Dated implementation reviews, data-path boundaries, distribution facts, and known limits. These are ClassEve’s internal notes, not independent audits, certifications, or universal performance guarantees.
Observed implementation behavior and the environments that can change it.
The Windows and Linux clients use their platform windowing and audio stacks rather than claiming identical behavior on every desktop environment.
Window managers, accessibility settings, and compositor behavior can affect overlays. A public-beta label is retained while those combinations are widened.
Facts intended to remain checkable across releases.
Current public binaries publish their exact version, byte size, and SHA-256 next to the download path.
Checksum sidecars and the public products.json registry provide a stable comparison point. A matching hash verifies the downloaded bytes, not the absence of every possible defect.
The reviewed range and what remains outside it.
Available platforms are separated from planned work.
Lven Instant currently publishes builds for Windows, Linux, and Android. macOS and iPhone builds are in internal testing and have no public release date.
The product registry records an exact version, byte size, download URL, and SHA-256 for every available platform. Planned platforms have no artifact or release date.
Implementation paths and the cases where they can fail.
Text delivery depends on the active application and operating system; one insertion technique is not represented as universal.
Windows can use UI Automation with a controlled paste fallback. On Linux, Lven Instant inserts through AT-SPI2 with a clipboard-and-paste fallback on both X11 and Wayland; Lven Cloud auto-pastes on X11 and leaves the transcript on the clipboard on Wayland. Protected fields and applications that block automation can reject insertion.
Recovery and network boundaries stated explicitly.
Lven Instant can need a network connection for model setup or repair and entitlement validation even though shipped-platform transcription is local.
The privacy claim is limited to the transcription data path; it is not presented as a promise that every application function works indefinitely while air-gapped.
Architecture and variables, without a universal benchmark claim.
Lven Instant performs speech recognition on the device, so its transcription path does not upload microphone audio for remote inference.
Actual completion time depends on the device, model state, audio length, and operating conditions. ClassEve does not publish one universal latency number for all hardware.
Documented architectural limits, not aspirational targets.