Definition
Plain language
Writing down exactly what you'll test and how, before you look at the results.
As stated in the literature
Time-stamped public specification of hypotheses, conditions, and analysis plan filed prior to data collection, limiting analytic flexibility and undisclosed post hoc choices.
Also called: pre-registered
Why it matters: It stops researchers from quietly reshaping their hypothesis to fit whatever the data happened to show, which is one of the main ways false findings enter the literature.
For example, a team publicly posts that they will compare two prompting conditions on 500 questions using one specific measure, before they run a single trial.
Heard on the show
“… logged tool calls, fit a relationship between agent aggression and scientific output, and then pre-registered four predictions, with probabilities attached, for later scoring. …”Episode 123 — Five Identical Worlds, One Swapped Model: What Happens When AI Agents Run for Fifteen Days