On the night of 15 March 2026, a self-built personal agent system produced, in a fifty-two-minute span, a six-character phrase that did not exist anywhere in its own recorded corpus. The system, a store of 1,749 recorded decisions connected by a 24,622-link association layer feeding a cloud language model, had been assembled and recorded by a single operator over fourteen days. This paper dissects that night quantitatively: a day-level density curve of an association layer grown 155-fold in three build episodes across three days, the final episode executed in ten minutes thirty-five seconds; a second-by-second reconstruction of the fifty-two-minute window (52 min 07 s, from first successful recall-based answering to the novel phrase); and a composition-level non-containment verification of both novel phrases against the complete pre-emergence corpus (character n-grams, n = 2…5: components common, compositions absent). The operator's own verification, performed at the moment and recorded in the store, is re-verified by machine nearly seven months later. We claim no mechanism, no causality, and no property of any model; only a dissection: one night, closed corpus, timestamped pipeline, falsifiable non-containment checks. Single nights cannot prove theories. They can show exactly what happened, and that is what we contribute. Includes both English and Chinese versions.
No takes yet. Share an insight, caveat, or question.
Baofeng Zhao (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: