factory-kit v0.3.0 — verifiability and scoring standards

v0.2.0 could lint conventions but couldn’t measure how well a project conformed — or tell a contributor which evals must run for a given change. v0.3.0 closes that gap: the standard is now scoreable, and the score guards the PR boundary.

npx @nonlinear-labs/factory-kit add-ci

What’s in this release

• Banded conformance score — factory-kit-check now renders a pass / warn / fail verdict instead of a false-precision number, with severity-aware coverage disclosure: the scorecard names every critical-class pitfall that isn’t machine-checked yet, so the coverage ceiling is honest, never hidden.
• factory-verification skill — the four-tier eval spectrum (CLI rule → test → agent → human gate) and the evals-graduate-downward promotion pipeline: checks start expensive and human, and earn their way down to cheap deterministic rules.
• verification-engineer subagent — designs the verification strategy for a change: blast radius → required eval tiers → known gaps. Sister to code-reviewer (which finds defects); this one decides what proof is required.
• Delta-gated GitHub Action — add-ci drops a conformance workflow into any repo. It fails only on newly-introduced critical findings: new debt blocks, pre-existing debt is reported but never punished, so the gate adopts cleanly on old codebases.
• Dogfooded — the conformance gate runs on factory-kit’s own PRs.

Getting started

New here? The Getting Started guide takes you from install to your first command in under two minutes.

What’s next

Hosted factory-kit — the read-only control plane that measures these patterns across your live infrastructure — is in private beta. Join the waitlist for first access.