1. Information we receive
- Submission content: pasted or locally extracted legal text, citations, quotations, research questions, selected jurisdiction, opaque document identifiers, file types and sizes, page or location ranges, and any facts or names contained in the text.
- Generated research data: extracted citations, source matches, candidate cases, snippets, source excerpts, rankings, audit labels, scores, and packet selections held in your browser during the page session.
- Technical and usage data: IP address, request time, route, browser and device information, response status, security signals, and service diagnostics that hosting or network providers ordinarily process.
This beta has no user accounts, payment flow, advertising tracker, or application document database. The site does not intentionally log request bodies. A process-local abuse limit temporarily holds a truncated hash derived from an IP address and a request count for approximately ten minutes; individual server instances may end sooner.
Temporary queue metadata. Our shared Upstash queue holds random job identifiers, job type, ordering and timing information, lease and cancellation status, and an authentication secret for browser-held progress. It does not hold your documents, filenames, research question, extracted claims, provider responses, or results. Tickets are valid for at most 24 hours; completed and cancelled tickets become eligible for cleanup after ten minutes. Inactive waiting tickets are invalidated after approximately three minutes when the queue is next checked. Expired records are removed during subsequent queue activity; the shared metadata store also expires after 25 hours without activity. These validity periods are not promises of immediate physical deletion. Shared provider-quota counters remain while the queue is in use, and provider logs or backups may follow the provider's retention practices.
2. How a research run is processed
Document intake. When you select a supported file, locally served code reads it in your browser and extracts text before submission. For a PDF page with a missing or suspicious text layer, the browser may render that page into temporary memory and run a locally served Tesseract.js OCR engine and English language model. The original binary, rendered page image, and filename are not intentionally uploaded or sent to an OCR provider. The application sends the extracted or OCR-recovered text, an opaque document identifier, file type and size, and available page or character ranges to our Vercel-hosted function. Your browser uses the opaque identifier to display the filename locally with returned findings. Removing a file before starting a run excludes it from that submission.
Citation audit. Our Vercel-hosted function receives the extracted or pasted submission and detects citations across it. Depending on the available retrieval path, CourtListener may receive bounded text containing citations or individual citation searches; historical citation details may be sent to the Caselaw Access Project; and the official Florida Courts opinion archive may receive a bounded Florida case-name or docket search. TypeSafe receives isolated propositions and context, case metadata, and retrieved opinion evidence needed for Jev to classify support and attribution. The workflow attempts eligible claim-bearing citations within operational limits; missing sources, interrupted work, and provider limits can prevent a completed model audit. Each result identifies its status.
Authority research. CourtListener receives a bounded version of your research question for meaning-based search, focused searches derived from its issues and case names, and may receive related-case identifiers. Each search retains the selected court filter; when no specific jurisdiction is selected, an unambiguous state named in your request may supply that filter. TypeSafe receives the research query and candidate metadata/snippets for relevance screening. Candidates proposed for the results also undergo an automatic source-audit attempt, for which TypeSafe receives retrieved opinion evidence and related metadata. Missing sources or provider limits can prevent a completed audit; each result identifies its status.
Broader-web citation checks. For unresolved or conflicting case identities, TinyFish Search may receive the extracted case name and reporter citation. This path does not send the uploaded file, its filename, the full submission, or the claim being evaluated. Case names can still identify people or organizations mentioned in a submission. Search queries and unverified result links may be displayed with the audit; these results do not by themselves verify a case or establish fabrication. Repeated references to the same case share a search result within a run.
Queue and automatic continuation. While waiting, your submission and a signed progress checkpoint remain in this browser tab's memory, not in a server-side content database. The checkpoint can contain retrieved opinions and model responses. When the next processing slot becomes available, the browser automatically resubmits the original text and checkpoint over HTTPS so completed provider work can be reused. The signature detects modification; it is not encryption. Keep the tab open. Cancelling, closing, or reloading the tab discards browser-held progress; abandoned queue tickets expire. No background completion or recovery after closing the tab is promised.
Results return to your browser over HTTPS. The “Copy packet as Markdown” control writes only the authorities you select to your device clipboard when you click it. Packet selections are not intentionally stored by the application after the page session.
3. Why we process information
We process information to provide requested research, retrieve sources, rank and audit candidates, show progress and counts, prevent abuse, secure and troubleshoot the Service, comply with law, and enforce the Terms. We do not sell submission content or use it for advertising. LegalJev does not keep a training dataset or train or fine-tune models with your submissions. Processing by our AI and search providers is governed by their policies and our applicable agreements; this is not a guarantee about every provider’s independent practices.
4. Service providers and legal-data sources
- Vercel hosts the site and functions and processes network, request, security, and operational information. When Jev is routed through Vercel AI Gateway, Vercel also receives the model inputs and responses described above before forwarding them to or returning them from TypeSafe. Vercel Web Analytics separately processes the limited public-page traffic information described below.
- Upstash stores temporary scheduling and authentication metadata for the shared queue, not submission content or research results.
- TypeSafe AI provides Jev model processing. TypeSafe’s own policy and our applicable agreement govern its processing and retention; review those policies before submitting information.
- CourtListener / Free Law Project provides case search, metadata, citation resolution, and opinion passages.
- Caselaw Access Project may provide bounded historical case and opinion evidence.
- Florida Courts may receive a bounded case-name or docket query and provide an official Florida DCA opinion PDF when the general providers do not supply one.
- TinyFish provides broader-web discovery for public case-name/citation queries when case identities cannot be resolved or conflict. Search leads are not automatically accepted as opinion evidence.
Linked courts, government sites, and other sources have their own practices. Providers may keep technical logs or input under their policies, contracts, security needs, and legal obligations. Do not treat this Notice as a promise that third parties retain no data.
5. Retention
LegalJev does not keep saved copies of your documents, pasted text, research questions, audit results, or generated packets. There is no application content database or saved account history, and the app does not intentionally log submitted text. We cannot restore previous work from a saved history. Copies you make, such as clipboard exports, remain under your control.
Automatic continuation keeps a temporary checkpoint only in browser memory and in server memory during each processing batch. It is not written to browser local storage or an application content database. The separate scheduling metadata described above is retained temporarily by the shared queue.
This no-saved-content design does not eliminate temporary processing or technical records. Original selected-file data and extracted text may exist in your browser memory during the page session; submitted text and metadata exist in server and provider memory while a request runs and are returned to your browser. Hosting and provider logs, backups, security records, and legally required records may remain under each provider’s retention practices. We keep any operational information we control only as long as reasonably needed for the purposes above, legal compliance, disputes, and security.
6. Cookies, analytics, and browser signals
This beta does not set application cookies or use advertising trackers. We use Vercel Web Analytics for aggregate visits to the homepage, Terms, and Privacy pages. Vercel may process visit timestamps, referring sites, browser/device information, approximate location, and a short-lived visitor hash. Vercel describes this service as cookieless; traffic statistics are separate from the no-saved-submission design described above.
We strip query strings and fragments from tracked page URLs and do not send custom analytics events, uploaded documents, filenames, pasted text, research questions, or audit results. Analytics is disabled on local and preview deployments and when the browser sends Do Not Track (DNT: 1) or Global Privacy Control. Content blockers may also prevent collection. Because no sale or cross-context behavioral advertising occurs, there is no sale/share opt-out flow.
7. Security and confidentiality limits
We use HTTPS and scoped server-side credentials, keep provider API keys out of browser code, and parse supported files locally before sending extracted text. Browser-local OCR can still produce materially incorrect text, including incorrect names, quotations, reporter citations, and pinpoint pages. Local extraction does not make the submitted text confidential. No parser, transmission, storage, provider, or security measure is perfect. We cannot guarantee complete extraction, confidentiality, security, deletion, privilege, or that unauthorized access will never occur. The safest control is not to select or submit sensitive material to this public beta.
8. Your choices and requests
You choose whether and what to submit. You can omit names and identifying facts, ask the browser to stop a running request, close the page, or avoid the Service. Cancellation cannot recall content already sent to a provider, and in-flight provider work may finish. Because the beta has no accounts or application content store, we generally cannot retrieve a past submission for you. For access, correction, deletion, or privacy questions about operational information we control, contact the service administrator through the channel by which you received access. We may need to verify your request and may retain information where law permits or requires. You may also contact a listed provider about data it controls.
9. Children and international use
The Service is not directed to children and is limited by the Terms to users 18 or older. Do not submit information about a child. The Service and providers may process information in the United States and other locations where they operate; those laws may differ from the laws where you live.
10. Legal disclosures and changes
We may preserve or disclose information when reasonably necessary to comply with law, legal process, or valid government requests; protect rights, safety, and service integrity; investigate abuse; or support a merger, financing, reorganization, or service transfer subject to applicable safeguards. We will post revisions here with a new effective date. Material changes affecting a research run will be paired with updated notice or renewed acknowledgment where appropriate.