Repo risk ratings

Patterns that keep showing up

• Engine quality and product quality are different things. The independent security study found that engine classes separate cleanly on every axis while products inside a class do not — and the worst findings were product-level decisions (a Privileged: true default, a frozen engine pin, nested virtualisation left on) rather than engine bugs.
• Pin policy is an operator variable, and it dominates. Engine-side patch latency was effectively zero for coordinated disclosures; the observed delay ranged from zero days to 471 days to "opaque", and every day of it came from downstream pinning. Whatever you choose, decide how you receive engine updates.
• The projects with the best documentation of their own limits are the ones worth reading. nono publishes a security model that enumerates what it does not protect; byre's README says outright that a container is not a microVM; Fence documents that its network allowlisting is content-blind. That honesty correlates with the quality of everything else.
• Abandonment is the most common failure, and it is measurable. The reliable signals are last-commit date, release count, maintainer count and whether the issue tracker has recent non-bot activity — not stars.
• Beta labels are usually accurate. microsandbox calls itself beta and has the bug list to prove it; MXC says its profiles are not security boundaries; Gondolin documents the trust assumptions it relies on. Take the label seriously and integrate as an option rather than a dependency.