Privacy Policy

Form setup and host security

Focusing an inquiry field requests a form token. This setup sends no contact fields or prepared notes.

The host may ask the browser to set a temporary security cookie before setup can continue. This cookie is separate from website counts.

The form reads only the host cookie instruction, then retries setup once. It does not run the returned script.

The cookie carries no inquiry text. Form setup tokens and cookie values are excluded from website counts.

If setup fails, your fields stay on the page. Copying the inquiry uses this device and does not send an email.

Business inquiries and optional notes.

The short form sends your business name, work email, decision category, deadline and general description to LLM Squared only when you choose Send inquiry and confirm consent. There is no document-upload field. Prepared notes are attached only when you separately select that option. We use the request to reply about scope; submitting it creates neither a booking nor an engagement.

The website uses TLS in transit. Bluehost processes the request and mail transport; our business email provider receives the team message and visitor acknowledgment. The website keeps a private retry record outside the public document root. These records become eligible for deletion after 90 days and are removed on the next inquiry or measurement cleanup. Email copies, provider logs and backups have separate retention. Ask info@llmsquared.com to access, correct or delete a request, quoting its reference. Do not email confidential sources.

The form sets an essential, session-only, Secure, HttpOnly, SameSite=Strict cookie for request validation. It is scoped to the form API, carries no note text and is unrelated to analytics. Short-lived keyed abuse counters limit repeated submissions. They use email/IP fingerprints separately from analytics, with eligible expired counter files removed on subsequent cleanup after two days.

First-party counts and your choice.

We count approved event names and fixed page paths on this server: page views, sample views, requested exports, download clicks, scope requests and accepted inquiries. These are aggregate activity counts, not unique visitors or completed sales. They contain no source text, form fields, query strings, email addresses, inquiry references, visitor IDs or cross-site identifiers. Hosting can still keep normal request logs.

No advertising pixels, session replay or third-party analytics are added. Browser counts honor Do Not Track and Global Privacy Control. Use the footer control to turn browser counts off on this device. That preference is saved locally. Server acceptance counts remain part of handling an inquiry. Aggregate files become eligible for removal after 180 days, on subsequent cleanup.

Actual booked calls, qualified opportunities, paid reviews and repeat reviews are recorded by the team from business correspondence and accepted engagements, not inferred from button clicks. A tool-run click does not prove a completed task. We do not enroll an inquiry in a marketing list.

Optional browser-agent methods.

Supported WebMCP preview browsers may expose the public sample and structured quote calculation while the corresponding page is open. The calculation uses the existing engine on this device and returns the complete supplied input and result to the invoking assistant. Use public or synthetic inputs only. Member information, account numbers, SSNs, card data, credentials and confidential contracts are refused. The adapter has no access to preparation notes or inquiry submission. Your assistant provider may process tool arguments and results under its own policies; a local calculation does not remove that exposure.

Anonymous website count choice.

Evidence Desk processing.

The local Claude evaluation plugin launches a stdio server on the computer running Claude Code. It reads no source files, opens no port, makes no network request and logs no input. The optional export helper reads only a user-selected JSON file and writes the packet to a chosen new folder. Users control retention and deletion of those files.

Claude or another approved assistant provider may process and retain conversation and tool content under that provider's settings and terms. Local calculation does not remove provider exposure. Review redaction before using any assistant.

The new remote Evidence Desk server is not operating as a verified public service. Its tested design accepts declared or redacted inputs, holds them in request memory and returns the prepared file to the calling client. The code has no input store.

Member data, account identifiers and credentials are outside this route. The heuristic screen may miss restricted information, so supply only permitted declared inputs. The assistant provider separately processes the conversation and tool response.

Application audit logging is disabled by default. If enabled, it emits tool name, UTC time and status only. Short-lived IP-derived counters support rate limiting. The approved host, platform metadata and retained-log deletion schedule must be verified before remote release.

Review limits and responsibilities are described in the method. Read the method.

Read the current server route and limits. Support: info@llmsquared.com.

Evidence Desk browser trial.

The browser trial runs the declared-input engine on your device. Entered metadata stays in page memory; it is not sent to the inquiry service or included in website counts. The page loads first-party code and starts the engine only when you prepare a file.

There is no autosave or source retrieval. Clear the page to discard in-memory inputs; closing the page also loses the prepared result. Downloads remain on your device and follow your institution's handling rules. Pattern screening does not detect all sensitive data; supply declared or redacted metadata only.

You may choose a saved Evidence Desk JSON file to reopen. The browser reads and validates it locally, then recalculates from the supplied inputs. No file is uploaded, no source is retrieved, and no approval is carried forward. Per-review drafts remain in page memory until cleared or closed.

The PowerOn project-scoping page keeps its selected workflow and readiness options in browser memory. Its acceptance example uses only the supplied illustrative cases. The JSON controls request a local download; neither control submits the packet. Existing first-party website counts exclude field values and supporting records.

The PowerOn acceptance bench compares declared selections against its illustrative fixtures and retains attempts in page memory. A changed rule queues those attempts for another review. JSON, CSV and printable HTML exports preserve the earlier inputs, comparisons and blank institution decision. Closing or clearing the page removes unsaved attempts. The optional reopen control reads one exported illustrative JSON file on this device, checks its original cases and blank review fields, and recalculates comparisons. It replaces current page memory only after you choose that separate action. File names, contents and attempts are excluded from website counts and inquiry submission.

Complete-example aggregate views

The complete example at /example/ uses the existing file_view event. The count contains only the event name and fixed canonical page route. Query strings, field values, evidence, notes and exported records are excluded.

Local-tool isolation, the website count preference and browser GPC or DNT signals stop this request. An example view is a distribution signal, not an inquiry, booked call or paid engagement.

First-party count details

Counts store a daily total for an allowed event and canonical page. Call-to-action counts may include a fixed location: hero, navigation, card, final or sticky.

Events include page views, CTA clicks, form starts, inquiry sends and downloads. Exports are counted only where page isolation permits it.

The readiness check makes no network calls during use. Its completion count records only an explicitly attached summary arriving at the inquiry page.

Counts exclude field values, source text, query strings, file contents, exported records and visitor identifiers. Counts honor GPC, DNT and the browser opt-out.

Rate limiting uses the existing restricted server-side route. Aggregate count files stay outside the public document root.

Readiness files on your device

The readiness check reads selected JSON files locally. Downloads can contain scope answers, matching record choices, preparation limits and blank review fields.

Reopening restores selections and requires a new summary. Clearing the page removes its in-memory selections. Downloaded files remain on your device.

No file, selection or export is uploaded or counted during the check. Only your explicit summary handoff can place a summary in an inquiry draft. Sending that draft requires separate consent.

Comparing an earlier readiness check reads the selected JSON file locally. Both preparations and their fixed selections stay in page memory until cleared or closed. Comparison downloads retain both checks, changed selections and blank review fields. Neither filenames nor file contents are uploaded, counted or attached to an inquiry.

Choosing Continue with this checklist places only charter, fixed scope answers, agent choice and matching record statuses in this tab's session storage. Earlier checks transfer their eight fixed selections. The register reads and removes the transfer. Its checklist stays separate from AI uses, examiner answers and evidence decisions. Save the checklist separately from the register file. Reopening it on the readiness page restores choices and requires fresh preparation. Clearing the checklist or closing the tab discards the local context. No choices enter website counts or an inquiry.

Vendor change workpaper

The vendor change workpaper runs in your browser. It sends no sources, excerpts, answers or counts. Closing the tab loses unsaved work. Explicit JSON and CSV downloads stay under your control. A JSON import is read locally; it does not send a file to us.

Open the local change workpaper