A listener names a year. The cast is dropped into it and starts talking.
Name 415 and they land in Alexandria, in the month Hypatia was killed. Change the subject mid-scene and they get pulled somewhere else, still carrying what they just experienced. That second part is the hard one.
It did not start here. The first version was a cyberdrama with no time travel in it, running on a custom GPT that held about three acts. I wanted a listener to be able to leave one story for another. My routing kept failing, and residue from the last story carried into the next.
A second GPT I had built was already doing that deliberately. Before it generated, it looked back for a story I had read earlier, ranked how relevant it was, and used it to frame the current one. A current event could land against something I looked up several sessions before.
I fixed the routing. By then I could see the better idea. Carry-over stops being a leak once it is ranked. Make the cast time travelers and it becomes the mechanic. A listener changing the subject reads as interference with their equipment. The obstacle became the differentiator.
Interactive stories have always made one bad trade. Hand-written branching buys choice by exhausting an author. Finite, expensive, still on rails. Freeform AI gives infinite choice and no world worth choosing in. No cast, no truth, no consequence. This engine refuses the trade.
The biggest risk in this medium is a listener derailing the story mid-generation. Rupture turns that into the signature: the interruption is absorbed in character, not a break.
The hardest problem, made a feature