York Studio AI Discovery Lab
The lab for AI-mediated discovery.
Public experiments on the systems quietly reshaping what people and agents discover.
Methods first. Results second. Commercial claims last.
Current study / 001
Protocol in preparationThe AI Answer Variability Study
A fixed prompt panel. Repeated comparable observations. No deliberate optimisation change. The study asks how much movement remains—and how large an effect a realistic measurement programme could distinguish from that background.
Research commitments
The method is part of the product.
A public chart is not reproducible research unless a reader can understand how it came to exist.
Precommit the question
Define the research question, sampling frame, outcomes, exclusions, and stopping rule before reading the result.
Preserve the evidence
Keep the answer, source, prompt version, surface, method, timestamp, and collection settings attached to each observation.
Publish the boundary
Disclose missing data, surface limitations, uncertainty, competing explanations, and results that do not resolve the question.
Correct in public
Version material changes, maintain a correction path, and leave the history visible when the interpretation changes.
Research pipeline
Questions worth measuring.
The pipeline is a direction of travel, not a claim that collection has begun. Priorities may change when evidence or customer questions change.
Answer variability
How much do repeated answers move under a null condition?
Protocol draftSource substitution
When an answer stays similar, how often do its cited sources change?
PlannedSurface divergence
How different are comparable API and authorised consumer-surface observations?
PlannedIntervention testing
What effect sizes can realistic prompt panels detect after a controlled content change?
PlannedLab updates
See the evidence arrive.
Join the Briefing for protocol changes, study releases, and honest post-mortems.
Follow the Lab