Skip to content

Invaris v0.2.0

Choose a tag to compare

@akkaraponph akkaraponph released this 14 Nov 15:01
· 69 commits to main since this release
dfd7a56

Invaris v0.2.0 adds campaigns. A campaign runs several mutation and guardrail experiments together from one frozen copy of the source, repeats them, and decides on the whole set with an explicit acceptance policy.

  • Run invaris campaign with an invaris.campaign/v1alpha1 file that references existing v1alpha3 mutation and v1alpha2 challenge harnesses. Reports use invaris.evidence/v1alpha4.
  • Repeats keep every attempt and are classified CONSISTENT, MIXED, or INCOMPLETE, with no voting.
  • The mutation score and the guardrail proof rate count consistent experiments once each, are null when nothing is eligible, and are never combined.
  • A typed acceptance policy decides ACCEPT (exit 0), REJECT (1), or REFUSE (2). Incomplete or contradictory evidence refuses before any reject.
  • --only runs a recorded subset, --format text prints a summary, and --parallel N runs attempts concurrently (default 1). Harnesses may sit in subdirectories, and optional tools record tool versions.
  • Archives include the CLI and generic command provider. Verify downloads with SHA256SUMS.

Compatibility: run, check, challenge, and mutate keep their inputs, exit codes, and report versions. The v1alpha2 and v1alpha3 evidence schemas now require the reason and provenance keys added since 0.1.0. In challenge, a guard that writes into its workspace now yields INCONCLUSIVE, and rejected structured output from a guard exiting 0 or 1 is reported as malformed_evidence. See the changelog and the Campaigns guide.

Commands and evaluators run with the user's permissions. Disposable workspaces and campaign snapshots are not security sandboxes. See the threat model.