Status: DRAFT — pending owner sign-off. Drafted 2026-07-19 to meet the dated obligation to classify the system before full high-risk application of Reg. (EU) 2024/1689 on 2026-08-02. This is an owner decision artifact behind a review gate: it takes effect when the accountable owner reviews it, replaces this banner with a
Reviewed <date> by <owner>line, and re-commits.
Regulation: (EU) 2024/1689 (the AI Act). Full high-risk application Aug 2, 2026; GPAI obligations live since Aug 2025; Annex III conformity deadline Dec 2, 2027. Under the portfolio's responsible-technology governance policy, the obligation is the decision artifact, not certification: silence is non-conformant, a written classification is conformant.
- Accountable owner: Chelsea Kelly-Reif
- Re-run trigger: any material change to the LLM extraction seam, matching behavior, or intended use; and on each AI Act enforcement phase-gate.
constituent-reconciler is a minimal-risk AI system under Reg. (EU) 2024/1689. It is not an Annex III high-risk system, not a prohibited practice (Art. 5), and not a general-purpose AI model (GPAI) — its optional extraction seam uses hosted or local models (Bedrock/Ollama; the DV pack forces NoOp), it does not place one on the market. Training compute = 0 (no model is trained or fine-tuned), so the 10^25-FLOP GPAI threshold is not in scope.
The nearest Annex III categories are 5 (essential services) and 8 (justice-adjacent use by legal-aid orgs). The tool reconciles duplicate constituent records for nonprofit/legal-aid operators. It does not evaluate any person's eligibility for a service, score or profile constituents, or make or feed any listed decision: it proposes candidate record matches, and a human reviews every uncertain match — nothing merges silently (the merge-review invariant is a design gate). The optional LLM seam performs field extraction from unstructured text only; in the DV pack it is forced to NoOp.
Not triggered in normal operation: operators interact with a review UI, not a conversational AI; extraction is a batch tool. Revisit if a user-facing AI interaction surface ships.
Reclassify before shipping any of: (1) auto-merge of uncertain matches without human review; (2) use of match output as an input to eligibility, service-access, or case-priority decisions about a person; (3) constituent scoring/profiling features; (4) training or fine-tuning a model.