The purpose of an evidence check is to make a requirement inspectable. It is not to persuade a language model that the application looks complete. Begin with a source-backed requirement and ask what evidence would demonstrate the requested information. Then locate that evidence in the controlled document set, recording the document identifier, revision and exact location.
Keep administrative completeness separate from technical adequacy. A report may be present, signed and correctly referenced while its assumptions still need specialist review. Conversely, a technically sound report may be unsuitable for submission because it describes an old design or omits a required declaration. Your check record should state which question it answers.
Write the checking instruction narrowly
Use explicit tasks: compare the site address across the application form and drawing title blocks; identify references to superseded drawings; find whether a supplied report includes its stated appendix. These are more testable than “check compliance”. When a task involves interpretation, performance assessment or legal applicability, route it to a suitably qualified reviewer.
Supply the tool with the relevant source material and request evidence quotations or precise pointers. Instruct it to return “not located in the supplied material” when appropriate. That phrase is more honest than “does not exist”, because the tool may have received an incomplete extract or failed to read a drawing layer. A failed search is not proof of absence.
Use three finding categories
Missing information
The expected document, field or evidence location has not been found in the reviewed material.
Conflicting information
Two identified sources disagree about a shared fact, dimension, reference or design condition.
Authority-specific question
The applicability, interpretation or acceptable evidence route needs confirmation from the responsible body.
Attach a proposed next action, owner and evidence reference to each finding. Avoid closing an issue with “AI checked” or “looks fine”. A useful closure records what changed, who reviewed it and which revision was accepted for the stated checking purpose. If the team decides a finding is not applicable, retain the reason and its supporting source.
Require citations that actually resolve
Generative systems can invent section numbers, merge quotations and cite real documents for propositions they do not support. Open the cited source and compare the relevant text. Check that a passage belongs to the correct edition and jurisdiction. Where a requirement refers to a standard, verify the specified edition and any applicable national provisions rather than substituting a familiar international reference.
The European Commission Joint Research Centre’s Eurocodes resources illustrate why this matters: structural design standards sit within a wider implementation context, including national choices. Merely detecting the word “Eurocode” in a report cannot establish that the design basis is appropriate. A document-control check can identify the stated basis and flag inconsistency; specialist assessment remains a separate responsibility.
| Check | Useful tool output | Human follow-up |
|---|---|---|
| Shared project facts | Side-by-side values with page references | Confirm the authoritative value |
| Referenced attachments | List of references not found in the inventory | Locate or request the actual files |
| Rule applicability | Source passage and explicit uncertainty | Obtain qualified or authority clarification |
Assess the tool, not just the result
The National Institute of Standards and Technology’s AI Risk Management Framework offers a general structure for considering AI risks. Its Generative AI Profile addresses risks associated with generative systems. These resources are not permit rules, but they support a sensible habit: define the use case, understand likely failure modes and retain proportionate oversight.
Test false positives and false negatives separately. A system that flags every page may appear diligent while wasting review time. A system that returns a polished clean report may silently miss tables, annotations or scanned appendices. Keep a small test set with known issues and compare tool performance after updates or configuration changes.
Protect the information used for checking
Construction files can contain personal data, signatures, security-sensitive layouts and commercially confidential information. Before using an external service, review permissions, processing arrangements, retention, training use and access controls. The General Data Protection Regulation and European Data Protection Board guidance are relevant starting points when personal data falls within their scope; they do not replace an organisation-specific assessment.
Use only the material necessary for the task and remove irrelevant personal details where appropriate. Do not assume that pasting text instead of uploading a file removes confidentiality obligations. This website does not receive or analyse project documents: its contact tool simply constructs a local text draft. The finished checking record should remain in your controlled project environment, connected to the same revision trail as the underlying evidence.