EVIDENCE & CONTROL RESEARCH
We are developing a research programme on evidence, uncertainty and human control. From a persuasive answer to a decision that can be checked.
The study below is a planned pilot. We share a draft protocol for consultation.
THE RESEARCH AGENDA
Reliability depends on more than the model. Our proposed work examines how evidence, controls and people shape the outcome.
When does a source actually support an answer?
We examine the relationship between claims and evidence, contradictions and justified decisions to leave a question unresolved.
Unsupported claims · evidence alignment · unnecessary refusals
Planned outputEvaluation protocol and comparative tasks
Can a person meaningfully intervene?
We examine whether people understand what they approve and whether they can stop, correct or limit an AI system’s actions.
Unintended actions · unnecessary blocks · effective intervention
Planned outputOversight scenarios and requirements
What happens when a previously accepted fact is wrong?
We examine how errors persist, how corrections affect dependent conclusions and how the origin of information remains traceable.
Error propagation · retraction · currency of information
Planned outputMemory and correction test protocol
What does a proof of execution establish?
We distinguish checking that a record remains intact from verifying that a specified computation was performed.
Scope of proof · verification cost · information disclosed
Planned outputVerification scenarios and a description of what each proof establishes
FROM QUESTION TO STUDY
The proposed pilot compares ways to check whether source material supports AI answers.
A planned comparison of AI answers and methods for checking whether a cited passage actually supports a claim.
METHOD BEFORE CONCLUSION
We will define what is being tested, what it is compared with and what would count against the hypothesis before interpreting outcomes.
Record model and source versions, conditions, scoring criteria and changes to the protocol. Describe disagreements between assessors.
Report false refusals, time and resource use alongside benefits. Keep feasibility, observed improvement and causal evidence distinct.
Identify authors, reviewers, funding and relevant relationships when established. A technological affiliation is not independent validation.
RESEARCH INTO PRACTICE
Research questions guide the design of controls: what the gateway checks, when a result should be withheld and what evidence should be retained.