Research
The adubio platform is grounded in applied research into how AI systems reason, where that reasoning fails, and how assurance professionals can evaluate it systematically.
Qlogue publishes research, benchmarks and practitioner explainers across three interconnected areas. The work is designed to be useful to practitioners in model risk, internal audit, regulatory assurance and AI governance — not only to researchers.
Synthetic Financial Institutions
Controlled synthetic institutions — Largebank002, UKDomesticBank001 and others — provide a reproducible environment for evaluating AI assurance tools. Because the ground truth is known, it is possible to measure whether a system correctly identifies unsupported conclusions, missing evidence and reasoning gaps.
Benchmark Design
The UKDomesticBank001 benchmark separates participant materials from evaluator ground truth, allowing controlled comparison of AI assurance outputs against a known standard. The design is intended to support reproducible evaluation across different models and configurations.
Reasoning Engineering
Reasoning engineering addresses how AI systems construct, represent and communicate chains of inference. Qlogue's work in this area focuses on the conditions under which AI-generated reasoning is traceable, verifiable and defensible — and the conditions under which it is not.
The adubio workflow
adubio is designed to bring assurance forward — capturing questions, evidence and observations progressively as the audit takes place, rather than reconstructing them retrospectively.
Record as you go
Auditors log questions, evidence requests and emerging observations during fieldwork, linked to the relevant audit objective, requirement or control.
Share and respond
The business receives observations as they emerge and can agree, partially agree or disagree, provide explanations and upload further supporting evidence.
Verify continuously
adubio checks assertions against supporting documents, applicable policies, methodologies and regulatory requirements. It identifies contradictions, unsupported claims and missing evidence. Both the auditor's observations and the business's responses are subject to evidence-based verification.
Resolve and record judgement
The auditor reviews the evidence and responses, then accepts, amends, withdraws, escalates or leaves an observation unresolved. Every decision retains its rationale and history. Agreement is not a prerequisite for an auditor to reach an independent conclusion.
Generate the workpaper
Once fieldwork reaches the appropriate stage, adubio synthesises the controlled record into a structured draft workpaper, preserving the underlying evidence, citations, challenge history and professional judgement.
Independently review and amend
The generated workpaper passes through adubio Analyse, which tests its reasoning, traceability, evidential sufficiency and consistency. The auditor reviews the results, makes amendments and records any remaining follow-up actions.
Approve and issue
Audit leadership reviews the final workpaper and approves formal findings for communication to stakeholders. The complete audit log remains available as a controlled record of how conclusions were developed, challenged and resolved.
Less retrospective reconstruction. Fewer avoidable disagreements. Greater transparency for both the business and the auditor. Continuous assurance without compromising independent judgement.
Interested in how adubio applies this research in practice?