Safety model¶

Status: proposed. Requires review by a qualified botanist/pharmacologist/toxicologist before Phase 1 ships.

Why safety is a first-class design topic¶

Mis-identification can poison. Many edible and medicinal plants have toxic look-alikes.
Traditional use is not medical advice. Dose, preparation, interactions with medication and contraindications matter and are rarely documented.
Over-collection can destroy species. Publishing locations of threatened medicinal plants invites stripping.
AI hallucination is most dangerous exactly here.

Always show top-k candidates with calibrated confidence, never a single confident answer from a weak match.
Below a confidence threshold: "We can't identify this reliably." Offer re-capture guidance (leaf, flower, bark, whole plant).
Never say "safe to eat/use". At most: "Reported as traditionally eaten. Do not eat anything you haven't confirmed with an expert."

Each plant has toxicity_level (none_known | caution | toxic | deadly) and lookalikes[] with distinguishing features.
Toxic/deadly warnings are shown before use information, not after.
Known dangerous look-alike pairs trigger an extra confirmation step.

Use-claims are shown as "Traditionally used by [community/source] for…" — never as treatment.
Persistent, readable disclaimer on medicinal content; link to seeking qualified/registered health care; emergency guidance (poison centre contact — to be verified with relevant bodies).
No dosage generation by AI. Dosage appears only as an attributed, safety-reviewed claim, and we default to not showing it until reviewed.
No product sales, no health-condition-targeted recommendations to individuals.

harvest_sensitive species: public location is generalised to a coarse region (e.g. quarter-degree grid or district); precise location is private to the contributor and never leaves their device unless consented.
Photos' EXIF GPS is stripped on upload by default.
Species with protected status display sustainable-use guidance and legal notes.

Claim states: unreviewed → reviewed → flagged. Medicinal and food-preparation claims on plants with toxicity_level ≥ caution cannot be published to public until reviewed by a qualified reviewer.
AI-generated or AI-assisted claims are always labelled and always enter as unreviewed.

Strong guidance for children: never taste/pick unknown plants; adult supervision cues.
Pregnancy and medication-interaction warnings where documented.

Retrieval-grounded only. The assistant answers solely from approved claims; if none support an answer, it says so.
Citation required. Every sentence about uses/safety links to a claim.
Refusal behaviours: won't give dosages, won't diagnose, won't advise stopping medication, won't suggest combining plants.
Evaluation suite of adversarial prompts (poisonous plants described as edible, dosage requests, "is this safe for my baby?") run in CI against any prompt/model change. See AI agents.
Prompt-injection hygiene: user-contributed text is data, never instructions.

A safety review queue staffed by qualified volunteers.
A fast-path "flag this" on every claim; flagged claims are hidden pending review for toxicity-related flags.
Incident process: if harm is reported, claim hidden, root cause analysed, public post-mortem.
Partnerships to pursue: SANBI, university pharmacology/ethnobotany departments, poison information centres.

Authoritative, openly licensed toxicity data for South African flora?
How to calibrate confidence honestly across wildly different species frequencies?
What liability exposure exists for a free educational app, and what disclaimers/structures mitigate it? (Counsel.)