Building traceable agent foundations for restricted-domain financial and corporate-filing analysis.
Problem
Audit agents must answer complex questions without inventing facts, crossing evidence boundaries, or hiding how a conclusion was produced.
Approach
Combine canonical fact and narrative layers with deterministic validators, task specifications, evidence-bearing answers, and programmatic evaluation.
Role
Research and engineering lead on the UKRI project
This work treats evidence as a system contract rather than a prompt instruction. Public AuditOps components focus on corpus ingestion, canonical facts and narratives, deterministic validation, benchmark generation, and traceable answer objects.
The restricted UKRI programme includes additional private data, modelling, and evaluation work. Those results are deliberately kept separate from claims about the public codebase.