YOUR RESEARCH · YOUR LOCAL LIBRARY

Privacy and limits.

Your board library lives in your browser. Research requests use external services. Here is what gets stored, what gets sent, and where this beta stops.

Stored in this browser

Boards, evidence, immutable revisions and local review notes use IndexedDB. There is no account, cloud backup or cross-device sync. A fresh browser or device has an empty library. Browser storage is not encrypted by Tide. A small localStorage journal stores only board IDs and expected revisions so a confirmed deletion can finish after an interrupted page reload.

Sent for research

Your question, requirements, selected public URLs and necessary page excerpts go to this site's server. Telem searches, routes and qualifies the public pages. The selected public URLs (never your question or requirements) are also sent to Tavily Extract, whose page text is the quoted evidence. The configured model provider extracts facts and drafts decisions. Not all data stays on-device.

Save, export, clear or delete

Successful operations save automatically. The workspace shows storage errors and offers a recovery download when a write fails. Concurrent tabs use expected local revisions, so an old tab cannot silently overwrite a newer board. Use Reload saved to reconcile.

Export a JSON archive before clearing site data. Clearing browser data removes all local boards. Delete board asks for confirmation and removes that board, its full local revision records and local review notes. Downloaded exports, copies in another browser and the agent vault are separate; deleting here does not remove them.

All successfully saved revisions remain locally until you delete the board. A bounded archive includes up to the latest 20 full revision records, with a 1 MiB maximum. Earlier local records are retained. A current-only recovery download omits prior full revision records. Keep exports private: they contain your question, source excerpts and observed answers.

Historical evidence is not a fresh check

Imported archives are unverified historical records, even if they contain a signed snapshot. Imports may have no snapshot at all. A live rebuild fetches permitted sources and creates a new board while preserving the old archive. Server signatures protect observation integrity; they do not encrypt data or establish truth.

A passage shows what a page said when observed. Cached, failed, partial or contradictory reads may leave freshness unknown. Model-inferred dependencies can be incomplete. Review notes are local annotations; to correct the computed graph, describe the correction in requirements and rebuild. A changed requirement is your input, not evidence of a remote page change.

Direct Tavily extraction uses basic depth, one attempt and at most five selected URLs. The evidence view shows returned provider request IDs, batch credits when reported, partial content and errors. A reported zero is an observation, not a promise of free usage. If Tavily fails, any Telem text used instead is labeled degraded. Both services reading one URL still represent one source. Observation time does not establish page publication or provider cache age; upstream cache freshness and complete provider-side redirect chains cannot be independently checked.

Supported public sources

Only HTTPS public text pages on these exact operator-approved hosts are accepted. Each research mode has its own list:

General

www.federalreserve.govwww.newyorkfed.orgfred.stlouisfed.orgwww.stlouisfed.orgwww.bls.govwww.bea.govwww.census.govwww.nber.orgwww.imf.orgwww.worldbank.orgwww.oecd.orgwww.bis.orgwww.ecb.europa.euwww.gold.orgen.wikipedia.orgwww.britannica.comapnews.comwww.bbc.comwww.npr.orgwww.who.intwww.cdc.govwww.nih.govpubmed.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.govwww.nature.comarxiv.orgourworldindata.orgwww.pewresearch.orgwww.gallup.comwww.brookings.edunews.stanford.edu

Technical

docs.telem.aideveloper.mozilla.orgdocs.python.orgdocs.github.comnodejs.orgnextjs.orgreact.devvercel.comdocs.netlify.comdevelopers.cloudflare.comwww.postgresql.orgredis.iosupabase.comupstash.com

Finance & Economics

www.federalreserve.govwww.newyorkfed.orgfred.stlouisfed.orgwww.stlouisfed.orgwww.bls.govwww.bea.govwww.census.govwww.nber.orgwww.imf.orgwww.worldbank.orgwww.oecd.orgwww.bis.orgwww.ecb.europa.euwww.gold.orgen.wikipedia.orgwww.britannica.comapnews.comwww.bbc.comwww.npr.org

Health & Science

www.who.intwww.cdc.govwww.nih.govpubmed.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.govwww.nature.comarxiv.orgourworldindata.orgen.wikipedia.orgwww.britannica.comapnews.comwww.bbc.comwww.npr.org

Work & Society

www.pewresearch.orgwww.gallup.comwww.brookings.edunews.stanford.eduwww.nber.orgwww.bls.govwww.oecd.orgourworldindata.orgen.wikipedia.orgwww.britannica.comapnews.comwww.bbc.comwww.npr.org

Private documents, logins, paywalled pages, signed URLs, PDF/video uploads, arbitrary internet crawling and medical, legal or financial advice are outside this beta. We do not silently follow outbound links. Discovery returns candidates for you to choose; a search snippet alone cannot support a fact.

Bounded public usage

Each admitted research operation allows at most 5 source pages, 2 search queries with at most 2 providers, 6 model attempts and 45 seconds. A board holds at most 10 facts and 6 decisions. Failed attempts count. Creating, discovering and refreshing each consume shared allowance; remaining allowance is not currently displayed.

This deployment admits at most 40 operations per signed anonymous visitor per day, 12 per minute, and 150 per day globally. These are operation limits, not a dollar billing guarantee. A quota response shows Retry-After when provided. Browser requests are never automatically retried. The Telem adapter may retry a transient failure once within the same operation deadline; direct Tavily extraction is attempted once.

A signed anonymous HttpOnly, SameSite cookie identifies the visitor quota. Production cookies use Secure. Shared Redis stores expiring pseudonymous quota counters, not your boards or source text. If the quota service is unavailable or public access is disabled, paid operations fail without a simulated substitute. Clearing cookies does not remove the global cap.

Logs, agents and the Simulation Lab

Application logs are intended to contain request IDs, coded errors and bounded timings, without raw questions, source documents, credentials or provider errors. No analytics or marketing tracking is added by this application.

The local agent connector keeps a separate file vault. Browser and agent libraries share only through explicit JSON export/import. Host permissions remain in effect. Native client compatibility requires separate verification.

Simulation Lab is finite authored teaching data, visibly labeled synthetic. It runs entirely in the client, makes no live-model or source-fetch calls, and never stands in for failed live research. There is no background polling or scheduled monitoring; real refresh happens only when you ask.