Methodology
How Symmathy minimizes AI hallucinations
AI summaries of medical literature can be confidently wrong. That's the failure mode we're built to avoid. Here's the stack of techniques Symmathy uses — and the techniques we deliberately don't claim to use.
Cite this methodology
Symmathy (2026). Symmathy Search & Ranking Methodology (v1). Zenodo. https://doi.org/10.5281/zenodo.20579117
Concept DOI (all versions): 10.5281/zenodo.20579117 · This version (v1): 10.5281/zenodo.20579118
1. Retrieval-Augmented Generation (RAG) with hard grounding
The model never answers from its own training data.
Every Scan, Compare, Discover, and Ask response is generated against text we just retrieved — either the PDF you uploaded (with page markers) or abstracts pulled live from our 10 source databases. If the retrieved text doesn't contain the answer, the model is instructed to say "Not stated in this paper" rather than guess.
2. Page-level citation on every factual claim
Inline page numbers like "HR 0.79 (p.4)".
Click any page badge in a Scan summary and the extracted source text from that page expands inline. You can confirm the AI's claim against the verbatim text without opening the PDF in a separate viewer.
3. One-click verification
A second model audits any claim on demand.
The "Verify" button on every key finding runs an independent pass that re-reads the source and returns one of four verdicts: supported, partially supported, contradicted, or not in paper — each with a verbatim evidence quote.
4. AI vs Extracted badges
We visually distinguish interpretation from extraction.
Every block carries either an Extracted badge (verbatim or near-verbatim from the paper, with citation) or an AI interpretation badge (model's synthesis, may smooth over nuance). You always know which is which.
5. Cross-source consensus
Discover runs a 2-pass cross-check (when enabled): the same corpus is analyzed by two differently-prompted reasoning passes and only claims that appear in both passes are kept. This filters out single-pass hallucinations.
6. Source attribution at the corpus level
Symmathy doesn't have a single training-data blob it might leak from. Every external abstract carries its source database (PubMed, Europe PMC, Cochrane, ClinicalTrials.gov, J-STAGE, LILACS, DOAJ, HAL, OpenAlex, or CrossRef), original venue, year, and a working URL back to the publisher. You can always click through and read the original.
7. Coverage & the paywall question
Symmathy queries open metadata indexes that contain the abstracts of essentially every indexed paper — including paywalled ones in NEJM, JAMA, Lancet, BMJ, Nature, Cell, etc. The abstract is what we feed the model in Discover and Search, so contradictions, effect sizes, and headline conclusions from paywalled trials are visible.
For full text, we automatically check Unpaywall and PubMed Central on every result. When a legal open-access copy exists (author manuscript, institutional repo, OA preprint), a Free full text badge links straight to it.
What we deliberately don't do: pay aggregators or scrape behind paywalls. If the paper you need is paywalled and has no OA version, download the PDF through your institutional access and upload it to Scan — the full-text analysis pipeline runs identically on user-supplied PDFs.
What we don't do (yet)
Verbatim-quote enforcement — we don't yet string-match every AI quote against the source to reject any quote that doesn't appear literally. It's on the roadmap.
Auto-verify-everything — verification today is one click away on every claim, but it's not run automatically before display. Unverified claims are not visually flagged.
Multi-model consensus on every output — only Discover runs a second pass; Scan, Compare, and Ask currently use a single model call.
The honest bottom line
No tool that uses LLMs can guarantee zero hallucinations. We engineer the failure rate down with retrieval, citation, verification, and visual honesty about confidence. The final check is always you, reading the cited source. That's why every finding has a page number you can click.