The normative-force ladder
Every provision is graded from a stated value to a duty backed by a consequence, by who it binds, how firmly and with what penalty: the Abbott and Snidal legalization framework, applied provision by provision.
Not our opinion of good governance. Theirs.
UNESCO, the OECD, UNDP, the G7 Hiroshima process, the EU AI Act, NIST and the UN digital compacts, at the versions named beside them. The roster is configuration, not code, so it moves as the frameworks do.
Eight dimensions, read at once.
Transparency. Accountability. Privacy. Safety. Human autonomy. Inclusivity. Fairness. Environmental sustainability. Each one gets a coverage verdict, a depth stage and the binding force behind it, scored on evidence rather than intent.
Ask it why. It cites the paragraph.
Every finding can be questioned in plain language, and every answer comes back with the framework text it rests on.
8
governance dimensions
per strategy
43
frameworks indexed
config/frameworks.yaml
96.4%
of citations confirmed at source
showcase runs, 8 countries
1,320
tests passing
CI, 80% coverage
The verdict is not the model’s opinion.
Four rules sit under every reading. They run in code, the same way every time, which is what makes a score arguable on the evidence.
Verdicts the language model sets
Coverage and depth are computed in code from counted, classified provisions. The language model is shown the verdict and explains it; it cannot set one, and cannot raise one.
Depth stages
Unaddressed, Emerging, Delegated, Operationalized, Institutionalized. A fixed scale means two analyses of two countries are actually comparable.
Citations confirmed at source
Every quoted excerpt is checked against the passage it cites. One that cannot be confirmed is marked unverified, with the reason, rather than quietly kept.
Nothing is asserted. Everything is traced.
A score you cannot audit is an opinion with a number on it. Each of the four stages below writes down what it did, so the brief at the end comes apart line by line: back through the reasoning, back to the paragraph it came from.
01
Ingestion
The document is split on its own headings, paragraphs intact. Nothing is summarised away before it is scored, and nothing is scored out of context.
02
Retrieval
Each passage is embedded and ranked per governance dimension, so a reading sees the paragraphs that bear on it and none of the ones that do not.
03
Analysis
Eight readings against the frameworks that govern them. Deterministic guardrails sit under each verdict, so it is not one the model can talk itself into.
04
Brief
The findings become a document a minister can act on: what is covered, what is missing, what to do first, and the citation behind every line.