Unquoting converts fragments into instructions

This proposed hypothesis asks whether ordinary editing can turn quoted story material into instructions for a language model, creating a self-reinforcing loop. It would replay edits from one shared parent text. The same earlier words would appear either as clearly quoted data or as ordinary conversation, while model history and human continuity vary separately. The proposed signal is a history effect at the model step that travels with context marking the words' role and fades after a verified boundary between instruction and data. The key test asks whether separately measured human scope conversion and model instruction-following predict held-out changes. If they do, known channels suffice; if they do not, the loop warrants further testing.