Public administration

Context & background

Public administrations prepare decisions on the basis of heterogeneous documents: budget data, expert reports, statements, statutory requirements. From them arise decision papers that must be politically carried and legally reviewed.

The demands on transparency and traceability are rising — through freedom-of-information rights, public attention and the EU AI Act, which requires logging (Art. 12) and transparency (Art. 13) for decision-preparing systems. At the same time resources stay scarce. Administration needs tools that bring speed without giving up provability.

The problem

Decision papers are built from scattered, inconsistent documents that are merged by hand. In the process the link between a statement and its source is often lost — the rationale is in the end not traceable down to the source.

Generative AI aggravates the problem rather than solving it: a model that 'phrases' papers can invent numbers and statements for which there is no evidence. In an administrative act that must withstand legal review, that is untenable.

Concrete failure points

• Decision papers are built from scattered, inconsistent documents.
• The rationale is often not traceable down to its source.
• Transparency demands rise, resources don't.
• Generative phrasing can smuggle in unbacked statements.

The causal chain

From scattered files to a contestable resolution runs a clear chain:

1. 1
   Scattered, inconsistent documentsare merged manually into a paper
2. 2
   Manual mergingsevers statement from source
3. 3
   Missing source linkmakes the rationale untraceable
4. 4
   Untraceable rationaleleads to contestable resolutions

How cNode solves it

cNode ingests the relevant sources deterministically and links them in the knowledge graph — nothing is retyped, nothing freely interpreted. Every figure stays unambiguously linked to a source and a period.

The figures needed for the decision are derived deterministically and by rules, not freely guessed. Via a fixed seed the derivation is reproducible: the same case file provably yields the same result.

The rationale follows the compute path and keeps an audit trail per data point. Every statement in the paper can thus be traced back to its source — the basis for transparency toward the council and the public.

The language model phrases the paper in legible administrative language but invents neither numbers nor rationales. The system runs in Frankfurt or fully on-prem, so sensitive data never leaves the premises.

The deterministic process

1. Ingest relevant sources deterministically and link them in the graph.
2. Rule-based, reproducible derivation of figures via a seed.
3. Reasoning along the compute path, with an audit trail per data point.
4. Processing in Frankfurt or fully on-prem.
5. Paper phrased by the LLM; the numbers come from the engine.

The outcome

The outcome is decision papers that withstand scrutiny and real transparency toward the council and the public: every statement is backed, every derivation reproducible. Administration gains speed without giving up traceability — with full data sovereignty through EU hosting or on-prem operation.

Methodology & verifiability

Methodologically, the paper rests on a deterministic, rule-based derivation from linked sources, not on free text generation. The audit trail per data point and reproducibility via a fixed seed make every statement auditable — the prerequisite for an administrative act that withstands legal control.

The logging maps to EU AI Act Art. 12/13; processing takes place in Frankfurt or fully on-prem up to air-gapped, as a processor arrangement under GDPR Art. 28. Customer data is not used for training, and the system is model-agnostic.