Public beta status

Ódinn Forge v0.4.0-beta.3 is a prerelease public beta. The local operator workflow and release automation have substantial verification, but Ódinn is not yet a stable release.

How to read product claims

Verified local behavior

This label covers the supported single-operator path on an owner-controlled machine: onboarding, the CLI, the loopback gateway, local console, durable jobs and records, policy-gated local tools, audit verification, release archives, checksums, and installer rollback.

“Verified” describes current beta evidence. It does not mean safety-critical, enterprise-supported, or immune to defects.

Provider- or platform-dependent

Live model accounts, quotas, model availability, OAuth and device flows, local model servers, websites, browser engines, external authentication, and operating-system behavior depend on systems outside Ódinn’s local verification boundary.

Experimental and off by default

The seven Labs—Proof, Sentinel, Capability Tokens, Rewind, Capsules, Counterfactuals, and Darwin—are implemented local slices that require explicit per-feature enablement. Read Labs before turning one on.

Third-party extension, MCP, Agent SDK, and Skills SDK packages also begin disabled and untrusted.

Enabled by default and bounded

Automatic improvements watches repeated audited failures and may apply only predefined reliability tuning. It cannot invent changes, edit source code, change credentials, grant permissions, install packages, or disable safeguards. Applied tuning keeps a restore point.

Explicitly unsupported

• Hostile-code containment by forked workers.
• Public exposure of the single-user gateway.
• Operating-system isolation between mutually hostile remote users.
• Guaranteed replay or rollback of external effects.
• Deterministic behavior from remote services or nondeterministic models.

Stable-release gates

The live cloud-provider gate is complete for one OAuth path and one API-key path. Stable v0.4.0 remains blocked on:

• installation, onboarding, restart/recovery, and rollback on a real Windows host;
• three independent users exercising the daily-use workflow over multiple days;
• an exact-candidate security review and explicit maintainer go/no-go decision.

Hosted Windows CI and synthetic provider tests are useful supporting evidence, but they do not replace those real-world gates.

For the project’s source-grounded evidence, see the repository’s stable-exit plan and surface matrix.