Paper 3 · the method · published 20/10/2026
Controlled Research at a High-Street Practice.
That artificial intelligence lowers the cost of analysis while judgement remains necessary is, by now, unremarkable. What is less common is an audited account of exactly where the machine was wrong, what caught it, and what would have been printed if nothing had. This paper is that account.
Why this one matters.
The one for other practitioners. It publishes the error ledger rather than the success story: twenty-one corrections, nine of which would have put a wrong headline into print, each attributed to whatever actually caught it. It ends with a checklist another practice could use.
Two of the findings.
A small practice built a register-scale study in a fortnight
Between 16 and 31 August 2026, an AI assistant directed by the practice wrote a fifty-three-file pipeline of some ten thousand lines and walked the register for 119,220 London food companies, living and dead.
Twenty-one corrections were made before publication, and nine would have put a wrong headline into print
A wrong estimator, a branch that could never be reached, two observation dates, two postcode districts never collected, a table counting rows instead of premises.
The paper carries 5 findings in all, with the method, the sources and every correction made before publication. It also carries the full ledger, four of the errors in detail, the controls that caught each one, and a checklist another practice could use.
Read it.
The whole paper is on this site, free and ungated. If you would rather have the typeset PDF, we will email it to you.
Roseworth Research is the research arm of Roseworth Limited, a Finchley accountancy practice. We act for businesses in this sector and have a commercial interest in the subject. The study uses public data only and no client information of any kind.