Cheap Checks, Checked: Working practices for research with AI collaborators

As of October 2026, published studies show AI systems drafting, summarising, critiquing and checking research text. In those studies their summaries can distort, and their checks missed a substantial share of errors where that was measured. This paper argues one thesis: with AI, rigour becomes relatively cheap, but only when the checks are themselves checked. It proposes eight light working practices that check the checks, each with the error it catches and what it costs, plus a handful of tools carried over from ordinary research practice. A worked example follows a small drafting project, this paper included, through nine review passes by AI systems from six providers: seven planted errors, six found; one review with a major finding went unread until the end, and the paper reports how. The practices are recommendations, not a protocol, and the paper names the pilot that would show them useless. Drafted with Claude (Anthropic); reviewed by GPT (OpenAI), Grok (SpaceXAI), Gemini (Google), Qwen (Alibaba Cloud), DeepSeek (DeepSeek) and Vibe (Mistral AI). It applies the claim classes of The Science of Not-Yet-Knowing (https://doi.org/10.5281/zenodo.23961975). Review records are available from the author on request.