Selection model requirements
CA-R2 requirements record. Updated: 2026-10-05. Defines what a selection capability must support, without selecting a final algorithm or product. Algorithm choice (leximin family vs quota-fill vs other authorised rules) remains open for CA-R3 assessment; Panelot/OpenDLP/Sortition Foundation evidence is used as evidence, not adoption.
The governing separation
GOVERNANCE DECISION TECHNICAL EXECUTION
which population, which strata, → apply an authorised selection
which quotas, which weights, specification reproducibly,
which data may be used auditable end-to-end
The specification is a governance artefact: authored by an authorised body, versioned, publishable, and signed off before any run. The run is a technical event: it must be impossible to execute without referencing an authorised specification.
Population and frame
- Eligible population — definable as a rule (residency/citizenship/age/exclusions), attributable to authority, versioned.
- Invitation population — the bounded subset of the frame actually approached; recorded with size and method.
- Invitation method — postal/register draw/digital/mixed; invitation weighting (e.g. deprivation-weighted, CA UK SRC-0038) is a governance parameter.
- Response pool — the opt-in set; funnel counts (invited/registered/selected/seated) reportable.
Selection specification contents
A selection specification must record:
- strata definitions and their political rationale;
- quota targets per stratum;
- permitted data fields and their legal basis (special-category prohibition vs consent model — Brussels SRC-0032 vs UK SRC-0122);
- attitudinal criteria if any — explicitly flagged as contested governance (CA UK used climate attitude; Brussels prohibits special-category strata);
- oversampling rules;
- reserve production (count and ordering rule);
- the selection algorithm identifier and parameter values;
- publication status (what is published, when).
Fairness concepts to support
- Equal invitation chance — the corpus ideal at invitation stage (SRC-0120); weighting is a declared exception, not a silent feature.
- Probability of final selection — demonstrably governed by the specification, not operator whim; within-pool fairness is what the algorithm optimises (leximin evidence, SRC-0078).
- Self-selection distortion — always present post-invitation; the record must make the funnel visible rather than claim it away.
- Oversampling — must be declared in the specification with rationale (CA UK +5 seats model).
Reserves, withdrawal, replacement
- Reserves — produced at selection time under the specification; reserve ordering/exhaustion rules declared.
- Withdrawal — a member status transition with a reason class (not free-text narrative).
- Replacement — executed under a declared protocol (same-strata redraw, reserve ordering); every replacement is a logged event with provenance. Replacement is a manipulation surface — operator discretion here must be minimal and auditable.
- Duplicate prevention — deduplication rules across respondent pool and reserves; a person must not be seated twice or sit in two overlapping roles.
Audit and reproducibility
- Witness/audit procedure — a ceremony model (witnesses, published seed or entropy artefact, signed run record); Decidim’s die-roll+witness pattern exists for proposal draws (SRC-0083); Panelot’s public computed lottery is the member-selection benchmark (SRC-0078–SRC-0080).
- Reproducibility — given the specification + seed + input pool digest, an independent party must be able to re-run and verify the output. Re-running must produce the same result; re-running must never silently replace a published result.
- Random seed/control — entropy source recorded; seed custody (who generated, who witnessed) recorded; post-hoc seed choice prohibited by construction (commit-before-draw where possible).
- Operator discretion — the corpus’s position: operators should be unable to choose members. Any discretionary parameter must appear in the specification.
- Published methodology — specification + method + outcome summary publishable without exposing member data.
Privacy constraints
- Strata data minimisation — collect only strata needed by the authorised specification.
- Special-category data — representable as prohibited, permitted-with-consent, or required per jurisdiction; the legal basis is a specification field (SRC-0032, SRC-0122).
- Privacy-preserving audit — the audit trail proves the draw’s integrity without publishing identifiable respondent data; small-cell publication risk is a recorded open problem.
- Retention — respondent, reserve and member data have distinct retention rules; reserves and non-selected respondents are the most sensitive (data held for people who never served).
AI boundary
AI must remain excluded from participant selection. No model may choose, rank, filter or weight persons. A future AI layer may explain a published specification; it may never execute or alter one (CA-D003, CA-D007).
Open questions
- Whether a public audit ceremony is required or strong reproducibility suffices.
- Reserve-list publication vs confidentiality.
- How to express “equal probability” when frames exclude unregistered people — declared coverage gap vs aspirational claim.
- Maximum tolerable operator discretion (near-zero per corpus; formal bound unresolved).
Sources
research/methods/civic-lottery-methods.md,research/methods/stratification-and-fairness.md,research/theory/selection-bias-and-stratification.md,research/software/opendlp.md,research/software/panelot.md- SRC-0017, SRC-0021, SRC-0030–0032, SRC-0038, SRC-0043, SRC-0047, SRC-0048, SRC-0073–SRC-0080, SRC-0083, SRC-0109, SRC-0120, SRC-0122, SRC-0123