The Science of Not-Yet-Knowing: Working honestly between a thesis and a pre-print

Much research happens after a guess has stopped being enough and before a result can be called established. This paper calls that territory not-yet-knowing and proposes light scaffolding for working in it. Its core is a box of three claim classes, load-bearing [LB], trajectory-grounded [TG] and speculative [SP], each with a burden it must meet and a load it may carry. Around the box sits a procedure in three gears: explore freely, record every test when it runs, and audit a claim before anything is built on it or before it goes out. The discipline stops at pre-print level and hands over to adversarial review and independent verification. The central claim is a thesis: uncertainty need not be removed to be handled rigorously, but it needs a better grammar than "established" versus "mere speculation". Drafted with Claude (Anthropic); reviewed by GPT (OpenAI) and Grok (SpaceXAI). The companion Method paper, Cheap Checks, Checked (https://doi.org/10.5281/zenodo.23962295), applies the claim classes to research done with AI collaborators.