AI rubric mapping, explained
Rubric mapping is the unglamorous middle step of any case: you have a symptom in the patient's own words, and you need to find the corresponding entry in a repertory whose phrasing was set in the nineteenth century by Kent, Boenninghausen and Hering. "AI rubric mapping" is the 2026 term for the cluster of statistical-retrieval techniques that compress that step — semantic search over a rubric corpus, large-language-model paraphrasing of patient speech, and ranked-list suggestion inside a case-taking interface. None of these techniques selects a remedy. Each one changes only the lookup.
What "AI rubric mapping" actually means in 2026
The term resolves to two distinct techniques that are increasingly bundled behind a single button.
Semantic search embeds each rubric in the repertory into a high-dimensional vector space using a sentence-transformer model, embeds your query into the same space, and returns nearest-neighbour rubrics with a similarity score. Similia ships this pattern on the Free plan over its classic repertory.
LLM paraphrase asks a large language model to rewrite the patient's natural-language phrasing into a candidate rubric phrase that more closely matches the repertory's idiom, then runs a conventional lookup against that phrase. Most modern stacks chain the two: an LLM rewrites the query, semantic search retrieves a candidate set, and a learned re-ranker orders the result before display.
Older expert-system implementations relied on hand-curated synonym tables and Boolean operators. The 2026 statistical approach differs in that the mapping is learned from data rather than authored by a clinician — this is the source both of its recall gains and of its failure modes. If you learned repertorisation on RADAR or MacRepertory you will recognise the user-facing surface (a rubric list, a remedy chart), but the retrieval underneath is no longer rule-based.
How rubric mapping fits into the case
Repertorisation is the third step of the classical case method as taught after Kent: receive the symptom, evaluate its hierarchy (mental-general, physical-general, particular, modality), and locate the rubric that matches at the right level of generality. Automated rubric mapping operates inside that third step and does not displace the first two. You still have to decide whether "anxiety before the patient's wedding" is a mental-general anxiety rubric, a circumstantial situational rubric, or a particular cardiac-region rubric — the AI returns candidates at several levels and lets you pick.
| Mapping approach | Input | What is computed | Failure mode |
|---|---|---|---|
| Classical Boolean lookup | Hand-typed rubric phrase | Exact-string match against the repertory index | Misses any variant phrasing |
| Synonym-table expansion | Practitioner phrase plus curated synonyms | Boolean OR over a fixed synonym list | Limited to the synonym author's vocabulary |
| Semantic search | Practitioner phrase | Vector-cosine nearest neighbours over rubric embeddings | Returns plausible-but-wrong rubrics for rare symptoms |
| LLM-paraphrased lookup | Patient's verbatim words | LLM rewrites to repertory idiom, then semantic search | LLM may hallucinate a rubric that does not exist in the repertory |
| Hybrid (rewrite plus semantic plus reranker) | Patient's verbatim words | Two-stage retrieval with a learned reranker | Reranker bias concentrates around training-set common rubrics |
What rubric mapping does and does not do
The technique compresses the lookup. It does not select a remedy, it does not weight rubrics by importance, and it does not perform anamnesis. You still have to evaluate the symptom's hierarchy, weigh the rubric against the rest of the case, and choose the simillimum from the candidate remedies. Vendors who describe AI rubric mapping as "case analysis" elide that distinction; the claim is marketing, not methodology.
The failure modes of statistical rubric mapping fall into three groups.
Plausible-but-wrong candidates. Semantic search returns a rubric whose embedding is near the query but whose clinical meaning differs — most often for rare symptoms where training data is sparse.
Hallucinated rubrics. An LLM paraphrase invents a rubric phrase that reads convincingly but does not exist in the underlying repertory. Constraining the LLM's output to the rubric index mitigates the risk but does not eliminate it.
Anchoring drift. A practitioner who scans a ranked list adopts the top candidate disproportionately often — a documented bias in the clinical decision-support literature.
Each failure mode has a corresponding mitigation — confidence thresholds, retrieval-augmented constraint, explicit "no match" outputs — and each mitigation is a question to ask the vendor.
A practitioner workflow check
Before adopting AI rubric mapping into routine case-taking, run a short calibration. Take five recent cases, hand-map their rubrics as usual, then re-map the same cases with the AI feature and compare the candidate lists. The questions worth asking are concrete: did the AI surface a rubric you missed; did it suggest a rubric you would have discarded; how often did the top-ranked candidate match your final choice; and how often did it propose a rubric whose phrasing did not appear in your repertory at all? The third number — top-1 agreement with your own mapping — is the closest available proxy for retrieval quality on your specific case profile.
If the calibration yields agreement above roughly half of cases on top-1 and useful surfacing on a further quarter, the feature is plausibly worth the subscription tier. Otherwise, you are paying for a faster path to a list you will mostly discard. You can run the five-case calibration end-to-end inside the Similia Pro trial, which unlocks the full AI pipeline for 14 days with 1,000 onboarding credits.
Open questions
Whether the top-1 agreement rate generalises across practitioner styles is unknown. Practitioners trained in the Kentian symptom-hierarchy school use a rubric vocabulary that differs systematically from those trained in the Sankaran sensation method; the LLM paraphrase step may narrow that gap by rewriting vital-sensation descriptions into more conventional repertory idiom, but no published evaluation has settled it. A second unresolved question concerns confidence display: whether showing a similarity score on each candidate worsens anchoring drift or acts as a calibrating signal you can read against.
References
Kent, J. T. (1900) Lectures on Homœopathic Philosophy, Lancaster, Examiner Printing House.
Hering, C. (1879) The Guiding Symptoms of our Materia Medica, Philadelphia, Estate of Constantine Hering.
Boenninghausen, C. von (1846) Therapeutic Pocket-Book, Münster.
Boger, C. M. (1931) A Synoptic Key of the Materia Medica, Parkersburg.
Vithoulkas, G. (1980) The Science of Homeopathy, New York, Grove Press.
Sankaran, R. (1991) The Spirit of Homœopathy, Bombay, Homœopathic Medical Publishers.
Sutton, R. T., Pincock, D., Baumgart, D. C., Sadowski, D. C., Fedorak, R. N. and Kroeker, K. I. (2020) An overview of clinical decision support systems: benefits, risks, and strategies for success, npj Digital Medicine, 3, 17.
Reimers, N. and Gurevych, I. (2019) Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks, in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing.
Topol, E. J. (2019) High-performance medicine: the convergence of human and artificial intelligence, Nature Medicine, 25, 44–56.
Similia (2026) Knowledge base — AI credits, AI processing, Subscription FAQ, https://similia.crisp.help/, fetched 2026-04-22 and 2026-05-16.
Verdict
Ready to act on this?