Ignat v0.2 — a rigor pipeline for AI-assisted science

Ignat — our open-source toolkit of Claude Code skills for scientific work — just reached v0.2, extending it from review into a full rigor pipeline: is it well-argued?is every fact real?which claims can we prove outright?

New in v0.2:

  • /groundcheck — an AI-slop gate. Where /review judges the whole paper, /groundcheck walks every load-bearing claim and asks one question: is this grounded in a verifiable source? It catches the five things models hallucinate most — vibe citations, ungrounded numbers, entity/property confusion, dangling cross-references, and unverified “first/novel” claims — and proposes a concrete fix for each.
  • /formalize — turn the core claims of a paper into machine-checked theorems (Lean 4 / KeYmaera X), with honest scoping: every claim is partitioned into machine-checked, pen-and-paper, or unproven. A sound proof of a physically empty statement is worse than an honest “not provable today”.
  • Two new agentspaper-architect (manuscript strategy: scope discipline, novelty-honesty, what goes in main vs SI, journal fit) and verifier (the dedicated citation/number fact-checker).
  • A best-practices guide distilling the discipline behind the tools — starting from the one rule we keep relearning: an agent’s “PASS” is a hypothesis, not a verdict.

MIT licensed. Contributions welcome.

Ignat v0.2 release →