Risk matrix
CA-R2 requirements record. Updated: 2026-10-05. Risks drawn from the accepted corpus — each cites its evidence basis. No numerical risk scores are assigned: the corpus does not justify quantification. “Impact” is qualitative and refers to institutional legitimacy damage.
| Risk | Stage | Impact | Mitigation requirement | Residual/open question | Evidence basis |
|---|---|---|---|---|---|
| Manipulated selection parameters | Selection | catastrophic — silent capture of membership | specification is a signed governance artefact before any run; runs cannot proceed without it; full published methodology | how unauthorised-parameter changes are detected post-hoc | SRC-0078, SRC-0109, SRC-0123 |
| Biased invitation frame | Frame | high — structural exclusion before selection | record frame type + coverage gap publicly; frame choice is a declared governance decision | no frame is complete; coverage-gap disclosure standard undefined | SRC-0120, SRC-0021 |
| Self-selection distortion | Response | high — respondent pool skews educated/engaged | publish funnel; stratification + overweighting to correct; never claim pool = population | correction is partial by nature; acceptable residue undefined | SRC-0120, SRC-0038 |
| Inappropriate strata | Selection spec | high — engineered composition | strata + rationale + legal basis in published specification; contested criteria flagged (attitudinal) | who adjudicates strata legitimacy disputes | SRC-0109, SRC-0038, SRC-0032 |
| Privacy leakage | All stages | high — member/respondent harm + chilling effect | data minimisation; special-category handling per jurisdiction; small-cell suppression in publication | small-cell publication risk unresolved corpus-wide | SRC-0032, SRC-0122, SRC-0031 |
| Participant targeting / lobbying | Membership, deliberation | high — members steered by interests | member contact protection; lobbying disclosure rules; protected deliberation option (Brussels model) | balance transparency vs protection per institution | SRC-0030, SRC-0101 |
| Duplicate participation | Registration/selection | medium — gaming, seat double-occupancy | deduplication rules across pool/reserves/members; identity-verification policy declared | verification strength vs privacy trade-off | corpus methods records |
| Replacement manipulation | Replacement | high — backdoor membership control | replacement only under declared protocol; every event logged with provenance; reserve ordering rules published | max replacement rate before re-draw needed is undefined | SRC-0073, SRC-0074 |
| Agenda capture | Agenda/remit | high — output steered before deliberation starts | agenda authority declared; proposer/approver/exclusion-reasons recorded; member/public channels where the model grants them | least-studied power in corpus; no settled safeguard | SRC-0070, SRC-0100 |
| Expert/evidence capture | Evidence | high — evidence framing steers judgement | expert selection rationale + disclosures + member-requested evidence rights + contestability procedure | balance assessment remains a human judgement — by whom? | SRC-0038, SRC-0100, SRC-0129 |
| Facilitator bias | Deliberation | high — least auditable corpus stage | facilitation method + assignment recorded; oversight mechanism; session structure documented | measurement of influence unsolved | SRC-0106, SRC-0115 |
| Inaccessible participation | Membership/deliberation | high — seating ≠ participation | provisions per INCLUSION-ACCESSIBILITY.md; exclusion events recorded | enforcement of provision delivery | SRC-0038, SRC-0101 |
| Digital exclusion | All member-facing stages | medium-high — excludes offline population | offline/postal/telephone parity for every required interaction | cost of offline parity at scale | SRC-0088, SRC-0120 |
| Opaque voting | Decision | high — unverifiable outputs | decision record contract per DECISION-AND-DISSENT.md; method + electorate + result preserved | secret-ballot verification without identity linkage | SRC-0038, SRC-0116 |
| Suppression of dissent | Decision/publication | medium-high — false consensus impression | dissent/minority objects are first-class; support levels published per recommendation | minority-statement trigger thresholds | SRC-0017, SRC-0047, SRC-0111 |
| Misleading claims of representativeness | Publication/communication | high — legitimacy fraud (esp. open assemblies presented as random) | legitimacy basis is a declared property; mass input carries bias label; self-selection never renderable as sortition | enforcement is reputational not technical | SRC-0066, SRC-0106 |
| Ignored recommendations | Response | high — hollow institution | response obligation declared + tracked; no response is a visible state; deadlines computable | tracking is not enforceability | SRC-0027, SRC-0043 |
| Misleading implementation claims | Implementation | high — declared ≠ delivered | response/commitment/implementation are separate objects; progress evidence linked, self-report labelled | who verifies implementation evidence | SRC-0027, SRC-0043 |
| Concentration of power in permanent members | Permanent institutions | medium-high — incumbency capture | rotation + bounded terms + repeat-service rules; memory in records not tenure | optimal rotation fraction unknown | SRC-0021, SRC-0118, SRC-0109 |
| Capture in delegated/nested systems | Multi-level | medium-high — delegate class emerges | delegation events auditable; fresh-draw fractions preserved; descriptive-claims not composed upward | delegation vs fresh-sortition ratio is constitutional, not technical | SRC-0111, SRC-0125 |
| Confusion between AI suggestions and human decisions | Any AI-assisted surface | medium-high — authority laundering | AI artefacts labelled and provenanced as advisory; never actions of authority (CA-D007) | interface conventions for advisory vs authoritative | CA-D007, THEORY-TO-PRACTICE |
Cross-cutting observations
- The three highest-leverage attack surfaces are all pre-deliberation: frame, specification, agenda. Controls there are worth more than all post-hoc transparency.
- Permanent institutions trade capture risk upward: from manipulation of a single draw to incumbency of a rotating class — the mitigation set differs fundamentally from one-off assemblies.
- The weakest-controlled risks are judgement-capture risks (facilitator bias, evidence balance) because the corpus offers no measurement instrument — they require governance process, not mechanism.
Sources
research/methods/deliberation-risks-and-safeguards.md,research/methods/stratification-and-fairness.md,research/theory/legitimacy-and-accountability.md,research/theory/THEORY-TO-PRACTICE.md- SRC-0017, SRC-0021, SRC-0030–0032, SRC-0038, SRC-0043, SRC-0047, SRC-0070, SRC-0078, SRC-0088, SRC-0100–SRC-0102, SRC-0106, SRC-0109, SRC-0111, SRC-0115, SRC-0118, SRC-0120, SRC-0122, SRC-0123, SRC-0125