ScaleVerifier v0.2.0 — import, mine, build

ScaleVerifier v0.2.0 turns existing Claude Code and Codex history into a day-one local eval pipeline.

Highlights

• vf import discovers both local history sources, distinguishes rich Codex sessions from prompt history, and incrementally indexes changed files.
• vf mine produces auditable useful, correction, execution, preference, recovery, and low-value labels without calling a model.
• vf build compiles top execution-verifiable sessions into task, Git environment, Dockerfile, and verifier bundles.
• vf watch keeps the index and mining results current for nightly workflows.
• Every bundle includes sandbox-policy.json and a non-root container image contract.

Safety boundary

Autonomous agents may work only on the ephemeral container copy. Host bind mounts, Docker socket access, privileged mode, production write credentials, and host-side agent execution are outside the contract. V0.2 rejects legacy benchmark --agent; a safe container orchestrator is the next runtime milestone.

Behavior changes

• The default local store is now ~/.scaleverifier (override with SCALEVERIFIER_HOME or --home).
• vf is the primary CLI; scaleverifier and sv remain aliases.

Verification

CI passes on Python 3.9, 3.11, and 3.13. The generated bundle Dockerfile was also built and launched under non-root, network-disabled, capability-dropped smoke-test settings.