The assistant
A way to ask the platform a question in plain words. Every number in an answer is one a platform tool returned, under the asker's own access, and each answer carries its citations. It does not speculate, calculate new figures, give legal views or recommend enforcement.
How an answer is produced
- 1
Pre-check
The question is screened against the refusal categories below, and personal data is redacted before it is logged or sent anywhere. - 2
Intent
A rule parser identifies what is being asked — a total, a change, a source breakdown, a facility summary, open flags, a submission's status, a standards gap, a method explanation, a lineage record, or where to find a view. A model may propose an intent; anything outside that vocabulary falls back to the rules. - 3
Tool
A whitelisted tool answers it, inside the asker's own row-level-security session: an operator's assistant cannot see another operator's facility, because the database will not return it. - 4
Template
The answer is written from a template around the tool's figures, with a citation for each — the view and, where there is one, the lineage record. - 5
Numeric check
Every number in the answer must be one the tool returned, formatted as the template formats it. An optional rephrasing that loses or invents a number is discarded and the template answer shown; if that failed too, nothing numeric would be shown. - 6
Log
The redacted question, the intent, the citations and the outcome are logged, and purged after the retention period.
What it refuses
Each refusal says why, and where the answer can be found instead if there is one.
- Forecasts
- It reports the record and the stored pathways; it does not predict.
- Compliance verdicts
- Whether a facility complies is a regulator's determination, not a lookup.
- Enforcement advice
- It does not recommend action against an operator.
- Its own calculations
- Every number comes from a platform tool. It does not add, scale or convert figures itself.
- Questions about people
- It answers about facilities, figures and the method, not individuals.
- Changes to data
- It reads; it cannot file, decide, edit or publish anything.
- Anything outside the platform
- General knowledge, news and opinion are out of scope.
- Attempts to rewrite its instructions
- Recognised and refused before anything else happens.
Models and data
By default no language model is involved: the rule parser and templates answer everything.
A deployment may add a model to propose intents and rephrase answers — one hosted inside its own environment, or an approved external service under recorded terms. What leaves the platform is the question text and, for rephrasing, the template answer with every number replaced by a placeholder. No figure, tool result or platform record is ever sent to a model.
How it is tested
A fixed set of questions is put to it weekly, as the API answers them, including cross-operator questions it must refuse.
Release thresholds: numeric fidelity and citation presence 100%, correct refusals at least 95%, intent accuracy at least 90%.