment terms | A cancellation changes a quantity and a deletion indicator, so this rule stayed idle. |
You met the limits in chapter 4 as settings the owner and finance own. The rules engine reads all five before every action and writes each check into the record. They also made the level review safe: when J. Ridley moved lines under $25,000 to A3 after six weeks, the threshold, the budget and the floor stayed where they were, so the reach of that change stayed small.
DEC-2026-09-21-0417 holds eight parts; each answers one question a reviewer asks.
| Part | What the record holds | What a reviewer checks it against |
|---|---|---|
| 1 · Signal | Exception message 0034921 for material 7741-X at plant 1210, read Monday 06:00 | The message in the planning suite |
| 2 · Reasoning | Version 3, released 18 September 2026 by J. Ridley; the facts as read at 06:01: projected excess 11.4 weeks · line value $8,420 · window closes Thursday 24 September · supplier class standard | The rule text of version 3; the values the sources held at 06:01 |
| 3 · Recommendation | Drafted 06:06: cancel line 00020 of order 4500067123; three conditions matched; confidence 0.91 | The three conditions against those facts |
| 4 · Limits checked | Value under $25,000 passed · daily budget 29 of 60 · confidence floor 0.80 passed | The limits set in version 3 |
| 5 · Approval | D. Okafor opened the queue at 08:50 and approved it as drafted at 09:12, after 3 h 06 min in the queue; session and role recorded | The named approver; the level in force, A2 |
| 6 · Action | Purchase order change posted through the released interface for this agent | The write route opened after the test write |
| 7 · Acknowledgement | Order 4500067123 item 00020 carries the deletion indicator; message 20260921-114 stored as returned | The document in SAP S/4HANA itself |
| 8 · Outcome | Projected cover 6.2 weeks at day 14; the objective was met | The objective on the agent; the cover on day 14 |
Showcase estate. A rejected draft and a run at A4 leave the same eight parts.
Chapter 5 showed this record as the run wrote it. An auditor reads it the other way, from the outcome back to the signal. Each part is checked against something outside Juno: the document in the ERP, the rule text, the values the sources held that morning.
The seventh part carries the most weight, because it is the ERP’s own word: the document number and the message SAP returned, stored as received. A reviewer opens the order and finds the same item.
Open-source parts, named layer by layer, so an architecture review has something to check.
Read downwards, the picture answers the security committee. A person asks or approves at the top; the agents layer decides and executes; the meaning layer supplies the one definition of a purchase order line that chapter 6 put at the centre; the rows stay in your systems until a query reads them. Read upwards, it answers the architect: a new source joins at the bottom and reaches every screen through one mapping and one set of access rules.
| Layer | What it is built from | What that gives you |
|---|---|---|
| Interface | React, Node.js and Nginx | One interface for registering, asking, reviewing and approving. |
| Models | Gemini, Claude, OpenAI, Llama or another, interchangeable | Your choice of cloud model at design time; a small model inside your estate at run time. |
| Services | Spring Boot and FastAPI microservices, in containers | Each service scales on its own; an update leaves the rest running. |
| Storage and query | MariaDB, an ArangoDB graph database, a federated query engine, RabbitMQ and object storage | The ontology as a graph, the registry and the queue; your sources read in place. |
| Run time | Kubernetes, Docker and Linux | One deployment model across the four cloud targets. |
Containers on managed Kubernetes, with the cloud’s own storage and identity.
The same containers and deployment model, under your tenancy.
The same again, inside your account’s network and key management.
Your own data centre on Linux, for estates that keep everything inside.
Identity, role-based access and the asking person’s rights apply to every answer and every action.
| What buyers and auditors ask for | Framework, with its date | The Juno control |
|---|---|---|
| Human oversight. People can oversee the system, interpret its output and override it. | EU AI Act, Article 14 on human oversight; the Act came into force on 1 August 2024. | The level, set per agent: a named approver at A2, exceptions at A3, audit at A4. |
| Logs. Events are recorded automatically across the life of the system. | EU AI Act, Article 12 on record-keeping. | Eight parts for every run, written as it happens, kept for the period you set. |
| The ability to stop. A person can halt the system at any moment. | EU AI Act, Article 14, which asks that overseers can stop a system. | One step down, at once, by the owner or a manager; the daily action budget holds the reach of a run. |
| Accountability. Trained overseers are named, the system is watched, logs are kept for at least six months. | EU AI Act, Article 26 on the duties of deployers. | A named owner on every agent, a level register with the evidence behind each change, and manager reports per 1,000 decisions. |
| Risk management. Risk is identified, measured and managed while the system runs, inside a system a third party can audit. | NIST AI Risk Management Framework 1.0, 26 January 2023: govern, map, measure, manage. ISO/IEC 42001, published 18 December 2023. | Simulation on past data, accuracy testing, outcome checks, drift detection, and a version history on every governed object. |
Most supply chain planning agents sit outside the Annex III list of high-risk uses in the EU AI Act. Regulation (EU) 2026/1744, published 24 July 2026, moved the stand-alone high-risk duties to 2 December 2027. Buyers and auditors ask for the Article 12, 14 and 26 pattern all the same, and Juno follows it for every agent. Your own counsel makes the legal reading; dates checked 19 September 2026.
At design time: the names of your tables and columns, your vocabulary and the sentence the planner typed. Order rows, prices, stock figures and names stay inside your estate.
Yes. An owner or a manager lowers its level in one step and the change applies at once. The agent carries on at the lower level, drafting for a person.
The one you choose: Gemini, Claude, OpenAI, Llama or another. A swap leaves the planning tables, the ontology and the audit trail as they are.
You have seen one agent built and run. What else can be built this way, and where should you start?
Five end-to-end chains, the use cases the market is deploying now, and the map of everything else.
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.