Governance authority versus software permission
CA-R2 requirements record. Updated: 2026-10-05. This is a foundational separation for the whole requirements layer.
political/institutional authority ≠ technical access permission
(who is *entitled* to decide) (who is *able* to act in software)
The corpus shows legitimacy-critical actions routinely performed by people with system access but contested authority (commissioners setting remits, SRC-0100; facilitators steering rooms, SRC-0106). Software must therefore record both who acted and under whose authority — and treat divergence between them as a first-class auditable condition, not an edge case.
Illustrative mappings
| Actor | Technical permission they need | Institutional authority they hold |
|---|---|---|
| Facilitator | edit session agenda text, manage breakouts | not the authority to define the assembly’s agenda (that belongs to the declared agenda authority, AGENDA-AND-REMIT.md) |
| Administrator / operator | execute a selection run | only to run an already-authorised selection specification — no authority to choose strata, weights or members (SELECTION-REQUIREMENTS.md) |
| Receiving authority | post a response | may respond, accept, reject, modify — may not rewrite the assembly’s recommendation text (MANDATE-AND-RESPONSE.md) |
| Secretariat | publish records, maintain tracking | maintains records; does not own the judgements within them |
| System superuser | technical omnipotence for administration | zero constitutional authority — a superuser must not be able to silently alter membership, outcomes, or provenance |
Requirements
- Dual attribution: every consequential action records the acting account and the authority under which it was performed. Authority may be a rule, an office, an authorised specification, or a named body’s decision — but it is always recorded.
- Authority objects: authorisations (selection specifications, remit decisions, agenda approvals, response obligations) are recorded artefacts with their own provenance, versioned and attributable — not implied by role assignment.
- Divergence flagging: an action whose actor lacked recorded authority must be representable and flaggable. Software flags the divergence; it does not adjudicate legitimacy.
- No silent elevation: technical access must never confer institutional authority as a side effect. Superuser/maintenance access is recorded as such and produces audit-visible events for every consequential object touched.
- Read-only primacy for receiving authorities: the receiving authority’s relationship to the recommendation object is respond/annotate — never edit.
- Witness model: operations whose legitimacy is procedural (selection runs, result closure) support independent witnesses whose attestations are part of the record.
- AI boundary: automated/AI outputs may only ever attach as suggestions with clear provenance — never as actions of authority (CA-D007). An AI artefact must be visually and structurally distinguishable from an authoritative record.
Audit separation
- The audit trail must be append-oriented: corrections append; they do not overwrite.
- Audit records must be independently verifiable by a party without operator cooperation (published specifications + seeds + signed run records for selection; versioned objects elsewhere).
- The authority layer (who is entitled) is reviewable independently from the permission layer (who can technically act) — different auditors, different questions.
Open questions
- Whether authority declarations expire or require reconfirmation between cycles.
- How a violation flag is resolved institutionally (who hears it, with what effect on the record’s standing).
- Granularity: per-action authority vs per-cycle blanket authorisation.