invAIriant keeps architectural invariants from drifting under AI-assisted change.

Your coding agent runs the audit and gates the merge on real findings. Claude Code · Codex · Cursor.

No evidence. No finding.

An LLM can propose anything; only cited, verified findings count. invAIriant runs a named set of review lenses over a bounded change, then makes every candidate survive adversarial evidence checks before it becomes a finding. Lenses discover, evidence verifies, severity gates.

Status: early. Usable now as an agent skill; the CLI and CI support are intentionally small and deterministic.

Start

1. Add the skill to your agent — this is the audit:

mkdir -p .claude/skills && ln -s "$PWD/skill" .claude/skills/invairiant
# Codex / Cursor: see skill/README.md

2. Run it in the agent: /invairiant audit-pr → a paste-ready PR comment.

3. Gate CI (optional — the deterministic CLI seatbelt):

pip install invairiant
invairiant ci-gate report.json      # exits 1 on an open S0/S1 finding

Worked examples

The low-latency-runtime example shows the core rule: a passing summary does not erase a verified finding. Five more worked audits in case-studies/, and a runnable demo in examples/refundpilot-demo/.

Learn more

• How it works — the pipeline and the evidence rules: docs/methodology.md
• The skill — every command and scope: skill/SKILL.md
• Lenses — 28 across 7 packs, pick 4–6 by risk: docs/lens-taxonomy.md
• The CLI: docs/cli.md · Gate PRs in CI: docs/github-action.md
• Walkthrough — a full run with real output: docs/demo.md

Not this

Not a linter, scanner, or proof of correctness; it turns their output into evidence and gates on verified findings. Not "AI, wander the repo and tell me what you think": it audits a bounded scope (a PR, a range, a module, an ADR) and refuses one it can't bound.

---

Apache-2.0 · © 2026 mindicator · Contributing · Roadmap