Research / architecture

Current research

CA-A1 — selection contract

Type: architecture Updated: 2026-10-05 Scale: cross-jurisdictional Path: ../architecture/SELECTION-CONTRACT.md

architectureselectionaudit

Sources: SRC-0077SRC-0078SRC-0083SRC-0109SRC-0123SRC-0144SRC-0145SRC-0152

Selection contract

CA-A1 architecture record. Updated: 2026-10-05. Three distinct layers. Architected for an external engine first — no new selection algorithm is designed.

Layer 1 — Governance specification (human/institutional authority)

The SelectionSpecification (CA-OBJ-008) is the authorised artefact. It records:

The spec is locked before any run. A run may only execute against a spec in authorised/locked state. Post-authorisation change → new version, old version preserved.

Layer 2 — Selection execution (deterministic engine)

Engine receives:

Engine returns:

Contract ID: CA-CONTRACT-SEL-001 (see SELECTION-ENGINE-ADAPTER.md).

Layer 3 — Audit ceremony

The AuditBundle (CA-OBJ-010) records:

Modelled on the corpus’s strongest evidenced practice — Decidim’s public die-roll + witness attestation (SRC-0083) — lifted into a first-class artefact rather than a blog post.

Rules that bind all three layers

  1. Spec authority is institutional; execution authority is administrative; audit authority is independent-of-operator. Three different actors may hold them — the record must capture all three (CA-REQ-080/081).
  2. The engine never sees identity fields it doesn’t need — permitted-fields list is enforced at the pool-export boundary (privacy by construction).
  3. Re-running a published selection is a verification event, never a write — CA-REQ-015.
  4. Reserves come from the same witnessed run or a declared follow-up run under the same spec — replacement manipulation requires forging the bundle.

Sources