Trust & methodology

From a question to an answer

LYVRO routes a question to a bounded selection of official providers, returning at most 20 deduplicated sources. PubMed / NCBI and Europe PMC supply literature metadata and available abstracts. This is not an exhaustive search or a systematic review. LYVRO synthesizes supplied evidence using an external AI processing service. Conversational research can include bounded passages from openly licensed PMC articles through Europe PMC, FDA label sections and ClinicalTrials.gov structured records. Coverage reflects selected passages actually supplied, not a complete full-document review. Unavailable or unlicensed full text remains abstract-only; LYVRO does not bypass paywalls.

Different source types, different roles

ClinicalTrials.gov records are sponsor-submitted trial registrations, not proof of efficacy. FDA drug labeling, Drugs@FDA, Orange Book and recall records retain their specific dataset attribution. They describe different regulatory purposes and are not interchangeable with clinical studies. Structured official records can support explicitly attributed facts about their own content, with passage references. They cannot establish comparative clinical efficacy or recommendations. Preprints and FAERS remain excluded from synthesis.

FAERS contains spontaneous adverse-event reports. These reports do not establish causation or incidence: duplicates, incomplete reporting and reporting biases exist. LYVRO excludes patient demographics and narratives from its normalized records and excludes FAERS entirely from synthesis. WHO population data remains pending verification of a stable replacement API contract; no WHO site scraping occurs.

Trace claims to records

Structured claims refer to source IDs. LYVRO checks identifier formats, source references, duplicate records, provenance requirements and publication-status constraints before accepting a result. Source metadata comes from retrieval, not from the model. Click a compact citation to open the source inspector and inspect available sections, limitations and provenance.

A valid reference does not prove that a source supports a particular interpretation. No automated check establishes medical truth, patient applicability or scientific correctness. Formal certainty grading is not performed; source counts and limitations are available on demand, not numerical confidence.

Publication integrity and updates

PubMed publication types and linked notices can identify retractions, corrections, expressions of concern and superseded versions. Affected records remain visible. Retracted, superseded, expression-of-concern and identifier-mismatched records are excluded from new synthesis. Corrections must be inspected; they may change the interpretation.

A bounded Crossref metadata check compares DOI and normalized title for up to five sources. Failures and mismatches are shown, not silently treated as verified. Crossref can expose update relationships deposited by publishers, but LYVRO does not perform exhaustive reverse-notice discovery or crawl publishers. No reported notice does not mean no notice exists. Indexing does not establish peer review.

Crossref metadata documentation · NCBI E-utilities documentation

Uncertainty, disagreements and freshness

Disagreements are AI-identified accounts with links to the relevant retrieved sources and explicit limitations. Different populations, endpoints or dates may explain a difference; they do not necessarily establish contradiction. No identified disagreement does not establish consensus.

Publication dates, retrieval timestamps, metadata checks, hashes and available update relationships are visible. Saved results are snapshots and are not continuously rechecked. Older evidence is not automatically invalid. Check current guidelines and original notices before acting on time-sensitive evidence. Missing or insufficient results do not mean a proposition is false; provider outages are reported separately.

Privacy and audit boundaries

Do not enter identifying patient information. LYVRO sends queries to the selected official retrieval providers and to an external AI processing service with retrieved evidence for synthesis; only DOI metadata goes to Crossref. API keys stay on the server. Completed research and follow-up conversations share one private evidence corpus in authenticated History. Follow-ups read the existing corpus first and run a bounded provider search when the question asks for fresh or uncovered evidence. Opening History restores stored content without rerunning synthesis; saved sources and workspace notes use Firebase Data Connect. Access requires account ownership. API responses and private data are not cached by the service worker; offline mode only provides a static screen.

The answer audit panel records process facts such as retrieval and validation. It does not expose hidden reasoning, prompts or credentials. Evidence concern drafts download locally: a submission and review backend is not connected yet.

Your review remains essential

Read the original evidence, assess methods and applicability, and use clinical judgment. LYVRO makes no claim of regulatory approval, medical certification or clinical validation. Automated citation checks are safeguards, not a guarantee of safety or correctness.