AI repertorization — capabilities and limits

AI repertorization is the newest layer on the oldest workflow in homeopathic software: a model reads case material — typed notes, a photograph, live consultation audio — and proposes repertory rubrics that the practitioner then accepts, switches, or rejects. The proposal step is genuinely new; everything around it is not. Symptom selection judgement, rubric verification, and the materia-medica confirmation read remain the practitioner's work, and a model that proposes rubrics does not change what a repertory is — an index of the tradition's own literature.

The capability map

Implementation details matter more than labels. Similia's live audio pipeline extracts at most one symptom per two-minute cycle (roughly thirty rubrics per hour at maximum), prioritises strange-rare-peculiar symptoms over modalities, sensations, locations, causations, and concomitants in that order, and deliberately avoids overly generic rubrics, preferring entries with fifteen to a hundred remedies. Those design choices encode a defensible reading of Hahnemann's §153 — weight the striking and peculiar — but a different vendor's pipeline with different cutoffs is a different clinical instrument under the same marketing phrase.

The four failure modes that matter

Plausible-but-wrong rubrics. A model can propose a rubric that reads convincingly but does not match the case's meaning, or that exists in a different repertory than the one selected. The mitigation is procedural and non-negotiable: open the proposed rubric, read its sub-rubrics, and verify against the patient's actual report.

Entrainment. Live suggestion changes the consultation's attention economics: rubrics appearing on screen pull the practitioner toward the model's reading and away from listening. Treat live suggestions as a parallel notetaker, review the full transcript afterwards, and re-derive the case hierarchy before locking the analysis.

Generic-rubric drift. When input is vague, extraction pipelines drift toward polychrest-heavy generic rubrics, which add noise to any analysis. Good pipelines suppress this by design; the practitioner's counterweight is to feed the model characteristic material, not disease labels.

Privacy and consent gaps. AI repertorization sends clinical material to inference providers. The floor is separate AI-processing consent captured before use, named subprocessors, zero-retention terms, and no training on submitted patient data. Similia documents all four — OpenAI for notes and photo analysis and Deepgram for audio under signed Business Associate Agreements with zero-retention processing, plus the audio recording itself never being saved. A practitioner pasting case notes into a general-purpose chatbot has none of these protections.

What AI repertorization does not change

No AI layer touches the core method. Rubric proposals are retrieval and classification over the tradition's existing literature; symptom-selection judgement, hierarchy weighting, and the materia-medica confirmation read remain entirely the practitioner's domain.

The economics are usage-based. AI features consume metered credits on top of subscription: Similia grants 100 credits per Pro renewal, deducts by token usage for text and image calls and by transcription minutes for live audio, and sells additional packs from $14.99 USD for 200 credits. Heavy live-audio use should be modelled in minutes, not months.

The workflow gatekeeper is still the practitioner. Every shipping implementation routes proposals through accept-switch-reject controls; an AI layer that auto-committed rubrics without review would be a defect, not a feature.

Adoption posture for 2026

Adopt notes-to-rubrics and OCR first, where review is naturally unhurried. Trial photo analysis on visible local signs where the extraction task is narrow. Treat live audio as a beta-grade parallel notetaker with mandatory transcript review. Refuse any AI feature — from any vendor — that cannot document consent capture, subprocessors, and retention terms in writing. A free trial makes the cost of this caution zero, and a clinician can run an anonymised case through notes, photo, and live-audio pipelines on Similia Pro with the review discipline above before committing.

AI repertorization in 2026 is a competent proposal engine wrapped around an unchanged clinical method — valuable where its proposals are reviewed, hazardous where they are trusted.

References

Hahnemann, S. (1842) Organon of the Medical Art, sixth edition, §153; translation by W. B. O'Reilly (1996), Redmond: Birdcage Books.

Similia (2026) "Using AI: Case Notes to Rubrics", Similia Help Centre, https://similia.crisp.help/en/article/using-ai-case-notes-to-rubrics-r3ldji/, fetched 2026-05-17.

Similia (2026) "Using AI: Photos to Rubrics", Similia Help Centre, https://similia.crisp.help/en/article/using-ai-photos-to-rubrics-wxscaq/, fetched 2026-05-17.

Similia (2026) "Using AI: Live Audio Mode (Beta)", Similia Help Centre, https://similia.crisp.help/en/article/using-ai-live-audio-mode-beta-1lvfreq/, fetched 2026-05-17.

Similia (2026) "Is my patient data secure?", Similia Help Centre, https://similia.crisp.help/en/article/is-my-patient-data-secure-sxmdfd/, fetched 2026-05-17.

Similia (2026) "Similia Pricing and Subscription FAQ" (AI credits section), Similia Help Centre, https://similia.crisp.help/en/article/similia-pricing-and-subscription-faq-jizl70/, fetched 2026-05-17.