Evidence strategy · AI evaluation
AI-assisted research workflow
I built the bibliography and technology benchmark that grounded an AI-assisted research workflow, then validated its outputs against the source evidence.
The research question
Which parts of the generated guidance were reliable enough to support research work, and which still required expert review?
Fluent output could still omit conditions, overstate a source or combine claims with different levels of support. The review therefore had to make the path from claim to evidence visible.
How I approached it
From question to evidence
- 01 / Ground
Build the evidence base before evaluating the output
I assembled and organised the bibliography and technology benchmark used to ground the workflow.
- 02 / Structure
Define what a defensible answer needed to contain
I separated technical capability, research evidence, limitations and open questions so they could be reviewed consistently.
- 03 / Validate
Trace claims back to the material reviewed
I corrected mismatched or incomplete references and documented where stronger evidence or researcher judgement was still needed.
How I organised the evidence
A review built around provenance, coverage and claim strength
I reviewed the answer and its evidence as separate objects. That prevented a well-written response from receiving more confidence than its sources justified.
- 01
Provenance
Where the claim came from
Each important statement was checked against the relevant scientific source or primary technical documentation.
- 02
Coverage
What the answer included or omitted
I looked for missing conditions, limitations and steps that could change how the guidance was used.
- 03
Claim strength
What the evidence actually allowed us to say
Claims were distinguished as supported, incomplete or requiring stronger independent validation.
What I delivered
A grounded evidence base, benchmark and validation record
I delivered a consolidated source base and a traceable review of the generated material, showing what could be retained, corrected or investigated further.