EU AI Act Classification — Sprout
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. Per
STANDARDS/RESPONSIBLE-TECH-FRAMEWORK.md §Governance,
the obligation here is the decision artifact, not certification: silence is non-conformant, a
written classification is conformant.
- Author / accountable owner: Chelsea Kelly-Reif
- Last regenerated: 2026-06-22 · Re-run trigger: any material change to model, prompt, architecture, or intended use; and on each AI Act enforcement phase-gate.
Classification (the explicit line)
Sprout is a minimal-risk AI system under Reg. (EU) 2024/1689. It is not an Annex III high-risk system, it is not a prohibited practice (Art. 5), it is not a general-purpose AI model (GPAI) — Sprout uses a foundation model in an opt-in cloud seam, it does not place one on the market. Training compute = 0 (no model is trained or fine-tuned), so the 10^25-FLOP systemic-risk GPAI threshold is not engaged. The only obligations that attach are the Art. 50 transparency duties, which are met by disclosure + the model card.
Why — step by step
1. Is it a "prohibited practice"? (Art. 5) — No
Sprout does no subliminal manipulation, social scoring, biometric categorisation, emotion inference, or untargeted face-scraping. It answers houseplant-care questions from a cited corpus. Not prohibited.
2. Is it "high-risk"? (Art. 6 + Annex III) — No
Annex III enumerates eight high-risk domains: biometrics; critical infrastructure; education/vocational training; employment/worker management; access to essential private/public services (incl. credit, benefits, emergency dispatch); law enforcement; migration/asylum/border; administration of justice and democratic processes. Sprout touches none of them. It is a consumer information tool about houseplants. It is also not a safety component of a product covered by the Annex I harmonisation legislation. Not Annex III; not high-risk.
Note on the toxicity feature: answering "is this plant toxic to my cat?" is adjacent to safety, but it is consumer horticultural information, not a regulated safety component, medical device, or an Annex III service. Sprout's design reinforces this: it never certifies a plant safe, routes ingestion questions to a vet / poison-control line, and labels every answer "not veterinary advice." The feature lowers risk; it does not create a high-risk classification.
3. Is it a GPAI model? (Art. 51–55) — No
A GPAI model is a model placed on the market that displays significant generality. Sprout places no
model on the market. In its default mode it ships a deterministic retriever + extractive generator
(HashingEmbedding + BM25 + ExtractiveGenerator) — no foundation model at all. In its opt-in cloud
seam it is a downstream deployer calling a third party's hosted Claude model; the GPAI obligations
fall on that upstream provider, not on Sprout. Not a GPAI provider.
4. Training compute / systemic-risk threshold — 0 FLOP
Sprout trains and fine-tunes nothing. There is no training run, so the 10^25-FLOP systemic-risk presumption for GPAI (Art. 51) is structurally inapplicable. The environmental footprint of the default mode is effectively zero (offline, no GPU inference); the opt-in cloud seam's per-token inference is recorded in the model-card CO2 row (NIST AI 600-1 Risk 5).
The obligations that do apply: Art. 50 transparency (minimal-risk)
For a minimal-risk system, the AI Act imposes voluntary codes plus the Art. 50 transparency duties where AI interaction or AI-generated content is involved. Sprout meets them by disclosure, not by a new mechanism:
| Art. 50 duty | How Sprout meets it | Where |
|---|---|---|
| Inform users they are interacting with an AI system | "Reference implementation" banner on the UI; README states it plainly; the assistant is self-evidently an AI chat tool | README, UI banner |
| Label AI-generated / machine-processed content | Every answer carries a disclosure string ("Answers are drawn only from a dated, cited plant-care corpus. This is not veterinary advice.") in EN and ES; every claim carries an inline citation + "as of" date | PromptConfig.disclosure_by_lang, Answer.disclosure |
| Make limitations understandable | Model card states intended use, out-of-scope use, and what the system cannot do; honest refusals are first-class output | docs/cards/model-card.md; refusal path in answer.py |
| Accessibility of the disclosure | Disclosure rendered in the WCAG 2.2 AA UI and the non-chat transcript view | Accessibility audit / ACR |
Machine-readable content labeling (Art. 50 §2, deep-fake/synthetic-media marking) is not engaged — Sprout generates extractive text grounded to a cited corpus, not synthetic media; it carries human-readable disclosure regardless.
Net obligations and conformance
| Question | Answer |
|---|---|
| Prohibited (Art. 5)? | No |
| High-risk (Art. 6 / Annex III)? | No |
| GPAI model provider (Art. 51–55)? | No (deployer of a third-party model in an opt-in seam) |
| GPAI with systemic risk (≥10^25 FLOP)? | No — training compute = 0 |
| Limited-risk transparency (Art. 50)? | Applies — met via disclosure + model card |
| Conformity-assessment package (Art. 17/18/47)? | N/A — only required for Annex III high-risk |
| Overall classification | Minimal-risk |
Conformance statement. No conformity assessment, CE marking, EU database registration, or fundamental-rights impact assessment is required. Sprout nonetheless ships the transparency artifacts (model card, disclosure strings, citation provenance, accessibility statement) and the broader responsible-AI audit set voluntarily — the point of the project is to demonstrate the discipline a buyer audits to, even where the law does not compel it.
Cross-references
- AI risk register (NIST AI RMF MAP):
ai-risk-register.md - ISO 42001 SoA:
iso42001-soa.md - Model card:
docs/cards/model-card.md - Methodology + current framework versions:
STANDARDS/RESPONSIBLE-TECH-FRAMEWORK.md