E2EM 0.1.3

E2EM 0.1.3 removes the permission gates from app connection and warning handling. Python, Node and Dart SDKs register new apps automatically on first use; the setup screen no longer asks for Connect app approval. Existing credentials are preserved, including concurrent app registrations.

Complete, current warning results are advisory and proceed without a confirmation prompt. The Rust send guard and SDK send-flow examples still hold stale drafts, review results, errors, cancellations and incomplete coverage.

Guided onboarding is available on every desktop platform. Install the package, open E2EM Setup from the application menu, and choose Set up E2EM. The setup screen installs and verifies Gandalf, starts the background runtime, enables startup at login, and confirms a real model assessment before showing Ready. Progress, offline operation and update preferences are available in the same screen. Applications connect automatically once setup is complete. Windows also offers setup from the installer's finish screen. No Python installation or terminal commands are needed.

Gandalf is the default local model for the shared runtime, with all 40 default policies, optional conversation and custom policy strings. Normal setup downloads verified model assets; offline installers include the same weights. SDKs can select registered custom models and owner-authorized applications can install compatible model manifests from HTTPS URLs.

CPU/RAM admission checks now reject model startup before downloading or loading weights when available memory is insufficient, and cap inference at two available CPU threads. Setup reports the resource diagnostic so it can be retried after memory is freed.

Native inference now prefers a usable accelerator: CUDA on Linux, DirectML on Windows, and CoreML on macOS. Installers include the platform's native provider libraries; Linux also bundles pinned CUDA dependencies and licence notices. Initialization or execution failures in automatic mode rebuild a CPU session when the CPU memory budget permits. Slow accelerator startup also falls back to CPU within the original startup deadline. Explicit CPU and strict GPU modes are available in the worker's device diagnostics.

Installation guide · SDK guide · Model format and controls

Choose the normal installer for your platform, or its -offline- counterpart to include Gandalf:

Platform	Asset ending
Windows x64	x86_64-pc-windows-msvc.msi
macOS Apple Silicon	aarch64-apple-darwin.pkg
macOS Intel	x86_64-apple-darwin.pkg
Debian / Ubuntu x86_64	x86_64-unknown-linux-musl.deb
Fedora x86_64	x86_64-unknown-linux-musl.rpm

Installers are unsigned. Model descriptors are separately Ed25519 signed; assets are size/hash checked and candidates pass native smoke checks before activation. Model checks run every six hours when idle, retain the previous working version and support rollback. Offline setup disables network access. Applications connect automatically through the SDK, without a permission prompt or setup-screen approval.

Gandalf 0.0.1 is distributed under the owner-authorized MIT licence with upstream attribution, base-model licence and pinned provenance retained. The FP32 ONNX export passes comparison against the original checkpoint. Supporting all policies does not establish their accuracy: published policy thresholds are retained, other wordings use fallback thresholds. Missing required conversation and evidence exceeding 512 tokens produce an indeterminate review outcome. No per-category quality guarantee, block enforcement or sandbox broker is claimed.

Python wheel/source, Node archive, Dart SDK source, Rust source and native C SDKs accompany the runtime. The Dart SDK supports Dart and Flutter desktop applications. Registry publication is not configured. Linux inference needs glibc 2.28+ and libstdc++; the daemon itself is musl-static. CPU inference needs no GPU or Python model packages.

Release publication requires native inference, installer install/upgrade/remove and guided onboarding gates on Linux, Windows and both Mac architectures. Reports and SHA256SUMS accompany the assets. Checksums establish byte integrity; model signatures do not sign the installers. release-provenance.json identifies the exact source commit and build run.

E2EM 0.1.2

E2EM 0.1.2 now includes guided onboarding on every desktop platform. Install the package, open E2EM Setup from the application menu, and choose Set up E2EM. The setup screen installs and verifies Gandalf, starts the background runtime, enables startup at login, and confirms a real model assessment before showing Ready. Progress, offline operation, update preferences and application connection are available in the same screen. Windows also offers setup from the installer's finish screen. No Python installation or terminal commands are needed.

Gandalf is the default local model for the shared runtime, with all 40 default policies, optional conversation and custom policy strings. Normal setup downloads verified model assets; offline installers include the same weights. SDKs can select registered custom models and owner-authorized applications can install compatible model manifests from HTTPS URLs.

CPU/RAM admission checks now reject model startup before downloading or loading weights when available memory is insufficient, and cap inference at two available CPU threads. Setup reports the resource diagnostic so it can be retried after memory is freed.

Native inference now prefers a usable accelerator: CUDA on Linux, DirectML on Windows, and CoreML on macOS. Installers include the platform's native provider libraries; Linux also bundles pinned CUDA dependencies and licence notices. Initialization or execution failures in automatic mode rebuild a CPU session when the CPU memory budget permits. Slow accelerator startup also falls back to CPU within the original startup deadline. Explicit CPU and strict GPU modes are available in the worker's device diagnostics.

Installation guide · SDK guide · Model format and controls

Choose the normal installer for your platform, or its -offline- counterpart to include Gandalf:

Platform	Asset ending
Windows x64	x86_64-pc-windows-msvc.msi
macOS Apple Silicon	aarch64-apple-darwin.pkg
macOS Intel	x86_64-apple-darwin.pkg
Debian / Ubuntu x86_64	x86_64-unknown-linux-musl.deb
Fedora x86_64	x86_64-unknown-linux-musl.rpm

Installers are unsigned. Model descriptors are separately Ed25519 signed; assets are size/hash checked and candidates pass native smoke checks before activation. Model checks run every six hours when idle, retain the previous working version and support rollback. Offline setup disables network access. Application enrolment requires the user's approval in the setup screen.

Gandalf 0.0.1 is distributed under the owner-authorized MIT licence with upstream attribution, base-model licence and pinned provenance retained. The FP32 ONNX export passes comparison against the original checkpoint. Supporting all policies does not establish their accuracy: published policy thresholds are retained, other wordings use fallback thresholds. Missing required conversation and evidence exceeding 512 tokens produce an indeterminate review outcome. No per-category quality guarantee, block enforcement or sandbox broker is claimed.

Python wheel/source, Node archive, Dart SDK source, Rust source and native C SDKs accompany the runtime. The Dart SDK supports Dart and Flutter desktop applications. Registry publication is not configured. Linux inference needs glibc 2.28+ and libstdc++; the daemon itself is musl-static. CPU inference needs no GPU or Python model packages.

Release publication requires native inference, installer install/upgrade/remove and guided onboarding gates on Linux, Windows and both Mac architectures. Reports and SHA256SUMS accompany the assets. Checksums establish byte integrity; model signatures do not sign the installers. This release's installers and SDKs have been rebuilt and replaced to include guided onboarding; release-provenance.json identifies the exact source commit and build run.

Rebuilt from 46e1b9c4b533; all platform qualification results.

E2EM 0.1.1 — developer preview

E2EM runtime and SDK developer preview. Chat message assessment, authenticated local IPC, and embedded Rust/C APIs. Named categories are accepted for reporting; the bundled backend evaluates email patterns and holds unavailable checks for review. No model or GPU required.

Start here: Installation guide · SDK guide · Changelog

Choose an installer for your computer:

Platform	Asset ending
Windows x64	x86_64-pc-windows-msvc.msi
macOS Apple Silicon	aarch64-apple-darwin.pkg
macOS Intel	x86_64-apple-darwin.pkg
Debian / Ubuntu x86_64	x86_64-unknown-linux-musl.deb
Fedora x86_64	x86_64-unknown-linux-musl.rpm

Installers are unsigned. Application enrolment and user-session startup remain explicit. Installing E2EM alone does not change other apps. The bundled check provides personal email-pattern warnings. Other named categories are reported as unevaluated without a supplied backend; no contextual model, block enforcement, browser transport, or sandbox broker is shipped.

Python wheel/source, Node .tgz, Rust source and platform C SDK archives are included. Registry publication is not configured. Linux C SDK libraries use the Ubuntu 24.04 runner baseline; Linux runtime installers are musl-static.

Native service checks and installer install/upgrade/remove gates passed before publication. Platform reports and SHA256SUMS are attached; checksums establish byte integrity, not publisher authenticity. Historical resource evidence is documented separately in the source tree.

Licensed under MIT. Security boundaries.