Skip to content
gluon

Every answer explained

Citations, traces and an audit log you can verify.

Each answer cites its sources on the page they came from and shows how it was reached. Every decision lands in an append-only, hash-chained audit log that you can verify yourself.

Does the staff regulation apply to contractors?

Only chapters 5, 6 and 8, through the hiring agreement. Their manager or principal at the bank makes sure they know them.1

Personeelsreglement 2026 · page 3 · 1.2 Voor wie

Where do I send a question about the staff regulations?

To the HR department at the head office, Hanzelaan 240, 8099 KX Zwolle.1

Personeelsreglement 2026 · page 3 · 1.3 Gegevens van de werkgever

Where is my leave recorded?

In Workday, the bank's HR system, where leave, requests and personal data are recorded.1

Personeelsreglement 2026 · page 3 · 1.4 Begrippen · Workday

How many hours is a normal working week?

36 hours. Up to 40 can be agreed when the work requires it, always in writing and with your consent.1

Personeelsreglement 2026 · page 3 · 2.1 Arbeidsduur

Do I have fixed working hours?

Not in an office role: you agree a working pattern with your manager. The contact centre and Operations work to rosters.1

Personeelsreglement 2026 · page 3 · 2.2 Werktijden

Where do I send a question about the staff regulations? To the HR department at the head office, Hanzelaan 240, 8099 KX Zwolle.

Example: a sample staff handbook. Point at a passage to ask about it. The page is a real capture of the document inspector.
Personeelsreglement 2026Page 3 of 6
Page 3 of the sample staff handbook. The framed passage says the head office is at Hanzelaan 240, 8099 KX Zwolle, and that correspondence about the regulations goes to HR at the head office.
Every stage kept
The original, the PDF twin, the parse and the chunks are stored, so a job can restart from any stage.
Masked before it is stored
The source’s masker runs over every chunk, and anything masked makes a redacted twin.
How this was reachedEvery answer shows its work.Each answer has a trace: the plan, the sources read and their age, the model that answered and the checks that passed.

A verifiable trail

Gluon writes every decision to an append-only, hash-chained audit log.

1 hash chain for every decision

Append-only. Search and export the log and verify the chain yourself, without asking us.

Investigate and verify the audit trail
$ gluon audit-verify --help
Verify the audit log's hash chains against every external checkpoint; names the first broken entry

Usage: gluon audit-verify

Options:
  -h, --help  Print help

$ gluon audit-verify
ok: 25 entries checked

The command and its verdict as the manual shows them. It reads the log and its external checkpoints and changes nothing. gluon audit-verify

The hash chainEach record carries the hash of the one before it.
  1. Writer node-1, entry 23chat.ask
    Chains to
    3f9a1c07d2e4
    Hashes to
    8b21e6f0a94d
  2. Writer node-1, entry 24chat.ask
    Chains to
    8b21e6f0a94d
    Hashes to
    c5e07a3d61f8
  3. Writer node-1, entry 25audit.payload.read
    Chains to
    c5e07a3d61f8
    Hashes to
    2d7c4e9b0a13

Illustrative hashes and writer, in the form of the Integrity line in the audit log’s event detail.

Streams to the SIEM you choose

3 SIEMs the log streams to

Microsoft Sentinel, Splunk or Elastic.

Audit log
  • Microsoft Sentinel
  • Splunk
  • Elastic
Verifiable
Search and export the log and verify its hash chain.
In your SIEM
The log streams to Sentinel, Splunk or Elastic.
Logboek Dataverwerkingen
Processing steps follow the Dutch logging standard.
With a reason
Reading a payload or a chat asks for a reason, and legal holds apply.

Controls

The audit log and the officers’ view ask for a reason before they open a payload.

Audit logSearch and export the audit log and verify its hash chain. Reading a payload asks for a reason.

A picture of the dialog Open the chat: it asks for the reason for opening this chat, which is required and kept in the audit log with your name, before the chat can be opened.

An officer reads a chat only with a reason, and legal holds apply.

When people delete a chat, it is hidden, not erased.

A picture of three rows, an action, a source and a model, each with its switch: the model is switched off and its reason is recorded.

Switch something off at once.

Any action, source or model can be switched off with the kill switch, with the reason recorded.

Down to the passage

The Evidence Engine records which source applied, which version, which permissions were checked and what was left out.

DocumentsCheck every chunk an answer may cite.The document inspector shows each page with its chunks boxed, so you can see what was read and how it was split.
CitationsA citation opens on the page it came from.Word, PDF and PowerPoint keep their layout, so a citation shows the passage on its page, with the source’s trust level.
LineageWhere each field in an answer came from.The source, the record, the transformations and the mapping confidence of one value.

Drift

Sources change, rules change and models change. Drift measures retrieval, answer and citation accuracy, abstention and conflicts per deployment, each number with its own trend, never averaged into one.

Night 18: a source was updated. Example: one deployment over 30 nights, illustrative figures, not a customer’s.
The five measures per night, in the example
NightRetrieval accuracyAnswer accuracyCitation accuracyAbstentionConflicts
192%Too little evidence94%4%1
293%Too little evidence93%4%0
392%Too little evidence94%5%1
494%Too little evidence94%4%0
593%88%93%4%0
693%89%95%3%1
792%88%94%4%0
894%90%94%4%1
993%89%93%5%0
1094%89%94%4%0
1193%90%95%4%1
1292%89%94%4%0
1393%88%94%3%0
1494%90%93%4%1
1593%89%94%4%0
1693%90%95%4%0
1794%89%94%4%1
1891%87%81%9%7
1992%88%83%8%4
2093%89%88%6%2
2193%90%92%5%1
2294%89%94%4%0
2393%90%94%4%1
2494%90%95%4%0
2593%89%94%4%0
2694%90%94%3%1
2794%91%95%4%0
2893%90%94%4%0
2994%90%95%4%1
3094%91%95%4%0

Each number keeps its own trend.

A number without enough evidence says so. Nothing is folded into one score that hides where it went wrong.

  • 5 measures, never averaged into one

    Retrieval, answer and citation accuracy, abstention and conflicts, per deployment and Capability, each with its own trend.

    Read Drift and send the report
  • Every night, the golden questions again

    And whenever the setup changes, through the real model.

    Edit the golden set
  • 1 report for the people who pay for the use case

    A PDF report from the console, with each measure and its trend.

Golden runs
The golden questions of a use case, through the real model.
Live sample
A share of live answers checked by an evaluation model.
Matched answers
Live questions close to a golden one, scored against its expected answer.
The report
A PDF for the people who pay for the use case.

From a golden question
to a report.

Golden questions run again every night and when the setup changes, a sample of live answers is checked by an evaluation model, a reviewer labels what was set aside, and a PDF report goes to the people who pay for the use case.

  1. Golden questions and live answersEdit golden cases as a form in the designer. They are tests in the package, written into the open proposal.
  2. A person reviews what was set asideAbstentions, conflicts and disagreements go to a reviewer. The labels calibrate the judge.
  3. Five numbers per deploymentPer deployment and Capability, over time. Too little evidence says so instead of showing a number.
  4. A report for the people who payA PDF goes to the people who pay for the use case.

Each number,
and where it comes from.

MetricsRetrieval, answers, citations, abstention and conflicts.Per deployment and Capability, over time. Too little evidence says so instead of showing a number.
Per sourceWhich source supports a claim, and which disagree.Claims supported and unverified per source, and conflicts per pair of sources, with how many reached an answer.

October 2026 · Admin consoleRetrieval, answer and citation accuracy, abstention and conflicts, per deployment and over time, with a PDF report for the people who pay for the use case. Drift is in preview: built and tested, still being verified on real deployments.

Questions

Can we verify the audit log ourselves?

Yes. You can search and export it and verify its hash chain, without asking us.

What does a decision trail show?

Which source applied and which version, which permissions were checked, and what was left out: versions by name, withheld sources only as a count.

Does the log leave our environment?

No. It stays in your own environment under your retention rules, and streams to the SIEM you choose.

Can an officer read anyone’s chat?

Only with a reason, and legal holds apply. When people delete a chat, it is hidden, not erased.

Can we switch something off at once?

Yes. Any action, source or model can be switched off with the kill switch, with the reason recorded.

Does Drift see who asked?

No. What it sets aside for review keeps counts and ids only, never who asked. A reviewer sees the text only when they may read every record the answer used.

Which model judges the answers?

An evaluation model you choose through Gluon Gateway, calibrated by the labels your reviewers give.

Is Drift finished?

It is in preview: built and tested, still being verified on real deployments.

Where it is used