From PubMed to a brief you can audit.
Curata is an AI-assisted orientation service for new clinical literature. Its workflow is designed to keep the source visible, separate selection from writing, and stop synthesis sections that have not cleared the current verification gate.
Last updated July 2026
At a glance
A staged pipeline, not one prompt
Search, triage, classification, ranking, writing, and verification are separate stages. Different checks answer different questions, and deterministic code handles the parts that should not be left to a language model.
Search a defined PubMed topic
Each topic has a maintained PubMed query. For actively followed topics, a scheduled monthly run searches records entered in PubMed during the preceding 30 days, requests results in publication-date order, and fetches the available citation and abstract data. The current retrieval cap is 120 records per topic and run.
Triage for clinical relevance
An Anthropic model scores the retrieved papers for relevance to the topic's clinical audience. Protocols, animal or in-vitro work, narrow methodology papers, and editorials are intentionally deprioritised. Up to 36 candidates continue to detailed classification.
Label the evidence, then rank in code
The classification stage is instructed to assign one label from a fixed study-design list, a 1–5 quality tier, and a 0–100 clinical-relevance score. The final order is then calculated in code: 60% study-design weight and 40% relevance, with publication year and relevance used as tie-breakers. Papers below the relevance floor of 40 are not carded, and an edition contains at most 12 cards.
Write one abstract-grounded signal per card
The card's clinical signal is generated from the PubMed abstract supplied to the model, not from an assumed full text. PMID values are checked against the candidate set; duplicates and unknown identifiers are dropped. Citation details and the PubMed link remain attached to the card.
Build a source-marked synthesis
Selected titles, abstracts, and signals are used to write short topic sections. Every clinical claim must carry an inline marker, and every marker maps to one PMID in that edition. Code strips markers tied to out-of-edition PMIDs, removes unused source declarations, and fails sections that contain undeclared markers or no valid cited source.
Verify with code and an independent model family
Programmatic checks cover PMID integrity, marker completeness, and numeric marker-to-abstract binding. Google Gemini then checks the section against the supplied source abstracts for statistical semantics, unsupported claims, attribution, and qualitative marker binding. A separate Anthropic judge adjudicates two subjective flag types; if that judge is unavailable or uncertain, the flag stays in place.
Regenerate or repair, then check again
Structural or deterministic failures trigger a fresh generation rather than an in-place edit. Confirmed subjective issues can enter a narrow repair pass using the relevant source abstracts. Repaired text does not clear itself: it must be verified again.
Apply the honesty gate
A synthesis is displayed only when every section is cleared by its latest verification, or has been explicitly accepted by a human reviewer. Current verification records are bound to the exact section content; if the text changes afterward, the prior result no longer clears it. A still-flagged topic remains hidden for manual review.
Transparency
What the labels do—and do not—mean
Evidence label
A controlled description of the study design reported in the abstract, such as systematic review, randomised trial, cohort study, or case series.
Quality tier
A simple 1–5 hierarchy derived mainly from study design. It is not a full critical appraisal, risk-of-bias assessment, certainty rating, or GRADE judgment.
Relevance score
An automated estimate of usefulness to a practising clinician in the topic's audience. It is one input to selection, not a claim about scientific truth.
Considered studies
The pipeline retains the selection funnel: how many records matched, how many were classified, and which retrieved studies were not selected for cards.
Cadence
What “monthly” means in practice
The scheduled curation run begins on the first day of each month in the Europe/Zurich time zone. It uses a rolling 30-day PubMed entry-date window, so it is not a claim that every paper carries a publication date in the same calendar month. A first edition may also be requested when a new account chooses a topic; already-prepared topic editions can be reused across readers.
A topic can produce fewer than 12 cards—or none—when the search is quiet or candidates do not clear the relevance and validation rules. Curata does not fill an edition merely to meet a quota.
Limitations
What this process cannot guarantee
- PubMed is the search source. Papers outside PubMed, delayed indexing, imperfect topic queries, and records beyond the retrieval cap can be missed.
- Most generated copy is abstract-grounded. An abstract can omit methods, subgroup detail, caveats, adverse events, and results that appear in the full paper.
- Automation can still be wrong. Independent checks and fail-safe gates reduce known error classes; they do not make model output infallible.
- This is not a systematic review. Curata does not claim exhaustive recall, duplicate screening by human reviewers, formal risk-of-bias appraisal, or guideline-grade certainty assessment.
- Clinical context still matters. A short signal cannot account for an individual patient, local guidance, contraindications, or the complete evidence base.