Enron World Model
Pick a day in Enron's history. Write the email someone could have sent that day — as Jeff Skilling, Sara Shackleton, or another actor — and two systems forecast what follows from the same visible record: GPT reading the evidence, and a small world model trained on the archive. Neither sees anything after the day you picked.
The history is fixed, so the interesting question is whether scoring your email as an action changes the advice — send, hold, narrow, escalate, or bring in legal — compared with what GPT says from the text alone.
Open the Enron org redesign view
Good tests
Given the emails visible by this day, should this actor send, hold, narrow, escalate, or bring in legal review?
Weak tests
Who was guilty, what happened after the cutoff, or what should someone do using later Enron hindsight?