In public administration the decision is rarely the hard part. The rule is written down, often as arithmetic a caseworker could do by hand. The hard part is assembling the facts from six registries, then proving months later which facts were used.
What one eligibility decision costs
Take a housing allowance revision. Employer-reported pay arrives in a monthly declaration whose payment period is not the period the money was earned in. Residence is inferred from three registers and asserted by none. The caseworker is not exercising judgment, they are reconciling: six screens, six identifiers, two systems that disagree about which person is which. The backlog is not the difficulty of the rule. It is the cost of the join.
Under load someone proposes a proxy: take the annual figure the tax system already holds, spread it evenly across the periods, raise the difference automatically. In at least one documented national scheme that is where the programme became unlawful: the statute defined entitlement on actual income per period.
A data engineering shortcut replaced a statutory definition with an approximation, and nobody signed it off, because it never looked like a decision.
The shortcut a model repeats
Point an agent at those registries and it makes the same substitution, faster and without a meeting. The statute carries three definitions of income and the warehouse conforms them into one. None of it is in the schema. It lives in the statute and in the caseworkers, so the agent joins on the key that looks right, which in France may be one a decree reserves for other purposes.
What the ontology declares
Four things, before anything runs. The decision, one sentence with a verdict and an appeal path; no appeal path means you are building reporting. The legal basis per field, because it decides what may be joined to what. The objects, with the test this sector adds to every other: if a citizen can contest it separately, it is an object. A household change is contested on different evidence than an income figure; as two attributes on one row they cannot be separated in an appeal. Time as part of identity: every fact carries the period it is about and the moment the administration knew it.
Declared that way the agent stops improvising. It reads across the registries in place, in the vocabulary the statute uses, and assembles in seconds the case that cost an hour of screens. The ontology is an input, not an output. Nothing generates it.
Reproducing a decision under appeal
Natural language finds the case and drafts the letter. It must not sit in the correctness path of the verdict. In Monarch a known decision runs as a registered transform: SQL over the base tables, rules in plain code, no model call. A saved decision stores the verdict, the rules that fired and the scope, so re-opening it re-derives and flags any change. The claim is narrower and stronger than “deterministic AI”: the verdict is computed by code a court can be shown, and the model chooses which code runs and explains it afterwards.
Staging is the other half: a write is never executed inline. It is held with a dry-run diff under the permissions captured at proposal, then committed on human approval or rolled back. French law lets a person subject to an algorithmic decision require the degree of the processing’s contribution, the data and their sources, the parameters applied to their situation, and the operations performed.[1] You cannot answer those four about a token stream. You can answer all four about a proposal a named officer approved under a recorded legal basis.
Where this is the wrong tool
If the decision is genuinely discretionary, this automates the cheap part and adds an explanation burden to the expensive one. If the register underneath is wrong, a decision system makes its errors faster and harder to argue with: they now arrive with official-looking provenance.
One condition throughout: every substitution is declared in the ontology and appears in the explanation given to the citizen. A proxy printed on the decision letter gets discussed before it ships.
Code des relations entre le public et l’administration, articles L311-3-1 and R311-3-1-2. ↩︎