Testability, Reproducibility, and the HMRC Audit Trail

HMRC's VAT431 reclaim process requires the claimant to justify every line item. For contractors and developers, any HMRC enquiry into VAT treatment requires the business to demonstrate — for each supply — why the rate applied was correct and which HMRC rule supports it.

A position justified by 'the AI said so' provides no defensible basis. A position where every line item is referenced to a specific HMRC Notice section, reviewed and confirmed by the responsible party, and documented in an audit trail is in a different category entirely.

VATBuild is designed to produce the second kind of record — not as an afterthought, but as a structural requirement. Every classification is made within the framework of the rules engine; every classification is confirmed by the user; every decision is stored with its timestamp, HMRC reference, and the reason it was made.

Because VATBuild's classifications are grounded in explicit rules rather than probabilistic AI outputs, the system can be regression-tested against a library of known-good invoice test cases. When any component of the system is updated, the test suite confirms that existing correct classifications have not changed.

This is not possible with a generic AI. There is no stable input-output relationship to test against — the same question may produce a different answer on re-run, or between model versions. There is no baseline, no regression suite, and no way to confirm that an update has not changed the model's behaviour on a specific edge case.

VATBuild maintains a comprehensive test suite covering the full range of VAT route types — new build, conversion, long-empty properties, change in number of dwellings, developer schemes, contractor DRC scenarios, and ineligible cases — each with a verified correct classification profile. Every system update is validated against this matrix before deployment. If a classification changes unexpectedly, the test suite catches it before users are affected.

The VATBuild output is designed to be reviewed by a third party — an accountant, a solicitor, or an HMRC officer — without any knowledge of how the system works internally. The audit trail does not require technical expertise to follow.

For each line item, the record shows: the supplier name and VAT registration number, the invoice reference and date, the supply description, the VAT amount charged, the classification outcome, the applicable HMRC Notice section and rule reference, whether the AI's initial classification was overridden by the user, and if so, why.

This is a different class of deliverable from an AI-generated summary. It is a structured, evidenced document that stands on its own merits — one that can be submitted to HMRC, reviewed by an accountant, or retained as the permanent record of a compliance decision made at a specific point in time.

The record is stable: the classification applied today is retrievable and explainable in four years' time if HMRC raises an enquiry — not dependent on a model version that may no longer exist, a conversation that was never saved, or a probabilistic output that could not be reproduced.

VATBuild reports are structured for third-party review from the outset. An accountant reviewing a VAT431C reclaim can see exactly which line items are included, why each was classified as reclaimable, what HMRC rule applies, and whether any were overridden by the claimant.

HMRC reviewing a contractor's invoicing position can see the project profile, the DRC assessment, the applicable Notice 708 sections, and the classification decision for each supply — all without needing to understand how the software works or rely on the software vendor's explanation.

This is not a feature added on top. It is the point of the architecture. Trustworthy, reproducible, auditable, and defensible: these are the four properties that construction VAT compliance requires and that generic AI cannot provide.