Week 7 Lab: Continuation-swap lab

A lab where pairs swap half-finished conversations and try to tell a model's continuation from a human one, then check their reasoning

15 April 2027

Before the lab#

Nothing to bring. You’ll work from a shared set of conversation openings, each with two possible continuations already prepared — one written by a person asked to continue it naturally, one generated by a model given the same opening.

In the lab#

Work through the set in pairs, recording a guess and a one-line reason before checking the answer — the reason matters more than the guess, since “it felt off” isn’t evidence and “it hedged on a claim a person would have just stated” is. Some pairs finish confidently wrong on several in a row; that’s the expected result on at least one or two of the set, not a sign you did the lab badly.

The debrief compares which features people cited across the room, and whether those features actually correlated with the right answer or just felt diagnostic in the moment.

Afterwards#

Keep your hit rate and your reasoning notes. Week 12’s synthesis lecture comes back to this exercise directly, and it’s a stronger argument with your own numbers in it than with a general claim about how good models have gotten.