Product methodology
How Answerlog makes research approachable.
Answerlog is an orientation layer for curious cooks and bakers: it searches connected research indexes, turns available titles and abstracts into a cautious explanation you can read or listen to, and keeps every source record visible. Topic radar is a separate way to find what is new.
1. Research discovery
A question is sent to the Answerlog server and translated into searches of the source lanes selected for its topic. When model-assisted localization is configured, a non-English question can also be converted into a bounded English scholarly search query while the original question is preserved for display. OpenAlex supplies a broad, multidisciplinary index; Semantic Scholar supplies wide coverage of food chemistry, culinary technique, and sensory studies; PubMed supplies nutrition, food-safety, and fermentation-microbiology records; and Crossref supplies publisher-deposited, DOI-centered records from food-science journals.
The source lanes are discovery routes, not mutually exclusive academic taxonomies. The same work can appear in more than one index, and coverage, update speed, abstract availability, and metadata quality differ by provider. A missing paper is not evidence that the paper or result does not exist.
2. Normalization and safety checks
Records are normalized and deduplicated by DOI or a normalized title. Records marked as retracted by a provider are excluded when that signal is available. Answerlog preserves publication dates, venue information, open-access status, provider identity, and source links when those fields are supplied.
Preprint-style records are clearly labeled as preprints. Answerlog does not independently verify peer review for any provider. Venue, publication type, or preprint information is descriptive metadata, not a manual quality judgment; PubMed or journal indexing by itself does not prove that an item passed peer review.
3. Plain-language synthesis
The answer is limited to the titles, abstracts, and metadata returned for that question. Crossref records without a usable abstract are not used as claim-level synthesis evidence. Answerlog does not currently claim to analyze the complete full text of every work. It is designed to distinguish a cautious bottom line, supporting context, uncertainty, and practical limits.
Abstract text is used transiently to build the answer and is not displayed as a substitute for the publisher's record. Answerlog stores the resulting explanation, short abstract-derived paper summaries, language presentations, and source metadata, but does not retain newly retrieved abstract text in full in its application database.
If the operator configures OpenAI, the server may use a model to produce schema-constrained output from that evidence packet and translate the presentation into a reader-selected language. Without an OpenAI key, it uses a conservative deterministic overview and clearly labels any displayed sentence as an extractive abstract excerpt rather than an AI summary or translation. In both modes, the original paper title and linked evidence set remain visible and generated prose is not treated as a new source.
4. Evidence display
Each answer shows the selected works used, their original titles, dates, venue or source, authors when available, and a reachable external record. When a usable abstract was supplied, the card can also show a short abstract-only summary and a translated display title. These derived fields are not a review of the full paper. Counts describe the displayed source set; they are not votes and do not measure study quality, effect size, or scientific consensus.
Paper filters and comparison views operate only on metadata that the connected services supplied. Publication date, source, preprint status, open-access status, and citation counts can be incomplete or uneven across providers. A “journal record” label describes the record type; it is not a quality score. Answerlog does not claim to assess agreement between studies unless a future workflow performs that analysis explicitly.
5. Saved papers and collections
Readers can save paper metadata to a device-scoped library, organize papers into collections, record a reading status, and add a private note. Saving a paper does not save or redistribute its full text or abstract. Citation exports are generated from the available metadata and should be checked before formal use.
6. Versions and updates
Answerlog is request-driven. Opening or rechecking a question can request a refresh after the saved answer's freshness window; following alone does not schedule research checks. When a fresh provider check succeeds, Answerlog stores a dated version and compares the short answer, explanation, evidence strength, and limitations with the previous version. A source set can change while the answer remains classified as having no material change.
Opening Home requests a separate shared food-and-baking feed. Answerlog searches focused baking and cooking terms, deduplicates the returned records, and groups them by the UTC calendar date used by each provider's latest-record query: an index or registration date when that provider supplies one, otherwise a publication date. It shows today's usable records when available and otherwise yesterday's. This is a bounded discovery feed from connected indexes, not a complete list of every paper published worldwide. A short server cache limits repeated calls to public research services, and an older snapshot is labeled as the latest available fallback if every provider is temporarily unavailable.
A saved topic radar uses a separate latest-paper search. The first successful check establishes a baseline, so older records are not mislabeled as new. Later checks compare stable provider and DOI/title identifiers with that baseline and report records first discovered after the reader's saved “seen through” point. Marking a radar read advances only to the displayed check snapshot, so papers arriving afterward remain unread. Radar results are discovery leads, not a relevance judgment or a newly synthesized answer; readers should open the linked abstract or paper before relying on one.
“Last checked” means the service successfully searched at least one requested provider at that time; it is not the newest paper's publication date. A recent saved response is labeled as cached. If all requested providers are temporarily unavailable, an answer from the last seven days can be returned and is labeled as a saved fallback rather than a fresh check; older answers are not used for that fallback.
If a followed answer materially changes after a reader-initiated refresh, Answerlog can show a browser notice while the app is open and browser permission is enabled. This beta has no background scheduler, service-worker push subscription, or email delivery.
7. Listening to an answer
Listen controls use speech synthesis supplied by the reader's browser or operating system. The voice is a spoken rendering of the on-screen answer, not a separate podcast or additional analysis. Voice availability and pronunciation vary by device and language.
Known limitations
- Abstract-only analysis can miss methods, caveats, subgroup effects, and corrections found in full text or supplements.
- Search relevance, provider overlap, indexing delay, language coverage, and abstract availability can bias the retrieved set.
- Metadata can be incomplete or wrong, including venue and retraction fields.
- Neither deterministic nor model-assisted synthesis replaces domain-expert appraisal.
- The service does not currently register protocols, assess publication bias, or apply formal risk-of-bias tools.
Corrections
If a citation is broken, a record is misclassified, or an explanation overstates the evidence, use the in-product feedback control. Feedback is stored for operator review; it does not automatically revise an answer, and a response is not guaranteed. Any accepted correction should appear in a new answer version instead of silently rewriting history.