Waking Dormant Cross-Domain Knowledge in Language Models: What Works, What Doesn't, and How to Judge It Without an Answer Key

Technical note, companion to "The Activation Gap: Cross-Domain Knowledge in LLMs Is Inert Under Default Prompting" (Delic 2026, doi:10.5281/zenodo.20313594). Large language models hold knowledge from many fields but rarely connect it to a problem in another field unless prompted to. This note reports what we learned trying to turn that dormant knowledge into a practical aid for scientists working on open problems, without an answer key to check against. (1) Two keys open dormant knowledge: naming the right donor field, and a faithful, jargon-free description of what is observed. (2) For most recently solved cross-domain problems we screened (6 of 7), the description alone was enough. (3) Some cases genuinely need a nudge, and in one the knowledge was absent rather than dormant. (4) Judging is the hard part: an independent, agreement-weighting rater looked promising in small early tests, but did not hold up in a pre-registered blind test adjudicated by five AI judges from different developers. We report both. (5) The diverse-lineage judge panel agreed strongly and identified every correct item; we propose such panels, combined with automated re-description of the problem, as the next method to test. Small samples throughout; exploratory findings are labelled as such. The blind-panel amendment, verdicts and analysis were timestamped with OpenTimestamps before the answer key was opened.