privacy
The benchmark runs in your browser. The app still stores pseudonymous operational analytics for page and benchmark flow health, and it only stores a redacted benchmark snapshot when you explicitly create a share link, save a private result, or send feedback with trace opt-in.
Operational analytics record page views, benchmark flow steps, shares, and CTA clicks with random session and visitor IDs.
Share links store a redacted benchmark snapshot only when you explicitly ask for a shareable result.
Private saved results store the same redacted benchmark snapshot plus hashed return/delete keys only when you explicitly opt in.
Accuracy feedback can include the same redacted calibration trace only when you opt in on the feedback form.
If the result-page pricing holdout is enabled, its answers are stored as anonymous analytics events and deleted under the same 90-day analytics window.
Admin exports are generated on demand for operators and are not stored server-side after the response finishes.
Operational analytics. 90 days. Pseudonymous page, benchmark, share, and CTA analytics stay in the app database for this window before automatic pruning.
Feedback and opt-in traces. 90 days. Feedback votes, match outcomes, optional notes, and any opt-in redacted traces follow the telemetry retention window and are pruned automatically.
Share metadata and redacted share payloads. 30 days. Share slugs, timestamps, and the redacted benchmark payload expire automatically when the share retention window ends.
Calibration evidence. 30 days. Public calibration evidence is rebuilt from redacted feedback traces and shared runs inside the shorter eligible retention window.
Admin exports. 0 days in-app, 30 days for saved files. Admin exports are generated on demand and not retained server-side; any downloaded working copy should be deleted on the manual file-retention window.
Current disclosure policy: 2026-03-12-disclosure-v1
Monetized CTA lane. Referral or sponsor CTAs stay below in a labeled module. They cannot change model order, runtime order, upgrade guidance, or confidence wording.
If a labeled referral or sponsor CTA is shown, its analytics stay separate from benchmark scoring and are used for offer auditing rather than personal profiling or recommendation ranking.
A redacted benchmark snapshot with browser, device class, rounded viewport, and hardware-fit notes.
Benchmark timings, score breakdown, model recommendations, and runtime recommendations.
A private return-link key and delete-link key stored as server-side hashes.
No email address is stored because email delivery is disabled in this environment.
Saved results are kept for 14 days, then pruned automatically.
Delete it any time with the private delete link shown after saving or included in the email copy.
Email delivery is currently unavailable, so only the private return link is offered.
Full user-agent strings are redacted.
User-agent hints are removed.
Viewport sizes are rounded before storage.
Whether the recommended model class actually matched real local AI use.
How close browser-side speed estimates are to native local runtimes.
How confidence should change across browsers and hardware families.
Public calibration evidence only replays aggregated summary counts plus redacted feedback traces inside the shorter eligible retention window.
Internal replay inputs are allowlisted before trust or calibration views are rebuilt; free-form notes do not appear on public evidence surfaces.
Raw admin exports stay behind the admin token and should be treated as short-lived working copies, not a second permanent database.