One decision point
Every question and every action passes the same four checks, in the same order, before AI answers or acts.
Who is asking
Choose who asks. The four engines run the same checks in the same order, and each person gets what applies to them.
“Which protocol applies on ward 4?”
The ward 4 protocol applies, in the version deployed to your ward. You can read §1 and §2. §3 is for physicians, so it is withheld.
The ward 4 protocol applies, in the version deployed to your ward. You can read all three sections.
The ward 4 protocol applies. The agent acts for the nurse, so it gets §1 and §2, like the nurse. §3 is withheld.
- §1 AdmissionEveryone on the wardIncludedWithheldRequest access
- §2 MedicationEveryone on the wardIncludedWithheldRequest access
- §3 PhysiciansPhysicians onlyIncludedWithheldRequest access
Every access on the recordNEN 7513
| Who asks | §1 Admission | §2 Medication | §3 Physicians |
|---|---|---|---|
| A nurseWard 4 | Allowed | Allowed | Withheld |
| A physicianWard 4 | Allowed | Allowed | Allowed |
| An agentActing for the nurse | Allowed | Allowed | Withheld |
- Who may see it
Permission Engine
§3 withheld§1 and §2. §3 is for physicians only, so it shows as withheld.
All threeAll three sections, §3 included.
§3 withheldThe nurse’s rights, never more: §1 and §2. §3 shows as withheld.
How the Permission Engine works
The same permissions for people and agents, checked on every request. Rights from your source systems apply, with your own policies and markings on top. The same decision point serves chat, API, MCP and the designer.
- Rights from your source systems are mirrored, with their age checked.
- Your policies and markings apply on top, and sensitive fields are masked.
- An agent gets the rights of the person it acts for. Never more.
- What applies here and now
Applicability Engine
Ward 4The version deployed to ward 4, not another site’s.
Ward 4The version deployed to ward 4, not another site’s.
Ward 4The version deployed to ward 4, not another site’s.
How the Applicability Engine works
Relevant is not the same as applicable. Your model records status, version, scope and validity. Gluon uses them to hand AI only what applies to this person, this team and this moment.
- The right scope: knowledge is deployed per group, department and team.
- The right status: proposed and draft knowledge never reaches an answer.
- The right moment: validity dates are part of the model.
- What the record says afterwards
Evidence Engine
LoggedWho asked and what was read, on the record as NEN 7513 asks.
LoggedWho asked and what was read, on the record as NEN 7513 asks.
LoggedThe agent, the nurse it acts for and what was read, on the record.
How the Evidence Engine works
Every answer and action can be explained. Each answer shows how it was reached, and every decision lands in an append-only, hash-chained audit trail that you can export to your SIEM.
- Which source applied, and which version.
- Which permissions were checked.
- What was left out: versions by name, withheld sources only as a count.
A worked example on the product as it is, with invented names and figures. Not a customer story.Read the example in full
What we believeWe’d rather say “we don’t know” than invent an answer.
Actions
Before an agent sends, pays or changes anything, Gluon gives a verdict and records why. Try it with the finance agent of the worked example.
The amount does not count here: a bank change is blocked whatever the amount.
Within the mandate. The invoice of EUR 60,000 is queued for payment.
Payment of EUR 40,000 put on hold until a second approver signs.
Refused. Only the supplier team changes bank details, after a call-back check.
Switched off. No action runs while the kill switch is on, whatever the amount.
Switched off. No action runs while the kill switch is on, whatever the amount.
- ActionActionActionActionAction
- Queue an invoice for payment · EUR 60,000Queue an invoice for payment · EUR 40,000Change a supplier’s bank accountQueue an invoice for payment · EUR 60,000Change a supplier’s bank account
- RuleRuleRuleRuleRule
- Procedure 2026 · §3.1Procedure 2026 · §3.1Procedure 2026 · bank detailsKill switch · onKill switch · on
- MandateMandateMandateMandateMandate
- Finance agent · up to EUR 10,000Finance agent · up to EUR 10,000Finance agent · not for bank detailsNot checked · the action is offNot checked · the action is off
- ApproverApproverReasonReasonReason
- None needed · within the mandateFinance controller · askedSupplier team only, after a call-back checkStopped at once, without a new version · who and why auditedStopped at once, without a new version · who and why audited
- Verdict
- AllowedApproval requiredBlockedSwitched off
Recorded in the audit log · append-only, hash-chained
A worked example on the product as it is, with invented names and figures. Not a customer story.Read the example in full
On the real screens
Why access was granted or refused, what an action changes before it runs, and the switch that stops it.
Why someone can or cannot see it.
Every check that passed or refused, and what was withheld or masked. An officer can view the answer as the person would.
Explain a person’s accessWhat an action changes, before it runs.
The parameters, the checks that passed, the approval it needs and what it sets in motion.
Preview an action and decide an approvalStop any action at once.
Without a new version. Who did it and why are audited.
The kill switchQuestions
How is Gluon different from RAG?
RAG finds passages that look like the question. Gluon first checks which sources have authority, which version applies and whether this person may see it. Retrieval then works only on that context.
Is Gluon a language model or a search engine?
Neither. Your model writes the answers and your search finds the documents. Gluon decides what the model may use and what your agents may do, and records each decision.
Do agents get more rights than people?
No. An agent gets the rights of the person it acts for, never more, checked on every request.
Who approves an action?
The approvers your model names for the action, with quorum and expiry. Gluon asks them and records their decision.
Can we stop an action at once?
Yes. The kill switch stops any action at once, without publishing a new version.
Does Gluon run long processes itself?
No. Gluon decides and executes single actions. Long-running processes belong to the agent platform you use, which receives an approved action.
Next steps
The manual walks through every screen for the roles it is for, with the screens as the product draws them.
- Explain a person’s accessWhy someone can or cannot see a record.
- Review segments and policiesWho gains access.
- See how an answer was madeEvery step an answer took.
- Preview an action and decide an approvalWhat it changes.
- The kill switchStop any action at once.
- ActionType in GOLEffects, approvers and risk tier.