trust · transparency

Transparency Report — Q2 2026 sample

This is what the quarterly report looks like that every partner organisation receives automatically.

section b

What's in it

Per quarter, per client, automatically generated.

  1. 01

    AI calls per class — how many on L0 cloud, how many on L2 local?

  2. 02

    Provider breakdown — which model, how often?

  3. 03

    Overridden suggestions — how often does the human deviate from the proposal?

  4. 04

    Identified bias signals — where measurable

  5. 05

    Incidents + response times — what went wrong, how quickly resolved?

  6. 06

    Changes in the model mix — which models were added or removed

section c

Sample — fictive figures Q2 2026

These numbers are illustrative, not real client data.

Caveat Numbers below are fictive, for illustration.

Calls per data class

  • L0Public: 82,341 (89%)
  • L1Confidential: 7,218 (8%)
  • L2Secret: 2,940 (3%)
  • L3Special: 0

Provider breakdown (L0 + L1)

  • Anthropic Claude 4 (Sonnet): 54%
  • OpenAI GPT-4 (EU): 28%
  • Self-hosted Qwen 2.5 (L1 fallback): 18%

Human overrides

  • Total proposed actions: 1,847
  • Executed as proposed: 1,210 (66%)
  • Adjusted by human: 521 (28%)
  • Rejected: 116 (6%)

Incidents

  • 1× short Anthropic EU outage (4 min), automatic failover to Azure EU.
  • 1× classification error (initially L1, after review L2) — data not forwarded, incident logged.

Experimental

Bias monitoring

This bias monitoring is in development and warrants caution in interpretation. We publish it anyway, because transparency about uncertainty is also accountability.

  • Under-representation of sources from southern Dutch local press in "media tab" detected → curated sources extended per 2026-05-12.

section d

How this comes about

Every AI call, every override, every classification is logged real-time. At the end of each quarter the platform generates this report automatically. No manual compilation.

Clients also receive the report in machine-readable format (JSON/CSV) for their own analysis or supervisory reporting.

section e

Why we publish this

  1. 01

    Accountability before request

    A regulator or citizen with a doubt doesn't need to ask for the report. It's already there.

  2. 02

    Self-correction

    Visible bias signals force us to expand sources, review models.

  3. 03

    Field example

    If more AI vendors did this, the public sector would need to be much less cautious.

section f

Download

PDF of this sample report — available on request.

Machine-readable (JSON/CSV) — monthly available for client deployments (not quarterly).

The report your organisation would receive is available from the first full quarter after onboarding.

Request sample PDF