How PsychHypo works.

PsychHypo retrieves papers from two literature databases simultaneously: OpenAlex (https://openalex.org) and Europe PMC (https://europepmc.org).

OpenAlex is a comprehensive open-access bibliographic database with over 250 million scholarly works across all disciplines, including extensive coverage of psychology and psychiatry journals not always indexed in PubMed (such as APA journals, social psychology, developmental psychology, and educational psychology). We chose OpenAlex because it provides broad psychological science coverage with a clean, free, well-maintained API.

Europe PMC indexes peer-reviewed biomedical journals, PubMed Central, and preprints. It provides strong coverage of biomedical psychiatry, neuroscience, clinical psychology, and translational research.

For each query, PsychHypo searches both sources in parallel, deduplicates results by DOI and title, and ranks them by citation count and recency before sending the top 15 papers to the hypothesis engine. This hybrid approach captures both biomedical and broader psychological science literature.

Given a user query, the engine builds a structured search across OpenAlex and Europe PMC, retrieves the most relevant recent papers, and passes them to a large reasoning model with a deliberate prompt structure.

The model is instructed to produce falsifiable hypotheses, supporting citations drawn only from the retrieved set, an experimental design appropriate to the construct and population, a confidence rating, and the single most important risk or gap.

PsychHypo runs on Claude Opus 4.7 — the strongest currently available reasoning model — chosen for its capacity to hold complex methodological context across long inputs.

PsychHypo is not a literature search tool. If your goal is exhaustive coverage of a topic, use OpenAlex or Europe PMC directly. The engine retrieves only what it needs to ground its hypotheses.

PsychHypo is not a fact-checking tool. Cited papers are real and matched by retrieval, but the engine's interpretation of those papers is generative. Read the citations yourself before building a study around them.

PsychHypo is not a replacement for primary literature. It accelerates hypothesis generation; it does not replace careful reading.

Sparse-literature queries: niche topics with few indexed papers will produce weak hypotheses or no output. The engine will refuse to generate rather than fabricate.

Hyphenated and compound terms: search syntax can break on unusual punctuation. Phrase your query in natural language.

Newly emerging subfields: if the literature is younger than the engine's training cutoff for context but newer than current indexing, retrieval may underweight the most cutting-edge work.

Confidence ratings: HIGH / MEDIUM / LOW reflect the engine's internal assessment of evidence alignment, not the probability the hypothesis is true. Treat them as a starting point, not a verdict.

Over time, PsychHypo tracks which hypotheses generated by the engine were tested by users and how those experiments resolved. This longitudinal data — connecting hypothesis generation to real research outcomes — is unique to PsychHypo. As we accumulate it, the engine's confidence calibration will become empirically grounded, not just based on the strength of literature evidence.

You can see your own outcome data on the /outcomes page. Logging status updates (especially terminal outcomes like 'confirmed', 'falsified', 'mixed', or 'abandoned') contributes to this longitudinal record.

This is a long-term project. We are at the beginning of it.

Your queries, uploads, and saved hypotheses are private. We never share your data with third parties. We never use your uploads to train AI models. We never expose your work to other users.

You own everything you create. Full deletion is supported at any time from your profile.