A grounded, evaluated, multilingual houseplant-care assistant — and the public evaluation harness that holds it to account. Sprout answers only from a versioned, cited corpus, never certifies a plant "safe," abstains when uncertain, and works in English and Spanish with enforced parity.
- Architecture — how the pipeline (guards → retrieve → extractive generate → citation guard → confidence) guarantees 100% groundedness by construction. See Architecture and the Threat model.
- The eval harness is the headline — five suites (groundedness, safety, calibration, refusal, multilingual), deterministic checks blended with an LLM judge whose model differs from the answer model. See the Evaluation report.
- Responsible by construction — see the responsible-tech audits, the model card, the data card, and the accessibility conformance report.
Sprout is an independent, personal open-source project (Apache-2.0). The bundled corpus is synthetic and CC0. This is not veterinary or medical advice.
Standards conformance and per-repo metric values live in the Roadmap; the cross-cutting rigor is recorded through the public controls, targets, and evidence in this site.