Mafalda Pereira
Back to workProject 01 · Case study 01 of 06

Evidence strategy · AI evaluation

AI-assisted research workflow

The workflow could produce a fluent answer. I checked whether the bibliography and technology benchmark actually supported it, and recorded where researcher judgement was still needed.

Abstract workflow connecting research evidence, evaluation and documentation.
Visual note

The review moved from generated answer to cited source and back again.

My role
I compared generated guidance with the bibliography and primary documentation. When a citation did not support the sentence, or an important limitation disappeared, I logged the mismatch.
Study scope
Bibliography, technology benchmark and evidence validation
Methods
Structured literature review · Benchmarking · Evidence validation

The review problem

Which parts of the generated guidance were reliable enough to support research work, and which still required expert review?

A polished response could still omit a condition, stretch a source or merge claims with very different levels of support. I treated the answer and its references as two things to inspect, not one.

What I checked

Three passes through every answer

  1. 01 / Sources

    Start with the references, not the prose

    The bibliography and benchmark came first. They gave me a stable base against which to check each answer.

  2. 02 / Gaps

    Look for the conditions that disappeared

    I checked technical capability, research support, limitations and open questions separately. Missing context was often more revealing than an incorrect sentence.

  3. 03 / Boundary

    Mark where automation had to stop

    Mismatched references were corrected, incomplete claims were qualified and questions needing a researcher were left visible rather than smoothed over.

One answer, three checks

Generated guidance was checked against its source trail.

This reconstruction shows the three-part check used throughout the review. It is a simplified view of the working method, not a client document.

Project 01
Bibliography and primary documentation stayed beside the generated answer.

Review criteria

Three questions kept the review honest

The same short checklist was applied throughout. It was simple enough to use repeatedly, but strict enough to catch answers that sounded stronger than their sources.

Answer audit
  1. 01

    Source

    Can I find the claim in the cited material?

    Important statements were traced to scientific literature or primary technical documentation, not accepted from the citation alone.

  2. 02

    Missing context

    What would change how someone used the answer?

    Conditions, limitations and omitted steps were recorded whenever they could alter a research decision.

  3. 03

    Confidence

    How far does the available evidence let us go?

    Each claim was marked as supported, incomplete or in need of independent validation.

Result

The source base, benchmark and a correction log

The team received the consolidated sources together with a review showing which passages could stay, which needed correction and which questions remained open.