Every clinical fact has a source.
Repertify.ai is built so nothing clinical is invented by the AI. The language model reads, extracts, and writes — but the remedies, rubrics, grades, dose ranges, and red-flag rules all come from the public-domain and peer-reviewed sources listed below.
Repertory
Which symptoms point to which remedies, at what grade. These are the canonical rubric trees the scoring engine reads.
Materia Medica
Three authoritative materia medica voices, cross-validated. The MMMatchAgent compares each candidate remedy against all three sources — convergent agreement raises the essence score; disagreement is surfaced to the practitioner. No commercial homeopathy software does this automatically.
Dose & potency
Everything on the prescription card — potency, dose form, repetition rule, stop rules, follow-up window — is derived from these texts, with the specific citation shown per prescription.
- Hahnemann S. — Organon der Heilkunst, 6th edition
- Kent JT. — Lectures on Homoeopathic Philosophy, Lect. 33–35
- Vithoulkas G. — Levels of Health (B. Jain, 2010)
- Vithoulkas G. — Science of Homeopathy, Ch. 14–15
- Boericke W. — Pocket Manual
- Phatak SR. — Materia Medica of Homoeopathic Medicines
Remedy classification (151 polychrests + long-tail inference)
Each remedy is classified by action duration (short / medium / long), inherent sensitivity, tissue depth, first-dose cautions and known incompatibilities — drawn from Kent, Vithoulkas, Boericke and Phatak. Remedies outside the curated table are classified programmatically from Latin-name patterns (–inum → nosode, –metallicum → metal, –acidum → acid, –muriaticum/–nitricum/–carbonicum → mineral salt) before falling through to safe defaults.
Safety (48 rules — red-flag screen, drug interactions, screening triggers)
Every turn starts with a safety pass. 48 rules across four tiers — emergency (18) blocks the remedy panel entirely; urgent (14) triggers same-day referral; contraindication (8) downgrades homeopathy to supportive only; interaction (8) adds drug-coordination footnotes to the prescription.
- AYUSH-style integrative pharmacovigilance
Scoring
Kentian + Boenninghausen weighted sum-of-grades, with a rubric-size penalty so unspecific rubrics (e.g. "Mind — anxiety", 200+ remedies) cannot drown out characteristic ones. PSRC and mental dimensions are weighted higher than particulars; concomitants are weighted per Boenninghausen.
Confidence
Four explicit axes — Levels Covered (Kent's three-legged stool), PSRC (Organon §153), Essence Match (Vithoulkas), and Top-remedy Clarity. No black-box percentage. The tier action (PROCEED / PROBE / DEFER / BLOCKED) is computed deterministically from the four axes.
Language models
Used only for natural language — reading the case, extracting rubric candidates, classifying signals, writing the chat reply. The model cannot output a remedy that isn't in the corpus above, cannot invent a rubric grade, and cannot override the deterministic scoring or safety tier.
- DeepSeek-v4-pro
- Claude Opus
What the engine is NOT doing
- Not searching the web at runtime
- Not inventing remedies, rubrics, or grades — the model is constrained to the corpus
- Not learning from your patients — your cases stay in your account
- Not replacing your clinical judgement — every prescription card ends with the practitioner-CDS disclaimer
Where you see this in the app
- Every prescription card carries a Rule basis section listing the specific authorities used for that dose (Hahnemann §246, Kent Lect. 35, Vithoulkas, etc.).
- Each remedy card shows the matching rubrics with their repertory paths and grades — so you can trace any ranked remedy back to the rubric that put it there.
- The confidence panel shows four explicit axes (Levels, PSRC, Essence, Clarity) instead of a single opaque percentage.