Selection technology deep assessment
CA-R3 record. Updated: 2026-10-05. Compares OpenDLP, sortition-algorithms, Panelot against CA-R2 selection requirements.
Layer separation (important)
Four distinct layers are frequently conflated:
- Selection algorithm — the fairness/optimisation mathematics (leximin etc.).
- Civic-lottery workflow — respondent pool, targets, run, replacements, member handover.
- Governance specification — the authorised artefact saying which strata/quotas/weights (CA-REQ-010/011).
- Audit ceremony — seed capture, witness record, reproducibility bundle, publication (CA-REQ-013/014).
The corpus shows algorithms 1 is solved and published; 2 exists in one early platform; 3 and 4 are under-served everywhere and are mostly institutional-method shaped.
Comparison
| Aspect | OpenDLP | sortition-algorithms | Panelot |
|---|---|---|---|
| Layer | workflow (2) + engine via dep | engine (1) | workflow+engine, hosted |
| Stratified model | targets with min/max categories | category quotas input | leximin per published method |
| Optimisation objective | via lib: maximin/leximin/nash/diversimax | the family itself | leximin (Nature paper lineage) |
| Randomness/seed | not documented publicly | caller-supplied seed semantics (Python random) | opaque |
| Reproducible runs | run records persisted (targets_used, remaining_ids) | deterministic given inputs+seed | claimed by method; unverifiable hosted |
| Witness/public lottery | not evidenced | n/a (library) | not evidenced |
| Audit artefacts | DB records; report CSV in dev history | none emitted | unknown |
| Reserves/replacement | first-class operation | remaining-pool support | hosted feature, undocumented |
| Privacy model | respondent CSVs to operator; engine sees anonymised features | anonymised feature matrix | inside hosted service |
| Input/output | CSV in; panel + report out | Python API / tabular | CSV in/out (per practitioner material) |
| Algorithm transparency | transitively full (open lib) | full source + paper | method published, implementation not |
| Licence | Apache-2.0 | GPL-3.0 | unasserted (no public source) |
| Self-host | yes (Docker, docs) | library — trivially | no |
| Ops complexity | moderate (Flask+Celery+Redis+PG) | none (in-process) | none (SaaS) |
| Pinned ref | HEAD 4d73880 (2026-10-05), no tags | v0.12.14 (2026-10-01) | n/a |
Findings
- The algorithm layer does not need rebuilding. A published, maintained, peer-reviewed implementation exists (sortition-algorithms 0.12.14) and is the same code running the field’s reference practice.
- The workflow layer exists but is young. OpenDLP (pre-release, no tagged versions) covers respondent→selection→replacement with real persistence; its seed/witness export story is undocumented.
- Panelot remains
unresolved: no public source or licence; credible only as a Sortition-Foundation-operated service. - The missing layer is thin and ours-shaped: packaging spec + seed + witness + output into a published, independently re-runnable artefact (CA-REQ-013/014) is a small amount of engineering but a large amount of method — it is not a reason to write a new algorithm.
- stratification-app / groupselect (GPL/MIT siblings) show the SF ecosystem consolidating selection surfaces; OpenDLP is the strategic one.
Answer: does Cicada need its own selection algorithm?
No — evidence does not justify it. CA-R2 requirements are satisfiable by the published engine. What no existing system supplies is the governance-specification artefact and the witness/reproducibility bundle — thin layers Cicada likely must own regardless of engine choice. Engine selection between adapt (wrap sortition-algorithms) and integrate (call OpenDLP’s workflow) is a CA-A1 decision, not needed now.