Software — sortition-algorithms
Classification: selection engine library — the algorithmic core used by Panelot and OpenDLP. Updated: 2026-10-05. CA-R3 record.
Identity
- Project:
sortitionfoundation/sortition-algorithms— “A package containing algorithms for sortition — democratic lotteries” (SRC-0145). - Publisher: Sortition Foundation; algorithm family developed with Harvard/Carnegie Mellon academics and published in Nature (SRC-0077, SRC-0078).
- Distribution: PyPI package
sortition-algorithms(v0.12.14).
CA-R3 pinned assessment (2026-10-05)
- Pinned source: release v0.12.14 (2026-10-01); HEAD
edc26e93a93c; PyPI 0.12.14 (SRC-0145). Pinned into OpenDLP backend assortition-algorithms==0.12.14(SRC-0144). - Licence: GPL-3.0 — copyleft; embedding in a combined Cicada process has distribution implications; use as a distinct service/process boundary is the natural containment (LICENSING).
- Maintenance: active; recent release cadence; small maintainer pool (Sortition Foundation).
- Deployment model: Python library — runs in-process or behind a thin service; no standalone server component.
- Algorithms: maximin, leximin, nash, diversimax and legacy methods with quota-fulfilment reporting (SRC-0077) — matches the published selection-optimisation literature.
- CA-REQ coverage (relevant): CA-REQ-013 partially (deterministic execution given identical inputs/spec — the engine half of the run record), CA-REQ-011 input side (targets/categories); reserves via remaining-pool handling in calling systems. Not covered: the governance specification artefact, the witness/seed ceremony, published-methodology rendering — those layers sit outside the library (they live in OpenDLP / operator process).
- API/integration seam: Python API + CSV-compatible tabular inputs; consumed today by OpenDLP, Panelot (historically) and stratification-app.
- Audit/provenance: deterministic given inputs+seed at the library level (per published method); the library does not itself emit a ceremony bundle — caller must construct it.
- Identity/privacy: operates on anonymised feature columns — the strongest privacy posture in the selection toolset (respondent rows need not carry identity fields to the engine).
- Operational burden: low — a library, not a platform.
- Extension model: library API; algorithm selection configurable.
- Accessibility/offline: headless; runs fully offline — directly relevant to offline-auditability.
- Material gaps: no run-ceremony packaging (seed capture, witness log, signed output); no respondent lifecycle; input validation is caller responsibility.
- Adoption risks: GPL-3.0 copyleft boundary must be architected deliberately at CA-A1; small maintainer concentration.
- Ecosystem context: sibling components
stratification-app(v1.1.1, GPL-3.0 GUI),groupselect-app/groupselect-lib(v2.1.3/v0.4.0, MIT) show the foundation maintaining multiple selection surfaces (SRC-0156). - Provisional classification:
adapt— credible engine for Cicada selection runs behind a bounded wrapper that adds the governance-specification input and audit-artefact output layers;integratealternative is consuming it via OpenDLP.